<![CDATA[The reliability of three-phase power systems depends heavily on the ability to detect faults quickly and accurately. However, much previous research has faced challenges such as data leakage, biased evaluations due to random data splitting, and limitations in the utilisation of system dynamics-based features. This study aims to develop a machine learning-based fault detection model that is accurate, robust, and possesses high generalisation capabilities. The method employed is a quantitative approach using an Artificial Neural Network (ANN) model as a binary classifier (normal and fault). The research process includes data preprocessing, elimination of features that could potentially cause data leakage, feature engineering based on current, voltage, energy, and system imbalance, as well as scenario-based data splitting to ensure the validity of the evaluation. The model was trained using repeated cross-validation and optimised based on the Area Under the Curve (AUC) metric. The results show that the ANN model achieved high performance with an AUC of 0.987 and an accuracy of approximately 98.5% and remained stable under varying conditions such as noise and load changes. Compared to other models, Random Forest delivered the best performance, followed by Support Vector Machine (SVM) and ANN. This study addresses the gap regarding more realistic evaluation methods that are free from data leakage. Its novelty lies in the use of scenario-based splitting, comprehensive evaluation, and the finding that raw signal features make a dominant contribution to fault detection.]]>
