Offshore Wind farm cluster power forecasting: An ASTGCN framework with dual-layer Co-evolution of loss function and graph construction method

Graph neural networks show potential in offshore wind farm cluster power output forecasting, but are limited by two challenges: adaptive loss function selection across different wind farm groups and optimal graph structure construction under dynamic wind speed propagation. Traditional methods use fixed loss functions and static graphs, failing to capture spatiotemporal coupling and nonlinear power mapping, leading to subpar accuracy in complex environments. This paper proposes a novel Loss–Construction Dual-layer Co-evolution Framework (LCDCE) based on ASTGCN. The core of LCDCE is a synergistic evolution mechanism that dynamically couples the evolution of loss functions and graph construction strategies, enabling adaptive and joint optimization of the objective landscape and spatial-temporal dependencies. A global optimal model selection mechanism ensures generalizability, and a cosine annealing scheduler balances convergence speed and optimization. Experiments on real European offshore wind farm data show the framework reduces RMSE by 33.46% and MAE by 17.02% compared to conventional ASTGCN, proving the dual-layer strategy adapts to complex wind farm cluster for robust forecasting.