Abstract: A regional short-term load forecasting method based on virtual similar days and dual-stage attention-based long and short time pattern network（DA-LSTPNet） is proposed for the demand of the refinement of short-term load forecasting. To obtain the fine-grained real-time meteorological data matching the load, a meteorological virtual similar day containing fine-grained meteorological data is firstly obtained using gray correlation and correlation weighting method based on the coarse-grained meteorological data. Then, the maximum information coefficient（MIC） is used to analyze the nonlinear correlation between meteorological feature information and load. And the MIC-weighted selection algorithm for the load virtual similar day is constructed to solve the problem of excessive local similarity or even non-similarity caused by selecting historical days as traditional load similar days. Finally, in order to address the problem that the relationship between characteristic factors and local load fine-grained variation, the DA-LSTPNet is constructed to effectively explore the characteristics of long-term macroscopic and short-term local variation of load feature data for the day-ahead short-term load forecasting. Based on the actual load data of the power grid in a certain area of southern China, the various forms of simulation are used to demonstrate the higher prediction accuracy and universality of the proposed forecasting method.