Comparison of an autoencoder and an insulation forest model for the condition assessment of low voltage cables based on powerline communication data

The increasing use of decentralised energy sources and new load types in low voltage (LV) grids is leading to bidirectional power flows and higher loads on underground cables. These conditions accelerate ageing and increase the risk of failure. Nevertheless, LV cables are largely operated without monitoring. Power line communication (PLC), which is often used for smart metering and automation, provides a sensorless basis for condition assessment, as its signals reflect the electrical environment. However, the metadata obtained from PLC is high dimensional, unsynchronised and operationally variable, which makes conventional analysis difficult. This paper explores unsupervised machine learning (ML) methods for detecting anomalies in PLC data. Three approaches are compared: an autoencoder…