Performance of Alternate Machine Learning Configurations for the 0-120 Hour Prediction of Extreme Wind Gusts for Outage Management in the Consolidated Edison Company of New York Service Area

Electric utilities recognize the value of anticipating weather-forced outages on their distribution systems. Reliable assessments of the local probability of outages over time periods ranging from hours to 5 days ahead can shorten response times and reduce outage costs to both the utility and its customers by enabling utility personnel to be effectively positioned and prepared prior to occurrence of the outage. A number of weather factors can play a role in causing distribution system outages. These include high wind gusts, icing, heavy wet snow, lightning and even high soil moisture content. The Consolidated Edison Company of New York (Con Ed), the utility that serves New York City and several nearby suburban counties has teamed with the Atmospheric Science Research Center (ASRC) of the University at Albany and MESO Inc, in a project supported by the New York State Energy Research and Development Authority to develop and refine tools to predict the localized probability of extreme weather conditions that are associated with the occurrence of distribution system outages. The first phase of the effort is focused on the deterministic and probabilistic prediction of extreme wind gusts over the 0-120 hour ahead look ahead period.

The initial focus is on the forecasting of wind gusts at 37 Automated Surface Observing System (ASOS) and New York State (NYS) Mesonet sites within and near the Con Ed service area. These sites provide high quality wind data at a high spatial resolution (average distance between stations ~ 20 - 30 km) for the training of machine learning (ML) models and the evaluation of the forecasts. The prediction system is based on the use of output data from a small set of Numerical Weather Prediction (NWP) systems (i.e. Global Forecast System (GFS), Rapid Refresh (RAP) model, the High Resolution Rapid Refresh (HRRR), and in-house 1 km Weather Research and Forecasting (WRF) model run) as input to several configurations of machine learning algorithms and ultimately the construction of the final forecast from a composite of the these NWP-ML ensemble members. The forecast target variable is the peak wind gust at the ASOS and NYS Mesonet sites over several time scales (e.g. 15 minutes, hourly, daily). A key issue is the identification of the best ML approach for this application from the almost-infinite set of possible options.

This presentation focuses on an analysis of the variations in the performance of both deterministic and probabilistic wind gust forecasts associated with changes in the type (e.g. neural networks, decision tree algorithms) and configuration (e.g. hyper parameter optimization and feature selection, global vs. local target variable training) of the ML algorithms. The training and evaluation periods for these experiments are from the start of 2017 through mid-2019. The experimental analysis also considers the role of feature importance (e.g. which NWP variables have the most input on forecast performance) and a comparison of the performance of the ML-based predictions with forecasts produced by physics-stochastic hybrid algorithms (such as the one used to supply the maximum surface gust forecasts in the operational HRRR graphics) that estimate wind gust values directly from NWP variables.