Offshore wind power forecasting with wind-regime clustering and multi-scale feature learning

Short-term forecasting of offshore wind power is critical for grid stability, yet it remains a persistent challenge due to the inherent volatility and noise in both meteorological and power data. To address this challenge, a deep learning approach is introduced that integrates a wind-regime clustering strategy with an architecture combining an Inception network, a Bidirectional Long Short-Term Memory network, and a Multi-Head Self-Attention mechanism. Wind-direction-based pre-classification and K-means clustering are employed to organize historical data into distinct meteorological regimes, which are subsequently processed by the predictive model. The Inception network extracts multi-scale temporal features, the Bidirectional Long Short-Term Memory network captures long-range dependencies, and the Multi-Head Self-Attention mechanism emphasizes critical feature interactions. Experiments using real offshore wind farm data demonstrate that the approach achieves high accuracy and strong robustness, effectively supporting the scheduling and integration of offshore wind power. • Proposes a wind-regime clustering framework that structures data into meteorological patterns. • Employs Inception and Bidirectional Long Short-Term Memory for hierarchical feature learning. • Uses Multi-Head Self-Attention to reinforce key relationships among meteorological variables. • Demonstrates improved forecasting accuracy and robustness on real offshore wind farm data. • Provides a scalable solution that supports grid management and renewable energy integration.