Probabilistic Estimates of Surface Ozone Concentration Derived Using an Ensemble of Model Configurations, Direct Sensitivity Calculations, and Bayesian Model Averaging

An ensemble of deterministic simulations is frequently used to create probabilistic estimates that account for uncertainty. A challenge with applying these approaches for simulations of ozone concentration is that chemical transport models require significant input data and computational resources to complete a single simulation. This research offers a computationally efficient approach to create an ensemble of model runs based on a set of inputs that is a more realistic characterization of the uncertainty in the meteorological simulation, boundary conditions, and emissions fields. An ensemble of Community Multiscale Air Quality Model with Decoupled Direct Method (CMAQ-DDM) simulations is used to generate large member ensembles while avoiding the computational cost of running the regional air quality model multiple times. Different approaches for generating the ensemble members are explored by allowing changes in the spatial variation of the boundary conditions and emissions values, rather than by adjusting them by a constant factor over all grid cells and hours. The Bayesian Model Averaging (BMA) statistical technique is used to weight each individual ensemble member based on how closely they match observed values. The final predictions provide a probability distribution of ozone concentration at any given location and time, rather than a single "best" estimate. These methods are applied to 2002 daily ozone data from a set of Air Quality System (AQS) monitoring stations in the South East to select a set of weighted ensemble members that have minimum spread but still capture the variability in the observations. Using additional observations, we evaluate the resulting ensembles and find that they provide a high level of reliability, resolution, and sharpness.