3.4
Hidden Markov Model with Multivariate Normal Emissions Applied to Ozone Emission Data

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner
Tuesday, 25 January 2011: 4:15 PM
Hidden Markov Model with Multivariate Normal Emissions Applied to Ozone Emission Data
2A (Washington State Convention Center)
Jeremy Troisi, Purdue University, West Lafayette, IN; and J. Rounds

The Committee on Artificial Intelligence (AI) for the American Meteorological Society (AMS) has organized a contest for 2011. The goal of the contest is to predict concentrations of ozone ($O_3$) at five monitoring stations in Rome, Italy, based on meteorological and pollutant measurements. When $O_3$ is in concentrations above 80 \units { } these days are considered to have severe ozone levels. Known pollutant observations have been deliberately removed from the data set according to a pattern set by contest authorities. We apply a hidden Markov model with a multivariate normal emission distribution to the log transformed data to create a smoother that replaces the missing observations. We select best number of hidden states for the model with five-fold cross validation using a criteria of lowest RMSE of predicted $O_3$ and best area under the receiver operator curve (AUC) of predicted severe ozone days. We briefly discuss how our methods can be adapted to create an online filter that would be more useful to authorities in this area.