Air quality forecasting using artificial neural networks with real time dynamic error correction in highly polluted regions

Abstract

Air pollution is an important issue, especially in megacities across the world. There are emission sources within and also in the regions around these cities, which cause fluctuations in air quality based on prevailing meteorological conditions. Short term air quality forecasting is used not to just possibly mitigate forthcoming high air pollution episodes, but also to plan for reduced exposures of residents. In this study, a model using Artificial Neural Networks (ANN) has been developed to forecast pollutant concentration of PM10, PM2.5, NO2, and O3 for the current day and subsequent 4 days in a highly polluted region (32 different locations in Delhi). The model has been trained using meteorological parameters and hourly pollution concentration data for the year 2018 and then used for generating air quality forecasts in real-time. It has also been equipped with Real Time Correction (RTC), to improve the quality of the forecasts by dynamically adjusting the forecasts based on the model performance during the past few days. The model without RTC performs decently, but with RTC the errors are further reduced in forecasted values. The utility of the model has been demonstrated in real-time and model validations were performed for the whole year of 2018 and also independently for 2019. The model shows very good performance for all the pollutants on several evaluation metrics. Coefficient of correlations for various pollutants varies from 0.79–0.88 to 0.49–0.68 between the Day0 to Day4 forecasts. Lowest deterioration of performance was observed for ozone over the four days of forecasts. Use of RTC further improves the model performance for all pollutants.