Short-term load forecasting with deep learning: Improving performance with post-training specialization

Abstract

Load forecasting has increasingly relied on deep learning models due to their ability to capture complex non-linear relationships. However, these models require substantial amounts of data for effective training. Data sparsity during peak load periods can degrade the performance of deep learning models to the point that they under-perform much simpler models. To address this issue, this paper proposes a post-training specialization method in which several copies of the original deep learning model are retrained for specific forecasting tasks. Results indicate an increase in performance in all baseline models used in this paper, and the method can potentially improve the forecasting of current applications at a low computational cost.