Weather-based power outage prediction in New York City: An ensemble machine learning approach

Power outages pose significant challenges in urban areas, even during non-extreme weather conditions. We hypothesized that variations in weather conditions (e.g., temperature, precipitation, humidity, snow, wind, etc.) would be predictive of power outages. Our study examined the relationship between weather variables and outage occurrences in New York City, using historical data and machine learning techniques. The ensemble regression model in this research consisted of random forest, extreme gradient boosting regressors, support vector regression, and multilayer perceptron base models with gradient boosting regressor meta model. The model achieved moderate accuracy on the most predictive dataset (mean absolute error = 3.774, mean absolute percentage error = 34.073%, mean squared error = 24.874, and R2 = 0.824). Correlation analysis identified energy demand and temperature as the features most strongly associated with outage frequency, and that rolling averages tend to be better predictors than daily values. Principal component analysis revealed that days with extreme temperatures had more outages, and that temperature, humidity, and precipitation were the main drivers of variance. The limitations of this approach include unaccounted infrastructure factors, complexity of power failure causes, and timing discrepancies in outage reports. Exploring alternative modeling techniques and expanding the dataset to other geographic regions will further refine the findings. This analysis contributes to our understanding of how urban energy systems respond to climate variability and inform strategies for enhancing power grid resilience.