Optimization of support vector regression model based on outlier detection methods for predicting electricity consumption of a public building WSHP system

- Both Boxplot and LOF methods are able to find the potential outliers.
- The PCOut method shows poor performance on outlier detection.
- The SVR-based prediction model is optimized by Boxplot and LOF methods.
- Combining Boxplot and LOF methods can find the most possible outliers.

Analysis and application for real-time operational data of buildings is important for energy management. But original data inevitably contains a number of outliers and which usually lead to a significant negative impact on performance of data-based models. In order to eliminate the influence of outliers and improve the robustness of data-based models, this study employs three methods (Boxplot, local outlier factor (LOF) and PCOut) to identify the potential outlying observations in original data set. For purpose of evaluating the outlier detection performance of these three methods, four SVR-based electricity consumption prediction models, Original-SVR, BOX-SVR, LOF-SVR and PCO-SVR, are established. And the performance indexes (RE, RMSE and RSE) of the models are compared and analyzed. The results show that the accuracy of electricity consumption prediction is improved with the help of Boxplot and LOF methods for outlier detection, but PCOut method reduces the accuracy compared with the Original-SVR model. Further study indicates that these observations which repeatedly identified as outlying by Boxplot and LOF methods are the most likely to be abnormal, and when these samples are removed from training data set, the RMSE falls to 2.76 from 6.44 and the RSE falls to 0.11 from 0.58 during testing course.