Deep learning and transfer learning techniques applied to short-term load forecasting of data-poor buildings in local energy communities

Abstract

The use of deep learning for electrical demand forecasting has shown great potential in generating accurate results, but requires a large amount of data to train the models. However, the limited availability of electricity consumption information for new users or recently monitoring buildings in local energy communities make it difficult to achieve these good results. To address this challenge, this research study proposes a novel methodology based on the combination of the transfer learning (TL) concept and the temporal fusion transformer (TFT) architecture. This is an innovative approach used to transfer knowledge from a general pre-trained model based on the historical data of the community's buildings to new users, reducing the needs for large amounts of data for each new building. The results show that TFT provides more accurate load estimates than other state-of-the-art methods by reducing RMSE by more than 11%. Moreover, the TL approach improves the prediction of load demand of buildings with limited availability of historical data, reducing CV_RMSE, SMAPE, and WQLoss by more than 40% compared to models that do not use knowledge transfer. This is a significant improvement that facilitates the incorporation of new users to the community achieving accurate load estimations even with lack of data. Moreover, this approach leads to an optimized and cost-efficient planning and management of energy distribution in local energy communities.