Comparison of regression, neural networks, and fuzzy logic for forecasting ozone in a semi-arid coastal region of Texas

In this research Statistical Regression, Artificial Neural Networks, and Fuzzy Logic models have been developed to forecast daily maximum eight-hour ozone concentrations at CAMS04 and CAMS21 stations of Corpus Christi. Ideally, model should be developed using optimum quantity of data and input parameters. Findings of this study indicated that only two years of data would be sufficient to develop a model instead of using large quantities of data. Utility of SO2 as ozone predictor was checked as well. Statistical "t" test was used to assess whether SO2 was varying over the time. The results of study indicated that SO2 was not a significant ozone predictor compared to previous day eight-hour average ozone concentration. Usefulness of "Principal Component Analysis" (PCA) as a data clustering technique for the model development was analyzed and it was found that predictive capability of model did not improve much, however application of PCA led to more intuitive and manageable parameter sets. Statistical Regression, Artificial Neural Networks and Fuzzy Logic models were developed to forecast high ozone episodes of Corpus Christi. The predictive capability of developed models was compared for their accuracy in forecasting daily maximum eight-hour ozone concentration by different model evaluation statistics. Results of the study indicated that all three models were able to capture trends in ozone time series, but they were not able to forecast peak ozone values. Among these models, performance of Statistical Regression and Artificial Neural Networks seem to be better than that of the Fuzzy Logic approaches.