Statistical Solution to Forecasting Electric Power Outages Caused by Severe Storms

Power outages caused by severe storms have major impacts on the electric utility distribution operations. To facilitate emergency management by utility companies for such operations, there are several factors to quantify and forecast storm-related damage, including outages. At the substation area level, they include the number of outages or restoration jobs caused by the storm event, the number of customers interrupted, the time to recover some percentage (e.g., 95%) of the customers in the service territory, and how damage is accumulated through a severe storm. We modeled all of the relevant variables in one framework using hierarchical statistical modeling strategies so that the damage forecast can incorporate the uncertainties caused from various sources.

We previously reported on our efforts to predict outages in an electric distribution network using historical weather observations and calibrating the results to enable a forecast to be driven with the results of a 24-hour numerical weather prediction. We are building on that work to enable a complete statistical solution to the electric power outage forecast. In particular, we are utilizing the output of an operational configuration of the WRF-ARW model customized to the service territory of a major utility company in the northeastern United States at 2km resolution for 84 hours. The model also incorporates data assimilation of observations from a dense mesonet in the region including and surrounding the service territory.

We developed a Quasi-Poisson model to forecast total number of outages and customers interrupted for the entire storm events which may cover more than one day. Given the total number of outages forecasted, we used multinomial regression to forecast damage duration (i.e., time to recover 95% of customers). We classified the damage duration into three categories, less than one day, one day to two days, and more than two days, based on the storm events forecasted by WRF-ARW. In addition, to understand the rate at which damage may increase during a storm event, we fit the damage accumulation curve using a logistic function, where the total number of outages or customers interrupted and damage duration are the inputs. By forecasting all of the relevant variables, we can provide a comprehensive forecast picture of the storm event in terms of damage to the electric overhead infrastructure.