Forecasting Ozone Exceedances in the Mid-Atlantic using Ensemble Statistical Models

This work discusses the development, implementation, and evaluation of an ensemble-based air quality forecast tool that takes advantage of the flexibility, scalability, and speed of statistical models to produce probabilistic forecasts of ozone exceedances at over 40 locations in the Mid-Atlantic (Maryland, Virginia, and Washington DC). Statistical models were developed using bootstrapped regression trees with extreme-value based evaluation algorithms and trained on up to seven years of air quality and meteorological observations. These models are run operationally once a day using fields from the Short-Range Ensemble Forecasts (SREF) to produce a probabilistic forecast for the following day. Forecasts are disseminated through an easily interpretable web-based interface along with a real-time 14-day running evaluation. Beta-testing of the forecast tool during a 2011 air quality oriented field campaign (NASA DISCOVER-AQ) and the 2012 ozone season provided proof-of-concept, sparking great interest from state agencies. All the software and data used in this forecast tool are open-source or free-to-use, making it an attractive and inexpensive tool for a range of users and applications. The methods described in this work are easily expandable, allowing for growth and stability in a dynamic forecasting environment.