Enhancing Wind Power Forecasting via Multi-Farm Coupling Under Data Isolation: A Physics-Guided Personalized Federated Approach

Abstract

Accurate wind power forecasting is critical for grid stability and long-term sustainability, but mountainous wind farms face challenges from complex micro-meteorology, restricted communication, and non-IID data, exacerbated by data silos that prevent centralized learning. Most federated learning relies on data-driven averaging that ignores multi-farm coupling, or adopt complex local models that increase communication overhead. To address these, a physics-guided personalized federated approach is proposed to enhance wind power forecasting. Its core is a physics-guided aggregation mechanism that constructs a dynamic weight matrix from distance, elevation, and real-time wind direction to enable personalized aggregation capturing multi-farm coupling. The federated framework combines a shared CNN-LSTM with multi-head attention for regional patterns and a personalized layer for local microclimate. A risk-aware asymmetric loss is incorporated to penalize high-power errors, enhancing operational reliability under high-power conditions. Validation on mountainous wind farms for 3-day forecasting under typical and extreme scenarios across wet, dry, and normal seasons shows that the average R2 exceeds 0.95, and the average RMSE is reduced by more than 24% compared to baselines, achieving high accuracy under strict privacy preservation. By enabling multi-farm coupling under data isolation, this approach achieves high forecasting accuracy on the studied wind farms, showing promise for similar ones.