Data-Driven Modeling of Utility Outages Using Weather Radar Observations

Utility grid systems are particularly vulnerable to severe weather events that are occurring more frequently than ever. The outages resulting from these storms can be very costly in terms of lost productivity by customers, and costs to restore electrical services. Currently available tools for preparation and recovery from severe weather events have limited capability. Outage management systems tend to be passive, suffering from untimely and inaccurate outage reporting and lack of analytic capabilities for estimating weather impacts and dispatching repair crews in a preemptive manner. There is a strong need to develop better tools identifying significant weather events and predicting their impacts on the utility infrastructure that will help to minimize the costs and times associated with outage repairs and service restoration under these events.

We have developed a methodology to use high-resolution (time and space) weather radar observations of storm characteristics for grid component outage prediction and storm response management. Feely available datasets from the Next-Generation Weather Radar (NEXRAD) network are combined with detailed geospatially-references outage information from participating utilities to build statistical-predictive models of weather-related outages for different, specific types of grid components (e.g., overhead wires). Using a general linearized model (GLM) we relate several possible explanatory variables including: radar reflectivity, Doppler wind speed, precipitation accumulation, vertically integrated liquid from historical weather radar observations as well as tree coverage with component-specific outage information to build failure rate models. The model generally performs well when applied to a strong storm with significant outages. We have tested the performance of the model to the use of forecast weather variables, with their inherent uncertainty, through the addition of noise (as a proxy for the uncertainty in the forecast variables) to the input weather variables. We find that there are significant discrepancies in the magnitude of outages, but the general trends are still captured. We have also formulated the model to account for the history of storm conditions experienced by utility systems. In this case, we find an improved agreement between the predicted and actual outages.

Finally, we have developed an innovative Bayesian update approach to synthesize the statistical outage prediction model and known outage numbers to better predict the outages. By incorporating outage information as it occurs during the storm, the outage prediction is continually revised resulting in an improved estimate of the total outages. This algorithm has important potential applications for more timely restoration, especially for utilities with enhanced situational awareness via deployment of sensing devices such as smart meters, through efficient deployment of repair crews and resources to regions most likely to have experienced weather-related outages.