When it comes to enhancing the overall dispatch of the power system and assisting with the operation of the power grid, short-term load forecasting, particularly non-holiday load forecasting, is essential. This research provides a mixed model prediction method based on clustering algorithm to choose similar days in order to increase the prediction accuracy of non-holidays. The parameters that were significantly correlated with load were chosen using the Pearson coefficient and significance test. The best similar days were then chosen by a clustering method based on the main influencing elements, and the similar daily load set, historical load, weather parameters, and time factors were used as the model’s input characteristics. To finish the forecasting assignment, a generalized regression model is created for each corresponding daily load set. As a concrete example, the data set provided by a power supply bureau in southern China is used. The experimental findings demonstrate the improved prediction accuracy and quicker model training speed of the suggested strategy.