Abstract Fine particulate matter (PM2.5) is a growing environmental concern worldwide. This study investigates the variation of PM2.5 with meteorological parameters and air pollutants in 10 cities of Texas from 2013 to 2022 and forecasts daily average PM2.5 for seven days in advance. The PM2.5 forecast is made using the Deep Autoregressive Neural Network (DeepAR), 1-dimensional Convolutional Neural Network (1d-CNN), Long Short-Term Memory with Autoencoder (LSTM-AE), and Bi-directional LSTM with Autoencoder (Bi-LSTM-AE) models. In Texas, Houston recorded the highest PM2.5 concentrations (> 40 µg/m3), whereas the lowest was found in Amarillo (4 µg/m3). To start with pre-processing, feature engineering was experimented with for 22 predictors, employing variance inflation factors (VIF) to avoid multicollinearity. Then, the models were trained on data from 2013 to 2020, with 2021 and 2022 used for validation and testing, respectively. Model performances were evaluated based on root mean squared error (RMSE), and index of agreement (IOA). During testing, the DeepAR model outperformed other models with an average IOA of 0.835 and RMSE of 2.51 µg/m3. DeepAR was integrated into SHapley Additive Explanations (SHAP) analysis, where the model underscores observed PM2.5, Tmax, solar radiation (SR), Wind Gust (WG), NO2, SO2, and time features as the most influential predictors for forecasting. The study then analyzed health impacts attributable to PM2.5 using AirQ+, including all natural causes for 30 + years adults and ischemic heart disease (IHD) for adults aged 25 + years. Jefferson County was identified as a hotspot for IHD, with males experiencing excess cases at a rate of up to 77.43 per 100,000. Graphical Abstract The graphical abstract in this study presents three central inquiries: (1) how PM2.5 concentrations have varied spatially and temporally across Texas from 2013 to 2022, (2) how accurately various deep learning models can forecast daily PM2.5 concentrations across the state, and (3) how forecasted PM2.5 relates to potential public health risks. First, air pollutant and meteorological data were collected and pre-processed from 2013 to 2022. Then the modeling framework was developed, where deep learning models including DeepAR, 1D-CNN, LSTM-AE, and Bi-LSTM-AE were trained and evaluated for 7-day-ahead forecasting. DeepAR achieved the best performance, with a mean IOA of 0.835. SHAP analysis of the DeepAR model revealed that observed PM2.5, Tmax, solar radiation, wind gust, NO2, SO2, and temporal features are the most influential predictors. Finally, the health risk related to PM2.5 was assessed using AirQ + highlighting Jefferson County as a PM2.5-related IHD risk hotspot. Excess IHD cases reached 68.44 per 100,000 (females) and 77.43 per 100,000 (males), emphasizing the need for accurate forecasting to protect public health.

Forecasting PM_2.5 and Assessing Health Impacts in Texas Using Advanced Deep Learning Models

期刊论文
发表日期：

2025年4月30日

查看全文

Shihab Ahmad Shahriar,
Yunsoo Choi,
Rashik Islam,
Hadi Zanganeh Kia,
Seyedeh Reyhaneh Shams,
Ahmed Khan Salman

摘要

Abstract

Fine particulate matter (PM_2.5) is a growing environmental concern worldwide. This study investigates the variation of PM_2.5 with meteorological parameters and air pollutants in 10 cities of Texas from 2013 to 2022 and forecasts daily average PM_2.5 for seven days in advance. The PM_2.5 forecast is made using the Deep Autoregressive Neural Network (DeepAR), 1-dimensional Convolutional Neural Network (1d-CNN), Long Short-Term Memory with Autoencoder (LSTM-AE), and Bi-directional LSTM with Autoencoder (Bi-LSTM-AE) models. In Texas, Houston recorded the highest PM_2.5 concentrations (> 40 µg/m³), whereas the lowest was found in Amarillo (4 µg/m³). To start with pre-processing, feature engineering was experimented with for 22 predictors, employing variance inflation factors (VIF) to avoid multicollinearity. Then, the models were trained on data from 2013 to 2020, with 2021 and 2022 used for validation and testing, respectively. Model performances were evaluated based on root mean squared error (RMSE), and index of agreement (IOA). During testing, the DeepAR model outperformed other models with an average IOA of 0.835 and RMSE of 2.51 µg/m³. DeepAR was integrated into SHapley Additive Explanations (SHAP) analysis, where the model underscores observed PM_2.5, T_max, solar radiation (SR), Wind Gust (WG), NO₂, SO₂, and time features as the most influential predictors for forecasting. The study then analyzed health impacts attributable to PM_2.5 using AirQ+, including all natural causes for 30 + years adults and ischemic heart disease (IHD) for adults aged 25 + years. Jefferson County was identified as a hotspot for IHD, with males experiencing excess cases at a rate of up to 77.43 per 100,000.

Graphical Abstract

The graphical abstract in this study presents three central inquiries: (1) how PM_2.5 concentrations have varied spatially and temporally across Texas from 2013 to 2022, (2) how accurately various deep learning models can forecast daily PM_2.5 concentrations across the state, and (3) how forecasted PM_2.5 relates to potential public health risks. First, air pollutant and meteorological data were collected and pre-processed from 2013 to 2022. Then the modeling framework was developed, where deep learning models including DeepAR, 1D-CNN, LSTM-AE, and Bi-LSTM-AE were trained and evaluated for 7-day-ahead forecasting. DeepAR achieved the best performance, with a mean IOA of 0.835. SHAP analysis of the DeepAR model revealed that observed PM_2.5, T_max, solar radiation, wind gust, NO₂, SO₂, and temporal features are the most influential predictors. Finally, the health risk related to PM_2.5 was assessed using AirQ + highlighting Jefferson County as a PM_2.5-related IHD risk hotspot. Excess IHD cases reached 68.44 per 100,000 (females) and 77.43 per 100,000 (males), emphasizing the need for accurate forecasting to protect public health.

Forecasting PM2.5 and Assessing Health Impacts in Texas Using Advanced Deep Learning Models

摘要

Forecasting PM_2.5 and Assessing Health Impacts in Texas Using Advanced Deep Learning Models