Prediction of 24-hour-average PM2.5 concentrations using a hidden Markov model with different emission distributions in Northern California

Abstract

Prediction of air pollutant levels plays an important role in the regulatory plans aimed at the control and reduction of airborne pollutants such as fine particulate matter (PM). Deterministic photochemical air quality models, which are commonly used for regulatory management and planning, are computationally intensive and also expensive for routine predictions. Compared to deterministic photochemical air quality models, data-driven statistical models are simpler and may be more accurate. In this paper, hidden Markov models (HMM) are used to forecast daily average PM2.5 concentrations 24h ahead. In conventional HMM applications, observation distributions emitted from certain hidden states are assumed as having Gaussian distributions. However, certain key meteorological factors and most PM2.5 precursors exhibit a non-Gaussian distribution in reality, which would degrade the HMM performance significantly. In order to address this problem, in this paper, HMMs with log-normal, Gamma and generalized extreme value (GEV) distributions are developed to predict PM2.5 concentration at Concord and Sacramento monitors in Northern California. Results show that HMM with non-Gaussian emission distributions is able to predict PM2.5 exceedance days correctly and reduces false alarms dramatically. Compared to HMM with Gaussian distributions, HMM with log-normal distributions can improve the true prediction rate (TPR) by 37.5% and reduce the false alarms by 78% at Concord. And HMM with GEV distribution can improve TPR by 150% and reduce false alarms by 63.62% at Sacramento Del Paso Manor. Comparisons between different distributions used in HMM show that the closer the distribution employed in HMM is to the observation sequence, the better the model prediction performance.