Short-term electrical load forecasting based on multi-granularity time augmented learning

Electrical load forecasting is a core element reflecting the operating conditions of the electricity system and a key tool responding to the demand of the electricity market. Achieving accurate short-term load forecasts remains a challenge due to the dynamic and non-stationary characteristics of the load data. Previous studies have mostly analyzed electrical load transformations from a single perspective. This approach often overlooks the dynamic diversity across different frequencies and the comprehensive effects of multi-time scale and granularity information. Research in electrical load forecasting has frequently failed to fully integrate multi-granularity perspectives. In this study, we introduce a novel approach, multi-granularity time-augmented learning (MTAL), to enhance the precision of short-term electrical load forecasting. Since the degree of dynamic change of different granularity information is overly influenced by time features, we design a time-augmented block to learn time representation and apply it to all granularity information to represent multi-granularity electrical load more reasonably. Furthermore, we incorporate an attention mechanism into the model, which serves to mitigate information redundancy and bolster its generalization capabilities. We evaluated our method on a univariate electrical load dataset and a multivariate electrical load dataset, respectively, and compared its performance with existing forecasting models. Experiments demonstrate that the MTAL model performs well in capturing load variation information and achieves better performance in both univariate and multivariate short-term electric load forecasting tasks. Compared to existing methods, our proposed model improves the prediction accuracy by 10 \(\%\) $\%$ and reduces the computation time by 18 \(\%\) $\%$ .