Potential Use of Transport and Dispersion Model Ensembles for Forecasting Applications

Using ensembles to improve the simulation of plume dispersion is becoming a more common practice. One of the biggest challenges in creating ensembles is developing the appropriate member selection process to get the most accurate results, quantify ensemble uncertainty, and use computing resources more efficiently by avoiding the use of redundant model information. In this work, two reduction techniques are tested: one that is independent of observations and is based on the exclusion of redundant members by using uncorrelation as a measure of distance among the members and a second one based on measured data, which minimizes the mean square error (MSE) of the average of all possible model combinations. These techniques are applied to a 24-member ensemble, created by varying the boundary layer parameterizations in the meteorological WRF Model and dispersion Hybrid Single-Particle Lagrangian Integrated Trajectory model (HYSPLIT) simulating six releases of the Cross-Appalachian Tracer Experiment (CAPTEX). Applying the first technique produced results statistically comparable to the full ensemble for four out of six releases, while the second technique shows a similar or superior performance for all cases. Furthermore, to mimic a forecast application, the first day of the tracer release is used to select the ensemble members and the subsequent days are utilized as a forecast proxy to evaluate their performance. The reduced ensembles chosen by applying the technique based on the minimization of the MSE statistically perform similarly to or better than the full ensemble for the forecasting time periods. This suggests that when observational data are available, the application of ensemble reduction techniques provides a potentially promising tool for improving the dispersion forecast.