Abstract: The operation scenarios of integrated energy systems have extreme patterns and contain abnormal data, which sharply increases the difficulty of integrated energy load forecasting. This paper aims to improve the accuracy of forecasting integrated energy loads by recognizing extreme patterns, detecting abnormal data, and proposing a method of short-term load forecasting of integrated energy systems based on reconstruction error and extreme pattern recognition. First, by clustering integrated energy load data based on the smallest cumulative distance, extreme patterns of the system are found. Then, error reconstruction is performed using clustering error and the residual of the deep learning model to detect anomalous data. Finally, the improved Stacking integrated learning method is used to forecast integrated energy loads in extreme patterns. The proposed method is tested against previous methods on a typical integrated energy system. The experimental results show that the proposed method is effective in addressing the issue of forecasting integrated energy loads with extreme patterns.