Machine Learning Models Forecast PM2.5-Bound PAH Levels in Nagoya

Researchers have built an integrated statistical framework that links hourly air pollution and weather data to fine particulate matter (PM2.5) bound polycyclic aromatic hydrocarbons (PAHs) in Nagoya, Japan, and uses machine learning to predict short-term PAH concentrations. The study combined association rule mining with predictive modeling, evaluated with time-blocked cross-validation to avoid leakage from hourly autocorrelation.

Association rules identified recurring pollutant–meteorology co-occurrence patterns linked to PAH accumulation. Reduced wind speed emerged as the most consistent statistical condition associated with elevated PAH levels, while pressure, humidity, and temperature were associated with atmospheric conditions favorable to pollutant accumulation, such as reduced dispersion.

The framework also captured the daily rhythm of ambient pollutants. Black carbon and nitrogen oxides peaked during traffic rush hours, and photochemical oxidants peaked in the afternoon. Rule structures revealed both shared meteorological patterns and compound-specific association patterns across PAH species.

Among the predictive models, eXtreme Gradient Boosting and Random Forest achieved the best performance for high-molecular-weight PAHs, with R2 values of 0.81–0.93, RMSE of 0.02–0.11, and MAE of 0.01–0.06. Support Vector Regression and Multilayer Perceptron performed poorly by comparison. Pressure, temperature, and humidity dominated feature importance, consistent with the association rule findings.

The study included a health-risk assessment. Benzo[a]pyrene dominated carcinogenic toxicity among the PAHs considered, and estimated incremental lifetime cancer risk values for all population subgroups remained below the U.S. Environmental Protection Agency negligible-risk threshold.

The authors caution that the identified associations reflect statistical covariation within a limited 10-day observation period. They should be interpreted as site-specific descriptors for local forecasting rather than as evidence of universal atmospheric mechanisms.

The work demonstrates how combining association rule mining with machine learning can characterize hourly variability of PM2.5-bound PAHs and enable short-term prediction at a single urban site, while the performance figures and risk estimates remain specific to the observation window and location described.

Subject of Research: Statistical association patterns and machine learning forecasting of PM2.5-bound polycyclic aromatic hydrocarbons in Nagoya, Japan

Article Title: Characterizing statistical association patterns and local forecasting of PM2.5-bound polycyclic aromatic hydrocarbon concentrations in Nagoya, Japan

Article References: Wang, M., Ohura, T., Au, V. K.-M., Tsuchiya, N., & Kameda, T. (2026). Characterizing statistical association patterns and local forecasting of PM2.5-bound polycyclic aromatic hydrocarbon concentrations in Nagoya, Japan. Environmental Monitoring and Assessment, 198(11), Article 1259. https://doi.org/10.1007/s10661-026-16083-6

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16083-6

Keywords: PM2.5, polycyclic aromatic hydrocarbons, Nagoya, machine learning, association rule mining, air quality, black carbon, benzo[a]pyrene, cancer risk assessment, meteorology, XGBoost, Random Forest