Dr. Xiaobo Xue Romeiko

University at Albany, United States

Invited Speech: Comparing Several Machine Learning Models to Predict Power Outages

Biography:

Xiaobo Xue Romeiko, PhD
Associate Professor
Department of Environmental Health Sciences
School of Public Health
University at Albany, State University of New York, United States

Dr. Romeiko is an associate professor hosted at the Department of Environmental Health Sciences of the University at Albany. Dr. Romeiko gained her doctoral degree at the University of Pittsburgh and worked at the US Environmental Protection Agency prior to joining in the University at Albany. Dr. Romeiko’s research focuses on advancing quantitative approaches for assessing environmental health impacts of the interconnected food, energy and water systems.

Abstract:

Power outage (PO) became a pressing national concern due to its increasing frequencies and devastating consequences. Accurately predicting PO is urgently required for disaster preparedness and management. In order to accurately predict the numbers of PO interruptions and affected customers, this study built six different machine learning models based upon a comprehensive set of input variables reflecting infrastructure, weather, land use, soil and neighborhood characteristics. These six models included Artificial Neural Network (ANN), Random forest, eXtreme Gradient Boosting, Support Vector Machine (SVM), linear regression, and Poisson regression. This study also ranked the importance of input variables by using Morris Sensitivity analysis. The model comparison suggested that ANN model presented the highest predictive accuracy for numbers of interruptions. Differently, Xgboost model presented the highest predictive accuracy for the numbers of affected customers. Hurricane, rainfall, overall power generation, winter storm and grassland ranked as the top five features for the numbers of interruptions. Similarly, hurricane, rainfall and overall power generation remained the top three features for the numbers of affected customers. However, the fourth and fifth features for the numbers of affected customers were switched to flooding and population density. Overall, these findings suggested that the PO occurrence and coverage are complex interplay among extreme weather, power infrastructure, land use and population density.

Abstract

Dr. Xiaobo Xue Romeiko

Power outage (PO) became a pressing national concern due to its increasing frequencies and devastating consequences. Accurately predicting PO is urgently required for disaster preparedness and management. In order to accurately predict the numbers of PO interruptions and affected customers, this study built six different machine learning models based upon a comprehensive set of input variables reflecting infrastructure, weather, land use, soil and neighborhood characteristics. These six models included Artificial Neural Network (ANN), Random forest, eXtreme Gradient Boosting, Support Vector Machine (SVM), linear regression, and Poisson regression. This study also ranked the importance of input variables by using Morris Sensitivity analysis. The model comparison suggested that ANN model presented the highest predictive accuracy for numbers of interruptions. Differently, Xgboost model presented the highest predictive accuracy for the numbers of affected customers. Hurricane, rainfall, overall power generation, winter storm and grassland ranked as the top five features for the numbers of interruptions. Similarly, hurricane, rainfall and overall power generation remained the top three features for the numbers of affected customers. However, the fourth and fifth features for the numbers of affected customers were switched to flooding and population density. Overall, these findings suggested that the PO occurrence and coverage are complex interplay among extreme weather, power infrastructure, land use and population density.