A novel residential electricity load prediction algorithm based on hybrid seasonal decomposition and deep learning models

Residential electricity load prediction is of great significance for power system planning. With the increasing complexity and uncertainty of the power grid, traditional prediction models still have insufficient accuracy and neglect seasonal changes. In this paper, a data-driven multi-scale hybrid prediction model for residential electricity load is proposed, which integrates a convolutional neural network (CNN), long short-term memory (LSTM), and attention mechanism. The seasonal decomposition was applied to extract seasonal patterns of the electricity consumption data. The hybrid model integrates the parallel processing capability of CNN and the long time-series modelling capability of LSTM to capture the spatial-temporal characteristics of electricity load accurately. The attention mechanism is employed to calculate the critical weight to enhance the prediction accuracy dynamically. Finally, detailed comparison experiments show that the proposed hybrid model outperformed state-of-the-art algorithms. The MAPE of the hourly and daily prediction results of the proposed model are 2.36% and 0.76%, respectively.