Application of Spatio-Temporal Clustering For Predicting Ground-Level Ozone Pollution

Ground-level ozone is an air pollutant, and as such negatively impacts human health and the environment. The complexity of the physical process of ozone formation makes ambient ozone concentration difficult to predict accurately. In this chapter, clustering techniques and multiple regression analyses are used to construct a simply interpretable forecasting model. Time series of ozone and meteorological variables in the Dallas-Fort Worth area for 12 years at 14 monitoring stations were acquired and processed. First, K-means cluster analysis was performed on ozone time series to specify data-driven ozone seasons at each station. Next, spatial hierarchical clustering was performed to find ozone zones in the area during each ozone season recognized in the previous step. Finally, a multiple linear regression was executed with meteorological variables and ozone in each zone. For ozone forecasting, temperature, solar radiation, wind speed, and previous ozone values were used because ozone is temporally autocorrelated. Monitoring stations in each temporal and spatial cluster show consistent behavior, which makes ozone forecasting possible even when a station is off. Results show high accuracy of ozone forecasting coupled with ease of interpreting the link between meteorology and ozone behavior. Also, clustering results are useful to understand the temporal and spatial patterns of the ozone dynamics in the area.