4 Presentations

• 
  10:30 AM - 11:00 AM

  Comparison Between Different Hydrologic State Variables in Stochastic Dynamic Programming Applies to Hydropower Production

  • Pascal Côté, presenter, Rio Tinto
  • Quentin Desreumaux, Université de Sherbrooke
  • Marco Latraverse, Rio Tinto Alcan

  Many studies have shown that the role of hydrological information in Stochastic Dynamic Programming (SDP) for hydropower reservoir operation is crucial. Last year, the Quebec Power Operations group at Rio Tinto Alcan have developed a new model based on SDP to manage the Kemano hydroelecric system located in British Columbia, Canada and several hydrologic state variables were tested to calibrate the model. This system undergoes a very particular hydrological regime with large streamflow volumes due to significant snow cover during winter and we will show that the choice of hydrological state variable can largely affect management policies. Numerical results will be present.

• 
  11:00 AM - 11:30 AM

  A Stochastic Model for Power Generation RD&D and Capital Investment Planning Under Decision-Dependent Uncertainty

  • Nidhi R. Santen, presenter, Harvard University Belfer Center for Science and International Affairs
  • Laura Diaz Anadon, Harvard University Belfer Center for Science and International Affairs

  Effectively controlling air emissions from the fossil-dominated electric power generation sector is critical for managing long-term environmental challenges such as climate change. To address the dilemma of resolving increasing electricity demand with emission reduction goals, policy makers and other stakeholders are interested in determining cost-effective balances between near-term low-carbon capital investment (deployment) in the power generation sector, and “diverting” near-term effort to research, development, and demonstration (RD&D) that can aid in future emissions reductions at potentially lower-costs.

  Unfortunately, adequate decision-support tools to guide decision-makers on this challenging technology “deployment” versus “development” question do not exist. Long-term environmental challenges, such as climate change, provide special challenges in finding least-cost solutions because of the intra- and inter-temporal dimensions present in the problem. Of particular significance are: (1) the multiple technology categories one can choose to invest in for both development and deployment; (2) the very long lifetimes of electric power capacity investments; (3) the magnitude of uncertainties associated with the outcomes of the RD&D process; and (4) the decision-dependency of the uncertainties (previous investment decisions impact future technology possibilities). Computational limitations also result in existing tools being mismatched with the manner in which long-term strategic decisions about energy use and production are actually made—at multiple times throughout the planning horizon after learning between decision points. Most current models assume that decisions are made up front as “single shots,” under an assumption of perfect information about the evolution of technology costs. Existing models also lack suitable empirical data about technology costs and performance for supporting a good calibration.

  In this paper, we present a new numerical decision-support tool for studying optimal balances of intra- and inter-temporal capital and RD&D investments in the power generation sector under decision-dependent RD&D uncertainty and learning.

  Building on classical power generation capacity expansion modeling methods, this work endogenizes RD&D-based technical change, permitting the representation of simultaneous RD&D and capacity investment decision-making. The relationship between RD&D investments and technology costs, as well as the uncertainty surrounding those costs, are characterized using expert elicitation data, allowing for a novel, transparent, and consistent integration of technology specific cost-estimates from experts from different sectors. The new model is formulated as a formal sequential decision under uncertainty problem and solved using stochastic dynamic programming. The dimensionality burden of this large computational problem with multiple technology decisions, uncertainties, and decision periods is managed through the use of approximate dynamic programming (ADP) techniques.

  Stochastic dynamic programming is appropriate for the current problem as it: (1) allows integration of uncertainty analysis and learning in a sequential decision framework; and (2) provides an efficient means for modeling realistic decision-dependent uncertainties such as those uncovered through the expert elicitation process. A common method used to evaluate the effect of uncertainty on optimal decision-making is Monte Carlo simulation—valuable for understanding the range of possible optimal decisions under uncertainty, but ultimately using deterministically structured models that rely on assumptions of perfect foresight. Such analyses therefore do not reflect the fact that the optimal solution may change if one accounts for the ability to learn between decision points. Additionally, other stochastic optimization methods such as stochastic programming with recourse are useful for constructing stochastically structured decision models that consider both uncertainty and learning, but encounter severe dimensionality challenges when modeling decision-dependent uncertainties.

  In a series of illustrative numerical experiments using the new model, we show the optimal power generation RD&D and capital investment decisions under decision-dependent uncertainty and learning, and compare the decisions to those made under perfect foresight. Further, we explicitly show the value of incorporating uncertainty and learning into the sequential decision framework. We do so by comparing the optimal investment decisions with uncertainty-only considered through a traditional Monte Carlo analysis, with optimal decisions under uncertainty and learning through analysis with the formal stochastic model. Finally, we show how optimal RD&D and capital investment decisions change under different carbon policy regimes. The modeling tool and insights are useful for RD&D and capacity investment planners alike.

• 
  11:30 AM - 12:00 PM

  Objective-Based Generation of Scenario Trees

  • David Munger, presenter, CIRRELT - Polytechnique Montréal
  • Michel Gendreau, Polytechnique Montréal
  • Antoine Saucier, Polytechnique Montréal

  Several resolution methods for multistage stochastic optimization problems rely on a discretized representation of the probability space of the underlying stochastic process, known as a scenario tree. Methods proposed so far to construct scenario trees are mostly concerned with preserving the structure of information, that is, selected properties of the probability space and of the stagewise revelation of information. We propose to consider the complete structure of the optimal objective in order to locate where in the scenario tree it pays off more to refine the discretization, based on the variability of the objective itself rather than on that of the underlying random process. We explore methods that exploit the above ideas and present the results when applied on a toy problem for which an analytical solution can be found.

• 
  12:00 PM - 12:30 PM

  SMART-ISO – Stochastic Optimization for Nested Energy Markets

  • Warren Powell, presenter, Princeton University
  • Hugo Simao, Princeton University
  • Boris Defourny, Princeton University

  SMART-ISO is a detailed model of the PJM energy market, which includes a complete model of the grid, generators and loads in five-minute increments. SMART-ISO currently models day-ahead, hour-ahead and real-time markets, with careful handling of the timing of decisions and information. It includes a “best deterministic” lookahead policy for the day-ahead decision, with value function approximations to manage pumped hydro for the hour-ahead decisions. We will describe ongoing research into the algorithmic challenges of handling nested decisions in a stochastic environment. We will also report on the current status of model calibration, and early results of a major wind integration study using off-shore wind.