Prediction of 8Â h-average ozone concentration using a supervised hidden Markov model combined with generalized linear models

- An 8 h-average ozone prediction model is proposed.
- The model combines the key advantages of the hidden Markov models (HMMs) and the Generalized Linear Models (GLMs).
- Application to data from Livermore Valley, CA, demonstrates the potential of the proposed modeling strategy.

An ozone prediction model based on a supervised hidden Markov model (HMM) and generalized linear models (GLMs) has been developed and tested on data from Livermore Valley, CA. Hidden states in the supervised HMM are assigned to represent different ozone concentration ranges which make the parameters of the supervised HMM easy to be explained. Using the Viterbi algorithm (VA), not only the most likely state of 8 h-average ozone concentrations but also the relative probabilities of different concentration ranges can be obtained. Then, GLMs corresponding to different ozone concentration ranges are used to quantitatively predict surface ozone levels. Using the relative probabilities and ozone levels predicted by GLMs, an ozone concentration value in the most likely concentration range can be finally determined. In this paper, data from 8 ozone seasons spanning 2000 to 2007 are used to build the prediction model and data from 2008 to 2009 are used for validation. The results show that this model can be used to predict all ozone exceedance days correctly. Compared to the generalized linear mixed effects model (GLMM), which is also used to model grouped data, the true prediction rate (TPR) of the proposed model is higher by 27%. Compared to the prediction results using the supervised HMM alone, the mean absolute error (MAE) of ozone exceedance days predicted by the proposed model is reduced by 72%.