Anomaly Detection in Low-Voltage Grids with LSTM Autoencoders: A Study on Future Scenario Impacts

This paper presents an innovative approach using Long Short-Term Memory (LSTM) autoencoders for anomaly detection in time series net power data, specifically focusing on low-voltage (LV) grids. The rapid integration of distributed energy resources and new loads introduces significant operational challenges, such as reverse power flow and pattern changes. The data-driven approach, such as LSTM autoencoders, can detect significant pattern changes without requiring pre-categorized or labeled data. The time series net power data used for anomaly detection lacks labeled anomalies, so the model must operate in an unsupervised manner. The study highlights the model's ability to detect anomalies by considering future scenarios in the grid, such as unaccounted photovoltaic systems and new loads. Evaluations on real LV grid data, characterized by high Photo-voltaic (PV) power generation, show that this method detected 107 out of 128 anomalies. These results highlight the potential of LSTM autoencoders to enhance anomaly detection, providing a robust solution for managing and analyzing large energy usage datasets.