Electrical Load Forecasting Based on Multi-Task Deep Learning

Electrical load forecasting models are usually designed for single task, however there are certain correlations among multiple forecasting tasks. This paper proposes a novel electrical load forecasting model based on multi-task deep learning, which predicts daily and hourly load demand simultaneously and addresses the limitations of existing models in accommodating different load forecasting tasks. Firstly, the daily and hourly load features are extracted through the cascaded Temporal Convolutional Neural network, which fully leverages the intrinsic correlation between two types of load data. Subsequently, the Long Short-Term Memory network with Gated Residual Units is employed to implement the two forecasting tasks. Additionally, composite attention suitable for each task is generated by the Cross-Stitch module, and a connection of information between the two tasks is established. As a result, two load forecasting outcomes are output simultaneously. The proposed model achieves the MAPE of 1.39% for daily load forecasting and 1.72% for hourly load forecasting on the real load dataset, outperforming several existing forecasting models.