This work presents a novel learning algorithm for the operation policy of power systems trying to minimize the cost of fulfilling the energy demand. The algorithm improves upon the classical reinforcement learning methods by controlling the sampling variance in the estimation of the future cost spatial differences, together with parameter regularization and dynamic exploring techniques. The proposed strategy was applied to a case of what could be the power system of Uruguay by 2050 based strongly in hydro, wind and solar energies, including three lakes, four groups of battery banks, and the basin runoff of the two main Uruguayan rivers. The generation in the year 2022 in Uruguay was 43% hydraulic, 40% wind plus solar, 7% biomass and 10% based on fossil fuels. This composition prints a very relevant stochastic component that makes it difficult to apply machine learning techniques without the kind of algorihms proposed in this work.