Because the abnormal location of the power grid load is mismarked, the average relative error increases in the process of forecasting. This paper proposes to study and analyze the method of short-term load forecasting based on data mining. It preprocesses the load characteristic clustering and related data, marks the abnormal position of a power grid within a calibrated mining range according to the load fluctuation state, and defines the short-term load forecasting range. It also constructs a data mining power load forecasting model through the acquired real-time forecasting data and information, designs layered and staged load forecasting links, and corrects and calculates by adopting data mining. The prediction residual error is obtained. The stage prediction standard is further determined, and the prediction processing is realized. The final test results show that the data mining short-term load forecasting test group designed in this paper has a small fluctuation range, indicating that the prediction error has been greatly controlled in the practical application process. The prediction error is small, the accuracy is improved, and it has practical application value.