An uncertainty information-guided optimization method for economic scheduling of PV–BESS systems

With the increasing penetration of photovoltaic (PV) power in the power system, the uncertainty and fluctuation of its output exacerbate the difficulty of grid scheduling. Existing prediction models are insufficient in characterizing disturbances, leading to charging and discharging strategies of energy storage in day-ahead scheduling that are difficult to match with actual output, thereby causing power limit violations and operational risks. Although reinforcement learning shows potential in PV–BESS scheduling, it still suffers from limitations in uncertainty modeling, policy stability, and computational efficiency. To address the above issues, this paper proposes an uncertainty information-driven reservoir computing reinforcement learning framework. The core idea is to use PV interval prediction results as prior knowledge, incorporate them into reservoir construction, and combine them with compressed sensing theory to generate a sparse weight matrix, thereby embedding the uncertainty information of future disturbances at the structural level. During the proximal policy optimization process, the reservoir replaces traditional deep neural networks for state mapping, enabling reservoir states to not only reflect the current environment but also carry prior knowledge of future disturbances, thus providing structured guidance for policy and value learning. This method effectively addresses the problems of insufficient adaptability, large policy fluctuations, and high computational overhead of traditional models in uncertain environments. Simulation results show that under typical high-fluctuation scenarios, the proposed method significantly reduces operational risks and economic losses: the average violation amplitude decreases from 91.89% to 29.52%, the compensation cost for violations decreases by about 67%, and the operational revenue increases by 81.72%. Benefiting from the lightweight design, the number of model parameters and computational overhead are significantly reduced, further improving scheduling efficiency and deployment flexibility. In summary, this paper provides an efficient, robust, and secure solution for the economic scheduling of PV–BESS systems under uncertain environments.