Neuro-Fuzzy approach to forecasting Ozone Episodes over the urban area of Delhi, India

• Ozone episodes were occurred approximately 31% of the total observations for the year 2012 in Delhi urban area.
• Worst ozone episodes were observed in the summer seasons of Delhi.
• Neural–Fuzzy model has been found the better forecasting model for ozone episodes.

Tropospheric ozone is a major air pollution problem, both for public health and for the environment, which is not directly emitted by human activities. It is a secondary pollutant, which produces due to reaction with volatile organic compounds and nitrogen oxide concentrations, both emitted by anthropogenic activities. Therefore, ozone episode has been studied in Delhi, the urban area, where anthropogenic emissions are playing major role in ambient air pollution. The 8-hourly analysis of past data pattern in each season at different monitoring stations in Delhi suggests that the ozone episodes were occurred approximately 31% of the total observations and the summer seasons is the worst performer of ozone concentrations in Delhi. The correlation matrix shows that the dew point temperature, NO2, and relative humidity are the dominating variables for Ozone (O3) concentrations in Delhi. The present study aimed to analyze the ozone episodes in a year and to develop the artificial intelligence based forecasting methodologies over Delhi megacity. Further, the hourly ozone forecasting models have been developed through different modeling techniques, e.g., multiple linear regressions (MLR), artificial neural network (ANN) and artificial intelligence based Neuro-Fuzzy (NF) techniques. The air pollutants as well as meteorological parameters have been used to analyze the ozone episodes. The forecasted results from different models are compared with the observed values with statistical measures, e.g., correlation coefficients (R), normalized mean square error (NMSE), fractional bias (FB) and index of agreement (IOA). The performed statistical analysis has indicated that the artificial intelligence implementations have a more reasonable agreement with the observed values.