Offshore wind power forecasting based on phase-space guided representation and regime-adaptive differential attention mechanism

Offshore wind power forecasting is challenged by nonlinear dynamic evolution, multi-scale atmospheric fluctuations, and redundant environmental disturbances under complex marine conditions. This paper develops a phase-space guided forecasting model with a regime-adaptive differential attention mechanism to improve dynamic representation learning and disturbance suppression. Different from conventional purely data-driven approaches, the proposed model uses delay time and embedding dimension as nonlinear dynamic priors to guide temporal representation learning. A dual-scale encoding architecture is designed to capture long-term trends and high-frequency fluctuations, while a regime-adaptive differential attention mechanism suppresses redundant environmental disturbances under varying operating regimes. Evaluated on real measured data from offshore wind farms in Zhejiang, China, the results demonstrate that the proposed model consistently achieves the best overall forecasting performance across different prediction horizons. It achieves up to a 19.3% reduction in relative squared error (RSE) compared with the suboptimal benchmark model, maintaining smoother error growth and more stable residual distributions under multi-step forecasting scenarios. These findings suggest that incorporating nonlinear dynamic priors into deep temporal forecasting can improve the robustness of offshore wind power prediction under complex marine environments.