Transfer Learning-Based Model Training for Short-Term Load Forecasting

Abstract

The smart grid infrastructure has recorded extensive real-time electricity consumption data, particularly at the levels of distribution transformers and below for short-term load forecasting (STLF). However, training individual short-term load forecasting model (SLFM) for each STLF scenario at these levels substantially increases the computational costs. To address this challenge, this paper proposes a transfer learning-based model training method for STLF. The proposed method is rooted in transfer learning principles and tailored to the unique characteristics of the aforementioned levels, incorporating several key steps. First, an approach for extracting key peak and valley points based on peak width and peak prominence is proposed for simplifying the evaluation of load sequence similarity. Subsequently, these key points are clustered using a density-based spatial clustering of applications with noise approach to ensure proper alignment along the time axis. Secondly, temporal and distribution similarity metrics are introduced to establish a performance guarantee for the transferred SLFM. Subsequently, a hierarchical clustering method groups load sequences, utilizing temporal similarity to quantify distances among sequences and distribution similarity to optimize cluster number selection. To minimize generalization error and further reduce computational costs, a modified bagging method is proposed and applied during the transferred SLFM fine-tuning. Empirical evidence from a study conducted in Guiyang, China demonstrates that the proposed method maintains the SLFM performance without degradation and significantly reduces computational costs by a minimum of 92.23% across multiple scenarios.