The long-term operation planning of large-scale power systems, such as the Brazilian interconnected system, requires the complex coordination of hydro, thermal, and highly variable renewable resources. In this context, cascaded reservoirs act as system-wide batteries, coupling operational decisions across time through water storage and providing the flexibility required to accommodate uncertainty. Because future natural inflows are inherently uncertain, the resulting planning problem is theoretically formulated as a continuous stochastic program, referred to as the Continuous Problem (CP). However, because solving the CP directly is computationally intractable and the Stochastic Dual Dynamic Programming (SDDP) framework intrinsically requires a finite discretization of the random variables, the planning process relies on a Sample Average Approximation (SAA) to solve a Discrete Problem (DP), where the underlying uncertainty is represented by a fixed set of backward scenarios, which remains unchanged throughout the SDDP iterations.
Relying on a single discrete approximation introduces severe practical vulnerabilities. In these large-scale environments, achieving theoretical convergence under the traditional SDDP framework requires a massive computational effort. Due to strict institutional time limits, the algorithm is typically executed with restricted parameters and stopping criteria, which frequently causes the iterative process to halt prematurely before the operating policy reaches sufficient maturity. Moreover, the resulting policies may become excessively sensitive to the particular set of scenarios sampled during the approximation process, causing minor changes in the problem representation to produce noticeably different operational decisions. As a consequence, distinguishing numerical artifacts from genuine operational effects may require additional validation efforts, undermining the robustness and reproducibility of the decision-making process.
To structurally reduce the dependence on a single discrete approximation and better represent the underlying CP within acceptable computational times, the main methodological contribution of this work is a novel dynamic framework that iteratively changes the backward scenario discretization at each SDDP iteration. The theoretical foundation of this approach leverages Jensen’s inequality, ensuring that if the scenario tree is constructed through a quantization-like procedure, the resulting recourse function is a valid lower approximation of the CP over the entire domain of the state variables. Building upon these valid lower bounds, we establish the almost sure convergence of the proposed backward-resampling SDDP algorithm.
The proposed framework is validated using a tutorial case, a small-scale cascaded hydrothermal system, and a comprehensive large-scale application involving the long-term planning of the Brazilian Power System, explicitly modeling 162 individualized hydro plants and 119 thermal units under high-dimensional correlated inflows. Numerical results demonstrate that, under the same computational budget, the proposed approach significantly improves the approximation of the recourse function across iterations, yielding higher-quality operating policies while substantially increasing policy maturity. More importantly, it drastically reduces the sensitivity of operational decisions to the initial random seed, providing the robustness and reproducibility required for large-scale operational planning.