Previously, we determined the viability of this idea and validated it by building upon the efforts at the IBM Thomas J. Watson Research Center to implement and apply an operational mesoscale prediction system to business problems, dubbed Deep Thunder. We have continued this effort in the New York City metropolitan area in several key ways, working with the emergency management group at a major utility company in the northeastern United States. In particular, we have deployed a coupled modeling system for outage prediction, in which the weather component is derived from a configuration of the WRF-ARW (version 3.1) community numerical weather prediction (NWP) model. It operates in a nested configuration, with the highest resolution at two km resolution focused on the utility's service territory and that of its subsidiaries. The configuration also includes parameterization and selection of physics options appropriate for the range of geography within the service territory from highly urbanized to rural. It produces 84-hour forecasts, which are updated every twelve hours.
Based upon analysis of historical damage events and anecdotal evidence, it is clear that storm-driven disruptions of the overhead distribution network (e.g., poles and wires) are caused by physical interaction of the atmosphere directly with that infrastructure or indirectly via nearby trees. However, reliable modelling with sufficient throughput for operational utilization of such interaction is neither tractable from a computational perspective nor verifiable from observations. Given that and the relative uncertainty in the processes and data, we approached the outage prediction from a stochastic perspective.
Historical observations from a local, relatively dense mesonet operated by AWS Convergence Technologies were analyzed along with data from the utility concerning the characteristics of outage-related infrastructure damage from weather events, as well as the information about the distribution network and local environmental conditions. Data from this mesonet are also utilized to validate and improve the forecasts as well as model tuning. A Quasi-Poisson model was developed to forecast the number of jobs to be dispatched to repair outages for each substation area within the utility's service territory. Since these areas are coarser than the computational grid for the aforementioned WRF configuration, this approach is effectively a statistical upscaling of the weather model to enable estimates for each of the three full days covered by each WRF execution. This approach also incorporates uncertainties in both the weather and outage data. Therefore, the model provides probabilities of outage restoration jobs per substation, which is associated with visualizations that explicitly depict the uncertainty.
Based upon the analysis of the historical data, one of the key linkages between outages and severe storms is the magnitude of wind gusts. However, the NWP model does not produce a direct representation of gusts. Therefore, a post-processing model was developed using a Bayesian hierarchical approach, which links the physical model outputs with real observations to create a calibrated, statistical representation of gusts. This model post-processes the sustained wind forecast on the topologically regular output generated by WRF based on the gust observations from the mesonet over time.
We will discuss the on-going work, the overall approach to the problem, some specifics of the solution, and lessons that were learned through the development and deployment. In particular, we will present how the content is being used and how we are evaluating its quality with respect to specific weather events that affect utility operations. This includes the validation methodology used for the weather and damage forecasts and deployment of customized visualizations. We will also discuss the overall effectiveness of our particular approach for these and related applications, issues such as calibration of data and quantifying uncertainty as well as recommendations for future work.