Design of Electrical Load Prediction System Based on Deep Learning Algorithm

Abstract

In order to solve the problems of low prediction accuracy and long model time in current prediction algorithms, the author proposes the design of an electrical load prediction system based on deep learning algorithms. The author proposes an improved deep learning short-term load forecasting model based on random forest algorithm and rough set theory. The model is first based on historical data and uses the random forest algorithm to extract key feature quantities that affect load forecasting. Then, the key feature quantities and historical load values are used as input and output terms for training the deep neural network, and the prediction results are corrected through rough set theory. Finally, simulation verification is conducted through numerical examples. The experimental results showed that compared with the RF-DL model, the MSE index of the RFDL-RST model decreased by 30.187%, and the overall prediction results were closer to the true values. The MAE index also decreased from 5.76% to 4.02%. During special periods of significant load changes such as 07:00-08:00 (rapid increase in load) and 22:00-23:00 (rapid decrease in load), the prediction accuracy was greatly improved. In addition, compared with the DL-RST model, the MAE and MSE indicators of the RF-DL-RST model were reduced by 15.210% and 21.414%, respectively, and the DL training time of the RF-DL-RST model was shortened by 10.175%, indicating that simplifying the DL input feature quantity through the RF model can improve the load forecasting effect. The prediction accuracy of this model is higher than that of a single deep learning model and a model without prediction correction.