A Multivariate Load Forecasting Method Based on ESAM-MTL Model

Long-term dependencies between long time series data of multivariate loads make errors in multivariate load forecasting. Specifically, based on the self-attention mechanism, the efficient self-attention mechanism (ESAM) is proposed in this paper. Compared with the traditional self-attention mechanism, ESAM has lower spatiotemporal complexity and solves the problems of long training time and memory overflow to some extent. Then, by sharing information among multiple load prediction tasks using the Multi-Task Learning (MTL) architecture a multivariate load forecasting model based on ESAM-MTL was established. Case analyses use two datasets for multivariate load forecasting to illustrate the effectiveness of the proposed method.