Hyperparameter Optimization of Long Short Term Memory Models for Interpretable Electrical Fault Classification

Abstract

The reliability of the model significantly affects early detection and accurate classification of electrical faults. In this study, a Long Short Term Memory based fault classification model was developed for the Power System Machine Learning benchmark dataset, focusing on improving reliability by increasing interpretability. First, novel metrics are introduced to measure model interpretability. These interpretability metrics are uniquely defined based on the disentanglement of the fault classification factors. Subsequently, hyperparameter optimization was performed using multi-objective Bayesian Optimization to determine the optimal model architecture. The objective of optimization is to maximize interpretability and classification accuracy. The Pareto-optimal solution presents different model architectures with varying accuracy and interpretability trade-offs. Finally, the manifestation of interpretability in terms of subsequences is studied using Shapley Additive Explanations. The impact of class representation and architectural parameters on interpretability was also analyzed. Furthermore, the most accurate model in the Pareto front achieved highly competitive accuracy for the benchmark data.