Abstract
Addressing the problems of high randomness and low prediction accuracy in short-term power load forecasting, this paper proposes a multi-featured short-term power load prediction model based on the error optimal weighting method and the improved combination prediction model. Firstly, the combined algorithm of grey correlation analysis and radial kernel principal component analysis is used to deal with the multi-factor coupled input data. Secondly, the prediction results of the improved sparrow search algorithm optimized convolutional neural network and deep bidirectional gated recurrent unit combination model, convolutional neural network and long short-term memory network as well as selective attention mechanism combination model, and improved time convolutional network and channel enhanced attention mechanism combination model are processed by the error optimal weighting method to obtain the final prediction results. Then, the Bagging and Extreme Gradient Boosting combination model optimized by Bayesian theory is used to optimize the prediction error. Finally, a validation model is constructed, and by comparing with many mainstream algorithmic models and combining with various error data to verify that the strategy proposed has better performance and can improve the accuracy of short-term power load forecasting.