Optimizing the production of hydroelectricity is a sequential decision-making problem: at each time step, operators must decide how much water to release through the turbines to maximize benefits. Sounds simple? In reality, it’s a complex problem due to several factors:
- Trade-offs between immediate and future gains: Water is a limited resource, requiring careful planning to balance short-term benefits with future ones.
- Uncertainty: Decisions are made under uncertainty including variable hydrological inflows and fluctuating electricity market prices.
- Constraints: Operations must comply with a range of environmental regulations, physical limitations, institutional constraints, and legal obligations.
Few algorithmic solutions exist to solve this sequential decision-making problem. Stochastic Dual Dynamic Programming (SDDP), proposed by Pereira & Pinto in the 90ies is one of them.
Our work on the SDDP algorithm has focused on the integration of various hydrologic information in order to improve the performance of hydropower systems. This is motivated by the fact that a more exhaustive description of hydrological processes is expected to yield higher benefits as it should reduce the uncertainty associated with future inflows (runoff). The enhanced SDDP algorithm was tested with snowpack, soil moisture and sea surface temperature. Details on this line of work can be found on this page.
A related but distinct extension of SDDP addresses climate persistence—the observation that, in some regions, hydrologic conditions can deviate from long-term averages for several consecutive years. To model these low-frequency shifts in the hydrological regime, we introduced a hidden climate state into the SDDP framework. This climate state can take on discrete values such as dry, normal, or wet, and transitions between these states are governed by a hidden Markov model. We tested this enhanced version of the SDDP algorithm in the Senegal River basin, where streamflow exhibits clear regime-like patterns consistent with climate persistence. This version therefore provides optimal, climate tailored, release policies.
