A multi-objective deep reinforcement learning method for intelligent scheduling of wind-solar-hydro-battery complementary generation systems

Abstract

Renewable energy resources are rapidly developing to pursue carbon neutrality. However, integrating these sources poses challenges due to their randomness, volatility, and spatial–temporal mismatch of energy-electricity demand. One effective solution is the complementary operation and bundled external delivery of wind, solar, battery storage, and cascade hydropower. However, the uncertainty of wind and solar and the complexity of multi-energy coupling systems increase the difficulty of power scheduling, making it challenging for traditional methods to overcome these obstacles. Thus, this work presents an intelligent scheduling method based on multi-objective deep reinforcement learning (MODRL) for the wind-solar-hydro-battery complementary system (WSHBCS). Firstly, a multi-objective scheduling model is established. Meanwhile, this work proposes a MODRL framework to learn the scheduling strategy for multiple competing objectives. Then, the scheduling problem is developed into a Markov decision process (MDP), in which the state, action, and reward functions of the WSHBCS are designed accordingly. Finally, a multi-policy twin delayed deep deterministic policy gradient (MPTD3) method is put forth to achieve intelligent decisions in continuous action spaces. Simulation results indicate that the proposed intelligent scheduling method effectively achieves multi-objective scheduling for the WSHBCS, outperforming traditional heuristic optimization methods in terms of multi-objective optimization performance, uncertainty adaptability, and solution time.