Enhancing Air Quality forecasting with functional neural networks: A case study of PM2.5 in Seoul

Abstract

Reliable prediction of PM_2.5 levels is essential due to their substantial impacts on public health, the environment, and society. This is especially critical in regions like South Korea, where air quality is often compromised by elevated PM_2.5 concentrations resulting from domestic emissions and transboundary pollution. This study considers a Functional Neural Network (FNN) model that combines Functional Data Analysis (FDA) with deep learning techniques to predict PM_2.5 levels. The FNN model is applied to data from 13 monitoring stations in Seoul and compared with traditional multivariate-based neural networks (NN) and functional regression (FM) models. The enhanced predictive accuracy was observed from the FNN model by integrating dynamic temporal patterns in pollutant and meteorological trajectories as functional inputs. Additionally, this study proposes a model selection procedure within the FNN framework to identify a subset of functional inputs that significantly enhances prediction performance. Comprehensive comparison studies confirm that the proposed FNN, combined with the input selection procedure, offers a reliable tool for PM_2.5 prediction. This functional approach holds potential for supporting air quality management and protecting public health.