A synchronous prediction method for hourly energy consumption of abnormal monitoring branch based on the data-driven

Abstract

Data quality assurance of building energy consumption monitoring platform plays an important role in building energy consumption management. However, Data collected and transferred to the cloud platform is affected by many factors and shows some missing values and outliers in the time series. Combining with the correlation of smart meters on branches and regularity of building energy consumption characteristics, this paper proposes a synchronous prediction method for predicting building energy consumption in the secondary branch. In this model, synchronous data feature similarity (SDFS) model is used to find a similar energy consumption feature, Extreme gradient boosting (XGBoost) model is used to train and produce accurate prediction results which are compared with back propagation neural network (BPNN) and adaptive boosting (AdaBoost) and maximum distance outlier correction (MDOC) model can further correct the prediction results. Taking the daily energy consumption of the primary branch with a smart meter in the building energy consumption monitoring platform as the test object, the predition results of VRF energy consumption show that the MAE, MAPE, RMSE, and CVRMSE are 1.150, 0.142, 1.511, and 0.132 respectively, which is much lower than BPNN and AdaBoost. This study explores a novel feature mining method for historical data and an integrated model for outlier recognition and correction which significantly improves the accuracy of prediction results. Moreover, after correlation verification, the prediction model can be widely applied in building distribution system with sub-meter system which improves the data utilization rate of building energy management system.