Multi-step Prediction of Wind Power Based on Multi-mode Feature Screening and Enhanced DeepAR

With the increasing installed capacity of new energy units, accurate prediction of wind power output has become crucial for realizing large-scale wind turbine grid-connection. To address the issue of multi-step prediction of wind power, a model of wind power prediction is proposed based on Bayes-DeepAR. Firstly, a multi-mode feature screening strategy is constructed, integrating the Pearson correlation coefficient, F-test and mean impact value algorithm to screen out the feature variable set that has a greater impact on wind power. Secondly, a DeepAR prediction model is proposed grounded in embedded-layer, long-short-term memory network layer and Gaussian distribution layer. The Bayes approach is utilized to optimize the hyperparameters of the model, aiming to achieve multi-step prediction. Finally, based on the prediction results, a comprehensive analysis on the error characteristics of time series multi-step prediction is conducted, containing indexes such as the horizontal error, vertical error and error accumulation. Compared to other prediction models, the simulation results demonstrate that the proposed multi-step prediction model not only achieves point prediction and probability interval prediction of wind power, but also improves the prediction accuracy.