Short-term wind power prediction based on hybrid feature selection and multiscale temporal convolution transformer

Abstract Accurate short-term forecasting of wind power plays a vital role in enhancing wind energy usage and supporting its smooth integration into the power grid. To improve the precision of wind power prediction, this study presents a short-term forecasting method that integrates hybrid feature selection with a multi-scale temporal convolutional transformer (MSTCN-Transformer). First, the isolation forest algorithm is applied to conduct three-dimensional data cleaning on wind farm data, reducing the impact of outliers on prediction results. Then, a hybrid correlation analysis method is utilized to comprehensively select influential factors, thereby determining the features used as inputs to the forecasting model. To handle the non-stationary nature of wind power data, complete ensemble empirical mode decomposition with adaptive noise (CEEMDAN) is employed to break the data into several component subsequences. Finally, MSTCN-transformer is constructed to individually forecast all subsequences, and the eventual wind power forecasting result is acquired by aggregating the predictions of all subsequences. Experiments conducted on real-world data collected at a wind farm located in Inner Mongolia, China, demonstrate that the presented approach achieves superior predictive performance, with evaluation metrics RMSE, MAE, and R 2 reaching 1.968, 1.540, and 0.985, respectively.