Abstract: In the research of public building energy consumption, public building load is easy to produce abnormal data, so identifying and repairing abnormal load value is an indispensable data processing link. In view of the limitations of existing abnormal data identification and repair methods, this paper proposes a method of public building load abnormal data identification and repair based on MDLOF-iForest algorithm and M-KNN-Slope algorithm. Mdloof-iforest algorithm introduces Mahalanobis distance into the traditional LOF algorithm to improve the model's perception of the correlation between data features. Meanwhile, combining the advantages of MDLOF algorithm and iForest algorithm, it can quickly and accurately identify abnormal data. Secondly, M-KNN-Slope algorithm uses neighbors with similar load trend line characteristics of abnormal data and normal data to obtain the slope weighted average value of similar trend line, complete the repair of abnormal data, and reduce the dependence on sample data. Through the verification of load data of an office public building and a commercial public building in Guangxi, China from August to November 2024, the difference between 90% data and correct data is less than 10%, and compared with the general algorithm, M-KNN-Slope algorithm can obtain more data with error less than 5%. The Root Mean Square Error (RMSE) of XGBoost, LSTM, BP and SVM before and after repair was reduced by 5.02% to 14.40%, respectively. Mean Absolute Error (MAE) was reduced by 2.44% to 13.34%, respectively.

Key words: Public buildings, Load, Abnormal data, MDLOF-iForest, M-KNN-Slope