Runoff, wind power output, and photovoltaic (PV) power output in cascade hydro–wind–PV complementary systems are inherently uncertain, making long-term scheduling a high-dimensional continuous-control decision-making problem. To address this issue, this study proposes a long-term optimal scheduling method based on the deep deterministic policy gradient (DDPG) algorithm. First, a long-term optimal scheduling model for a cascade hydro–wind–PV complementary system is established with the objective of maximizing renewable energy accommodation. Second, the original optimization problem is formulated as a Markov decision process, and the multi-constraint scheduling task is transformed into a deep reinforcement learning problem. Then, the Actor–Critic architecture of DDPG is employed to iteratively update the continuous control policy, while experience replay and target networks are introduced to stabilize the training process and improve learning performance. Finally, a large-scale cascade hydropower system and its surrounding wind and PV plants are selected as a case study for validation, and the proposed method is compared with a deep Q-network (DQN) and proximal policy optimization (PPO). The results show that the proposed method can learn a stable scheduling policy within relatively few training episodes. Compared with a DQN and PPO, DDPG achieves better overall scheduling performance, with higher renewable energy accommodation, lower curtailment, and faster convergence in the considered case study.
Li et al. (Fri,) studied this question.