Exploring the significance of temporal, meteorological, and previous states parameters in PM2.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {PM}_{2.5}$$\end{document} concentration predictions: a neural network sensitivity study for Aguascalientes, Mexico

Air pollution is a critical issue in many developing cities, including Aguascalientes, Mexico, where PM2.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {PM}_{2.5}$$\end{document} concentrations pose significant public health risks. This study employs a feedforward neural network (FNN) as a base model to predict hourly PM2.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {PM}_{2.5}$$\end{document} concentrations, integrating meteorological, temporal, and previous state parameters. The model includes all available data to maximize predictive performance. Using techniques such as Sequential Backward Selection, sensitivity analysis, and Shapley additive explanations (SHAP), the study identifies the parameters that contribute the least to model performance. These include certain redundant temporal and meteorological parameters that can be excluded without significantly compromising accuracy. The key influential parameters identified were previous PM2.5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {PM}_{2.5}$$\end{document} levels, hour of the day, temperature, relative humidity, and wind speed components (WSx\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {WS}_x$$\end{document} and WSy\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {WS}_y$$\end{document}). The optimized model demonstrated moderate predictive performance, with a coefficient of determination (R2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\hbox {R}^2$$\end{document}) of 0.2292 and a Pearson correlation coefficient (r) of 0.7577. These findings highlight the potential for simplifying the model by excluding less relevant parameters, which can reduce computational costs while maintaining reliability. Additionally, the results underscore the importance of high-quality, continuous data and the integration of additional contextual parameters, such as emissions and traffic data, to further enhance model performance. This approach provides valuable insights into optimizing predictive models for air quality forecasting and informs strategies for data collection and model design in regions with limited monitoring infrastructure.