SciEPublish

MSTGraph-RAR: Multi-Step Offshore Wind Power Forecasting with Variable Graph Learning and Ramp-Aware Residual Correction

lssue 4, Volume 3
Download PDF
Cite This Article
Contents

Article Open Access

MSTGraph-RAR: Multi-Step Offshore Wind Power Forecasting with Variable Graph Learning and Ramp-Aware Residual Correction

Mingju Gong 1 Linlin Zhu 1 Yu Tian 1 Siyuan Yin 1 Yaming Wang 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: 02 September 2026 Revised: 10 October 2026 Accepted: 23 October 2026 Published: 05 November 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:18
Downloads:8
Mar. Energy Res. 2026, 3(4), 10024; DOI: 10.70322/mer.2026.10024
ABSTRACT: Accurate offshore wind power forecasting is essential for grid integration and secure operation, yet pronounced non-stationarity, multi-scale dynamics, and inter-variable dependencies make continuous multi-step forecasting challenging. This paper proposes MSTGraph-RAR for multivariate multi-step offshore wind power forecasting. Continuous series are segmented into patches, and the Multi-Scale Temporal–Variable Graph (MSTGraph) captures short- and long-range dynamics while learning scale-specific directed dependencies between power and meteorological variables. A Transformer encoder then generates base forecasts. Ramp-Aware Auxiliary Learning (RAL) uses future Down, Normal, and Up power-change states as auxiliary supervision to enhance state-sensitive representations, while Ramp-Aware Residual Correction (RRC) selectively refines forecasts through gated and bounded residual correction. Experiments on five real offshore wind farms show that MSTGraph-RAR achieves the best performance across all forecasting horizons and metrics. Compared with baseline models, MSE, MAE, and RMSE decrease by up to 36.13%, 24.80%, and 19.71%, respectively, while R2 increases by up to 0.13. Ablation and ramp-intensive analyses further confirm the effectiveness of the proposed components and improved tracking of rapid power variations. Overall, MSTGraph-RAR provides an effective approach for reliable offshore wind power forecasting under complex non-stationary conditions.
Keywords: Offshore wind power forecasting; Graph neural networks; Multivariate time series forecasting; Multi-scale feature learning; Wind power ramp events
TOP