Evaluation of Alternate Methods of Forecasting Meteorological Parameters for Input into an Electric Utility Outage Prediction Model (OPM)

In recent years, many electric utilities have seen an increase in the amount of electric service outages and the number of customers impacted by those outages. Several causal factors have been cited including the increase in extreme or severe weather due to the changing climate, aging infrastructure, and an increase in the number of customers served. In order to be better positioned to efficiently respond to the outages, there has been a motivation to develop machine learning type models that can predict the area and volume of outages one to several days in advance.

A key input into outage prediction models (OPM) are forecasts of meteorological parameters that are related to the cause of the outages. The critical parameters are typically different from the standard meteorological parameters that provided in public weather forecasts and typically in forecast model performance analyses. Some examples of such parameters are (1) maximum 4-hr average 10-m wind speed during an event, (2) average precipitation rate during the period of 4-hr maximum wind, (3) maximum hourly precipitation rate, (4) density of fallen snow, and (5) amount snow fallen with surface temperatures near freezing. Of course, the relative importance of the OPM input parameters will vary with the type of case (e.g. winter storm with frozen precipitation, large scale rain and wind event, convective scale wind and rain etc.) In order to optimize the performance of an OPM it is desirable to maximize the accuracy of the forecasts of the key parameters.

In this investigation, an evaluation of several alternative methods of producing the predictions of the key OPM input parameters over the state of New York for 373 outage events in the 2017 to 2021 period. The data from the 126 New York State Mesonet (NYSM) surface observing stations were used to evaluate the forecasts. The alternative forecasts were based on:

• the interpolated output from NCEP operational models (GFS, NAM, RAP and HRRR)
• the interpolated output from a custom configured 2 km WRF forecast
• machine learning models developed individually from each of the above NWP datasets
• machine learning models developed from the ensemble of all of the above models

A wide range of verification metrics were employed to evaluate the forecasts including (1) the temporal and spatial correlation of the observed and forecasted values, (2) event forecast metrics for threshold values of variables (e.g. hit rate, false alarm rate, CSI etc.), (3) standard metrics that measure the typical error of a forecast time series (bias, mean absolute error, root mean square error etc.).

The presentation will provide an overview of the experimental configuration, the forecast and observed datasets and results of the comparative forecast evaluation.