To address challenges in short-term load forecasting for distribution networks, challenges such as the impact of complex weather features and the difficulty in analyzing correlations between different feature sequences using traditional models, a method considering those issues is proposed. First, the method presents a multi-variable rapid maximal information coefficient (MVRapidMIC) to extract highly correlated weather feature sequences. Exploratory factor analysis (EFA) is then employed for dimensionality reduction on these sequences. Finally, the dimension-segment-wise (DSW) and two-stage attention (TSA) mechanisms are integrated with the Informer model to enhance the model’s ability to analyze correlations between different feature sequences. Simulation tests using historical load data from the DTU 7K 47-bus distribution system validate the forecasting accuracy, robustness, and timeliness of the method.