Machine Learning Applications in Weather-related Outage Forecasts

A weather-related outage forecast guidance system was built and recently implemented for the eight service districts of San Diego Gas and Electric (SDG&E).  The system is comprised of long- and short-range outage forecast guidance products.  The long-range forecast guidance product associates synoptic scale Global Forecast System (GFS) 10-day forecasts of 500mb height, mean seal-level pressure, 850mb wind speed and precipitable water with atmospheric modes.  The 110 mode synoptic climatology was established by applying self-organizing maps to the same four atmospheric variables of the North American Regional Reanalysis data across a period matching the 16 year utility outage data.  Additionally, a historical 2km resolution weather data set was constructed for SDG&E's territory using the Weather Research and Forecast (WRF) model across the same 16 years.  The short-range forecast guidance product uses 24 distinct Random Forest models to forecast both outage predictors and predictands.  These models were trained using the historical SDG&E outage, high resolution WRF and observed lightning data, all aggregated to the district level.  The operational short-range forecast guidance product uses GFS forecasts dynamically downscaled to 2km horizontal resolution using WRF twice per day.  Ultimately, the goal is to provide guidance to SDG&E meteorologists when forecasting the number of outages and customers impacted during severe weather events.