Multi-attention-powered learning genetic algorithm for real-world 3D wind farm layout optimization

Abstract

Wind farm layout optimization plays a crucial role in improving wind energy utilization, reducing construction and operational costs, enhancing the reliability and stability of wind farms, and promoting technological innovation in wind energy. However, this NP-hard problem is often approached in current research under idealized conditions, typically assuming a flat plane with no consideration of elevation. To address these limitations, we propose a 3D wind farm optimization layout framework that incorporates a 3D Gaussian wake model, accounting for spatial factors like terrain elevation to more closely reflect real-world engineering conditions. To handle the high-dimensional complexity of 3D wind farm layout optimization, we introduce a multi-head attention-based genetic learning algorithm, named ALGA, that learns and leverages successful evolutionary patterns within the population. This enables the calculation of attention scores for promising regions in the search space. By iteratively refining high-scoring regions, the population achieves greater vitality and has a stronger ability to escape local optima, optimizing continuously toward the best solutions while maximizing energy conversion efficiency and minimizing wake effects. Our study involves two cases: one with ideal terrain and four standard wind speeds, and another that simulates the real terrain and annual wind conditions of the Guishan wind farm project. Across total 24 experimental scenarios, ALGA achieves the highest energy conversion efficiency, outperforming seven other state-of-the-art algorithms.