Towards intelligent air quality forecasting using integrated machine learning framework with variational mode decomposition and catboost feature selection

Predicting air pollution is crucial in improving air quality (AQ), which consequently provides benefits to the ecosystems and human health. AQ predictions often make use of Machine Learning (ML) approaches; nevertheless, these methods are not without their limitations. The main contribution of this research is to develop an efficient framework using machine learning (ML) for forecasting daily air quality metrics for Sulfur dioxide (SO₂) and Nitrogen dioxide (NO₂) in Changping, China. The suggested ML method is based on a set of local weights and a kernel extreme learning machine (LWKELM) model integrated with an efficient feature selection, the Catboost method, to extract influential input variables. Additionally, the input variables, collected from 2013 to 2017, are decomposed using multivariate variational mode decomposition (MVMD), which enhanced the predicting accuracy. Furthermore, the interior search algorithm (ISA), a robust optimization strategy, is a possible way of optimizing the models' hyperparameters. The results of the developed model was compared with the four other reliable ML approaches, including the locally weighted linear regression (LWLR), gaussian process regression (GPR), KELM, and multivariate adaptive regression spline (MARS) models. Based on the results, the proposed model demonstrates superior performance across statistical metrics for both parameters (NO₂: R = 0.978, RMSE = 0.537 and SO₂: R = 0.974, RMSE = 1.965) compared to alternative models. The MVMD-LWKELM-ISA model delivers highly accurate one-day-ahead forecasts for SO₂ and NO₂ and stands out as the most effective and intelligent approach for forecasting these daily parameters.