Frequency-Aware Spatiotemporal Graph Modeling of Multi-Pollutant Dynamics in Industrial Air Quality Systems

Chia-Hui Liu^*, Chen-Chuan Cheng
Department of Electronic Engineering, National Formosa University, Yunlin, Taiwan
* Corresponding Author: Chia-Hui Liu. Email: email
(This article belongs to the Special Issue: Emerging Artificial Intelligence Technologies and Applications-II)

Computer Modeling in Engineering & Sciences https://doi.org/10.32604/cmes.2026.083813

Received 11 April 2026; Accepted 30 June 2026; Published online 27 July 2026

Download PDF

Abstract

Industrial air quality forecasting remains challenging due to nonlinear pollutant formation, localized emissions, meteorological variability, and nonstationary spatiotemporal dependencies among monitoring stations. This study proposes FFTGNet, a frequency-aware spatiotemporal graph neural network for multi-pollutant forecasting in industrial air quality systems. It integrates an FFT-guided dominant-period estimation and period-folding module with a temporal-to-spatial graph backbone composed of TemporalGLU and Chebyshev graph convolution. The frequency-guided module reorganizes input sequences into intra-period and inter-period representations, TemporalGLU adaptively filters nonlinear temporal fluctuations and short-term spikes, and ChebGCN propagates information across inter-station spatial dependencies. Experiments were conducted using five years of hourly observations from ten Photochemical Assessment Monitoring Stations in Yunlin, Taiwan, covering O₃, CO, PM₁₀, meteorological variables, and PMF-derived VOC source features. Compared with conventional baselines and recent spatiotemporal models, including MGCGRU-SAN and AirNN-Adapted, the proposed method achieves the most consistent performance across pollutants and forecasting horizons. For 1 h forecasting, it obtains R² values of 0.9509, 0.8574, and 0.8556 for O₃, CO, and PM₁₀, respectively. For O₃, it further maintains R² values of 0.8116 and 0.6797 at the 3 and 6 h horizons, while showing smooth degradation rather than abrupt performance collapse up to 24 h. Ablation results confirm the complementary contributions of frequency-guided period folding, gated temporal modeling, and graph-based spatial propagation. Computational analysis further indicates that FFTGNet remains feasible for server-side near-real-time deployment, with an inference latency of 71.7 ms per sample. These results demonstrate that coupling frequency-domain priors with spatiotemporal graph learning improves forecasting accuracy, stability, and applicability for industrial multi-pollutant air quality systems.