Machine Learning Based PM 2.5 and 10 Concentration Modeling for Delhi City

Abstract

The global decline in air quality, attributed to pollutants from various sources such as fossil fuel usage, industrial expansion, and heightened commercial activities, underscores the importance of monitoring and forecasting air quality levels. This study delves into 3 years of daily particulate matter data spanning the pre-COVID (2019), COVID-era (2020), and post-COVID (2021) periods across thirty-seven monitoring stations in Delhi. Prior to analysis, the dataset underwent preprocessing to address missing and outlier values. Analysis of the dataset aimed to discern pollutant trends across stations and timeframes, identifying influential factors such as air temperature, surface pressure, and precipitation for modeling particulate matter concentrations. An Artificial Neural Network employing backpropagation was utilized for modeling. Training the model with 80% of the dataset, the remaining 20% served as the test dataset. Validation of the model's performance utilized standard statistical metrics including R², r, root mean square error, and mean absolute error. Notably, the R² for the training dataset were 0.82 and 0.84 and r for training dataset were 0.90 & 0.91 for PM 10 and PM 2.5, respectively. While the R² for the test dataset were 0.78 and 0.79, r values for the test dataset stood at 0.88 for both PM 10 and PM 2.5. Furthermore, the model facilitated upscaling of observations to a spatial scale, broadening the scope of observations via simulations to enhance regional understanding.