Incluant une Journée industrielle de l'optimisation
HEC Montréal, 7  9 mai 2012
JOPT2012
HEC Montréal, 7 — 9 mai 2012
TB5 Énergie et Environnement 3 / Energy and Environment 3
8 mai 2012 10h30 – 12h10
Salle: Raymond Chabot Grant Thornton
Présidée par Michèle Breton
3 présentations

10h30  10h55
Optimal Carbon Sequestration when Reservoirs are Leaky
We provide a complete solution to a model of optimal carbon sequestration and storage in a pollutionemitting economy. Atmospheric pollution is kept below a certain threshold. The CO2 storage is leaking at a constant rate. The solution includes some situations very different from those previously described.

10h55  11h20
A MultiAgent Reinforcement Learning Approach to Develop the Water Value Function for a Multireservoir Hydroelectric Systems
We present the application of a new algorithm (MARLOMMR) used to establish the marginal value of water and value of waterinstorage for multireservoir hydroelectric power systems. The new algorithm relies on recent advances in the multiagent reinforcement learning (MARL) technique combined with function approximations and other approximation techniques. To develop optimal release policies within the proposed multiagent environment framework, decentralized agents, each representing a reservoir on a river system, cooperate via indirect communication channels. A second level of communication exists between different river systems to achieve the global goal of maximizing present revenue as well as future value of the water stored in the reservoir system. The advantages of the new algorithm are: less computational effort, flexibility and better representation of the operation of actual hydro system (e.g., the BC Hydro system). To validate the results of the new model, a stochastic dynamic programming algorithm (SDPOM2R) was developed and used to benchmark the MARLOMMR model. Both models use the same input data sets for the two largest reservoirs in the BC Hydro system, the Williston and the Kinbasket. We plan to compare the efficiency and accuracy of the MARLOMMR model with other models that are alreadyinuse or are underdevelopment at BC Hydro, such as the Reinforcement Learning Reservoir Optimization Model and a Stochastic Dual Dynamic Programming model.

11h20  11h45
Foresight Approach for International Environmental Agreements
In this paper, we consider a group of countries, where a subset of them (the signatories) agree to adhere to an environmental agreement to reduce pollution emissions, while the remaining countries are nonsignatories. We assume that the number of signatories evolves over time according to a discretetime replicator dynamics. In our model, both the stock of pollution and the number of signatories are state variables. We look for both Nash and Stacklberg equilibria strategies by using numerical methods.