On the solution variability reduction of stochastic dual dynamic programming applied to energy planning

In the Brazilian energy operation planning, Stochastic Dual Dynamic Programming (SDDP) determines hydrothermal planning decisions based on auto-regressive (AR) models for associated risk factors. In this work we show that using AR models to generate scenarios leads to an undesirable drawback on SDDP: the variability of the solutions increases with respect to changes in the AR initial conditions. We propose a modified version of the risk averse SDDP algorithm aimed at reducing decisions and marginal costs variability induced by the use of AR models. We show that it is possible to obtain results with less variability and with the same characteristics of the ones obtained by traditional approach. Moreover, we argue that the proposed approach is more flexible since it is not restricted to linear models as in the original SDDP algorithm.