Developing an interpretable wind power forecasting system using a transformer network and transfer learning

• Published in: Energy conversion and management 2025-01, Vol.323, p.119155, Article 119155
• Main Authors: Tian, Chaonan, Niu, Tong, Li, Tao
• Format: Article
• Language: English
• Subjects: administrative management; Attention mechanism; China; data collection; energy conversion; Interpretability; Transfer learning; Transformer; wind; wind power; Wind power forecasting
• ISSN: 0196-8904
• DOI: 10.1016/j.enconman.2024.119155

[Display omitted] •Propose a novel wind power forecasting system with higher accuracy.•Design a novel transformer integrating attention and stack deep learning model.•Develop a parameter-sharing-based transfer learning for accuracy improvement.•Adopt attention mechanism and feature selection to improve model interpretability. Accurate wind power forecasting is crucial for enhancing the stability and security of power grid operations and scheduling. However, previous studies have primarily focused on data preprocessing or model optimization, often neglecting the challenge of efficiently forecasting wind power for newly built wind farms with limited historical data. To address this issue, we developed a novel wind power forecasting system consisting of six modules that leverage a transformer network and a parameter-sharing transfer learning strategy, with a strong emphasis on model interpretability. In this forecasting system, the feature selection module and attention mechanism work together to identify key features from the input set and assign importance weights to each selected feature rather than treating all features equally. To validate the effectiveness of our proposed forecasting system, we conducted three simulation experiments using ten multivariate datasets from two wind farms in China. The results were compared against six benchmarks and various feature selection methods. Our findings demonstrate that the proposed wind power forecasting system outperforms all benchmarks. On average, across the three experiments, it achieved considerable performance improvements of 46.29% in mean absolute error and 31.02% in root mean square error compared to the worst-performing multi-layer perceptron. Additionally, the implementation of the transfer learning strategy markedly enhanced the forecasting system’s accuracy, leading to average reductions of 13.84% in mean absolute error and 7.77% in root mean square error.