In view of the non-linearity and time-varying of short-term load data in power system, a local instant learning algorithm based on variable correlation and a short-term load forecasting model based on least squares support vector machine are proposed. Mutual information is used to calculate the correlation of meteorological data, meteorological factors and other variables, and it is introduced into the training set of real-time learning algorithm to select the current power system load modeling neighborhood and improve the accuracy of short-term load model prediction. Similarity threshold is used to update the local model adaptively to enhance the real-time performance of the system load model. The load forecasting in Wancheng District of a City is carried out by using Matlab. The results show that the short-term load forecasting model based on instant learning algorithm has smaller error and higher prediction accuracy. This work is supported by National Natural Science Foundation of China (No. 71471060) and Scientific and Technological Project of Henan Province (No. 202102210134).