A Transfer Learning-Based Model for PM2.5 Prediction in Seoul

Predicting PM_2.5 concentrations using air quality data is often hindered by the presence of missing values, which can compromise accuracy and reliability. Traditional prediction models frequently suffer significant data loss during the handling of missing values, necessitating new approaches that address data quality issues while improving prediction performance.

This study proposes an efficient prediction methodology leveraging transfer learning to minimize the impact of missing values. A pre-trained model was constructed using integrated data from all monitoring stations in Seoul, followed by fine-tuning for specific monitoring stations to develop a PM_2.5 prediction model. Transfer learning is a machine learning technique that utilizes knowledge from previously trained models to enhance learning efficiency and performance in new tasks or domains. Unlike traditional approaches that require training from scratch, transfer learning reuses the weights and structure of pre-trained models, reducing training time and improving performance.

In this study, a deep neural network (DNN) pre-trained model was built using data from all monitoring stations in Seoul, and fine-tuning was applied using specific station data. The model was trained on six-hour average PM_2.5 data to predict the next six hours, effectively addressing missing values.

Preliminary results indicate that the transfer learning-based model effectively handles missing values and demonstrates improved prediction accuracy compared to independently modeled traditional approaches. By leveraging domain-wide information, the model compensates for the limitations of individual monitoring station data, achieving higher accuracy and reliability.

This study provides a scalable solution for addressing data gaps in air quality prediction and contributes to research on the health impacts of air pollution and urban air quality management. Future research will explore the application of this methodology to other pollutants and regions, further enhancing its generalizability and effectiveness.

Acknowledgement

"This research was supported by Particulate Matter Management Specialized Graduate Program through the Korea Environmental Industry & Technology Institute(KEITI) funded by the Ministry of Environment(MOE)"