Data Driven Machine Learning Model for Condition Monitoring and Anomaly Detection in Power Grids

The power system complexity and associated stability problems are greatly linked to the increasing penetration of unconventional energy sources and loads, such as renewable energies. The application of renewable for climate change, sustainability, and Net Zero come at the cost of deteriorated power quality, faults, instability, and disturbances in the power system. It gives rise to various problems such as equipment malfunctioning, power factor problems, transformer heating, inertia, voltage sags/swells, transmission lines overloading, etc. This requires and adjudicates the need for efficient monitoring and identification of faults and anomalies happening in the power system so as to accordingly mitigate these in a timely manner. The fault data however is not readily available and requires on-site inspection and accumulation. This paper thus aims at developing a synthetic database for various abnormal power system conditions captured from a well-known Kundr's two-area system. These include symmetrical and asymmetrical faults, frequency, and phase variations, as well as voltage amplitude disturbances (sag/swell). The synthetic database is then combined with artificial intelligence techniques to enable fault detection and identification featuring low linear complexity and small memory requirements. The paper includes a benchmark study for three unsupervised anomaly detection algorithms, evaluating their performance in terms of both Area under the ROC Curve (AUC) and the execution time. The results show that iForest and iNNE provide competitive results in detecting anomalies of all fault types, with iNNE providing significantly better execution time performance.