Comparative analysis of quantum machine learning and classical deep learning for power system fault classification under noisy conditions

Classifying faults in a power system is not only an important part of making sure the electrical grid is dependable, it is also a very challenging task since you have to come up with methods that are not only robust but also able to handle things like noisy, imbalance, and overlapping data sources from modern multi-source power networks. This paper is going a step further and doing a thorough comparative study of Quantum Machine Learning (QML) and Classical Machine Learning (CML) methods for detecting seven different classes of faults in three-phase power systems. It created a highly sophisticated dataset that is quite a bit more realistic because it not only takes into account random fault inception angles, varying fault resistance, and current transformer (CT) saturation effects, along with including sensor noise that is correlated and based on realistic Internet of Things (IoT) communication channels; but also imbalance class distributions have been introduced to mimic real-world Supervisory Control and Data Acquisition (SCADA) statistics. Seven different models have been rigorously tested: Random Forest (RF), Support Vector Machine (SVM), Deep Neural Network (DNN), a Classical Shadow Baseline consistent with the constraints of a quantum circuit, a Variational Quantum Classifier (VQC) that is Constrained Optimization BY Linear Approximations (COBYLA) optimized, a Quantum Support Vector Machine (QSVM) with a kernel based on fidelity, and a Quantum Inspired Random Forest (QiRF) that makes use of the WalshHadamard Transform for feature rotation. Experimental results show that traditional models perform around 98.99% weighted F1-score on the real dataset whereas quantum models only reach 93.95%. The QiRF model is the best hybrid model among all and it is able to handle a competitive F1-score of 97.8%.