Dual-Stage Adaptive Normalization and Dual-Residual Physics-Energy-Gating–Assisted Pyramid LSTM Network for Offshore Wind Power Forecasting

Abstract Offshore wind power has become a critical component of global clean energy; and its accurate forecasting directly governs grid reliability and security. However, complex marine meteorological conditions generate strongly nonstationary, stochastic, and multidimensional periodic data patterns that severely challenge high-precision forecasting. To address these challenges, this paper proposes a novel forecasting framework named “the dual-stage adaptive normalization and dual-residual physics-energy-gating–assisted pyramid long short-term memory (LSTM) network (DSAN-PLSTMResPEG).” The framework introduces three core innovations. First, we propose the dual-stage adaptive normalization (DSAN) mechanism, which integrates global feature normalization with Gaussian noise-enhanced instance-level reversible normalization to effectively mitigate strong data nonstationarity. Next, we design a pyramid LSTM (PLSTM) structure that utilizes dilated recurrent skip connections and hierarchical isolation to capture multidimensional periodic dependencies in the data. Further, to counter high signal stochasticity and enforce physical constraints, we propose a physics-energy gate (PEG), which leverages physical information and incorporates information decay compensation to unsupervisedly enhance perception of random events. PEG collaborates with a dual-residual mechanism to dynamically regulate deep-feature energy flow within the PLSTM, jointly mitigating vanishing gradients and feature-fusion bottlenecks. Finally, the proposed model is validated using operational data from two offshore wind farms in Fujian, China, and one European offshore wind farm. Extensive experiments across three prediction horizons demonstrate that DSAN-PLSTMResPEG achieves superior accuracy and generalization, significantly outperforming state-of-the-art baseline models on key metrics. Ablation studies further confirm the indispensable contribution of each proposed component.