Real time air quality forecasting using integrated parametric and non-parametric regression techniques

- A statistical model to provide real time hourly forecasts of NO2 is presented.
- Non-parametric kernel regression is applied in parallel with multiple linear regression.
- The model has low computational resources and requires simple input data.
- IA values of between 0.74 and 0.94 were obtained.

This paper presents a model for producing real time air quality forecasts with both high accuracy and high computational efficiency. Temporal variations in nitrogen dioxide (NO2) levels and historical correlations between meteorology and NO2 levels are used to estimate air quality 48 h in advance. Non-parametric kernel regression is used to produce linearized factors describing variations in concentrations with wind speed and direction and, furthermore, to produce seasonal and diurnal factors. The basis for the model is a multiple linear regression which uses these factors together with meteorological parameters and persistence as predictors. The model was calibrated at three urban sites and one rural site and the final fitted model achieved R values of between 0.62 and 0.79 for hourly forecasts and between 0.67 and 0.84 for daily maximum forecasts. Model validation using four model evaluation parameters, an index of agreement (IA), the correlation coefficient (R), the fraction of values within a factor of 2 (FAC2) and the fractional bias (FB), yielded good results. The IA for 24 hr forecasts of hourly NO2 was between 0.77 and 0.90 at urban sites and 0.74 at the rural site, while for daily maximum forecasts it was between 0.89 and 0.94 for urban sites and 0.78 for the rural site. R values of up to 0.79 and 0.81 and FAC2 values of 0.84 and 0.96 were observed for hourly and daily maximum predictions, respectively. The model requires only simple input data and very low computational resources. It found to be an accurate and efficient means of producing real time air quality forecasts.