PSEformer: Period-Series Embedding Transformer with Enhanced Input Embedding for Load Forecasting

Load forecasting is a key foundation for ensuring the stable and efficient operation of smart grids. However, existing Transformer-based time series forecasting methods have notable limitations: they overly rely on attention mechanisms to capture intra-series and inter-series dependencies, which results in low computational efficiency and inadequate performance in modeling periodical features of load series (such as daily/weekly periods) and deep cross-series relationships. To address this issue, we innovate by starting with the simple yet powerful modeling technique of input embedding and designing additional embedding mechanisms for load series—period embeddings (for representing periodic features such as daily/weekly periods) and inter-series dependency embeddings. This enables efficient and accurate modeling of both intra-series and inter-series dependencies. Our proposed Period-Series Embedding Transformer (PSEformer), through this joint period-series embedding mechanism, effectively enhances the model’s ability to capture intra-series and inter-series dependencies in load data. Experimental results show that, compared to baseline models, PSEformer achieves optimal forecasting performance on three real-world power load datasets.