Reinforcement learning-based particle swarm optimization for wind farm layout problems

Optimizing wind farm layouts is critical to maximizing wind power generation. The wake effect significantly impacts turbines located downwind, making farm layout a key determinant of power generation efficiency. Traditional algorithms often overlook the value of leveraging historical information, which can lead to entrapment in local optima. Our survey reveals that previous studies on wind farm layout optimization (WFLO) have not adequately integrated the historical data of particle swarm optimization (PSO) with reinforcement learning’s empirical pool, resulting in the loss of valuable information. Here, we present a novel approach that enhances algorithm development and exploration by utilizing historical data and integrating proximal policy optimization from reinforcement learning with an experience pool. This method markedly outperforms the conventional genetic PSO in terms of performance. Extensive numerical experiments across wind farms of various sizes and four distinct wind scenarios demonstrate the superior efficacy of our reinforcement learning-based particle swarm optimization (RPSO) algorithm compared to 12 state-of-the-art methods. Under four wind scenarios, the average power conversion efficiencies of RPSO for the three turbine scales reach 98.68%, 98.14%, and 97.33%, respectively, underscoring the high competitiveness of the proposed RPSO for WFLO in diverse wind conditions. • A novel combination of PPO combined with PSO for WFLO is presented. • Use RL’s experience playback to dynamically adjust PSO parameters. • A reasonable blend of RL’s experience pool and PSO’s historical information. • PPO significantly improves the convergence and performance of the PSO. • RPSO achieves state-of-the-art in WFLO under different scenarios.