A hierarchical multi-stage fusion deep learning framework for short-term wind power prediction

The Supervisory Control and Data Acquisition (SCADA) system encompasses a wealth of information with diverse physical significance, which can be categorized into homogeneous and heterogeneous data. The underlying differences and complementarities within and among these data types are crucial for enhancing the performance of short-term wind power forecasting. Accordingly, this paper introduces a deep learning architecture with a multi-stage information fusion mechanism for short-term wind power forecasting. The initial fusion module focuses on analyzing the internal correlations of homogeneous information and extracting valuable insights from historical data. Subsequently, the second fusion module emphasizes the disparities among heterogeneous information, highlighting the interplay and mutual enhancement of different physical data through spatial and temporal interaction mechanisms. This multi-layered, multi-dimensional information fusion strategy not only bolsters the model's ability to comprehend intrinsic data relationships but also enhances forecasting accuracy. Ultimately, a bidirectional long short-term memory network is employed for real-time and precise wind power forecasting tasks. Three experiments conducted on real SCADA data from wind farms substantiate the reliability and effectiveness of the proposed method. The approach outperforms 12 existing advanced methods in both single-step and multi-step predictions, significantly improving the accuracy of short-term wind power forecasting and providing robust decision support for wind farm operations and maintenance.