An Anomaly Detection Scheme based on LSTM Autoencoder for Energy Management

This paper proposes an anomaly detection scheme based on LSTM autoencoder for energy management, which is to prevent anomaly states before they actually occur. When the prognosis of an anomaly state is detected, the anomaly state can be prevented by taking appropriate measures. However, it is difficult to determine normal and anomalous data, since energy consumption varies greatly depending on weather, time, day of the week and season. Thus, this paper proposes an anomaly detection scheme using LSTM autoencoder to detect a data pattern that deviates from the normal data pattern and to determine it as an anomaly state. Experimental results show that this scheme can discriminate anomaly from the observed multivariate data and can be used to prevent fault and incorrect operation in advance.

I. Introduction (Heading 1)

Recently, anomaly detection technologies have been extensively studied in various domains including fault detection, intrusion detection, fraud detection, and event detection [1]-[6]. The purpose of anomaly detection is to find something that deviates from what is standard, normal, or expected. Thus, it needs to detect any data pattern indicating the prognostic symptoms of anomaly states during operation. If anomalies are detected in a timely manner, it can prevent performance degradation, improper operation, and accidents from proceeding any further. Therefore, it can help building managers expedite fault recovery and prevents further performance deterioration.