Deep Learning-Based Prediction of Particulate Matter and Application of Air Quality Health Impact Indicators in Industrial Complexes

Particulate matter (PM) has been classified by the International Agency for Research on Cancer under the WHO as a Group 1 carcinogen, posing a serious environmental health risk. Outdoor workers in industrial complexes face particularly high exposure compared to the general population. This study developed a deep learning–based prediction model to manage fine particulate matter exposure among industrial complex outdoor workers and evaluated the applicability of the Air Quality Health Index (AQHI) as a health impact assessment tool. Hourly air quality data collected from the Yeosu National Industrial Complex between 2023 and 2024 (PM2.5, PM10, O3, NO2, CO, SO2) were used to construct LSTM- and GRU-based prediction models. Model performance was assessed through input sequence length optimization and k-step-ahead forecasting to evaluate reliability across prediction horizons. Results showed that PM2.5 achieved optimal performance with 6–24 h input sequences, while PM10 reached its best performance with 336 h input sequences. In k-step-ahead forecasting, one-hour predictions yielded the highest accuracy, with reliable ranges up to 2 h for PM2.5 and 1 h for PM10 based on an R2 threshold of 0.7. Predicted AQHI exhibited patterns consistent with observed AQHI, confirming its potential as a health-oriented evaluation index. Furthermore, by introducing the k-step-ahead approach, this study mitigated error accumulation inherent in recursive multi-step forecasting and applied the Hong Kong AQHI methodology, thereby contributing to health risk assessment for outdoor workers in industrial complexes.