Traditional load forecasting algorithm can predict short-term load when there is no bad data in historical load sequence. Actual load data will inevitably produce errors during the process of monitoring, collecting and storing, if the traditional prediction algorithm is still used for load forecasting, it may cause large errors at some time nodes. In order to solve this problem, this paper proposes a prediction model which can not only identify and correct the bad data of historical load sequence and but also predict the load similarity and correct the load deviation. By using the actual load data to verify the model, the model can better avoid the interference of bad data, and effectively improve the prediction accuracy of the historical load sequence with bad data. This work is supported by National Natural Science Foundation of China (No. 61602295) and Natural Science Foundation of Shanghai (No. 16ZR1413100).