A deep residual network integrating entropy-based wavelet packet ensemble model for short-term electrical load forecasting

This paper designs a model for predicting 24-h ahead load using a combination of entropy-based wavelet packet transform (EBWPT) and a residual block-based framework. Given that entropy represents the average level of uncertainty or information content, Shannon's entropy is utilized to select nodes from the wavelet tree that possess higher levels of information. The wavelet packet coefficients of the best basis are extracted using EBWPT. These coefficients, which represent rich frequency-domain information, are fused with temporal features extracted from the raw data by a residual block. In each primary cell, short-term features, medium-term trends, and periodic and irregular temporal features are extracted separately. These features are then combined to obtain rich fused features. These rich features are used within the primary cell to predict the 1-h ahead. An ensemble model consisting of 24 primary cells provides a preliminary forecast for 24-h ahead. This preliminary forecast is subsequently fed into a deep residual block-based framework to provide the final prediction of the load for the next 24 h. Multiple experiments and comparisons with existing state-of-the-art methods demonstrate the higher accuracy and generalizability of the proposed model, and the maximum performance improvement is up to 19.3 % in term of the MAPE metric.