Quantification of ultra-short-term forecasting performance of wind power based on multivariate LSTM and logistic coupled models

The power optimization of wind farms and the optimal control of wind turbines require high-precision power ultra-short-term prediction for each wind turbine. In order to improve the performance of ultra-short-term prediction of wind power, this paper couples the LSTM model with the Logistic model and combines it with Graph Convolutional Neural Network (GCN) to construct the ultra-short-term prediction model of wind power based on Logistic-LSTM-GCN, and test and analyze the prediction performance of the model. Comparing the LASSO, XGboost, LSTM, GRU and TCN-LSTM models, the MAE and RMSE of this paper’s model are the lowest among all the models, which are 3.34% and 5.89%, respectively, and the R² is the highest, which is 79.76%. And the MAE and RMSE predicted by the model with inputs of four-dimensional spatio-temporal feature matrix are smaller than the model with inputs of one and two dimensions, and the R² value is larger than that of one and two-dimensional model. It indicates that the Logistic-LSTM-GCN model based on spatio-temporal information can extract the spatio-temporal information of wind farms more effectively, which improves the accuracy of wind cluster power prediction. In addition, with the increasing time step, the error indicators MAE, MAPE and RMSE are gradually increasing. Taking a time step of 4s for prediction, the prediction error of the model is minimized when considering multivariate variables such as wind speed, wind speed decomposition component, yaw error, wind direction, and rotor speed. This indicates that the multivariate LSTM, logistic and GCN coupled model can significantly improve the performance of ultra-short-term prediction of wind power.