Weather-Induced Power Outage Prediction: A Comparison of Machine Learning Models

• Main Authors: Garland, Jasmine, Baker, Kyri, Livneh, Ben
• Format: Conference Proceeding
• Language: English
• Subjects: machine learning; power outages; reliability; weather
• ISSN: 2474-2902
• DOI: 10.1109/SmartGridComm57358.2023.10333953

This study investigates power outage duration and occurrence for the state of California, Los Angeles County, and Alpine County using historical weather data. With extreme weather events on the rise, many areas have seen a subsequent increase in weather-related power outages. To better understand the relationship between these two variables, generalizable frameworks using publicly available data are investigated, using outage data down to a single customer. These models aim to predict power outage duration and occurrence, which would aid in power system preparedness and restoration efforts. The results of the models are also assessed not only in terms of prediction accuracy, but also in terms of reliability metrics. Out of the models considered to predict power outage duration (regression), the Multi-Layer Feed-Forward Neural Network (MFFNN) showed significantly better results for large regions, but showed mixed results for a small region. When considering power outage occurrence predictions (classification), models that incorporate boosting outperformed the more simplistic models. The results indicate that the optimal model is highly dependent on the characteristics of the region, and that these models can prove valuable for identifying regions that will have weather-related reliability impacts in the future.