Report No: ACS23320 . Building Climate Resilience into Power System Planning: The Case of Bangladesh . November 2017 . GEEES GCCRA . Document of the World Bank . . . Standard Disclaimer: This volume is a product of the staff of the International Bank for Reconstruction and Development/ The World Bank. The findings, interpretations, and conclusions expressed in this paper do not necessarily reflect the views of the Executive Directors of The World Bank or the governments they represent. The World Bank does not guarantee the accuracy of the data included in this work. The boundaries, colors, denominations, and other information shown on any map in this work do not imply any judgment on the part of The World Bank concerning the legal status of any territory or the endorsement or acceptance of such boundaries. . Copyright Statement: . The material in this publication is copyrighted. Copying and/or transmitting portions or all of this work without permission may be a violation of applicable law. The International Bank for Reconstruction and Development/ The World Bank encourages dissemination of its work and will normally grant permission to reproduce portions of the work promptly. For permission to photocopy or reprint any part of this work, please send a request with complete information to the Copyright Clearance Center, Inc., 222 Rosewood Drive, Danvers, MA 01923, USA, telephone 978-750-8400, fax 978-750-4470, http://www.copyright.com/. All other queries on rights and licenses, including subsidiary rights, should be addressed to the Office of the Publisher, The World Bank, 1818 H Street NW, Washington, DC 20433, USA, fax 202-522-2422, e-mail pubrights@worldbank.org. Building Climate Resilience into Power-System Planning: The Case of Bangladesh Source: NOAA November 2017 CONTENTS ACKNOWLEDGMENTS V ABBREVIATIONS VI 1 1. OBJECTIVE: BUILDING CLIMATE-AWARENESS INTO POWER-SYSTEM PLANNING IN BANGLADESH 1 1.1 DESCRIBING CLIMATE-POWER SYSTEM DEPENDENCIES ...................................................................................................2 1.2 ANALYZING UNCERTAINTY USING SLP AND RDM.........................................................................................................3 2 2. COUNTRY CONTEXT: AN EXPANDING POWER SYSTEM INCREASINGLY VULNERABLE TO CLIMATE EVENTS 3 3. METHODOLOGY AND MODELS IMPLEMENTED FOR BANGLADESH 7 3.1 UNCERTAINTY SURROUNDING KEY POWER-SYSTEM PARAMETERS ....................................................................................7 3.2 THE PROBLEM OF ASSIGNING PROBABILITIES TO SCENARIOS ............................................................................................8 3.3 MANAGING “DEEP” UNCERTAINTIES ..........................................................................................................................9 3.4 METHODOLOGY:DEFINITIONS AND KEY ASSUMPTIONS ................................................................................................12 4 4. KEY INPUTS AND ANALYTICAL FINDINGS FOR BANGLADESH 15 4.1 KEY INPUTS ........................................................................................................................................................15 4.2 DEMONSTRATION OF CLIMATE-AWARE FEATURES ......................................................................................................19 4.3 RISK MANAGEMENT USING RDM AND SLP ..............................................................................................................28 5 5. SUMMARY AND WAY FORWARD 38 5.1 SUMMARY OF MAJOR FINDINGS..............................................................................................................................38 5.2 WAY FORWARD ..................................................................................................................................................40 6 6. REFERENCES 44 7 7. APPENDIX 48 7.1 DATA ON DAMAGE AND OUTAGE IN POWER PLANT FACILITIES DUE TO FLOODING ..................................48 7.2 APPENDIX REFERENCES .....................................................................................................................................50 1.1.1 List of figures Figure 2.1. Generation capacity mix scenarios for 2041 (57 GW capacity) .................................................. 5 Figure 3.1. Procedure followed to implement hybrid SLP model and RDM ............................................... 11 Figure 4.1. Peak demand scenarios, 2016–41 (MW) .................................................................................. 17 Figure 4.2. Decision tree for SLP model used in step 3b (Figure 3.1) under the hybrid SLP ...................... 33 Figure 7.1. Flood damage and inundation depth (assumptions and historical datapoints) ....................... 49 Figure 7.2. Outage due to flooding and inundation depth (assumptions and historical datapoints) ........ 50 1.1.2 List of tables Table 4.1. Handling sources of uncertainty in the Bangladesh power-system-planning model ................ 17 Table 4.2. Model types used in this analysis............................................................................................... 19 Table 4.3. Cases constructed to answer Question 1, Part A ....................................................................... 21 Table 4.4. Key cost figures for Question 1, Part B (in millions of 2015 U.S. dollars) .................................. 21 Table 4.5. Capacity for power plants using imported coal across locations in GW (Case 1/Case 2) .......... 22 Table 4.6. Prioritization of sites for power plants using imported coal...................................................... 23 iii Table 4.7. Energy mix for cases 1 and 2 during 2035–41 (percent)............................................................ 24 Table 4.8. Cases constructed to answer Question 2, Part A ....................................................................... 24 Table 4.9. Key cost figures for Question 2, Part B (value in millions of 2015 U.S. dollars) ......................... 25 Table 4.10. Cases constructed to answer Question 3, Part A (first type of benefit) .................................. 26 Table 4.11. Cases constructed to answer Question 3, Part A (second type of benefits) ............................ 26 Table 4.12. Estimate of first type of benefits: Increase in operating costs as a result of considering flooding and climate change in operations (Question 3) (value in millions of 2015 U.S. dollars) .............. 27 Table 4.13. Cost categories used to estimate second type of benefits (Question 3) (value in millions of 2015 U.S. dollars) ........................................................................................................................................ 28 Table 4.14. Candidate strategies for RDM .................................................................................................. 30 Table 4.15. Scenarios retained for stochastic programming ...................................................................... 32 Table 4.16. Performance of strategies across all 486 scenarios (in millions of 2015 U.S. dollars) ............. 34 Table 4.17. Performance of strategies across nine discrete scenarios considered by SLP (value in millions of 2015 U.S. dollars) .................................................................................................................................... 35 Table 4.18. Capacity mix identified under the strategy with the minimum average regret (in GW) ......... 35 Table 4.19. Energy mix identified under the strategy with the minimum expected regret (percent) ....... 36 Table 7.1. Data on flooding damage and outage at a sample of coal power plant facilities ...................... 48 Table 7.2. Data on flooding damage and outage at a sample of natural gas power plant facilities .......... 48 Table 7.3. Data on flooding damage and outage at a sample of diesel power plant facilities ................... 49 iv Acknowledgments “Building Climate Resilience into Power-System Planning: The Case of Bangladesh” is the outcome of the collaborative efforts of the World Bank and John Hopkins University. The analytical work was done by Evangelia Spyrou (PhD candidate, JHU), with supervision by Professor Benjamin Hobbs (JHU). Hua Du (MS candidate, JHU), Adrien Vogt-Schilb and Mook Bangalore (World Bank) also contributed to the analytical work. The study was conducted under the guidance and direction of Debabrata Chattopadhyay and Neha Mukhi of the World Bank. The project team acknowledges the support and guidance provided by Marianne Fay in the preparation of the report. Michael Toman, Grzegorz Peszko, Marianne Fay, and Ana Bucher provided detailed comments on early drafts that substantially improved the content of the analysis and clarity of presentation. Financial support for the study came from the Pilot Program for Climate Resilience (PPCR), the Green Growth Knowledge Platform (GGKP), the Korean Green Growth Trust Fund (KGGTF), the Energy and Extractives Global Practice, and the Climate Change Group. v Abbreviations AR5 Assessment Report 5 BERC Bangladesh Energy Regulatory Commission BPDB Bangladesh Power Development Board CDD cooling degree days FOM fixed operation and maintenance (cost) GW gigawatt IEA International Energy Agency GDP gross domestic product JICA Japan International Cooperation Agency km kilometer kW kilowatt kWh kilowatt hour LNG liquefied natural gas m meter MW megawatt NLDC National Load Dispatch Center PGCB Power Grid Corporation of Bangladesh PSMP power-system master plan RCP representative concentration pathways RDM robust decision making SLP stochastic linear programming TEPCO Tokyo Electric Power Company, Ltd. WB World Bank Monetary values are expressed in U.S. dollars at the following rate: US$1 = Bangladesh Taka 80 vi 1 1. OBJECTIVE: BUILDING CLIMATE-AWARENESS INTO POWER- SYSTEM PLANNING IN BANGLADESH The objective of the analytical work reported here is to improve the analysis of power system resilience by developing and piloting an approach to incorporate climate-change considerations into power-system planning using Bangladesh and its power-system-planning models as a case study. A key part of that effort is to identify how different plans for generation investment might help to build climate resilience. This analysis considers only the physical impacts of climate change on power-system infrastructure, and assumes that Bangladesh adopts no new policies focused on climate mitigation (such as a carbon tax or requirements to import renewable energy from neighbors). The key questions addressed in the analysis are as follows: • Do flooding and storm risks eliminate a significant number of potential generation sites in Bangladesh? If a flood-protection standard exists, should it be incorporated in the planning model? • How will rising temperatures affect (a) thermal generation capacity and efficiency, and (b) electricity demand? • How vulnerable is the power-system plan if no flood-protection measures are taken? • How could different scenarios on climate and socioeconomic factors be integrated into the framework for power-system planning? • How does the optimal power plan, as identified by various alternative methodologies for considering climate resilience, differ from currently planned investments? Historical climate and climate change impacts (such as increased frequency and severity of flooding events), paired with uncertain socioeconomic trends in Bangladesh, can affect the long-term cost and reliability of the power system. Considering those climate risks and uncertainties in planning offers the opportunity to make investment decisions that perform better under a wider range of plausible futures and ultimately lead to better outcomes in terms of costs and reliability of the power supply. A goal of this report is to develop a methodology that would suggest a strategy recognizing those considerations. In order to address the questions posed earlier in this section, we first enhance the traditional least-cost power-system-planning model by adding a functional relationship between the climate variables (temperature and precipitation) and power-system technical and economic parameters. In the second task, we then use the enhanced model to explore how consideration of climate trends and scenarios together with flood-protection standards might change the identified power-system plan using two uncertainty-based planning approaches: stochastic linear programming (SLP) and robust decision making (RDM). A major conclusion of the analysis is that recognition of climate risk prioritizes investments in locations with lower flooding risk. The report, similarly to reports preceding it such as (Cervigni et al. 2015), also reflects on how the planning processes and tools need to evolve in order to mainstream climate resilience. We touch upon the 1 institutional capacity and institutional arrangements required to successfully mainstream climate change considerations into the power sector planning and operation. The analysis offers the Bangladeshi government and energy professionals in the World Bank an updated summary of the power situation in Bangladesh, while also providing a context to understand critical issues that need to be addressed through planning. It is expected that the discussions and supporting data and analyses in the report will inform future work across a wide spectrum of power sector activities —among them economic analysis, investment strategy, tariff reform, gas imports, cross-border power trading, and energy efficiency on both the supply and demand sides. 1.1 Describing climate-power system dependencies The first task in the analysis is to identify and quantify climate-dependent parameters of the current power-system-planning model. Traditionally adopted values of these parameters correspond to historical (or stationary) climate, or assume negligible impact of climatic conditions on power-system infrastructure. Based on a literature review, we decided to link two climate variables with parameters used in power- system planning: precipitation and temperature.1 We consider how precipitation might affect the profile of flooding risks, quantified as probability distributions of maximum annual water levels (stage), while for temperature we focus on how it affects demand for electricity and the availability of generation capacity. We assume that flooding profiles affect three parameters used in power-system planning: (a) the investment cost, namely the investment undertaken to protect a power plant against a flooding event having a particular frequency (or “return period”), (b) fixed operating costs, which include the cost of insurance against expected damage caused by flooding, and (c) plant unavailability (“forced outage rate”) which reflects on disruptions of normal operation due to facility flooding. This work relies on downscaled climate variable projections of Assessment Report 5 (AR5) provided by the (Bangladesh Climate Change Resilience Fund The World Bank Group International Center for Tropical Agriculture, n.d.), flooding projections provided by Fathom, a leading authority in modeling flood risk2 (Sampson et al., 2015), and processing of historical profiles based on results provided by (Hirabayashi et al. 2013). 1 For a more comprehensive overview of the effects of hydro-meteorological and climate parameters on the power system, see (Ebinger & Vergara, 2011). In this project, we analyze a subset of interactions between climate and the power system based on the relevance to the country of interest (Bangladesh), availability of data (e.g., if water temperatures were available, we could also include the impact of rising water temperatures on ability of power plants to satisfy their cooling needs), and our hypothesis on the significance of the effect for investment decisions (e.g., the effect of rising temperatures on efficiency is omitted since we assume it will not change the merit order, while the additional fuel needed will be provided by existing suppliers). 2 Fathom (http://www.fathom.global/), formed at the University of Bristol (United Kingdom) in 2013, was formerly known as SSBN Flood Risk Solutions. 2 1.2 Analyzing uncertainty using SLP and RDM Because climate projections are inherently uncertain, the second task is to find a way to appropriately reflect that uncertainty in decision models. The literature provides a variety of tools for risk management. These can be broadly classified into two categories (Ranger et al. 2010) according to the relative order in which projections and strategies are considered in the decision making process: (1) science-first methods, of which stochastic linear programming (SLP) is a popular example, require the determination of possible scenarios before the identification of a strategy because they employ future projections to inform the selection of a strategy and (2) policy-first methods, such as robust decision making (RDM), require the selection of alternative strategies upfront and they later use the future projections to assess the performance/vulnerability of alternative options. These methods can be deployed in planning models to address both climate and socio-economic or policy uncertainties. In a stochastic programming model (Birge and Louveaux 2011), uncertainties around climate as well as more conventional parameters (fuel prices and availability, economic growth, outages, etc.) are captured directly in the planning optimization by specifying probability distributions. The planning optimization then seeks an expected-least-cost generation and transmission plan that will deliver the best performance based on the specified expectations. Moreover, multistage stochastic programming fully models the adaptations that planners may make as they obtain more information—that is, as uncertainty resolves. Multistage stochastic programming therefore provides multiple plans contingent on resolution of any of the uncertainties. Before resolution of any of the uncertainties the decisions in the plans are identical; upon resolution of uncertainties, the plans become scenario-specific. RDM (Lempert et al. 2006) is an iterative, quantitative, decision-support methodology designed to help policy makers identify strategies that are robust, satisfying decision makers’ objectives in many plausible futures rather than being optimal in any single best estimate of the future or for a specified probability distribution of future scenarios. While traditional analyses begin by asking an unanswerable question: “What will the future bring?” RDM asks, “What are the strengths and limitations of our strategies, and what can we do to improve them?” We discuss the two methods in more detail in section 3 and our assumptions concerning the scenarios considered for uncertain factors in section 4. This work presents an opportunity to provide insights on relative strengths and weaknesses of both methods since we apply them on the same practical problem. Both SLP and RDM have been built into a least-cost planning model that was developed by the World Bank as part of its Bangladesh Analytical and Advisory Assistance activity (Pargal 2017). The model implementation was carried out by a research team from Johns Hopkins University and World Bank staff. 3 2 2. COUNTRY CONTEXT: AN EXPANDING POWER SYSTEM INCREASINGLY VULNERABLE TO CLIMATE EVENTS The challenges that Bangladesh faces in its effort to build sufficient power generation capacity are more serious than implied in existing planning reports, including the 2016power-system master plan (JICA and TEPCO 2016b). Some ambitious targets for expanding thermal capacity do not fully consider climate effects on power system infrastructure. Bangladesh, home to 161 million people, had 13,179 MW of electrical generating capacity in April 2017, about three-fourths of which operated on domestic gas extracted from onshore gas fields. The current level of generation is grossly inadequate, translating into less than 400 kWh per capita— one of the lowest in the world and about 2 percent of per capita consumption in the United States. This inadequacy is reflected in frequent rolling blackouts and other customer curtailments. Resources are managed inefficiently, as is evident from the fact that very expensive (19–31 U.S. cents/kWh) rented power plants running on heavy fuel oil or diesel are used even when gas-based capacity is available at less than 3 U.S. cents/kWh.3 Bangladesh announced in 2014 an ambitious program of development of the nation’s power system to increase capacity by an additional 10 GW by 2019 and achieve 100 percent electrification by 2021 at a cost of approximately $35 billion (Bdnews24.com Senior Correspondent 2014). The centerpiece of the program is a power-system master plan dating from 2010 (Ministry of Power, JIC Agency, and TTE Power 2011) that weighs options to increase capacity to about 57 GW by 2041. Recognizing the need for lower- cost electricity, the 2010 version of the master plan adopted a least-cost planning methodology that was carried over to the 2015 iteration of the plan. The least-cost planning analyses in 2010 and 2015 were led by the consulting division of Tokyo Electric Power Company (TEPCO) and funded by the Japan International Cooperation Agency (JICA).4 This analysis suggested a shift in the generation mix from natural gas to coal, reflecting the declining amount of natural gas available domestically and the low price of coal (relative to imported liquefied natural gas (LNG)). 3 World Bank analysis conducted using data obtained from the National Load Dispatch Centre. Natural gas has been priced below supply cost in the past, leading to inefficient consumption. However, the Bangladesh Energy Regulatory Commission has proposed price increases in the past two years (Quadir,2017). 4 JICA and the Asian Development Bank completed plans in 2005 and 2010. The current planning exercise began in 2015 and was completed in September 2016 (JICA and TEPCO 2016b) The results presented here are draft findings released by the JICA team in April 2016 in a presentation made to the World Bank (JICA and TEPCO 2016a) that was available when the present study was conducted. The PSMP generation plan is prepared on behalf of the Bangladesh Power Development Board, which owns majority of the nation’s generation assets (primarily gas -fired power stations). 4 Under the 2010 PSMP, installed capacity was to exceed 16 GW by the end of 2016 and 38 GW by 2030. Plants running on domestic and imported coal were expected to have capacity close to 20 GW by 2030 under the optimal plan (referred to as the Fuel Diversification Scenario).5 However, the implementation of the 2010 PSMP faced several challenges that have prevented any of the coal projects that it identified from being completed. First, the 2010 PSMP assumed exploitation of domestic coal. Yet as late as 2016 the national policy on coal had not yet been communicated (Rasel 2016), the construction of the coal mine at Phulbari encountered substantial social opposition (UN News Centre 2012). Both these developments make reliance on domestic coal highly uncertain. Similarly, construction of a coal power plant in Rampal has raised negative reactions for several reasons (Harvey 2016), notably (a) its proximity to the Sundarbans mangrove forest, (b) an impact assessment considered incomplete by opponents of the project, and (c) the amount of land required for the plant. Another challenge for the development of power infrastructure in Bangladesh relates to the level of investment required and the need for effective mobilization of joint public–private ventures (PRI of Bangladesh 2015). JICA’s most recent analysis initially6 examined five generation-mix scenarios (P1–P5) for 2041 (figure 2.1). These include a heavy reliance on domestic capacity, including 40 GW of coal and gas. Somewhere between 20 GW (P3) and 31 GW (P1) of new coal capacity would have to be installed in Bangladesh over the next 25 years, whereas only 250 MW of domestic coal capacity is currently operational. Figure 2.1. Generation capacity mix scenarios for 2041 (57 GW capacity) 100% 80% 60% 40% 20% 0% Coal Max P2 P3 P4 Gas Max Gas Coal Power Import/Renewable Energy Oil/Hydro/Others Nuclear Source: Government of Bangladesh (2016). Note: P3 is the plan recommended in the 2016 PSMP. Although environmental concerns hindered the implementation of the 2010 PSMP, the latest PSMP overlooks most climate-related risks. Bangladesh lies in Asia’s largest and the world’s most densely 5 2010 PSMP, tables 3-4. 6 In the Final Report 2016 (JICA and TEPCO 2016b), additional scenarios are considered which have lower share of thermal power plant capacity in 2041. 5 populated delta and is highly vulnerable to climate change (Adams et al. 2013). The country’s power sector shares in that vulnerability. Khan et al. (2013) discuss how a higher ambient temperature, a rise in sea level, increased salinity of water resources, drought, and frequent flooding may significantly affect power generation. They argue that climatic effects on the power system should be quantified, monitored, and reflected in national standards and infrastructure design. However, such effects are not captured as part of the least-cost PSMP methodology. Capturing the risks of flooding and other phenomena traceable to climate change using the PSMP methodology is not a simple task, in terms either of the additional data needed or the effort required to adjust the model. However, enhancing the methodology with climate information will help improve decision making, especially in a country like Bangladesh, where siting of power stations has faced public opposition owing to environmental concerns. If climatic factors suggest a reconsideration of siting decisions, and if politically acceptable sites cannot be found, the planner may choose to explore supply alternatives, such as renewables and power imports, at levels higher than those considered in the April 2016 PSMP (JICA and TEPCO 2016a). 6 3 3. METHODOLOGY AND MODELS IMPLEMENTED FOR BANGLADESH In this section, we discuss the sources of uncertainty surrounding key power-system parameters and the difficulties of assigning probabilities to those sources of uncertainty. We then review two methodologies for risk management provided in the literature and explain how we implemented them in this analysis. 3.1 Uncertainty surrounding key power-system parameters For purposes of power-system planning, uncertainties can be grouped into two broad categories: (1) policy and socioeconomic uncertainties and (2) climate uncertainties. For both types, devising accurate projections is challenging because power-system planning has a very long time horizon, and many uncertainties in power planning depend on broader uncertainties, both macroeconomic and social. Further, climate-change projections still come at rather low resolution (e.g., 100 x 100 km), and limited work has been done on processing climate data into climate risks of spatial and temporal resolution required for analyses in the energy sector. For example, while precipitation projections are available through various climate models, flood-risk projections derived from precipitation data are not readily available. Four factors from the first category (“policy and socioeconomic uncertainties”) are explicitly represented in this modeling exercise: demand growth, fuel price, supply of domestic fuels, and ability to import fuels.7 We identified these factors as uncertain based on review of recent power-system-planning exercises in Bangladesh (JICA and TEPCO 2016a). Demand is highly affected by the economic growth of the country under examination—in this case Bangladesh. Fuel prices are influenced by the balance of global supply and demand and by policies related to renewable energy and trade. In Bangladesh, the supply of domestic coal is considered uncertain because of long delays in already announced coal-mine development, public opposition to mine projects, and the absence of a current coal policy. Natural gas supply could be considered uncertain for similar reasons and because of uncertainty about available reserves. Projections for two climate variables (and associated power-sector climate indicators) were used in this analysis: temperature (and cooling degree days) and flooding (derived from precipitation projections).The uncertainty associated with the projections related to climate variables can be broken down into: (a) climate model uncertainty, (b) processing model uncertainty, and (c) impact function uncertainty. Uncertainty attached to climate projections by global circulation models has been a central concern of reports produced by the Intergovernmental Panel on Climate Change and many research groups. Even if the projections of various climate models are considered, epistemic uncertainty arising from our limited understanding of climate processes persists. Second, processing projections provided by global circulation models into information of finer spatial granularity or useful to the power-system-planning analyst information such as flooding risk may employ 7 Additional inputs that might be uncertain might be capital costs for new technologies, future renewable policy requirements. 7 statistical models or flooding models. (Trigg et al. 2016) discuss how state-of-the-art global flooding models provide divergent results because of differences in (a) input generation methodology (for example, one group of models relies on statistical analysis, while the second generates flows from a land- surface model), (b) model resolution, and (c) variations in the methodology used to simulate inundations (that is, in the description of the “computational hydraulic engine”, (Sampson et al. 2015)). Finally, the exact function describing the impact of the climate variable on the power system is not known. Limited anecdotal evidence and empirical research permit no more than estimates of linkages between climate variables with power-system parameters. 3.2 The problem of assigning probabilities to scenarios Assigning probabilities to any of the scenarios even when considering just one source of uncertainty is challenging, given the long-term horizon of power-system planning. However, this challenge has long been faced by power sector planners, and various ways to assess the robustness of plans in the face of deviations from assumed probabilities can be employed as we will discuss later in this section. Agencies such as IEA usually provide scenarios on projections of key parameters but with no probability attached to them. For example, for fuel price scenarios developed for the World Energy Outlook 2015 (International Energy Agency 2015b), we found only commentaries on the plausibility of some scenarios. For example, the “IEA New Policies” scenario was presented as more plausible while realization of the “IEA Current Policies” scenario was considered “extremely unlikely”. Uncertainty about demand, meanwhile, depends not only on the uncertainty of macroeconomic projections but also on technology trends, such as electric vehicles. Finally, for fuel availability, it might be possible to quantify uncertainty for some aspects (e.g., related to estimates of reserves), but not for others (such as delays or cancellations of mining projects due to public opposition). For climate projections, there is no consensus on an approach to assign probabilities (or weights) to projections provided by climate models. Several applications have given even weight to climate scenarios. However, there are limitations to this approach, such as omission of any information on the predictive skill of the model or ignorance of model interdependency (Knutti 2010). More elaborate approaches attempt to address the aforementioned limitations. For example, Rupp et al. (2013) assess the ability of CMIP5 projections to simulate temperature and precipitation over the Pacific Northwest region of the United States. Olson, Fan, and Evans (2016) employ statistical methods to derive model weights for regional climate model projections based on the “skill” of the models. Climate model genealogy may also indicate dependency of models. As Steinschneider et al. (2015) show, a different probabilistic description of the climate variable may be obtained when the genealogy is taken into account. Finally, Rasmussen, Meinshausen, and Kopp (2016) use a simple climate model to construct model surrogates to enhance climate projections with tails of the probability distribution that are not captured by the CMIP5 ensemble but might significantly affect the climate risk analysis. 8 3.3 Managing “deep” uncertainties Lempert et al. (2006) define deep uncertainty as “the condition in which analysts do not know or the parties to a decision cannot agree upon (1) the appropriate models to describe interactions among a system’s variables, (2) the probability distributions to represent uncertainty about key parameters in the models, or (3) how to value the desirability of alternative outcomes.” The uncertainties considered for Bangladesh can be characterized as deep ones. In particular, for climate parameters, we face epistemic uncertainty, given our limited understanding of the climate system. Based on that, Lempert’s conditions (1) and (2) seem to hold. On the other hand, epistemic uncertainty is not an issue for socio-economic variables. But despite our better understanding of these socio-economic uncertainties, it is very difficult to get multiple parties to agree on a unique probabilistic distribution of the respective scenarios (which is why scenarios provided for fuel prices are not accompanied by an estimate of probability). So even for socio-economic uncertainties, condition (2) seems to hold, and this category could also be classified as a deep uncertainty. The literature provides a variety of tools for risk management that, as briefly noted in section 1, can be broadly classified into two categories according to the characterization of the uncertainty: (1) science-first methods and (2) policy-first methods. Science-first methods usually require probabilistic characterization of the uncertainty prior to identification of a strategy. An example of a science-first approach is stochastic linear programming (SLP), which has been quite popular in the power-system community (Pereira, M.V.F. & Pinto 1991; Ho et al. 2016; Nunes et al. 2016). SLP employs probabilistic information to identify the optimal expectation strategy. In case uncertainty resolves within the horizon, stochastic programming is formulated with multiple stages. Then SLP suggests adaptive, scenario-specific strategies for the stages following the uncertainty resolution. Policy-first methods usually require that uncertainty be described as a range, with no probabilistic information. Alternatively, they may assign several probabilities to the same uncertainty. An example of the policy-first approach is robust decision making (RDM), which has been popular within the climate- change adaptation community (OECD 2015). RDM asks the modeler to first specify a candidate set of strategies and then tests the candidates against all plausible futures (or, with respect to our analysis, against combinations of scenarios). We reviewed several papers (Hallegatte and others 2012; Dittrich, Wreford, and Moran 2016; Watkiss et al. 2015) that compare and contrast the two methods, but there were no definite recommendations concerning the selection of one method over another. It seems that RDM is recommended for cases where no probabilistic information is available. However in reality, a mix of inputs is usually available—some with probability estimates and others without—and additional practical problems arise for both methods as discussed in the following paragraph. We therefore experimented with both methods. We refer to our implementation of SLP as a hybrid SLP because it has an additional step compared to its conventional application. The hybrid approach can be interpreted as SLP followed by a sensitivity analysis, with the 9 latter being implemented across the same set of scenarios that are considered by RDM, which usually has many more scenarios than the set implemented in the original SLP. Here we discuss in detail the reasons that led us to apply both methods and compare their results. First, for both methods, a crucial piece of information is missing: probabilities (in the case of SLP) and prespecified strategies (for RDM). Second, both methods are challenged by the high number of plausible scenarios. In SLP, the more scenarios considered, the greater the difficulty of obtaining a solution.8 In RDM, each strategy is tested against all plausible scenarios, leading to a linear increase in the number of simulations as the number of scenarios expands. Third, RDM is not completely “probability-free,” since a subjective judgment on the plausibility of each scenario is required in order to assess the performance of each strategy (Lempert et al. 2006). Fourth, the original papers introducing RDM do not describe SLP and RDM as being mutually exclusive. On the contrary, Lempert et al. (2006) suggest the use of a stochastic programming model to identify the first candidate strategy to be tested, since it will have the optimal expected “regret”—defined as the cost of a given strategy minus the cost of the best-performing strategy under the same case—contingent on the weighting across scenarios. Finally, a hybrid approach in which the strategy identified by SLP is tested against all plausible scenarios will reveal any vulnerabilities of the stochastic programming strategy traceable to scenarios omitted in the scenario reduction step. Figure 3.1 illustrates the implementation of the two methods, step by step. The left-hand side of the figure presents the steps in the hybrid SLP approach; the right-hand side, the steps for the RDM. The same type of information is used for both methods. 8 Higher number of scenarios increases the model size and subsequently the memory requirement and the computational time to solve the model. 10 Figure 3.1. Procedure followed to implement hybrid SLP model and RDM At left, hybrid SLP; at right, RDM Step 1: Deterministic runs over all plausible future states Step 2: Cluster strategies identified in step 1 based on capacity vector Step 3a: Draft one scenario for Step 3: Choose one strategy to each cluster and assign to it the represent each cluster probability of all states represented by the cluster Step 3b: Run the stochastic program and identify one strategy Step 4: Conduct vulnerability Step 4: Conduct vulnerability assessment by testing the assessment by testing the strategy stochastic strategy across all future of each cluster across all future states states Step 5: If the performance of the Step 5: If none of the strategies strategy is not acceptable, update shows acceptable performance, stochastic programming identify strategies that might offer (e.g.,identify cases in which better performance in vulnerable performance was weak and include areas and include them in the them as scenarios) candidate strategies set Note: Step 5 was not implemented in the analysis reported here but is included in the figure in the interest of completeness. In step 1, we run the model to identify the “perfect foresight” strategy for each scenario, i.e., realization of future. Put in simple terms, this step yields multiple investment plans, i.e., one for each state of demand, fuel price, and so on that the planner could implement if the uncertain factors were known with perfect certainty. In step 2, the information obtained in step 1 is used to devise strategies to be tested using RDM and scenarios to be considered using SLP. Among the various methods and heuristics that can be applied at this point, clustering and sampling methods are commonly used. For this analysis, as explained in section 4, k-means clustering based on investment decisions identified in step 1 was employed within RDM, and hierarchical clustering within the hybrid SLP approach. Future research could examine the performance of different clustering approaches and sampling methods and their impact on model performance (similar research has been conducted by (Bruninx & Delarue 2016)) . In step 3, the task is to select the scenario that will represent each cluster of scenarios for the hybrid SLP method. In this particular analysis, from each cluster of scenarios, the scenario with the smallest maximum 11 distance from the other scenarios in the same cluster was selected to represent the cluster. For RDM, the representatives of the clusters (“centroids”) calculated by k-means were used to define candidate strategies. For hybrid SLP, in order to identify a strategy, we ran a SLP in step 3b, weighting each representative scenario by the fraction of original scenarios in its cluster. Step 4 is a common step in RDM known as “characterizing vulnerabilities.” It aims to test the performance of candidate strategies across as many original scenarios as possible. In the case study, we tested the performance under all original scenarios considered. We also introduced the same step in the hybrid SLP to make the comparison fair. Step 4 is very important for the evaluation of any strategy and also for the cross-comparison of the methods. In it, we imposed the first-stage decisions of the investment plan (referred to as a “strategy”) before any uncertainty is resolved. Following resolution of uncertainty, we allowed the model to suggest the best-performing plan. That way, we made sure that learning and adaptation were taken into account in the same way for both methods. By its nature, multistage SLP accounts for learning and provides scenario-specific results for the scenarios modeled; whereas RDM must be enhanced to provide adaptive strategies. In practice, however, it is enough for both methods to specify the investment decisions that should be fixed at a predetermined level before the next planning cycle and the resolution of the uncertainty. That being so, we use the two methods accordingly, specifying the first stage investments and allowing for complete adaptation (recourse) in the second stage. Step 5 as shown in figure 3.1 was not implemented in this analysis, but it is included it in the figure for completeness. In case none of the strategies identified is considered adequate based on a criterion set by the planner, a new iteration is initiated. This is a common step in RDM, where the “patient rule induction method” (or PRIM) common in data mining is used to discover scenarios where performance was not acceptable. Under the hybrid SLP method, a similar step could be applied, in which case the SLP could be re-run with updated weights or with different in terms of size or composition set of representative scenarios. 3.4 Methodology: definitions and key assumptions 3.4.1 What counts as a “strategy”? The purpose of any power-system-planning exercise is to obtain a set of investment decisions, which combine to form a “strategy.” The important features of each decision must be specified. For this analysis, we decided that the year, location, and technology of each investment decision would be specified. Moreover, we chose to use absolute units (in MW) for the investments and not their relative contribution to the capacity mix, because the relative information might be misleading if different demand-growth scenarios were employed. The power-system-planning model is typically solved to specify investments over a specific time horizon. For this exercise, we assumed that investments that would come on-line within the next 10 years should be specified. We decided to focus on the first 10 years because the planning cycle (the time between two consecutive planning exercises) presently is 5 years, and we assumed that it takes at most 5 years for an 12 investment to come on-line.9 Investments that come on-line after 10 years are also provided by the planning model. However, given the scope of this exercise we do not include them in the definition of a strategy. A power-system planning strategy, then, was defined as the set of investments to be completed within the next 10 years. The set of investments was described with respect to time, location, and technology and measured in MW. The output of step 3b (for the hybrid SLP approach) or step 3 (for the RDM approach) is one or more strategies. 3.4.2 Uncertainty For each uncertain factor considered in this analysis, we have a set of discrete scenarios specifying the values of the uncertain factor for entire horizon. We combine the single-factor scenarios in any possible way to create multi-factor scenarios for the power-system-planning model. In other words, we consider all possible combinations among the 3 temperature scenarios, 3 demand growth scenarios, 3 fuel price scenarios, 3 domestic coal supply scenarios, 2 natural gas supply scenarios, and 3 flooding scenarios to create 486 (3*3*3*3*2*3) scenarios (more detailed information on scenarios in Section 4). In this analysis, we use the words scenarios, original scenarios, futures and states interchangeably to refer to the 486 scenarios. 3.4.3 Uncertainty resolution Because of computational limitations, all uncertain factors are considered to be fully resolved after the tenth year of the model (2025). That is, there is no residual uncertainty after that date. This may be a realistic time frame for resolution of the uncertainty surrounding some factors, notably the socioeconomic ones. For factors related to climate change, by contrast, the topic of the resolution of uncertainty has been debated for years (Loulou, Labriet, and Kanudia 2009). In the results reported here, we consider only one possibility for uncertainty resolution. However, in actual planning exercises, sensitivity analyses with respect to resolution can be conducted. 3.4.4 Brief description of the model Once the factors are specified, the model computes: • Investment and retirement decisions • Expected generation dispatch for all existing and new projects at 28 demand levels for each year • Expected unserved energy—that is, the share of demand that cannot be met owing to lack of capacity or outages • Unserved reserve capacity—that is, the share of the reserve requirement that cannot be met owing to lack of installed capacity (this effectively means that the system may be able to meet demand, but has less capacity to manage system disturbances). 9 Other assumptions on the duration of the first stage might be realistic as well. For example, another realistic assumption might be longer, to account for more time to attract investors and obtain authorization for construction or an assumption that takes into account differences in the construction time of different types of power plants. 13 The objective function of the SLP computes the expected value of the annualized capital costs of new generation projects, the operating cost of all existing and new projects and penalties for unserved energy and reserves. Three model versions are used for the implementation of the uncertainty methods.10 These versions are used to address questions discussed in section 4. They are: • Deterministic (single scenario, investments optimized). This version integrates one future scenario at a time and optimizes investments over the entire horizon (until 2041). This version of the model is run in step 1 as many times as the number of scenarios considered (486 scenarios in this analysis). • SLP (multiple scenarios, investments optimized). This version integrates multiple future scenarios and optimizes investments over the horizon, but in two stages. First-stage decisions (to 2025) are identical across scenarios. After 2025, uncertainties are assumed to be resolved, and the decisions can vary contingent upon the scenario realized. This model version is employed in step 3b. A reduced number of scenarios (nine) is considered in the second stage. • Deterministic (single scenario, investments after 2025 optimized). This version integrates one future scenario at a time and optimizes additions after 2025 until the end of the horizon (2041). Additions of new units up to 2025 are imposed at fixed levels specified by the strategy under examination. This model version is employed in step 4 as many times as the product of the number of scenarios considered with the number of strategies assessed (486 scenarios times 8 strategies, 3,888 runs total). 3.4.5 Evaluating the performance of proposed strategies To evaluate performance of a given strategy (in step 4), we used as a metric its “regret.” Regret is defined as the cost of the strategy under examination (an objective function estimated by step 4 for each state) minus the cost of the best-performing strategy for that state (an objective function estimated by step 1 for each state). Given the multi-year horizon of the power-system-planning problem, we employ net present value to aggregate costs from different years into the objective function. Discount factor is a key assumption in the net present value calculations. In Section 4, we report the average, minimum, and maximum regret for all strategies, including the one identified by the SLP exercise. In this analysis, we do not attempt to choose a particular strategy and that’s why we do not attempt to choose one metric or employ a method to aggregate multiple metrics into a single metric. For future research, though, methodologies that aggregate multiple objectives could be employed and it might worth to consider on top of cost metrics, other metrics such as emissions, levels of unserved energy etc. 10 All models used in section 4 are enhanced versions compared with the status quo because they account for interactions between climate and power-system parameters interaction. 14 4 4. KEY INPUTS AND ANALYTICAL FINDINGS FOR BANGLADESH In this section, we discuss the key inputs and model results for Bangladesh. Section 4 is meant to serve as a reference guide to future implementations of the methodology by practitioners in Bangladesh and elsewhere. After reviewing the assumptions made, we document the results of our effort to build climate awareness into the planning model. In particular, we consider how upgrading to climate-aware models might improve the accuracy of cost estimates and the efficiency of decisions. We then compare the two methods used to address uncertainty (SLP and RDM) discussed in section 3. 4.1 Key inputs 4.1.1 Key inputs: cost and potential The reference year for discounting of the objective function is 2015, and real 2015 U.S. dollars are used. The discount factor is 6 percent. We assume a weighted average cost of capital of 10 percent for new investment projects. The value of lost load is assumed to be 50 U.S. cents/kWh. With respect to power imports, Bangladesh already has a 500 MW interconnection with India. In addition, new interconnections with Bhutan, Nepal, and India may be possible following the SAARC Regional Trade study. Here, we consider them as possibilities, and use the following assumptions: • The transmission investment cost is $3,184/MW/km. • Interconnection capacity is constrained by an annual ceiling that increases from 1,500 MW (2020) to 13,500 MW (2041). • An upper bound of 8 GW per interconnection option is used. Price assumptions are as follows: • For India, we assume for each time block the most expensive type of power plant that will be on- line based on energy mix projections provided for the scenario “IEA New Policies” (IEA 2015a). In particular, we assume: high-speed diesel, 2,600 hours; natural gas, 1,828 hours; coal, 2,019 hours; nuclear, 2,313 hours. The price assumptions are: high-speed diesel, $223.4/MWh; LNG, $98– 172.32/MWh; coal, $53.6/MWh; and nuclear, $44/MWh. Note that we assume the variable part of the price (assumed to be 67% for HSD, 73% for LNG, 48% for NG, 46% for coal) will increase at the same rate as international prices for fuel. • For hydropower imported from Nepal, we assume $47/MWh and availability 50 percent of the year. • For hydropower imported from Bhutan, we assume a price of $37/MWh and $0.5 million/MW based on a review of existing agreements with India (Premkumar 2016); availability 44 percent of the year; and a production profile based on imports from Bhutan to India as reported in executive summaries over the past 10 years (Government of India, n.d.) 15 New coal capacity up to a limit of 30 GW is considered, with a capital cost of $2,032/kW for domestic coal (4 sites) and $2,622/kW for foreign coal (13 sites). Given the paucity of land in Bangladesh we do not assume any sites for coal or gas beyond what has been considered in the 2010 PSMP and the Ashugonj Power Station Company’s master plan. Investments in natural gas power plants are considered at a capital cost of $1,342/kW for combined cycle units and $1,012/kW for open cycle units across 95 locations.11 We assume that the new gas power plants can either be developed on land being considered for coal power plant development (that is, competing with foreign coal power plants for the roughly ~13,000 acres of land that in total are available for development) or on land already being used or proposed to be used for gas power plant development (~5,000 acres). Regarding investment in renewable sources, biomass is assumed to have a capital cost of $3,000/kW. However, its potential is capped at 274 MW in line with the estimate provided in (Government of Bangladesh 2015). Photovoltaic’s capital cost is set at $2,430/kW; this includes an estimate of $880/kW for land acquisition. Investments in wind farms are not considered since the estimated potential is of ~600 MW (Government of Bangladesh 2015). 4.1.2 Key inputs: scenarios considered for uncertain factors In Table 4.1, we provide the assumptions on the six uncertain factors: demand growth, fuel prices, coal and natural gas availability, temperature and flooding. We relied on (JICA and TEPCO 2016a) for scenarios on socio-economic uncertainties. For demand, three scenarios are modeled. Peak demand is expected to grow from the current level of about 9 GW to about 40–60 GW in 2041, depending on the scenario assumed for demand growth (figure 4.1). For gas supply, we drafted two scenarios: one in which no new domestic gas reserves or new infrastructure for LNG imports (apart from already planned infrastructure) was available and another in which all the sources mentioned in the last master plan were available.12 For domestic coal availability, we used the same scenarios as in (JICA and TEPCO 2016a). However, our scenarios for fuel prices differed slightly from those in the TEPCO analysis (JICA and TEPCO 2016a). For fuel prices, (JICA and TEPCO 2016a) considered four IEA scenarios but here we chose slightly different scenarios: • two IEA scenarios (also considered in (JICA and TEPCO 2016a) : the central scenario (“IEA New Policies”) provided in (International Energy Agency 2015b) and IEA 450, which is considered as plausible if the global community succeeds in keeping the increase in global average temperature below 2ºC by the end of this century. 11 Some of the 95 locations are quite close to one another. For future runs, we could aggregate them but we tried to keep as much spatial detail as possible since the flooding risk depends on the elevation of the terrain, which might change abruptly. 12 The LNG infrastructure development could also be modeled as a decision variable. However, most power- system-planning models assume fixed supply curves for fuels and focus on electricity investments. In future planning exercises for Bangladesh, it might worth to check the benefits of co-optimizing both decisions on fuel supply and electricity infrastructure. 16 • one scenario we drafted based on the latest fuel price projections by the World Bank (a low-oil- price scenario).13 Table 4.1. Handling sources of uncertainty in the Bangladesh power-system-planning model Uncertain factor Full set of Source of data on uncertain factor Scenarios modeled scenarios Demand 3 (JICA and TEPCO 2016a) 3 Fuel prices 5 (JICA and TEPCO 2016a)F1–F4, (The World F1, F3 ((The World Bank Group, 2017) Bank Group, 2017) Domestic coal availability 3 (JICA and TEPCO 2016a) 3 Natural gas availability 2 Based on (JICA and TEPCO 2016a) 2 Temperature/cooling 17 http://climatewizard.ciat.cgiar.org/wbclimat bcc-csm1-1, cesm1- degree days eanalysistool/ bgc, mri-cgcm3 Flooding 3 FATHOM and (Hirabayashi et al., 2013) 3 Figure 4.1. Peak demand scenarios, 2016–41 (MW) Base demand Low demand High demand 60,000 50,000 Peak demand (MW) 40,000 30,000 20,000 10,000 0 2016 2021 2026 2031 2036 2041 years Source: JICA and TEPCO 2016. Projections for two climate variables (and associated power-sector climate indicators) were used in this analysis: temperature (and cooling degree days) and flooding (derived from precipitation projections). For temperature projections, we considered three temperature scenarios based on clustering of the 17 scenarios available from (Bangladesh Climate Change Resilience Fund The World Bank Group International Center for Tropical Agriculture n.d.) for temperature and cooling degree days. For flooding, FATHOM provided us with projections using low, mean, and high projections for precipitation change from climate scenario RCP 8.5, the scenario with the highest radiative forcing among the scenarios considered in AR5.14 However, because the differentiation between the three flooding scenarios provided by FATHOM was negligible in terms of their impact on power-system parameters, we decided to retain only the high 13 Price forecast provided through 2030. After 2030, real prices were held constant for coal and considered to be increasing at the same rate for other fuels. 14 RCP 8.5 has also been described as “a high-emission business as usual scenario.” (Riahi et. al, 2011) 17 scenario and to create one additional scenario relying on projections published by (Hirabayashi et al. 2013). According to the latter, the return period of a 100-year event under historical climate conditions will be 5–25 years in Bangladesh by the end of the century, as projected by the median model of AR5 in the RCP 8.5 scenario.15 Using this result, we constructed a new scenario with modified flood profiles in order to project the historical 100-year event as a 20-year event for fluvial/pluvial flooding and as a 25- year event for coastal flooding.16 We combine the single-factor scenarios, described in Table 4.1, in any possible way to create multi-factor scenarios for the power-system-planning model. In other words, we adopt the perspective of a planner who does not identify the interdependencies among individual uncertain factors that may be present in some scenarios and does not comment if some combinations are impossible (e.g., high fuel prices and high fuel availability). That said, all possible combinations of scenarios are considered and we obtain 486 scenarios.17 For this analysis, we made a “naïve” assumption of even probabilities across 486 scenarios. 4.1.3 Key assumptions: impact of climate variables on power system With respect to temperature impact, we assumed a capacity derating of 0.4 percent for coal, 0.5 percent for combined cycle gas turbine, and 0.7 percent for peaking open cycle gas turbine for every 1oC above 27oC. The impact of cooling degree days on electricity demand is captured through empirical relationships provided in McNeil and Letschert (2008). These relationships rely on GDP and cooling degree days projections to estimate the penetration of air conditioners (AC units) in the residential sector and annual energy consumption per AC unit. With regard to flooding impact, the cost of filling material to elevate the power plant site was assumed to be $39/cubic meter, and land requirements were assumed to be 0.45 acres/MW for coal and 0.1 acres/MW for gas units. In case of flooding (that is, an inundation depth higher than the facility’s protection level), two consequences were modeled: • Outage: 1 day per decimeter up to 1.5 m, the whole monsoon period if between 1.5 m and 4.5 m, and the whole year if more than 4.5 m. 15 Figure 1a (Hirabayashi et al., 2013) 16 In our flood frequency analysis, we fit the Gumbel distribution to the data provided by FATHOM (inundation depth and exceedance probability), specifying the location and scale parameter of the Gumbel distribution at each power plant site. When constructing each new scenario, we assumed that the scale parameter is the one specified by fitting the distribution to historical data. We modified the location parameter to achieve the change described by Hirabayashi and others (2013). 17 However, if the planner identifies some of these combinations as less plausible or even impossible, the planner can assign a lower weight or eliminate them from the full set of scenarios. Similarly, if the planner identifies some combinations as more plausible, the planner can assign to them a higher weight when they are considered in the analysis. 18 • Damage of the facility: 32.5 percent of the capex at an inundation depth of 3.5 m, increasing at a rate of 6 percent per meter up to 7 m inundation depth and then 1 percent per meter reaching a maximum of 87 percent damage. The annual cost for insurance was assumed to be 2.59 times the expected annual damage cost, in line with premiums recorded during the sale of New York Metropolitan Transportation Authority (MTA) catastrophe bonds.18 4.2 Demonstration of climate-aware features A climate-aware model was developed as part of this project and used for the first time to address questions regarding the Bangladeshi power system’s resilience to climate change. In this subsection, we consider the results of activating climate-aware features under one particular scenario (a combination of the most plausible parameter values for each uncertain factor, such as base demand, base domestic coal supply, IEA New Policies for fuel, and high domestic gas and LNG supply). The results will help readers understand the basic mechanisms through which climate considerations affect the power system plan according to our model. Table 4.2 describes the different variations of model cases that were used to evaluate the value of climate- aware features. Table 4.2. Model types used in this analysis Case type Description Climate-ignorant (or status quo/stationary Does not account for any climate variables, and offers no model) (Case 1) protection against climate variations Climate-naïve (type of climate-enhanced model) Accounts for historical temperature and flood risks (Case 5) Climate-aware (type of climate -enhanced Accounts for future climate projections—temperature model) (Case 6) and/or flooding risk The model currently used in Bangladeshi planning studies does not explicitly take into account historical data or climate projections. Using the status quo model, a planner is aware that certain cost components are omitted from the objective function (the total cost the planner aims to minimize such as possible additional damages or other cost impacts of a changed climate. So, the planner expects that the climate- ignorant model will underestimate total costs. In omitting these cost components, the planner hopes that the investment plan identified by the status-quo model will not be significantly different from the optimal plan, and saves the immediate cost and time required to obtain the information needed to calculate the additional cost components and upgrade the model. In order to test the efficacy of this approach, we compare the performance of investment plans identified by the climate-ignorant model and the enhanced model by addressing the following questions: 1. Should an existing flood protection standard be incorporated in the planning model? 18 New York MTA catastrophe bonds offered three-year reinsurance protection for storm surge risks. The probability of the catastrophic event was estimated to be 1 in 60 years and the expected loss was estimated to be 1.71 percent. Investors asked for a 4.5 percent spread (Kusche and Coyne 2015). 19 2. What are the benefits of considering how resilient power systems with regard to varying climate patterns? Moreover, we use climate projections to investigate how the costs of a particular plan might increase under a high-climate-change scenario and whether a model with perfect foresight would lead to a lower- cost investment plan. In brief, we aim to answer the following additional question: 3. What are the benefits of considering climate change projections? In order to answer the questions posed above, multiple runs of the model are performed with different specifications. In the following paragraphs, we address one question at a time. For each question, we describe the specifications of the scenarios we considered and compare and discuss the benefits identified. Despite an obvious improvement in the accuracy of cost estimates, the value of using an enhanced model could be small if the investment plans identified by the status quo model and the new enhanced model are identical or very similar. In our analysis, we do not document the differences in the investment plan; instead, we estimate the value of using the enhanced model by comparing the cost of the two plans (using the same cost components). 4.2.1 Question 1. Should an existing flood protection standard be incorporated in the planning model? Question 1, part A: Experimental design used to address the question • To answer this question, we use three cases (table 4.3): Case 1 is the climate-ignorant planning model, where no climate considerations are taken explicitly into account; • Case 2 is an enhanced version of the climate-ignorant model that takes into account the construction costs a developer might incur to comply with a flood protection standard; and • Case 1* optimizes the approximate operations of the plan identified by Case 1, as well as the estimated construction costs of complying with a flooding standard. By comparing Case 1* with Case 2, we can quantify the benefits of considering flooding standards when optimizing investments (Case 2) compared to not considering those standards when optimizing (Case 1). Note that all three cases disregard climate change, and differ only in the time they consider flood protection standards (i.e., Case 1* considers flooding standard after having chosen an investment plan while Case 2 considers the flooding standard within the optimization problem that suggests an investment plan). 20 Table 4.3. Cases constructed to answer Question 1, Part A Investment Temperature Descriptive name decisions Flooding features features Case 1 Climate-ignorant Optimized OFF OFF Case 2 Climate-ignorant with Optimized Partially ON OFF flooding standard (flooding standard) Climate-ignorant Partially ON Case 1* in a model with flood Fixed from Case 1 OFF protection standard (flooding standard) Source: World Bank. Question 1, part B: Estimation of benefits of considering the existing flood protection standard We document the costs of all three cases in table 4.4. Table 4.4. Key cost figures for Question 1, Part B (in millions of 2015 U.S. dollars) Cost of unmet energy and Capital cost Fixed O&M Variable cost reserves System cost Case 1 39,127 8,254 58,987 — 106,369 Case 2 39,346 8,211 59,502 — 107,059 Case 1* 42,343 8,254 58,987 — 109,585 Source: World Bank. First, the cost estimate obtained by Case 1 underestimates the cost of the power system plan because it ignores the construction cost required to comply with the flood protection standard. In particular, capital costs were estimated to be $39.1 billion under Case 1 but this increases to $42.3 billion when the additional cost is considered. So, we underestimate the cost by $3.2 billion or 3 percent when we use the status quo model. This is essentially the capital cost required to construct the flood control embankments for investments identified by Case 1. The $3.2 billion increase corresponds to an 8 percent increase in capital cost payments in the objective function. This is a significant omission in the current PSMP. Second, the major advantage of an enhanced model that includes more cost components is its ability to find a plan that takes into account these additional cost terms in order to identify a different plan. So, Case 2 could lead to a lower cost than Case 1 by differentiating the investment plan. As results indicate, the model with the flood protection standard (Case 2) is able to find a different investment plan, that costs $2.5 billion less than the plan identified by Case 1. So, it is clear that a more cost-efficient investment strategy is identified if the flooding standard is considered. The significant benefits of considering flooding standards should encourage power system planners to include flooding code considerations in future PSMPs. Question 1, part C: How different are the investment strategy and estimated operation of the power system when the flood protection standard is considered? 21 In the discussion above, we identified potential benefits by considering the flood protection standard in the planning phase. We see these benefits because the investment strategies in cases 1 and 2 differ, which lead to different estimates of operating costs between cases 1* and 2. In the following paragraphs, we will discuss in more detail the differences in timing, location, or type of investment between the two plans. We focus the discussion on two particular time periods, as follows. Differences before 2032. In the early years of the model horizon, we do not observe any significant difference in the capacity mix in terms of technology. The capacity mix remains the same despite the consideration of the flood protection standard, since some locations eligible for the development of power plants would not incur high costs for protection against flooding. However, the investments under Case 2 differ from investments under Case 1 with respect to location. In particular, under Case 1 all locations considered for power plant development are assumed to have identical costs and are equally preferable under the optimization model given that no transmission constraints are taken into account. Under Case 1, the optimal investment strategy includes investments in locations with high flood risk in the short term (for example, the Payra and Matarbari units are committed as early as 2021 under Case 1). By contrast, under Case 2, the optimization model prioritizes investments in locations with lower flood risk (for example, capacity invested in Payra and Matarbari by 2021 under Case 1 shifts to Meghnaghat, Orion Dhaka and Cox Bazaar under Case 2). In table 4.5, we provide capacity levels across locations considered for imported coal in four future years. There, the behavior described above is observed: investments in locations with high flood risk (closer to the bottom of the table) are postponed to a later year, while investments in locations with lower flood risk are implemented earlier. Capacities that differ between the two cases are highlighted in bold. Table 4.5. Capacity for power plants using imported coal across locations in GW (Case 1/Case 2) Location 2025 2030 2035 2040 a 1 Chandpur 0/0 0/0 0/0 0/0 2 Meghnaghat 0/1.3 0/1.3 0/1.3 1.3/1.3 3 Bheramara 0.1/0.1 0.1/0.1 0.1/0.1 0.1/0.1 4 Orion Dhaka 0.5/1 0.6/1 1/1 1/1 5 Mawa 0.5/0.5 0.5/0.5 0.5/0.5 0.5/0.5 6 Cox Bazaar 0/4.9 0/7.6 5.8/7.6 7.6/7.6 7 Zajira 0/0 1.8/1.8 1.8/1.8 1.8/1.8 8 Khulna South 0/0 0/0.4 0/4.7 2.2/4.7 9 Rampal 0/0 0/0 0/1.2 1.2/1.2 10 Ashugonj 0.6/0 1.3/0 1.3/0 1.3/1.3 11 Payra 2.3/0 4.7/0 4.7/0 4.7/4.7 12 Matarbari 3.8/0 3.8/0 3.8/0 3.8/1.2 13 Chittagong 0/0 0/0 0/0 2.6/0 Total capacity using imported coal 7.9/7.9 12.8/12.7 19.1/18.1 28.1/25.3 Source: World Bank Note: Capacities that differ between Case 1 and Case 2 are highlighted in bold. a. Capacity is 0 GW at this site because it is used for a gas power plant. 22 Based on the results discussed, it seems that one benefit of the enhanced model is its spatial granularity. Enhancing the status quo model with location-specific assumptions is crucial to represent the impact of a location-dependent risk such as flooding. The location of investments is the aspect that changes the most in the short term. To conclude the discussion of location-dependent flood risk, we provide a list of locations considered for the development of power plants that use imported coal, ranked from lowest to highest flooding-related costs in table 4.6. It is interesting to note how the sites with the highest flood risk are expected to be commissioned the soonest based on current plans provided by the Bangladesh Power Development Board. This commission timeline might seem counterintuitive but it could be explained either by the low consideration given to flooding during site selection or by factors other than flooding that might render the development of power plants in lower-flood-risk sites challenging. Table 4.6. Prioritization of sites for power plants using imported coal Depth at rp = 200 years (m) Expected Location Historical Projected commission year 1 Chandpur 0 0 2 Meghnaghat 0.1 0.1 3 Bheramara 0.1 0.2 4 Orion Dhaka 0.1 0 2021 5 Mawa 0.2 0.2 2021 6 Cox Bazaar 0.3 0.3 7 Zajira 0.5 0.2 8 Khulna South 1.2 1.7 2021 9 Rampal 1.5 2.0 2020 10 Payra 3.8 4.2 2019 11 Matarbari 7.1 7.5 2022–23 12 Chittagong 8.5 9.0 2020 Note: The source of the expected commission year of the plants in rows 4, 5, 8, 9, 10, and 12 is Bangladesh Power Development Board (2016). The source of the expected commission year of the plant in row 11 is Coal Power Generation Company Bangladesh Ltd. (2016). Differences after 2032. After 2032, the capacity mix between cases 1 and 2 differs even in terms of technology. In particular, during the period 2032–34 in Case 2 the interconnection capacity increases and is higher than in Case 1. After 2034, we observe multiple shifts in the capacity mix, with higher levels of both interconnection capacity and combined cycle gas units and lower capacity levels for coal units. The shifts in capacity mix also lead to a change in the energy mix. After 2035, the share of imported energy from India more than doubles in some years (as results in table 4.7 demonstrate). 23 Table 4.7. Energy mix for cases 1 and 2 during 2035–41 (percent) 2035 2036 2037 2038 2039 2040 2041 Domestic coal 11/11 13/13 12/12 11/11 12/12 11/11 10/10 Domestic gas 19/19 17/17 16/16 15/15 14/14 13/13 11/11 Imported coal 66/62 66/59 68/61 69/62 70/63 71/64 73/69 Interconnection 5/8 4/11 4/11 4/11 4/11 4/11 6/10 Source: World Bank. 4.2.2 Question 2. What are the benefits of considering how a power system depends on climate patterns? Question 2, part A: Experimental design To answer this question, we use three cases (table 4.8): • Case 1 could be considered the status quo, where no climate considerations are taken explicitly into account in the planning model. This is the same as Case 1 considered under Question 1, above; • Case 5 is an enhanced version of the status quo model (the climate-naïve version explained in table 4.2). It takes into account how flooding might affect the power system plan, and includes construction costs that the developer might incur to comply with a flood protection standard. It also considers how flooding might require additional flood insurance and lead to increased outages, and how temperature highs and cooling degree days might affect the power system. It is termed climate-naïve because it only considers historical climate data rather than climate change scenarios. • Case 1*** optimizes the approximate operations of the plan identified by Case 1 by considering the construction costs of compliance with a flooding standard and the impact of flooding and high temperatures or cooling degree days on capacity availability and demand. By comparing the cost of this case with Case 5, we can see how much cost is reduced in Case 5 by explicitly considering how flooding and climate change affect the investment plan. Table 4.8. Cases constructed to answer Question 2, Part A Investment Temperature Descriptive name decisions Flooding features features Case 1 Climate-ignorant Optimized OFF OFF Climate-naïve ON ON Case 5 Optimized (all features) Historical Historical Climate-ignorant Case 1*** in a future with historical Fixed from Case 1 ON ON Historical Historical climate Source: World Bank. Question 2, part B: Estimation of benefits We document the costs of all three cases in table 4.9. 24 Table 4.9. Key cost figures for Question 2, Part B (value in millions of 2015 U.S. dollars) Cost of unmet energy and Capital cost Fixed O&M Variable cost reserves System cost Case 1 39,127 8,254 58,987 — 106,369 Case 5 40,136 8,489 59,357 — 107,981 Case 1*** 42,343 8,495 59,830 62 110,730 Source: World Bank. The benefit is again two-fold, as it was under Question 1, above: • First, Case 1 underestimates the cost of the power system plan because it ignores the construction costs required to comply with a flood protection standard, flood insurance expenses, and any operational impact caused by capacity changes, outages, and demand increase. In particular, capital costs were estimated to be $39.1 billion under Case 1 but this increases to $42.3 billion when the flood protection standard cost is considered. Then, flood insurance expenses increase the fixed operation and maintenance (FOM) cost by $0.24 billion. Finally, the operational expenses increase by $0.9 billion. So, the status quo model underestimates total costs by $4.4 billion. • Second, considering historical climate data might lead to a different plan, and suggest different timing, location, or type of investments as better performing. As Case 5 demonstrates, there is such a plan, more expensive by only $1.6 billion compared with Case 1. So, it is clear that an updated investment strategy is identified if historical climate information is taken into account in the planning phase. Comparing Case 1*** and Case 5, it seems that the planner could have saved $2.7 billion by using a model that integrates historical climate information instead of relying on a climate-ignorant model to determine an investment plan. 4.2.3 Question 3. What are the benefits of considering climate change projections? As discussed in Trigg et al. (2016), flooding models for Bangladesh, and their interactions with climate models, are still in the early stages of their development. The same limitation applies to the downscaling methods used for climate projection models. Even so, we estimate the benefits of considering the best information available at the moment. For demonstration of the climate-features, we ignore uncertainty in this question and assume a particular combination of the current projections for climate change will be realized. 19 Question 3, part A: Experimental design As previously discussed, considering climate change projections in planning models has two benefits: (i) more accurate estimated costs and (ii) more cost-efficient investment decisions that lead to lower-cost (improved) power system plans. For each type of benefit, we compare several different models. 19 One scenario for climate projections is considered in this case: the “high” climate change scenario provided by FATHOM for flooding and the temperature/CDD projections provided by model cesm1-bgc at Climate Change Knowledge Portal Climate Analysis Tool-Powered by Climate Wizard. 25 For the first set of benefits, for each of cases 1, 2, and 5 described above we include cost components that correspond to flood insurance; compliance with a flood protection standard; and increased operational expenses due to additional load, higher derating, or outages. This helps us calculate underestimated costs. Table 4.10. Cases constructed to answer Question 3, Part A (first type of benefit) Investment Temperature Descriptive name decisions Flooding features features Case 1 Climate-ignorant Optimized OFF OFF ON Climate-ignorant Case 1** in a future with climate Fixed from Case 1 Projections ON (no flooding Projections change standard) Climate-ignorant with Partially ON Case 2 flooding standard Optimized (flooding standard) OFF Climate-ignorant with flood Case 2* protection standard Fixed from Case 2 ON ON in a future with climate Projections Projections change Case 5 Climate-naïve Optimized ON ON (all features) Historical Historical Climate-naïve ON ON Case 5* in a future with climate Fixed from Case 5 Projections Projections change Source: World Bank. In order to calculate the second type of benefits, we compare the cost of the four cases provided in table 4.11. Table 4.11. Cases constructed to answer Question 3, Part A (second type of benefits) Investment Temperature Descriptive name decisions Flooding features features ON ON Case 6 Climate-aware Optimized Projections Projections ON Climate-ignorant Projections ON Case 1** in a future with climate Fixed from Case 1 change (no flooding Projections standard) Climate-ignorant with flood Case 2* protection standard Fixed from Case 2 ON ON in a future with climate Projections Projections change Climate-naïve ON ON Case 5* in a future with climate Fixed from Case 5 change Projections Projections Source: World Bank. • Case 1** optimizes the approximate operations of the plan identified by case 1, respectively, including all the ways that projected climate conditions might affect power system operations and costs in our model (the impact of flooding and temperature/CDD change on insurance cost, capacity availability and demand), except for the cost of complying with a flood protection standard, assuming that the power plant developers do not build any infrastructure to protect the plant against flooding. • Cases 2* and 5* optimize the approximate operations of the plan identified by cases 2 and 5, respectively, including all the ways that projected climate conditions might affect power system 26 operations and costs in our model (that is, the construction costs of complying with a flooding standard and the impact of flooding and temperature/CDD change on insurance cost, capacity availability and demand); • Case 6 uses the same features as Case 5 but instead of historical climate conditions considers projected climate conditions in deciding on investments. Question 3, part B: Estimation of benefits For the first type of benefits (more accurate cost estimates), we summarize in table 4.12 the differences between the pairs of cases compared. This analysis aims to identify the additional costs that a planner would incur by following the plan prescribed by the relevant model. We assume that when Case 1 is implemented, no protection against flooding is constructed. Based on that assumption, for all cases, the capital costs are accurately estimated. However, other cost components are underestimated too; for example, the insurance cost expense (included under FOM) has been underestimated by $54 million– $3.115 billion and the variable costs by $2–3.3 billion. Table 4.12. Estimate of first type of benefits: Increase in operating costs as a result of considering flooding and climate change in operations (Question 3) (value in millions of 2015 U.S. dollars) Capital Fixed Cost of unmet cost O&M Variable cost energy and reserves System cost Case 1**- Case 1 0 3,115 3,302 355 6,771 Case 2*- Case 2 0 221 3,041 10 3,273 Case 5*- Case 5 0 54 2,005 10 2,069 Source: World Bank. For the second type of benefits (cost savings expected through implementation of more cost-efficient investment plans), we assess the level of benefits that the planner might gain by upgrading from one of the following three models: (i) the status quo, climate-ignorant model (Case 1); (ii) a climate-ignorant model with a flood protection standard (Case 2); (iii) a climate-aware model that uses historical climate data (Case 5); to (iv) a climate-aware model using climate projections (Case 6). By construction, Case 6 will identify the best plan since it considers the projected climate conditions as part of the optimization model, assuming perfect foresight. For the other model versions, it is expected that the more climate features a model has, the better the power plan identified will perform under projected climate conditions. As the last column of table 4.13 indicates, the benefit of upgrading to a climate-aware model considering a climate change scenario varies between $0.2 billion and $3.3 billion (or equivalently 0.2–3 percent) depending on the model we depart from. As expected, the benefit of switching to another model diminishes as the sophistication of the existing model increases. 27 Table 4.13. Cost categories used to estimate second type of benefits (Question 3) (value in millions of 2015 U.S. dollars) Cost of Increase in unmet cost energy and compared to Capital cost FOM Variable cost reserves System cost Case 6 Case 6 40,886 8,621 60,358 5 109,870 -- Case 1** 39,127 11,369 62,289 355 113,140 3,270 Case 2* 39,346 8,432 62,543 10 110,332 462 Case 5* 40,136 8,543 61,362 10 110,050 180 Source: World Bank. 4.3 Risk management using RDM and SLP In the previous sections, the discussion focused on the climate features the enhanced model has. To clearly demonstrate the capabilities and the mechanism of the enhanced model, we only presented results for deterministic models. However, as discussed in section 3 and in the assumptions documented earlier in this section, parameters used in the power system planning problem are uncertain. To handle uncertainty within the power-system-planning, we implemented two methods: RDM and SLP (see detailed explanation in the implementation in Section 3). Here, we discuss results obtained through both methods. Our discussion is structed around five questions. Four out of the five questions (questions a–d) addressed in this part of the report are specific to Bangladesh; one (question e) is methodological: a. How will additions to Bangladesh’s power system over the next 10 years be affected by uncertainty? b. Which set of distinct strategies could adequately represent the range of possible and relevant decisions, and should be tested in RDM? c. What reduced set of scenarios could be used in the SLP model to adequately describe the range of plausible future conditions represented by the original (larger) set of scenarios? d. How do the strategies selected in (b) perform across all plausible future conditions? e. How do SLP and RDM differ in their implementation and identified solutions? Could they be combined to complement each other? 4.3.1 Question a. How will additions to Bangladesh’s power system over the next 10 years be affected by uncertainty? In order to answer this question, we performed 486 deterministic runs to explore how different near-term investment decisions would be if a planner had perfect knowledge of the outcomes of a particular scenario. Reviewing near-term decisions (capacity additions up to 2025), we realize that investments in most candidate plant locations are quite similar20 across the 486 scenarios. But for three candidate 20 The minimum and maximum investments observed in any of the 486 scenarios differ by no more than 20 MW. 28 investments, significant differences are observed: (i) additional investment in interconnection with India varies between 1 GW to 3.5 GW, (ii) investment in coal capacity at Cox Bazaar varies between 4.5 GW and 6.4 GW, and (iii) investment in coal capacity using domestic coal at Barapukuria varies between 0.4 and 1.2 GW. Based on these three differences, we will design alternative strategies to consider under RDM. 4.3.2 Question b. Which set of distinct strategies could adequately represent the range of possible and relevant decisions, and should be tested in RDM? In its first step, RDM requires the analyst (in collaboration with stakeholders) to identify candidate strategies. For the power-system-planning problem, it is challenging to decide on a small number of candidate strategies since it seems practically impossible for the decision makers to enumerate all combinations of candidates and come up with a short list of candidate strategies. For this reason, we use the results of the 486 deterministic runs. In particular, we exploit the fact that the results differ significantly for three candidates and we use a partitioning method that clusters the 486 scenario-specific solutions based on the investment levels for these three candidates. In this application, we used a partitioning method called k-means clustering to provide different levels of investment for the three candidates.21 We provide the results of the clustering in table 4.14. In column 1 of table 4.14, we provide a brief description of cases under each cluster. Descriptions of climate scenarios (flooding or temperature) are not provided in the description of the cluster, because there is little systematic difference between the clusters in those dimensions. Similarly, descriptions of demand scenarios are not provided in the description of five clusters.22 On the contrary, fuel prices along with domestic fuel availability vary among the clusters. This might lead to the conclusion that the latter uncertainties might be more relevant to near-term investment decision than climate or demand. 21 For the remaining candidates, we fix their investment levels at the average capacity of levels recorded in cases belonging to the same cluster. 22 This is not the case for the scenarios considered for the stochastic programming, which look beyond 2025. There, almost all high-demand scenarios belong to different clusters than the base and low-demand scenarios. 29 Table 4.14. Candidate strategies for RDM New capacity Scenarios clustered under this added Number of strategy by 2025 in GW Near term strategy (capacity future (Names refer to the scenarios (coal at additions up to 2025) scenarios in considered; coal and gas Barapukuria, coal the cluster scenarios refer to supply) at Cox Bazaar, interconnection) Moderate coal investment (low Low coal and (high gas-IEA 450) or domestic), moderate 36 0.4, 5.8, 2.6 (low gas-WB17-High demand) interconnection Low coal and low gas and no High coal investment (low 54 0.4, 6.2, 3.5 WB17 domestic), high interconnection WB17 AND (low coal-high gas) or High coal investment (moderate (base coal-low gas) or (low coal- 72 0.6, 5.8, 1.1 domestic), low interconnection low gas-not high demand) IEA New policies-High/base Low coal investment (high 54 1.0, 4.8, 3.5 domestic coal-High gas domestic), high interconnection Moderate coal investment (high IEA 450-high/base domestic coal- domestic), moderate 54 1.0, 5.3, 2.6 High gas interconnection WB17 and (high coal-any gas) or Moderate coal investment (high 81 1.0, 5.1, 1.0 (medium coal-high gas) domestic), low interconnection (no WB17-low gas- high/base coal) High coal investment (high or (IEA new policies-high gas-low 135 0.9, 5.6, 3.4 domestic), high interconnection coal) Source: World Bank. The seven strategies identified seem to be a satisfactory discrete representation of the decision space for the following reasons. First, for each of the three candidates we observe that investment levels under the seven discrete strategies have range similar to the range provided by the full set of 486 scenarios. In particular, interconnection varies between 1.0 GW and 3.5 GW across the full set of 486 scenarios, and the same range of levels is kept under the reduced set of 7 cases. Domestic coal investment varies between 0.4 GW and 1.2 GW across the full set of 486 scenarios but has a slightly smaller range (between 0.4 and 1.0 GW) under the reduced set of 7 cases. Similarly, coal development at Cox Bazaar varies between 4.5 GW and 6.4 GW across the full set of 486 scenarios, while the range shrinks slightly (4.8 GW to 6.2 GW) under the reduced set of 7 cases. Second, the discrete strategies chosen seem reasonable because they include both strategies with investment levels at the bounds of the range and strategies consisting of moderate levels of investment in three candidates. Note that the strategies of table 4.14 could also represent the perspectives of stakeholders with different beliefs about the future (in particular the ones described by the cluster). For example: • A stakeholder who considers a future with low gas supply, low domestic coal supply, and fuel prices in line with IEA scenarios (cluster 2) would propose large short-term investments in both imported coal and interconnection. 30 • Alternatively, a stakeholder who considers high domestic coal, high gas supply, and prices in line with the IEA new policies scenario (cluster 4) might favor a plan with small short-term investments in imported coal, large investments in domestic coal, and large investments in interconnection. Similar descriptions, to the two provided above, reflecting stakeholder’s view favoring a particular strategy can be provided for the remaining 5 strategies of table 4.14. To conclude, the strategies seem to satisfactorily cover the decision space since the combinations incorporate different levels of investment across the three options. 4.3.3 Question c: What reduced set of scenarios could be used in the SLP model to adequately describe the range of plausible future conditions represented by the original (larger) set of scenarios? Under question b, we outlined candidate strategies to consider under RDM. For stochastic programming, however, we need to come up with a set of scenarios that covers the range of uncertainty but is also small enough to be included in a single linear programming optimization problem. Based on considerations of execution time and available memory, we decided to limit the number of scenarios to nine. We used a partitioning method called hierarchical clustering to create nine clusters of the 486 scenarios.23 In contrast to question b – where we clustered the 486 scenarios based on total investments over the first 10 years, here we cluster the 486 cases based on investments over the entire planning horizon ending in 2041.24 We look into the entire horizon because two-stage stochastic programming needs to look beyond the first stage in order to adequately represent the distinction between “here-and-now” and “wait-and- see” decisions. After identifying nine clusters, we choose one of the scenarios per cluster to represent its cluster25—that is, we used the input vector (assumptions regarding temperature, demand, fuel prices, coal/gas supply, flooding) of this scenario to describe the cluster in the stochastic program. A weight proportional to the number of original scenarios within the cluster is assigned to the representative scenario, assuming equiprobable scenarios. We summarize the scenarios, along with their weight, in table 4.15. A conceptual decision tree is provided in figure 4.2, accompanied by a brief description. 23 The distance metric used for the hierarchical clustering is the Euclidean distance. 24 In particular, we sum up the investments for each candidate every 5 years for the first 15 years and then consider the rest of the years together. Then, we discard any candidates with a range lower than 200 MW (the “range” is defined as the difference between the maximum and minimum level of investment across the full set of 486 cases). Finally, we organize the 486 scenarios into 9 clusters based on the level of investments for the remaining candidates. 25 For this application, we choose the scenario with the smallest maximum distance from any scenario within the same cluster. 31 Table 4.15. Scenarios retained for stochastic programming Scenario selected to represent the Probability Cluster description cluster (%) Low/base demand and high demand—high Base demand—IEA 450—high 70.4 coal—high gas—IEA scenarios coal—high gas High demand—IEA 450—low coal— High demand—low coal—high gas 5.6 high gas High demand—IEA new policies— 5.6 High demand—medium coal—low gas medium coal—low gas High demand—IEA new policies— High demand—low coal—low gas 5.6 low coal—low gas High demand—IEA scenarios—high coal—low High demand—IEA new policies— 3.7 gas high coal—low gas High demand—IEA scenarios—medium coal— High demand—IEA 450—medium 3.7 high gas coal—high gas High demand—WB17 prices—high coal—low High demand—WB17—high coal— 1.9 gas low gas High demand—WB17 prices—medium coal— High demand—WB17—medium 1.9 high gas coal—high gas High demand—WB17 prices—high coal—high High demand—WB17—high coal— 1.9 gas high gas Source: World Bank. Note: The representative scenario also includes specific cost estimates for temperature/CDD and flooding, but these are not included for the sake of brevity. The discrete scenarios that were identified by the above procedure consist of (i) a central scenario with high weight (70 percent) that represents all cases with base/low demand and high demand with high fuel supply, and (ii) eight other scenarios with lower probabilities that describe distinct cases, all involving high demand but different values of the other variables. This discrete scenario set seems reasonable because we—as planners—would like to optimize the system’s plan for a central scenario, but at the same time take into account “extreme” cases that would challenge a plan based on the assumptions of the central scenario. In that way, we can make sure that near-term investment decisions consider low-probability cases that might put high stress on a plan based solely on a central scenario. Moreover, the inclusion of extreme scenarios provides us a way to estimate the range of costs and regret26 across scenarios using stochastic programming results. In that manner, we can quickly obtain an estimate of the “robustness” of the identified plan across scenarios. 26 Regret is defined as the cost of the identified plan minus the cost of the best-performing plan under the same case. 32 Figure 4.2. Decision tree for SLP model used in step 3b (Figure 3.1) under the hybrid SLP Source: World Bank. The decision tree provided in figure 4.2 illustrates the formulation of the model employed under step 3b. In particular, nine scenarios (presented in table 4.15) are considered as part of the stochastic programming formulation. The planner faces an uncertain future when attempting to determine the investment decisions for the first 10 years of the horizon. However, we assume that the uncertainty clears during system operation (for example, fuel prices are known when unit dispatch is decided). Moreover, we assume that the investments that come on-line after 2025 can be scenario contingent because the planner will have greater certainty regarding key parameters at that time. 4.3.4 Question d: How do the strategies selected in (b) perform across all plausible future conditions? According to table 4.16, the strategy identified through k-means, described as “Moderate coal investment (low domestic), Moderate interconnection,” has the lowest average regret and cost across the 486 scenarios among the strategies considered. This strategy also happens to have the lowest maximum regret. However, in case the planner is concerned with not only the average cost but also the maximum cost that could be incurred across all the scenarios, the strategy described as “High coal investment (low domestic), High interconnection” might be recommended. This strategy has an average cost just $20 million higher but its maximum cost is about $300 million lower. 33 Table 4.16. Performance of strategies across all 486 scenarios (in millions of 2015 U.S. dollars) Minimum Expected Maximum Near-term strategy according to: Maximum regret regret regret cost Stochastic model 622 46 267 131,327 Moderate coal investment (low domestic), moderate interconnection 436 31 231 131,145 High coal investment (low domestic), high interconnection 648 2 251 130,853 High coal investment (moderate domestic), low interconnection 1,140 52 600 131,987 Low coal investment (high domestic), high interconnection 1,043 51 424 131,598 Moderate coal investment (high domestic), moderate interconnection 981 47 408 131,735 Moderate coal investment (high domestic), low interconnection 2,200 29 980 133,047 High coal investment (high domestic), high interconnection 927 59 344 131,336 Source: World Bank. An interesting observation based on results from table 4.16 is that the near-term investment strategy identified by SLP is not the best-performing strategy on average. In theory, the strategy specified by the stochastic model should be the one with the lowest expected regret, at least for the set of scenarios considered under the SLP. This is indeed the case as results in table 4.17 indicate. However, when considering the full set of 486 scenarios, this is not the case (table 4.16); instead it is only the third best strategy in terms of expected regret. Therefore, it seems that organizing 486 scenarios into 9 clusters did not identify the best-performing strategy across all 486 scenarios. However, the best strategies outperform the SLP solution by only $16-$36 million. 34 Table 4.17. Performance of strategies across nine discrete scenarios considered by SLP (value in millions of 2015 U.S. dollars) Maximum Minimum Expected Maximum Near-term strategy regret regret regret a cost Stochastic model 493 75 148 129,332 Moderate coal investment (low domestic), moderate interconnection 343 76 245 129,130 High coal investment (low domestic), high interconnection 555 7 274 128,882 High coal investment (moderate domestic), low interconnection 1,104 54 707 129,972 Low coal investment (high domestic), high interconnection 827 57 208 129,612 Moderate coal investment (high domestic), moderate interconnection 904 50 194 129,753 Moderate coal investment (high domestic), low interconnection 2,116 30 839 130,984 High coal investment (high domestic), high interconnection 536 65 230 129,351 Source: World Bank. a. If a strategy has the lowest average regret, by construction of the regret cost, it also has the lowest average cost. In tables 4.18 and 4.19 we provide the capacity and energy mix for the strategy specified as “Moderate coal investment (low domestic), moderate interconnection.”27 Table 4.18. Capacity mix identified under the strategy with the minimum average regret (in GW) Combined Gas Combustion cycle gas Steam engine turbine turbine turbine Interconnection PV Biomass 2025 0.7–1.5 0.25–0.4 7.0–7.4 9.2–9.3 3 2030 0.7–1.5 0.25–0.4 6.6 14–18 3–4 2035 0.2–0.8 0.1–1.3 6–12 20–30 2.5–5.7 2040 0 0–4.7 2–17 25–38 2.5–7.7 0–8.5 0.–0.3 Source: World Bank. 27 We have also examined the capacity mix under a stochastic (SLP) strategy, which is very similar to the one reported in table 4.18. 35 Table 4.19. Energy mix identified under the strategy with the minimum expected regret (percent) Domestic Domestic Energy Imported coal gas efficiency Imports coal LNG Oil 2016 2–3 73 0–7 17–25 2020 2 58–60 11–14 11–12 14–17 2025 3 24–29 10–13 57–58 0–6 2030 6–15 13–21 4–11 54–73 0–3 2035 6–19 8–21 0–11 56–83 0 2040 5–18 6–16 0–4 0–19 58–83 0–13 Source: World Bank. 4.3.5 Question e: How do SLP and RDM differ in their implementation and identified solutions? Could they be combined to complement each other? It is worth understanding the value of individual techniques (namely SLP and RDM) and also if their comparison reveals additional insights. A full comparison of the two methods is out of the scope of this analysis but here we will discuss any insights obtained through application of both methods on the same problem. As mentioned earlier, recommendation of a particular investment strategy is out of the scope of the analysis. That’s why we did not attempt to choose one strategy to recommend to the power system planner. We discussed though under question d the performance of all 8 strategies (7 from RDM and one from SLP) across three metrics: average regret, maximum regret and max cost over the full range of scenarios. As an example, we could assume that a single metric is used: average regret. In theory, we would expect the SLP to recommend the strategy with the lowest average regret. However, this does not seem to be the case in practice. A strategy identified within our RDM implementation, has the lowest average cost over the full range of scenarios. Thus, it would appear that for this particular case and example, relying on strategies followed under the RDM offers a better-quality solution than using the SLP approach of relying on a small set of representative scenarios. However, the outcome of this particular case study cannot be generalized to a conclusion that would apply to all countries or even under other assumptions for Bangladesh. Nonetheless, we can draw several insights from our example: 1. SLP relying on a reduced set of scenarios might not identify the strategy with the lowest expected regret across all scenarios. Testing the performance of the strategy under examination, even if it has been identified through SLP, across the full set of scenarios might offer valuable information on the performance of the strategy across scenarios not included in the optimization problem. 2. In practical cases where many scenarios might be plausible, SLP’s performance is dependent on the quality of the scenario reduction or sampling method used. Thus, it might be valuable to apply more than one methods to identify a set of representative scenarios to consider under the SLP. 3. On the other hand, in practical cases with many plausible scenarios, implementation of RDM seems more difficult and execution time might be long. The number of runs the 36 planner has to do increases significantly and computational skills in fields such as parallelization of model runs might be needed to complete the task in a timely manner. For example, in this analysis the SLP solution was identified within 90 minutes. In the contrary, it took between 210 to 310 minutes to conduct the vulnerability assessment for each RDM strategy. Given the complementarities in strength and weaknesses of each method, as part of future research, it might worth to try different heuristics integrating aspects of both methods into a single method. As an example, we mention a heuristic that would have worked in this particular example. This heuristic relies on the reduced set of scenarios, identified for SLP to screen promising strategies for further vulnerability analysis within RDM. Note however that there is no guarantee that the described heuristic would have been successful for another example. Applying this heuristic to our case, we assume that we implement step 2 for both methods. So, we have identified the seven strategies described in table 4.14 (RDM) and we have also discretized the uncertainty space into 9 scenarios described by table 4.15 (SLP). Instead of running the SLP model or jumping into the vulnerability assessment of all 7 RDM strategies over the full set of 486 scenarios we could first test the performance of the 7 strategies across these 9 scenarios. Then, if any of the 7 strategies demonstrates unsatisfactory performance across a criterion of our choice, we could eliminate it. For example, in table 4.17 we could decide to eliminate the 4 strategies with highest maximum regret among the seven (colored with yellow). Subsequently, for the remaining three strategies (the ones that passed the first screening) we would have to conduct the full vulnerability analysis. Note that if the strategy that has the best performance according to the planner’s metric is not eliminated from the heuristic, time might be saved by skipping the vulnerability analysis of the eliminated strategies over the full set of scenarios (i.e., which involved more than 10 hours of computational time in this example). A thorough comparison of both approaches is a subject of future research. The insights from the current example might have limited applicability for the following reasons: 1. Only one planning problem is considered. It is possible that the structure of our problem might favor one method over another. For example, in our problem, decisions across plausible scenarios differed only along a few dimensions which may have made the selection of discrete strategies easier than the selection of representative scenarios. 2. Only one clustering method for scenario reduction is used in stochastic programming. It is widely known that different clustering and sampling methods might lead to different sets of representative scenarios, which could significantly influence the performance of the SLP. 3. No theoretical justification for the conclusions is provided that would guarantee applicability of the conclusions to other examples. 4. We did not complete all steps described in Figure 3.1 due to the absence of stakeholders to consult on what consists acceptable performance. Moreover, absence of stakeholders also limits the conclusions of the comparison since we might miss any important differences on the way both methods facilitate stakeholder engagement. 37 5 5. SUMMARY AND WAY FORWARD 5.1 Summary of major findings Bangladesh’s power system master plans of 2005, 2010, and 2015 reflect three separate analyses, undertaken over an 11-year span. These years saw significant shifts in the nation’s energy mix, including from gas to coal. While comprehensive, the plans for domestic generation to some extent overlooked climate-related risks, such as flooding and extreme heat waves. This present study represents a first attempt to explicitly integrate climate risks in models that underpin power system planning. The focus here is Bangladesh, but the implications are global. The study asks if there is a manageable way to undertake risk assessment in Bangladesh, a country known to be among the most vulnerable to climate change in the world. The findings are summarized below under two headings: methodology and significance of impact. 5.1.1 Methodology It is possible to integrate flooding and temperature projections within a power-system-planning framework. To achieve this integration, existing models and datasets need to be enhanced in two respects. First, relationships between climate-related variables and planning parameters need to be identified in order to create models of different climatic conditions. Second, the possible locations of future power plants need to be mapped as accurately as possible, because hazards such as floods differ significantly by location. Given that the relationships of climate-related variables to planning parameters, such as inundation depth and power plant damage, might differ by power plant design, planners would do well to engage power plant developers in efforts to understand risks and costs. Both the methods applied here, hybrid-SLP and RDM, offer a big-picture view of the future that will need to be filled in as data become available. Findings should be considered in the context they are provided: within constraints of available time, resources, and information. Modifying the standard least-cost planning model in order to manage risks is quite straightforward under either of the two methods. Using hybrid-SLP, the least-cost planning model remains the same but the decision space expands to include several scenarios, whose costs may be weighed against one another. In the case of RDM, the standard least-cost planning model is implemented iteratively across all future scenarios, with some of the decisions fixed to test the efficacy of various risk management strategies. The two methods might complement each other in the following way. SLP, and a small set of discrete scenarios, might help quickly identify promising strategies (and drop others). Then, RDM (in particular its vulnerability component) may be used to test the proposed strategies. When comparing various strategies across scenarios, it is important to clarify that the efficacy of a particular strategy is a distinct value, separate from the probability of its realization. In other words, the ideal option is not always the best option, given practical constraints. 38 5.1.2 Significance of impact Standard least-cost planning methods may underestimate capital costs if they omit the costs of protecting against floods. In the case of Bangladesh, a plan that ignores flood risk would underestimate capital costs by as much as $3 billion over the next 25 years. Although it is reasonable to assume that a significant portion of these costs would eventually be recognized at the project level, it is of paramount importance that they be recognized at the planning stage. This applies to not only flood risk, but any and all climate-related protection and resilience measures. Under one of several worst-case scenarios, failure to consider these costs overlooks a staggering $7 billion in needed funds. Candidate locations for power-plant development should be prioritized based on their flood risk profiles. A climate-aware plan that explicitly considers flood risk can optimize decisions regarding plant location and alternative generation technologies to save $0.2 billion–3.3 billion over the planning period. It is worth noting that of the plants being developed in Bangladesh today, most are on sites at high risk for flooding (which, by this study’s definition, would put them at a low priority for development). Uncertainties regarding fuel supply and fuel prices seem to introduce a higher degree of uncertainty to the generation mix than do uncertainties regarding demand growth and climate. (This conclusion follows from the clustering pattern of near-term strategies, which seems to be decided by assumptions regarding fuel availability and prices.) The best performing (with respect to average cost) investment plan in the present study seems to be in agreement with the PSMP released in September 2016 with regards to near-term investment decisions (up to 2025). However, several differences between the two plans may be observed: • The power system plan proposed in this report suggests the interconnection of 3 GW by 2025 (slightly more than the 2.5 GW proposed by the PSMP 2016) and estimates the share of imports in the energy mix at 10–13 percent. • The power system plan proposed here suggests slightly larger investments in coal capacity by 2025: 9 GW rather than the ~7GW of the PSMP 2016. However, the price of CO2 emissions was set at zero in our analysis, while in the PSMP 2016 a CO2 price of $125 per ton was used. • The locations recommended here for the development of coal power plants using imported coal differ from the locations suggested in the PSMP 2016. The PSMP 2016 considers development only at Matarbari and Payra. Both of these locations are identified as having high flood risk in our analysis, and they are chosen for later development in only a subset of 486 scenarios. This study prefers the lower flood risk of, for example, Cox Bazaar. However, in case other concerns, such as proximity to a tourist destination, eliminate a recommended location, we could exclude it and then examine if the model still prefers investments in coal or prioritizes investments in less-flood- prone areas and/or different technologies. • The power system plan proposed here is fully adaptive to conditions post 2025, assuming that uncertainty regarding fuel prices, supply, and demand, as well as climate change will be resolved. In case planners in Bangladesh adopt a regular planning cycle such that their confidence in near- term projections increases over time, a framework such as the one adopted here offers different recommendations after the resolution of uncertainty (2025 assumed here). For example, 39 depending on conditions, the share of power imports in 2041 varies between 1 percent and 20 percent, while the overall share of coal in the mix varies between 65 percent and 95 percent. 5.2 Way forward This exercise is a modest beginning to what is potentially a very large task, especially considering that in a number of other countries, the impact of climate change on the power sector might be significant. The adoption of its methodology in the regular planning cycle—for instance, as part of the next PSMP in 2020 or thereabouts—promises to be a major undertaking. It would require: • Developing an awareness of these issues, especially among utilities, planners, and policy makers; • Developing necessary tools and climate-related information for the energy sector; • Building capacity to undertake analysis in a meaningful way; and • Consciously paving the way for supply- and demand-side options that render the system more resilient to climate change. Each of these are indeed substantial tasks that go beyond the scope of this exercise, but we outline key issues relevant to potential follow-up activities. 5.2.1 Adoption of the methodology by Bangladesh’s power-planning body Building awareness. Although some of the issues around climate change are known in Bangladesh in qualitative terms, these are not receiving adequate consideration in sector planning. It is, for instance, well known that there are flooding risks at several plant sites. For one of the JICA plants, a site at an elevation of 11 meters is being considered, and at a high cost of several hundred million dollars. Recent heat waves have led to soaring peak demand. However, the risks of flooding and rising temperatures have not yet been addressed in a PSMP. The first task, therefore, is to raise awareness among all parties involved in the planning process, including domestic developers and civil servants, donor agencies, and investors. The present study might be a starting point for that discussion, and ways should be found to engage stakeholders in the collection and analysis of data. For example, when faced with uncertainty in various parameters, we have relied largely on our own set of assumptions. Better information—collected through a collaborative workshop, for example—would vastly improve forecasts. This in turn is important to creating a context for climate resilience. Exploring which supply options render the system more resilient is a first step toward reducing or eliminating those options that expose the system to high costs or the risk of supply failure. Building climate data. One of the constraints we faced in our present analysis is a lack of sufficiently downscaled and reliable data on flooding and temperatures. Inundation depth is the key parameter used in this study to model the impacts of climate change on Bangladesh’s power system. Yet FATHOM’s available flood level projections have been validated only in other parts of the world, such as the United Kingdom. Meanwhile, significant differences in the flooding projections of available global models have been observed in delta areas in Africa. Moreover, climate model projections do not list the extreme temperatures that have already been observed in the Bangladesh (none of the maximum temperatures provided for 2015–41 is higher than 41℃). Datasets are also incomplete; for example, it remains difficult to assess the quality of various renewable energy resources across a range of climate scenarios. While 40 these data issues did not make our illustrative analysis impossible, a full-scale PSMP must use a more complete dataset. Describing the power system’s interaction with climate change in quantitative terms. How climate variables relate to power system operations and costs needs to be described more accurately. This study represents a first attempt, using available empirical studies and anecdotal evidence. As a second step, we could engage with key stakeholders including BPDB (generation), Power Cell (policy issues), PGCB (transmission), BERC (regulatory matters), and JICA/TEPCO (implementation of PSMP) to discuss our assumptions and update our findings as needed. In the long term, planners would do well to develop models that more accurately project peak loads as contingent on temperature, humidity, and wind speed, and to provide a list of alternative power plant configurations and their costs. Finally, the impact of climate change on other power system components such as the transmission grid and distribution system is important to consider when building a climate-aware planning model. Building supply- and demand-side options more widely. The PSMP 2016 provides a comprehensive analysis of the resources available for electricity generation in Bangladesh. However, the plan simulates scenarios using a pre-decided contribution of each source in the generation mix. According to our analysis, it seems that under an optimal plan, the contribution of each resource far in the future (post 2025) might vary depending on fuel prices, fuel supply, and power demand. Also, the role of regional interconnection seems to be important. In the PSMP 2016 and the plan presented here, interconnection of 2.5 GW and 3.0 GW, respectively, is recommended by 2025, satisfying ~10-13 percent of demand. Then, under the PSMP 2016, the interconnection capacity reaches 9 GW by 2041, while under our plan, it varies between 3 GW and 8 GW. (We endogenously considered the costs of a transmission connection with an adjacent system, and adopted assumptions for profiles of hydropower generation based on historical data; under these assumptions, the optimal plan mainly developed interconnection with India but only rarely pursued interconnection with Nepal and Bhutan.) Given the objectives for renewable procurement from neighboring systems, further study of interconnection opportunities might be of benefit. This would require effort to look into cross-border trading options on a larger scale than previously attempted. Such a study might consider, among other things, the possibility of integrating Bangladesh’s power market into that of India, and eventually into a subregional (for example, the SASEC28) market and then a SAARC-wide electricity market.29 Building climate-aware planning tools. Although there is a growing volume of academic literature on the topic, there is no tool—not even a research-grade tool—available that can be used to perform the proposed stochastic programming analysis. Two solutions may be considered, neither specific to the Bangladesh application. First would be to develop a special purpose tool, similar to the GAMS30 based 28 South Asia Subregional Economic Cooperation: India, Bangladesh, Sri Lanka, Nepal, and Bhutan. 29 The Bangladesh Analytical and Strategic Advisory activity recently initiated a short desktop study of this very issue, titled Bangladesh: Scaling up Cross-border Power Trade and Developing a Regional Market, mimeo, September 2016. It is being prepared by Mr. Ravinder (former CEO of India’s Central Electricity Authority) under the guidance of Sheoli Pargal (TTL). 30 General Algebraic Modeling System (www.gams.com). 41 model we used for this analysis, and certainly a possibility for Bangladesh because it has been built around PSMP assumptions. Second would be to integrate climate resilience analysis into one of the existing power system planning tools, which would take more time and cost more, but would also provide a more permanent solution that could be adopted for other countries.31 RDM is, more often than not, an iterative methodology consisting of multiple implementations of a model that simulates the performance of candidate strategies. In this instance, the simulation model could be a standard tool used for power system planning, with investment decisions fixed at levels implied by the strategy. It should be noted that all three options (two for stochastic programming, and RDM, as noted above) must develop a least-cost planning methodology. The rigor of a planning methodology should not be compromised—for instance, a climate-resilient plan should still consider integrated capacity and dispatch optimization, reserve, fuel, transmission constraints, and so on. Climate resilience needs to be integrated into this framework, using quantitative data as available; planning such resilience should not evolve into a soft technique with an inadequate level of technical details. The chance that an electric utility planning body would accept a capacity plan underpinned by rigorous analysis seems much higher than for it to adopt a completely new technique. In the case of Bangladesh, we consider the model we have developed as a good starting point— until there is a standard tool (e.g., a beta version of PLEXOS) available for BPDB to consider. 5.2.2 Institutional capacity issues For the technical steps discussed above to be successfully implemented, it is important to address institutional issues, including a need to build generation planning capacity within Bangladeshi power supply bodies. To start with, Power Cell, which resides in the Ministry of Energy Mineral and Resources (MEMR), and the Bangladesh Power Development Board, have until now filled an executive role in power system planning. Over the past decade or so, PSMPs have been funded by the Asian Development Bank and JICA, and executed by consultants, including Nexant and TEPCO. The system today is sufficiently large to warrant a planning body of its own. Generation capacity rapidly increased in the last 7–8 years, by near 7 GW (or 140 percent), to cross 12 GW, and is expected to maintain a high growth rate for the next 25 years. Bangladesh faces several challenges around fuel supply, transmission, siting, technology, etc., along with the climate issues discussed here. All these issues are unique to its particular context. A planning unit that includes a highly skilled and experienced Bangladeshi engineering crew is needed to instill a clear understanding of these issues in the plan. Given the high growth of capacity and demand, it is also important that the plan be updated more often than every five years. Power Grid Bangladesh has already introduced a reasonable level of proficiency into transmission planning, and this needs to be replicated on the generation side by a core group of experienced engineers, modelers, and economists from Power Cell, BPDB, and BERC. 31 We have explored this option briefly with Glenn Drayton, CEO of Energy Exemplar, which has developed and marketed PLEXOS—one of the leading power system and market simulation tools. PLEXOS offers flexibility by making it relatively easy to add new variables and constraints. PLEXOS also includes a stochastic unit commitment model; this means that the two-stage stochastic model is already available, albeit in a different (short-term operational) context. 42 The planning endeavors need to start with a least-cost planning model, but as noted above should integrate climate data using a stochastic formulation to explore the many trade-offs that exist for a power plan in Bangladesh. Since there are off-the-shelf models and a dataset already available for the least-cost planning portion of this exercise (e.g., the Nexant/TEPCO model or the Bank’s model), the starting point is reasonably easy to find. Maintaining, updating, and extending the model would, however, require a wider group of actors to support the core planning team. The inputs required include: • PGCB transmission planning, and NLDC data, to integrate information on the transmission plan and operational dispatch (including constraints on generation dispatch) in the planning model; • Longer-term information on fuel, generation, and cross-border trade contracts, primarily from BPDB so that the commercial reality is fully reflected in the plan (and compared and contrasted with an economic outcome); • An active linkage with one or more climate data agencies who can support the planning team with accurate, downscaled climate data on coastal and riverine inundation, and long-range temperature, wind, and solar forecasts; and • Representation from BERC in regulatory and electricity market issues. The latter would be particularly important as Bangladesh expands its interconnection with India and potentially in future with Bhutan, Nepal, and Myanmar. As the discussion suggests, a major program of activities would be needed to form a planning unit, impart training, and develop required institutional linkages. Clearly, a significant share of such activities are needed regardless of climate-related issues. Nonetheless, planning for climate resilience is a prerequisite for any planning analysis in Bangladesh. We end this discussion with a set of concrete recommendations for next steps following the completion of the final report of this study, namely: • Hold a workshop to present the findings of this analysis, with a focus on comparing it against the PSMP and identifying some of the climate-induced vulnerability issues; • As a follow-up activity, form a core team of engineers, modelers, and economists from relevant stakeholders to discuss the detailed inputs and findings of the study, with a view to identifying inputs/assumptions that could be refined; • Undertake a training course on the specific model implementations covering both least-cost and robust decision making techniques, and also commercially available planning tools; • Hand over the Bank models and data (including climate data collected as part of the project) and engage with teams to help them update the analysis; and • Identify longer-term resource requirements to bolster and maintain planning work in Bangladesh. 43 6 6. REFERENCES Bangladesh Climate Change Resilience Fund. n.d. “Climate Change Knowledge Portal, Climate Analysis Tool.” World Bank Group and International Center for Tropical Agriculture. http://climatewizard.ciat.cgiar.org/wbclimateanalysistool/. Bangladesh Power Development Board. Annual Report 2015-2016. http://www.bpdb.gov.bd/download/annual_report/Annual percent20Report percent202015-16.pdf. Birge, J. and F. Louveaux. 2011. Introduction to Stochastic Programming. Springer. Bruninx, K. and E. Delarue, 2016. "Scenario reduction techniques and solution stability for stochastic unit commitment problems" 2016 IEEE International Energy Conference (ENERGYCON), Leuven, pp. 1-7.doi: 10.1109/ENERGYCON.2016.7514074 Cervigni, Raffaello, Rikard Liden, James E. Neumann, and Kenneth M. Strzepek, eds. 2015. Enhancing the Climate Resilience of Africa's Infrastructure: The Power and Water Sectors. Washington, DC: World Bank. Coal Power Generation Company Bangladesh Limited. 2016. Annual Report 2015-2016. http://www.cpgcbl.gov.bd/wp-content/uploads/2016/12/Annual-Report-15-16- compressed.pdf. Dittrich, R., A. Wreford, and D. Moran. 2016. “A Survey of Decision-Making Approaches for Climate Change Adaptation: Are Robust Methods the Way Forward?” Ecological Economics 122: 79–89. http://doi.org/10.1016/j.ecolecon.2015.12.006. Government of Bangladesh. 2015. “Scaling Up Renewable Energy in Low-Income Countries (SREP) Investment Plan for Bangladesh.” Sustainable and Renewable Energy Development Authority, Power Division, Ministry of Power Energy and Mineral Resources, Government of Bangladesh, Dacca. https://www- cif.climateinvestmentfunds.org/sites/default/files/meeting- documents/bangladesh_srep_ip_final.pdf. Government of India. n.d. “Monthly Executive Summary on Power Sector.” Ministry of Power, Central Electricity Authority, New Delhi. http://www.cea.nic.in/reports/monthly/executivesummary/2016/exe_summary-10.pdf. Hallegatte, S., A. Shah, C. Brown, R. Lempert, and S. Gill. 2012. “Investment Decision Making under Deep Uncertainty—Application to Climate Change.” Policy Research Working Paper, World Bank, Washington, DC. Harvey, C. 2016. “A new power plant could devastate the world’s largest mangrove forest,” 44 Washington Post, https://www.washingtonpost.com/news/energy- environment/wp/2016/07/18/a-new-power-plant-could-devastate-the-worlds-largest- mangrove-forest/?utm_term=.bdcb93eccc8e. Hirabayashi, Y., R. Mahendran, S. Koirala, L. Konoshima, D. Yamazaki, S. Watanabe, H. Kim, and S. Kanae. 2013. “Global Flood Risk under Climate Change.” Nature Climate Change 3 (9): 816–21. http://doi.org/10.1038/nclimate1911. Ho, J. L., B. F. Hobbs, P. D. Vallett, Q. Xu, S. Kasina, S. W. Park, and Y. Ouyang. 2016. Planning Transmission for Uncertainty: Applications and Lessons for the Western Interconnection. https://www.wecc.biz/Reliability/Planning-for-Uncertainty-Final-Report.pdf. IEA (International Energy Agency). 2015a. India Energy Outlook. World Energy Outlook Special Report. Paris: IEA. http://www.worldenergyoutlook.org/media/weowebsite/2015/IndiaEnergyOutlook_WEO 2015.pdf. ———. 2015b. World Energy Outlook 2015. London, United Kingdom: IEA. http://www.worldenergyoutlook.org/weo2015/. JICA and TEPCO (Japan International Cooperation Agency and Tokyo Electric Power Company). 2016a. “Survey on Power System Master Plan 2015: Draft Final Report.” April. http://powerdivision.portal.gov.bd/sites/default/files/files/powerdivision.portal.gov.bd/pa ge/4f81bf4d_1180_4c53_b27c_8fa0eb11e2c1/final report_2016.pdf. JICA and TEPCO (Japan International Cooperation Agency and Tokyo Electric Power Company). 2016b. “Power System Master Plan 2016.” Power Division, Ministry of Power, Energy, and Mineral Resources, Government of Bangladesh, Dacca. http://powerdivision.portal.gov.bd/site/page/f68eb32d-cc0b-483e-b047- 13eb81da6820/Power-System-Master-Plan-2016. Khan, I., F. Alam, and Q. Alam. 2013. “The Global Climate Change and Its Effect on Power Generation in Bangladesh.” Energy Policy 61: 1460–1470. Knutti, R. 2010. “The End of Model Democracy?” Climatic Change 102 (3): 395–404. http://doi.org/10.1007/s10584-010-9800-2. Kusche, P. and Coyne, L. 2015. " How to Utilize the Influx of Capital Markets Capacity , Including CAT Bonds" RIMS 2015, New Orleans Lempert, R. J., D. G. Groves, S. W. Popper, and S. C. Bankes. 2006. “A General, Analytic Method for Generating Robust Strategies and Narrative Scenarios.” Management Science 52 (4): 514–28. http://doi.org/10.1287/mnsc.1050.0472. Loulou, R., M. Labriet, and A. Kanudia. 2009. “Deterministic and Stochastic Analysis of Alternative Climate Targets under Differentiated Cooperation Regimes.” Energy Economics 45 31(Suppl. 2): S131–43. http://doi.org/10.1016/j.eneco.2009.06.012. McNeil, M.A. and V. E. Letschert. 2008. “Future air Conditioning Energy Consumption in Developing Countries and What Can Be Done about It: The Potential of Efficiency in the Residential Sector.” Lawrence Berkeley National Laboratory, University of California. Ministry of Power, JIC Agency, and TTE Power. 2011. “Power System Master Plan 2010.” General Economic Division, Dhaka, Bangladesh. Nunes, J. B., N. Mahmoudi, T. K. Saha, and D. Chattopadhyay. 2016. “Stochastic Generation and Transmission Planning Considering Future Renewable and Gas Opportunities in Queensland, Australia.” In 2016 IEEE Power and Energy Society General Meeting (PESGM), 1–5. Boston, MA: IEEE. http://doi.org/10.1109/PESGM.2016.7742027. OECD (Organisation for Economic Co-operation and Development). 2015. “Tools to Mainstream Adaptation into Decision-Making Processes.” In Climate Change Risks and Adaptation: Linking Policy and Economics. Paris: OECD Olson, R., Y. Fan, and J. P. Evans. 2016. “A Simple Method for Bayesian Model Averaging of Regional Climate Model Projections: Application to Southeast Australian Temperatures.” Geophysical Research Letters 43 (14): 7661–69. http://doi.org/10.1002/2016GL069704. Pargal, S. 2017. “Lighting the Way: Achievements, Opportunities, and Challenges in Bangladesh’s Power Sector. Unpublished report, World Bank, Washington, DC. May. Planning Commission. 2012. “Perspective Plan of Bangladesh 2010–2021: Making Vision 2021 a Reality.” Government of Bangladesh. http://bangladesh.gov.bd/sites/default/files/files/bangladesh.gov.bd/page/6dca6a2a_985 7_4656_bce6_139584b7f160/Perspective-Plan-of-Bangladesh.pdf. Premkumar, L. 2016. A Study of the India-Bhutan Energy Co-Operation Agreements and the Implementation of Hydropower Projects in Bhutan. New Delhi, India: Vasudha Foundation. http://www.vasudha-foundation.org/a-study-of-india-bhutan-energy-agreements-and- the-implementation-of-hydropower-projects-in-bhutan/. PRI (Policy Research Institute) of Bangladesh. 2015. “PRI Quarterly Policy Briefs on Bangladesh Economy.” Dhaka. Quadir, S. 2017."Bangladesh's proposed hike in gas prices sparks protests." Reuters, February 24. https://in.reuters.com/article/bangladesh-energy-prices/bangladeshs-proposed-hike- in-gas-prices-sparks-protests-idINKBN1631T3 Ranger, N., Antony, M., Dietz, S., Frankhauser, S., Lopez, A., Ruta, G., 2010. "Adaptation in the UK: A Decision-making Process". Policy Brief. Environment Agency 2010–09 Rasel, A. R. 2016. “Coal policy draft ready to be finalised.” Dhaka Tribune, Dhaka, April 4. 46 Rasmussen, D. J., M. Meinshausen, and R. E. Kopp. 2016. “Probability-Weighted Ensembles of U.S. County-Level Climate Projections for Climate Risk Analysis.” Journal of Applied Meteorology and Climatology 55 (10): 2301–22. http://doi.org/10.1175/JAMC-D-15- 0302.1. Riahi, K., Rao, S., Krey, V. et al. 2011. "RCP 8.5—A scenario of comparatively high greenhouse gas emissions." Climatic Change 109:33 https://doi.org/10.1007/s10584-011-0149-y Rupp, D. E., J. T. Abatzoglou, K. C. Hegewisch, and P. W. Mote. 2013. “Evaluation of CMIP5 20th Century Climate Simulations for the Pacific Northwest USA.” Journal of Geophysical Research 118 (April 2013): 1–23. http://doi.org/10.1002/jgrd.50843. Sampson, C. C., A. M. Smith, P. D. Bates, J. C. Neal, L. Alfieri, and J. E. Freer. 2015. A high- resolution global flood hazard model. Water Resources Research 51, 7358–7381. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4989447/. Senior Correspondent. 2014. “$24 Billion Will Be Invested in Bangladesh’s Power Sector, Says PM’s Energy Advisor.” bdnews24.com, November 25. http://bdnews24.com/business/2014/11/25/24-bn-will-be-invested-in-bangladesh-s- power-sector-says-pm-s-energy-advisor. Steinschneider, S., R. McCrary, L. O. Mearns, and C. Brown. 2015. “The Effects of Climate Model Similarity on Probabilistic Climate Projections and the Implications for Local, Risk-Based Adaptation Planning.” Geophysical Research Letters 42 (12): 5014–44. Trigg, M. A., C. E. Birch, J. C. Neal, and many others. 2016. “The Credibility Challenge for Global Fluvial Flood Risk Analysis.” Environmental Research Letters 11 (9): 94014. http://doi.org/10.1088/1748-9326/11/9/094014. UN News Centre. 2012. “Open-pit coal mine project in Bangladesh threatens human rights – UN experts.” http://www.un.org/apps/news/story.asp?NewsID=41398#.V7YquaKRZGQ. Watkiss, P., A. Hunt, W. Blyth, and J. Dyszynski. 2015. “The Use of New Economic Decision Support Tools for Adaptation Assessment: A Review of Methods and Applications, Towards Guidance on Applicability.” Climatic Change 132 (3): 401–16. http://doi.org/10.1007/s10584-014-1250-9. World Bank Group. 2017. “World Bank Commodities Price Forecast.” pubdocs.worldbank.org/en/926111485188873241/CMO-Januart-2017-Forecasts.pdf. 47 7 7. APPENDIX 7.1 DATA ON DAMAGE AND OUTAGE IN POWER PLANT FACILITIES DUE TO FLOODING In the following tables, we provide data we collected on damage and repair days caused by flooding in power plant facilities. The list is by no means comprehensive but we used the data to assess how realistic our assumptions are in Figures 7.1 and 7.2. Table 7.1. Data on flooding damage and outage at a sample of coal power plant facilities Row Power plant Flood event Inundation Repair time Damage (as % of depth (ft) (days) the capex) 1 Simhapuri, India 11/18/2015 4-5 18 <0.03% 2 6th street generating June 2008 n/a 29% station , Iowa (USA) 3 Prairie Creek, Iowa, USA June 2008 n/a 26% 4 Sutherland June 2008 2-4 14 0 5 Watson Hurricane 20 46 <18% Katrina, 2005 6 Watson Hurricane 20 120 <18% Katrina, 2005 Sources: Row 1 inundation depth (The Hindu Staff Reporter 2015), damage in the power sector (D. Bosco 2015). Row 2 damage at the power plant (Dewitte 2009) Row 3 damage at the power plant (Power Engineering website 2009) Row 4 (Alliant Energy Corporation 2008) Rows 5 & 6 repair days (Mississippi Power 2006) Rows 5 & 6 damage (Mississippi Power 2005) Table 7.2. Data on flooding damage and outage at a sample of natural gas power plant facilities Row Power plant Flood event Inundation depth Repair time Damage (as % of the (ft) (days) capex) 1 Linden, New Hurricane Jersey Sandy,2012 21 n/a 2 Kearny, New Hurricane Jersey Sandy,2012 12 n/a 3 Sewaren,New Hurricane Jersey Sandy,2012 63 n/a 4 Michoud, Hurricane Katrina, Louisianna 2005 <6 <81 1% 5 Sabine, Louisiana Hurricane Ike 2008 4 150 6 Rojana power plant Thailand, 2011 4-7 ~365 29% Sources: Rows 1,2,3 repair days estimate (Disavino 2012) and (U.S. EPA Air Markets Program Data). The latter is used for row 6 48 Row 4 (Smith 2006); Row 5 inundation depth (OASIS OATI 2008) Row 6 (Tristating 2012) Table 7.3. Data on flooding damage and outage at a sample of diesel power plant facilities Row Power plant Flood event Inundation depth Repair time Damage (as % of the (ft) (days) capex) 1 Hadramount Yemen, Wadi 2008 4% 2 Hadramount Yemen, Sahel 2008 29% 3 Yemen, Mahara 2008 58% Source: (Global Facility for Disaster Risk Reduction 2009) Reader can refer to Figures 7.1 and 7.2 to compare the assumptions used in this report against the data collected and reported in Tables 7.1- 7.3. Figure 7.1. Flood damage and inundation depth (assumptions and historical datapoints) 100 Damage (% of c overnight 90 80 construction cost) 70 60 50 40 30 20 10 0 0 50 100 150 200 250 300 350 400 Inundation depth (dm from grade level ) Historical examples Assumption used in this report FEMA HAZUS MR4 Sources: For historical examples Tables 7.1 -7.2 and FEMA HAZUS MR4 (DHS, n.d.) 49 Figure 7.2. Outage due to flooding and inundation depth (assumptions and historical datapoints) 400 350 Duration of outage (days) 300 250 200 150 100 50 0 0 20 40 60 80 100 120 Inundation depth (dm from grade level ) Historical examples Assumption used in this analysis Sources: For historical examples Tables 7.1 -7.2 7.2 Appendix references Alliant Energy Corporation, 2008. “Sutherland Generating Station Back to Powering Iowa Homes and Businesses After Historic Floods,” PRNewswire [Online]. Available: http://www.prnewswire.com/news- releases/sutherland-generating-station-back-to-powering-iowa-homes-and-businesses-after-historic- floods-57549947.html. Bosco. D, 2015. “Puducherry seeks Rs 182.45 crore immediate flood relief assistance,” The Times of India [Online]. Available: http://timesofindia.indiatimes.com/india/Puducherry-seeks-Rs-182-45-crore- immediate-flood-relief-assistance/articleshow/49938116.cms. Dewitte D., 2009. “Steam customers say Alliant offer not viable: Customers would have to bear the cost,” The Gazette, [Online]. Available: http://callcenterinfo.tmcnet.com/news/2009/01/07/3895228.htm. Disavino S., 2012. “Power company restarts a quarter of NJ plants hit by Sandy,” Reuters. [Online]. Available: http://www.reuters.com/article/us-storm-sandy-pseg-newjersey-idUSBRE8AI10Q20121119. Global Facility for Disaster Risk Reduction, 2009. “Damage, Losses and Needs Assessment October 2008 Tropical Storm and Floods, Hadramout and Al-Mahara, Republic of Yemen”. Missisippi Power Company,2005. “2005 Annual Report,” [Online]. Available: http://www.mississippipower.com/about-us/pdf/MPC_2005.pdf 50 Mississippi Power, 2006. “Mississippi Power’s Plant Watson Fully Restored, Ready for Summer Peak ing Season,” PRNewswire. [Online]. Available: http://www.prnewswire.com/news-releases/mississippi- powers-plant-watson-fully-restored-ready-for-summer-peaking-season-56140092.html. OASIS OATI, 2008. “Hurricane Ike made landfall at Galveston,” [Online]. Available: http://www.oasis.oati.com/EES/EESDocs/sept. 14 2008 -- hurricane gustavike restoration -- 1130 a.m. cdt.txt. Power Engineering, 2009 .“Prairie Creek Generating Station restarts after flood,”[Online]. Available: http://www.power-eng.com/articles/2009/01/prairie-creek-generating-station-restarts-after- flood.html. [Accessed: 01-Jan-2016]. Smith J., 2006. “Entergy New Orleans’ Michoud Power Plant Returns to Service,” Entergy. [Online]. Available: http://www-temp.entergy.com/News_Room/newsrelease.aspx?NR_ID=824. The Hindu Staff Reporter, 2015. “Simhapuri power plant partly submerged,” The Hindu. [Online]. Available: http://www.thehindu.com/news/national/andhra-pradesh/simhapuri-power-plant-partly- submerged/article7901978.ece TrisRating, “Rojana Industrial Park,” Credit News, 2012. [Online]. Available: http://www.trisrating.com/en/pdf/announcement/ROJANA870-e-Sfas98z.pdf. United States Environmental Protection Agency, “Air Markets Program Data.” [Online]. Available: https://ampd.epa.gov/ampd/. 51