Deep reinforcement learning real-time dispatch approach for cascade hydropower with hybrid pumped-storage mitigating photovoltaic uncertainties

The integration of high levels of photovoltaic (PV) generation into grids poses a significant challenge to stability due to its inherent variability. Since cascade hydropower (CHP) plants provide rapid response and large storage, and retrofitted hybrid pumped-storage (HPS) adds further flexibility, integrating PV into complementary systems can mitigate fluctuations. However, their real-time coordination is complicated by spatio-temporal hydraulic coupling and PV uncertainties. Thus, this study develops a model-free deep reinforcement learning (RL) real-time dispatch approach based on the soft actor-critic (SAC) method to coordinate power output of CHP and HPS online. A comprehensive reward function is designed to balance generation plan tracking, reservoir volume flexible tracking, and output smoothing. In addition, a framework combining Wasserstein distributionally robust optimization (WDRO) day-ahead model is proposed to generate robust generation plans and unit operation status schemes. WDRO is employed to reduce infeasible exploration risk and facilitate stability in the training process of SAC. A case study on the Yalong River basin CHP-PV complementary system demonstrates the effectiveness of the proposed approach. The SAC agent learns optimal policies directly from interactions with the nonlinear system environment, enabling it to make real-time decisions based on actual PV power and system states, thus avoiding online computational burdens of multi-step optimization. Compared to the method without WDRO guidance, the integrated power volatility is reduced by 93.5%. Compared to model-based methods, the constraint violation rate is reduced from approximately 2% to zero.