Skip to content
Open access

Explainable Machine Learning for Next-Hour Prediction Using Three Meteorological Variables and Historical PM2.5 Data

Nov 2026 · International Journal of Mathematics, Statistics, and Computing · Vol 4, pp. 219-245 · 0 citations

TL;DR

It is demonstrated that a minimal, readily obtainable feature set can support highly accurate and interpretable next-hour PM2.5 forecasting, offering a practical and scientifically transparent tool to support early-warning air-quality systems and evidence-based environmental health policy.

Abstract

Ambient fine particulate matter (PM2.5) remains one of the most consequential urban air-quality hazards, yet forecasting its next-hour concentration accurately and transparently remains challenging owing to the non-linear, multivariate nature of pollutant-meteorology interactions and the black-box character of high-performing ensemble models. This study proposes an explainable machine learning framework for next-hour PM2.5 concentration prediction that integrates Extreme Gradient Boosting (XGBoost) with a parsimonious set of predictors comprising three meteorological variables (air temperature, relative humidity, and wind speed) together with historical PM2.5 concentrations. Using 43,824 hourly observations from the Beijing PM2.5 dataset (2010–2014), lag, rolling-statistic, and cyclical calendar features were engineered, and model hyperparameters were systematically tuned through Bayesian optimization using the Optuna framework with the Tree-structured Parzen Estimator over 30 trials under a strictly chronological training validation testing split. Model interpretability was achieved through SHapley Additive exPlanations (SHAP) analysis. The tuned XGBoost model achieved a root mean squared error of 16.50 µg/m³, a mean absolute error of 9.89 µg/m³, a mean absolute percentage error of 21.75%, and a coefficient of determination of 0.9478 on the test set, corresponding to reductions of approximately 4.4%, 10.5%, and 12.4% in RMSE, MAE, and MAPE, respectively, relative to a non-optimized baseline model. SHAP analysis revealed that the one-hour PM2.5 lag feature overwhelmingly dominated the model's predictions, far surpassing the combined contribution of all meteorological variables, while wind speed and temperature retained secondary but physically meaningful influence. These findings demonstrate that a minimal, readily obtainable feature set can support highly accurate and interpretable next-hour PM2.5 forecasting, offering a practical and scientifically transparent tool to support early-warning air-quality systems and evidence-based environmental health policy.

Read PDF

Similar papers

Open access Oct 2026

Explainable Machine Learning for Air Quality Prediction, With Satellite and Weather Data

Air pollution is a major threat in current times and I consider PM2.5 as a pollutant because PM2.5 has negative impacts on human health. Hence PM2.5 prediction is crucial. The prediction of PM2.5 is a challenging task due to its non-linear relationship with various environmental factors and dust storm events. In this p...

Lopapurna Sahoo, Sambit Sahoo, S. Pradhan et al. · 0 citations
Open access Sep 2026

Next-Day PM10 Prediction Maps for South Korea Using Machine Learning with Meteorological and Transboundary Air Quality Data

An observation-driven framework for generating nationwide next-day PM10 prediction maps over South Korea using machine learning extends point-based machine learning to nationwide spatial PM10 mapping and provides an observation-driven complement to operational air quality forecasting and environmental management.

Yemin Jeong, Jaejin Kim, Jinsoo Kim et al. · 0 citations
Open access Oct 2026

Forecasting of PM2.5/PM10 Using Machine Learning: A Benchmarking Study Based on Open Air Quality IoT Datasets

The results indicate that model selection for PM forecasting should be validated across multiple evaluation windows rather than a single split, that conventional statistical baselines remain a necessary comparison point, and that reported gains from more complex architectures should be interpreted cautiously in light o...

Ioannis Psomadakis, Christos E. Christakis, Christina L. Metallidou et al. · 0 citations
Open access Oct 2026

Hourly Temporal Imputation of Five Key Meteorological Variables at 17 Iranian Airport Stations Using a Machine Learning Framework

Meteorological observations are fundamental to environmental modeling, yet data gaps frequently compromise their utility. This study benchmarked five machine-learning algorithms (Random Forest, Extra Trees, Gradient Boosting, Histogram-based Gradient Boosting, and Ridge regression) for hourly temporal imputation of fiv...

A. Rad, M. Nematollahi, S. Faridi · 0 citations
Open access Nov 2026

Predicting Pollution Before It Strikes: Machine Learning Models for Next-Day PM10 Forecasting Under Extreme Events in Portugal

Air pollution remains a major environmental threat, particularly during extreme events such as wildfires and dust intrusions. This study assesses the ability of two artificial neural network architectures to predict daily PM10 concentrations over the Portuguese Centre region using spatially continuous reanalysis data a...

André Brito, R. Durão, Ana Oliveira et al. · 0 citations

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.