A CNN-Sequence-to-Sequence network with attention for residential short-term load forecasting

Abstract

Residential short-term load forecasting has become an essential process to develop successful demand response strategies, and help utilities and customers optimize energy production and consumption. Most previous works focused on capturing the spatial and temporal characteristics of residential load data but fell short in accurately comprehending its variations and dynamics. The challenges come from the high non-linearity and volatility of the electric load data, and their complex spatial and temporal characteristics. To address these challenges, we propose a hybrid deep learning approach consisting of a Convolutional Neural Network and an attention-based Sequence-to-Sequence network. The model aims at capturing the spatial and temporal features from time-series data, the irregular load pattern, and the frequent peak consumption values to improve the overall quality of the forecasts. The proposed model is compared to several state-of-the-art approaches, and the performance is validated on the residential load data for a household in Sceaux, France. The results showed an improvement of 9.6% in the mean square error on different prediction time horizons. The proposed approach produced more accurate real-time forecasts and showed better adaptation at peak consumption instances.