SCALING UP ENERGY DATA ANOMALY DETECTION WITH TREE MODELS: AN ANALYSIS OF GENERALIZABILITY AND APPLICABILITY

Abstract

This paper discusses the presence of big data in the energy consumption inbuilding sensors. Leveraging on the availability of data, many methods for detecting and diagnosing faults have been established. Although many models were proposed for energy anomalies over the last few years, there has always been a lack of discussion on the generalizability of models in energy data anomaly detection. Therefore, this research will investigate how well the proposed method performs under different environments and different meter types to determine the generalizability and applicability of the model. This study has developed and trained a tree-based classification model to test and determine the accuracy in predicting anomalies data. The tree-based model was chosen as it was able to outperform the baseline models to achieve higher accuracy. From the research findings, the proposed model performed excellently with an AUC-ROC score of 0.983. Despite a change in environment and meter type, the model was still able to perform excellently, with a score higher than 0.90 in all instances. In addition, only 10% of the dataset is required to achieve stable performance. From the research, we can conclude that the tree-based classification model is a good model which can be used for different datasets and is suited for all building types. The model allows for building owners and facilities managers to detect anomalies earlier and take actions to prevent energy wastage and extra expenses. The automation of the process also helps to reduce the amount of effort put in by building stakeholders.