A data-driven framework for optimized forecasting of non-methane hydrocarbon concentrations in urban air quality

Abstract

Non-methane hydrocarbons (NMHC) represent significant air pollutants that have a major negative impact on urban life. Forecasting NMHC concentrations accurately plays an important role in controlling air quality as well as developing focused pollution control plans. This paper provides an optimized framework for NMHC level forecasting with machine learning models using metal oxide sensor-based measurements of air quality and some environmental factors. For this propose, six machine learning models, such as Extra Trees, Random Forest, K-Nearest Neighbors, Multi-Layer Perceptron, CatBoost, and XGBoost, were used, which are optimized the Black Widow Optimization Algorithm. This optimization involves with the adjusting of models hyperparameters, which strengthens the reliability and scalability of the machine learning model and enhance their forecasting performance. Furthermore, the performance of each model was evaluated by applying several evaluation indicators. According to the results, the Black Widow Optimization Algorithm-Extra Trees (BWOA-ET) model outperformed the other models during testing phase by attaining the best prediction accuracy with R² score of 0.948. Furthermore, the model showed a comparatively low prediction deviation with a MAE of 7.380 and a RMSE of 33.624. According to sensitivity and feature importance analysis, the most significant influential parameters on NMHC concentrations, were temporal factors of month with importance score of around 1.1 and sensor-based measures like PT08.S4 (NO2) and CO(GT) with importance scores of around 0.2. Overall, this work contributes to the knowledge in air quality monitoring based on a data-driven approach with regard to the prediction and understanding of the dynamics of NMHCs.