This paper proposes a solution for fault detection in power systems using machine learning algorithms, namely Random Forest and Support Vector Machines. A Phasor Measurment Unit (PMU) network is emulated in the IEEE 39-Bus New England Power System, and several fault types are simulated, including three-phase to ground, two-phase, two-phase to ground and single-phase to ground as well as line and load contingencies. The magnitude and phase of voltage and current, alongside with frequency, are measured from each PMU, and used as input to the machine learning models. Two scenarios were contemplated in this work, the first with a network of 14 PMUs, and the second with half that number, in order to verify the robustness of the aforementioned methods in relation to the number of PMUs present in the system. A Feature Importance analysis is also made, via Permutation Feature Importance, indicating which features contributed the most to the classification task at hand. Both algorithms reached a performance of around 93% accuracy and 0.94 F 1 -Score and the feature analysis method seems to be suitable for systemic visibility analysis. Future works are also discussed in this paper, briefly elaborating on the possibilities and immediate impacts of the addition of a feature engineering stage in this problem and on the application of the used algorithms on problems such as fault identification and location.

No abstract