SCIEPublish

RegimeMoE-PatchTST: Metocean-Conditioned Mixture of PatchTST Experts for Multi-Step Offshore Wind Power Forecasting

lssue 4, Volume 3
Download PDF
Cite This Article
Contents

Article Open Access

RegimeMoE-PatchTST: Metocean-Conditioned Mixture of PatchTST Experts for Multi-Step Offshore Wind Power Forecasting

Mingju Gong 1 Yaming Wang 1 Minghao Sun 1 Yu Tian 1 Linlin Zhu 1 Siyuan Yin 1 Yilin Zhang 1 Wei Xu 2,3,* Fei Su 1

Author Information
Other Information
1
School of Integrated Circuit Science and Engineering, Tianjin University of Technology, Tianjin 300384, China
2
Sea Island Environment Science and Technology Research Institute Co., Ltd., Tianjin 300000, China
3
School of Ocean Energy, Tianjin University of Technology, Tianjin 300384, China
*
Authors to whom correspondence should be addressed.

Received: 29 August 2026 Revised: 28 September 2026 Accepted: 30 September 2026 Published: 20 October 2026

Creative Commons

© 2026 The authors. This is an open access article under the Creative Commons Attribution 4.0 International License (https://creativecommons.org/licenses/by/4.0/).

Views:52
Downloads:23
Mar. Energy Res. 2026, 3(4), 10022; DOI: 10.70322/mer.2026.10022
ABSTRACT: Offshore wind power forecasting is challenging because rapidly varying metocean conditions induce strong non-stationarity and heterogeneous power-generation dynamics. This study proposes RegimeMoE-PatchTST, a metocean-conditioned mixture-of-experts framework for multi-step offshore wind power forecasting. For each historical window, four statistical descriptors of seven metocean variables are extracted to form a 28-dimensional condition vector. Three independently parameterized PatchTST networks serve as forecasting experts, while a gating network maps the condition vector to sample-dependent weights and softly fuses their multi-step predictions. The entire framework is trained end-to-end without predefined regime labels. Experiments on two offshore wind turbine datasets at 2-, 4-, 8-, and 16-h horizons show that the proposed model achieves the lowest mean squared error (MSE) in all eight turbine–horizon settings and reduces average MSE by 3.69% relative to PatchTST. Ablation results verify the importance of both metocean-conditioned routing and dynamic expert fusion, while regime-wise evaluation shows MSE reductions of 7.7% and 8.1% under high-wind and high-turbulence conditions. These results demonstrate improved adaptability and forecasting accuracy under heterogeneous offshore operating conditions.
Keywords: Offshore wind power forecasting; PatchTST; Mixture of experts; Metocean conditions; Multi-step forecasting
TOP