Building energy management through fault detection analysis using pattern recognition techniques applied on residual neural networks

In this paper a fault detection analysis through a neural networks ensembling approach and statistical pattern recognition techniques is presented. Abnormal consumption or faults are detected by analyzing the residual values, which are the difference between the expected and the real operating data. The residuals are more sensitive to faults and insensitive to noise. In this study, first, the experimentation is carried out over two months monitoring data set for the lighting energy consumption of an actual office building. Using a fault free data set for the training, an artificial neural networks ensemble (ANNE) is used for the estimation of hourly lighting energy consumption in normal operational conditions. The fault detection is performed through the analysis of the magnitude of residuals using peak outliers detection method. Second, the fault detection analysis is also carried out through statistical pattern recognition techniques on structured residuals of lighting power consumption considering different influencing attributes i.e. number of people, global solar radiation etc. Moreover the results obtained from these methods are compared to minimize the false anomalies and to improve the FDD process. Experimental results show the effectiveness of the ensembling approach in automatic detection of abnormal building lighting energy consumption. The results also indicate that statistical pattern recognition techniques applied to residuals are useful for detecting and isolating the faults as well as noise.