Studies on Machine Learning Techniques for Multivariate Forecasting of Delhi Air Quality Index

In this paper, an extensive study is conducted to determine the most promising machine learning (ML) model among seventeen ML models including linear regression, lasso regression, ridge regression, elastic net, decision tree, random forest, K-nearest-neighbor regressor, Tweedie regressor, extra trees regressor, support vector regression, multilayer perceptron, bagging regressor, extreme gradient boosting (XGB) regressor, Adaboost regressor, stochastic gradient descent regressor, gradient boosting regressor, stacking regressor employing XGB and lasso regression for multivariate forecasting of the air quality index (AQI) of Delhi. Twelve independent variables, namely, xylene, toluene, benzene, O3, SO2, CO, NH3, NOx, NO2, NO, PM10 and PM2.5 are used to predict the dependent variable, i.e., the Delhi AQI. In order to assess the true potential of ML models in multivariate forecasting of Delhi AQI, fifty independent simulations are conducted using each model employing different ratios in train and test samples. The obtained forecasting accuracies are analyzed using statistical tests to draw decisive conclusions. It is observed from the simulation results that the extra trees regressor model acquires the best rank among all the considered ML models in mean absolute error (MAE) and symmetric mean absolute percentage error (SMAPE) for multivariate forecasting of Delhi AQI. Additionally, in contrast analysis, the dimensionality of independent variables is transformed and reduced using principal component analysis (PCA) or independent component analysis (ICA), and using the best determined ML model, the transformed variables are modeled. Results show that though the application of PCA and ICA reduces the dimension, it results in poor forecasting accuracy.