Short-term load forecasting, Smart grid, Deep learning, Mask-guided multiscale interactive self-attention mechanism, Convolutional enhancement-fusion embedding, Medicine, Science

Abstract Accurate short-term load forecasting (STLF) provides important support for the economic and stable operation of the power system. Although various deep learning methods have achieved good results in STLF, they usually model load features only from a limited perspective, i.e., they do not uniformly utilize the three features of multivariate load data: the influence of covariates, multiscale features and local-global variations. The insufficient mining of these three features limits the improvement of prediction accuracy. To address the above problems, we design a novel STLF model called UniLF based on Transformer framework, which contains the proposed convolutional enhancement-fusion embedding method to capture the correlations between load and covariates for embedding, the proposed feature reconstruction-decomposition block to distill multiscale features as well as more detailed local-global variations from 2D space and the core mask-guided multiscale interactive self-attention mechanism to further realize the enhanced interactions of scale features and temporal features. Experiments conducted on three load datasets from Australia, Panama and Austria show that UniLF achieves superior forecasting accuracy with competitive practical efficiency under different prediction lengths, providing a new solution for STLF.