| dc.description.abstract | There is no world without energy. Dependence on energy continues to dominate everything that we do. It is known that any failure in the production of energy can directly affect thousands of lives. Because of this, data is closely monitored and collected. Our research intends to apply Automated anomaly detection to energy performance data to detect degradations in energy consumption. We use a multivariate time series dataset from a six year period of time at a Combined Cycle Power Plant. Anomaly detection is a data analysis method with the purpose of identifying points in data that do not follow the intended behavior of the dataset. These points can be caused by error, technical faults, or bugs that have potentially devastating impacts if unnoticed. Anomaly detection has become more flexible as more methods of data processing, feature analysis, and detection have become available. Although the techniques of anomaly detection have drastically improved in recent years, there has been little research done on automated anomaly detection. Building an anomaly detection system requires an expert to manually select features such as data pre-processing methods and feature analysis methods in order to construct an anomaly detection pipeline that is suitable for the dataset. This method is very costly and can be done with a machine learning approach. With the incorporation of machine learning, Automated Anomaly Detection has the ability to build an optimal pipeline according to the types of the dataset. Instead of wasting time and money manually building possibly unreliable anomaly detection algorithms, the process is simplified by just feeding in the desired dataset to detect anomalies. The system would process the dataset and get information by running the system on the dataset. According to the information, the system would pick different components and build pipelines to check for the accuracy until the best optimized pipeline is generated for the dataset. Our Anomaly detection system construction is built by Time Series Outlier Detection System (TODS) which implements modern machine learning principles to construct an optimal anomaly detection system for our time series dataset. TODS focuses on data processing, time series processing, feature analysis, and detection algorithms to construct an outlier detection system. Utilizing automated anomaly detection methods in monitoring energy consumption can help energy consumers quickly find and fix the degraded parts to minimize consumption and provide more safety to users. The requirement of an expert in energy performance is no longer needed. The cost of constructing anomaly detection pipelines for industries would significantly be lower as well. | |
