Location
Hilton Hawaiian Village, Honolulu, Hawaii
Event Website
https://hicss.hawaii.edu/
Start Date
3-1-2024 12:00 AM
End Date
6-1-2024 12:00 AM
Description
Severe weather conditions are known for causing forced outages in the electric distribution grid. Recent research efforts were aimed at predicting outages using weather and historical outage data. This paper studies the sensitivity of different Machine Learning (ML) algorithms to the inclusion of weather parameters from adjacent geographic areas and data availability. We analyzed the ability of different ML algorithms to predict electric grid outage State of Risk (SoR). The selected algorithms are trained and tested on actual utility company data. The findings indicate that a bigger size of the training dataset improves the performance of all models, which is measured by the Receiver Operating Curve, Average Precision, and F1 Score. Conducted experiments suggest that at least two years of training data is required to achieve satisfactory performance. Also, we investigate a statistical significance in models’ performance with the inclusion of weather in adjacent geographic areas.
Recommended Citation
Baembitov, Rashid; Kezunovic, Mladen; Saranovic, Daniel; and Obradovic, Zoran, "Sensitivity Analysis of Machine Learning Algorithms for Outage Risk Prediction" (2024). Hawaii International Conference on System Sciences 2024 (HICSS-57). 7.
https://aisel.aisnet.org/hicss-57/es/resilient_networks/7
Sensitivity Analysis of Machine Learning Algorithms for Outage Risk Prediction
Hilton Hawaiian Village, Honolulu, Hawaii
Severe weather conditions are known for causing forced outages in the electric distribution grid. Recent research efforts were aimed at predicting outages using weather and historical outage data. This paper studies the sensitivity of different Machine Learning (ML) algorithms to the inclusion of weather parameters from adjacent geographic areas and data availability. We analyzed the ability of different ML algorithms to predict electric grid outage State of Risk (SoR). The selected algorithms are trained and tested on actual utility company data. The findings indicate that a bigger size of the training dataset improves the performance of all models, which is measured by the Receiver Operating Curve, Average Precision, and F1 Score. Conducted experiments suggest that at least two years of training data is required to achieve satisfactory performance. Also, we investigate a statistical significance in models’ performance with the inclusion of weather in adjacent geographic areas.
https://aisel.aisnet.org/hicss-57/es/resilient_networks/7