Assessing the accuracy of various statistical models for forecasting PM\(_{2.5}\): a case study from diverse regions of Gandhinagar and Ahmedabad

PM\(_{2.5}\) is the most hazardous air pollutant due to its smaller size, which allows deeper bodily penetration. Three diverse regions from Gujarat, India, namely Sector 10, Maninagar, and Vatva, which have green space, high population concentration, and industries, respectively, were chosen to forecast PM\(_{2.5}\) concentration for the next day. Four statistical models, including Multiple Linear Regression (MLR), Principal Component Regression (PCR), Simple Exponential Smoothing (SES), and Autoregressive Integrated Moving Average (ARIMA), were chosen to forecast PM\(_{2.5}\) levels. For this study, data of various pollutants and meteorological parameters were collected from February 2019 to September 2023. Analysis of the seasonal patterns of PM\(_{2.5}\) revealed elevated concentrations during post-monsoon and winter, in contrast to reduced levels during summer and monsoon. Statistical analysis revealed that the concentration of PM\(_{2.5}\) in Sector 10 is much lower than in the other two regions. The analysis of the test results, utilising various accuracy measures like RMSE, MAE, MAPE, IA, and others, indicated that Sector 10 achieved the highest precision in its results. While assessing the models’ accuracy on the test data, the ARIMA model demonstrated the highest level of precision. The average RMSE, MAE, and MAPE values for the ARIMA model were 12.63, 8.59, and 0.24, respectively. In the comparison of the performance between these statistical models and the neural network-based Multilayer Perceptron (MLP) model, it was observed that the statistical model demonstrated superior performance over the MLP model.