Short-term residential load forecasting via transfer learning and multi-attention fusion for EVs’ coordinated charging

Accurate load forecasting plays a crucial role in the optimal scheduling of electric vehicles’ (EVs) coordinated charging. Although many load forecasting methods have emerged in recent years, these methods face two significant challenges: effectively capturing the impact of special events on load and requiring a substantial amount of historical data for model training. To better apply day-ahead load forecasting (DALF) to the optimal scheduling of EVs’ coordinated charging, we propose a Transformer-based network architecture combining transfer learning and multi-attention fusion. The core idea of this algorithm comprehensively considers the dependencies between special events, seasonality, and load through multi-attention fusion. Simultaneously, by introducing time series decomposition block (TSDBlock), the load data is decomposed into seasonal and trend components to effectively extract crucial information, enhancing the performance of load forecasting models. To address data scarcity issues, we introduce transfer learning, selecting the best-performing model on the target task as the source model to avoid negative transfer effects. Ultimately, the experimental results show that our proposed method achieves the best forecasting performance in the datasets in five different regions. Especially in the non-working days, its performance is outstanding. • Introduced a Transformer-based network with transfer learning and multi-attention fusion, enhancing short-term load forecasting performance. • Proposed EPSAttention and CSPBlock to capture complex dependencies. • Addressed data scarcity through strategic transfer learning, selecting the source model based on performance. • Demonstrated outstanding forecasting performance across five different regions, particularly excelling in non-working days.