You are here:
A FAST BAYESIAN METHOD FOR UPDATING AND FORECASTING HOURLY OZONE LEVELS
Citation:
Sahu, S. K., S. C. Yip, AND D. M. HOLLAND. A FAST BAYESIAN METHOD FOR UPDATING AND FORECASTING HOURLY OZONE LEVELS. ENVIRONMENTAL AND ECOLOGICAL STATISTICS. Chapman and Hall Limited, London, Uk, 18:185-207, (2011).
Impact/Purpose:
Accurate, instantaneous and high resolution spatial air-quality information can better inform the U.S. public and regulatory agencies about air pollution levels that lead to adverse health effects. The most direct way to obtain accurate air quality information is from measurements made at surface monitoring stations across the United States (U.S.). However, many areas of the U.S. are not monitored and typically, air monitoring sites are sparsely and irregularly spaced over large areas. As the need for spatial prediction has become reality in the regulatory environment, it is now important to develop computationally effcient models to combine air monitoring data and numerical model output, in a coherent way for better prediction of air pollution over short time periods.
Description:
A Bayesian hierarchical space-time model is proposed by combining information from real-time ambient AIRNow air monitoring data, and output from a computer simulation model known as the Community Multi-scale Air Quality (Eta-CMAQ) forecast model. A model validation analysis shows that the interpolated maps are more accurate than the maps based solely on the Eta-CMAQ forecast data for a two week test period. The out-of sample spatial predictions and temporal forecasts also outperform those from regression models with independent Gaussian errors. The method is fully Bayesian and is able to instantly update the map for the current hour (upon receiving monitor data for the current hour) and forecast the map for several hours ahead. In particular, the eight-hour average map which is the average of the past four hours, current hour and three hours ahead is instantly obtained at the current hour. The exact Bayesian method proposed here is preferable to more complex (possibly more accurate) models since iterative methods such as the Markov chain Monte Carlo (MCMC) are not required to obtain the fitting and forecasting results.