The reliability of power distribution networks is critical to maintaining the continuous delivery of electricity to all consumers and sectors of the economy. Fault localization plays a key role in minimizing downtime and repair costs. In recent years, a growing number of studies have leveraged machine learning techniques to enhance the precision and the speed of fault detection and localization. However, a major challenge remains: the scarcity of comprehensive datasets that capture the complexities of power system faults. This study addresses this limitation by developing an enriched dataset and proposing methods to mitigate overfitting, a phenomenon in which a model captures noise or patterns specific to the training data, impairing its performance on unseen data. By augmenting existing data with simulated fault scenarios, our framework improves robustness and generalizability, and reduces the risk of overfitting of machine learning models.