Real-time Anomaly Detection of Power Demand using Probabilistic Forecasting

Abstract

A significant percentage of total energy consumption is represented by residential electricity usage, which has been steadily increasing due to the ongoing digitalization of our world. As household electricity consumption grows, also intensifies the dependency on electrical systems, making it critical to ensure an uninterrupted supply of electricity. Continuous electricity provision plays a vital role in maintaining a secure energy supply. However, in abnormal situations such as disasters or crises, simultaneous high electricity usage can risk grid security. This can lead to overloading of the network and, if appropriate measures are not taken, result in a significant power outages. It might be possible to detect such problems through real-time monitoring, enabling grid operators to intervene in a timely manner. In this thesis, a prediction-based real-time anomaly detection method has been developed. Firstly, household electricity consumption will be forecasted using probabilistic methods like Quantile Regression and XGBoost-Quantile. By predicting various quantiles through probabilistic methods, potential consumption points will be identified. These predicted values and real-time consumption data will be inputs to the anomaly detection strategies. Subsequently, the developed unsupervised real-time strategies detect anomalies in real-time by comparing these inputs and interpreting the pattern differences. Additionally, a deterministic method, XGBoost, will be used to predict consumption, and the results of anomaly detection strategies based on this deterministic prediction will be compared. Aggregated residential data from 2019 and 2020 will be used as test data. Artificial anomaly scenarios will also be added to the test data to evaluate performance under both anomalous and non-anomalous conditions. Furthermore, anomaly detection strategies and each load forecasting method using metrics will be evaluated.