Hybrid Deep Learning Model for Ultra-Short-Term Load Forecasting with Adaptive Decomposition and Enhanced <i>Q</i> -Learning

Aiming at the problems that the forecasting accuracy of a single model deteriorates due to load fluctuation when it is used for ultra-short-term load forecasting (LF), and that most hybrid models cannot fully consider the spatio-temporal characteristics of load, a hybrid model based on deep learning is proposed for ultra-short-term LF. First, Gaussian white noise is added to the adaptive noise complete set empirical mode decomposition algorithm for optimization, and the actual load is decomposed to obtain a series of subsequences. Then, the maximum information coefficient is used to analyze the correlation between the subsequence load and meteorological factors, and the strongly correlated variables of the load are obtained. Next, bidirectional gated recurrent units and deep residual graph convolution models are used to predict each subsequence separately. Finally, the Q-learning algorithm improved by utilizing the action-selection mechanism is used to weight and combine the mixed prediction results to obtain the prediction results of ultra-short-term load. By selecting the electricity user load in the cement industry as the dataset for experimental analysis of the proposed method, the results show that its RMSE, MAE and MAPE are 1.761[Formula: see text]MW, 1.458[Formula: see text]MW and 1.047% respectively. The proposed method performs better in ultra-short-term LF.