Classifying electric service disruptions under deep uncertainty: A comparative analysis of algorithms and their policy outcomes.

As machine learning finds increasing applications in energy policy, it is critical to evaluate model accuracy alongside interpretability and equity. We leverage classification of electricity disruptions within residential advanced metering infrastructure (AMI) data as a testbed for evaluating four models: logistic regression, a support vector machine (SVM), a histogram-based gradient boosting (HGB), and a neural network. We find that the HGB approach consistently outperforms logistic regression and marginally outperforms the SVM and the neural network across multiple evaluation metrics. Despite similar accuracy across models, derived duration and frequency metrics show that models prioritize different communities for investment, reflecting the challenges of decision-making under deep uncertainty. Depending on the reliability metric, up to 36 communities (over 100,000 households) are sensitive to the choice of classifier. These results highlight how the investment pathways for outage mitigation vary with algorithmic choice, emphasizing the need for both interpretable and equitable models.