Short-term wind power prediction based on anomalous data cleaning and optimized LSTM network

Abstract

Wind power prediction values are often unstable. The purpose of this study is to provide theoretical support for large-scale grid integration of power systems by analyzing units from three different regions in China and using neural networks to improve power prediction accuracy. The variables that have the greatest impact on power are screened out using the Pearson correlation coefficient. Optimize LSTM with Lion Swarm Algorithm (LSO) and add GCT attention module for optimization. Short-term predictions of actual power are made for Gansu (Northwest China), Hebei (Central Plains), and Zhejiang (Coastal China). The results show that the mean absolute percentage error (MAPE) of the nine units ranges from 9.156% to 16.38% and the root mean square error (RMSE) ranges from 1.028 to 1.546 MW for power prediction for the next 12 h. The MAPE of the units ranges from 11.36% to 18.58% and the RMSE ranges from 2.065 to 2.538 MW for the next 24 h. Furthermore, the LSTM is optimized by adding the GCT attention module to optimize the LSTM. 2.538 MW. In addition, compared with the model before data cleaning, the 12 h prediction error MAPE and RMSE are improved by an average of 34.82% and 38.10%, respectively; and the 24 h prediction error values are improved by an average of 26.32% and 20.69%, which proves the necessity of data cleaning and the generalizability of the model. The subsequent research content was also identified.