Boosting PM_2.5 short-term prediction by coupling graph attention and koopman networks

Abstract

Accurately predicting PM_2.5 concentrations remains a major challenge due to the complex and nonlinear nature of its formation and transport processes. Traditional models often struggle to capture the spatiotemporal dynamics of PM_2.5, particularly under varying meteorological conditions. In recent years, Koopman operator theory has attracted increasing attention for its ability to transform nonlinear systems into linear representations, thereby enhancing model interpretability and stability. However, most existing Koopman models primarily focus on temporal dynamics and overlook the spatial correlations. To address this limitation, we propose a novel framework called Graph Attention Physics-Constrained Learning (GAPCL). This method combines Graph Attention Networks—designed to model spatial dependencies between PM₂.₅ monitoring stations—with the Physics-Constrained Learning framework grounded in Koopman theory. The model employs an attention mechanism to dynamically weight PM_2.5 monitoring stations, uncover spatial relationships, and integrates graph topology with Koopman eigenfunctions to represent the spatiotemporal evolution of complex nonlinear dynamical systems. The effectiveness of GAPCL was validated using hourly data from 2019 to 2021 in the Beijing-Tianjin-Hebei region. The results demonstrated that the model achieved superior prediction accuracy, particularly for short-term forecasts. Compared to the Spatial Physics Constrained Learning (SPCL), GAPCL improved the RMSE, MAE, IA, and r by an average of 12.07%, 11.47%, 11.16%, and 11.11%, respectively. Additionally, the mean RMSE of GAPCL was 3.26% lower than the next-best SPCL and 26.64% lower than the worst-performing Long Short-Term Memory Network. The GAPCL model significantly improves outlier prediction accuracy by dynamically weighting PM_2.5 stations to better capture spatial dependencies.

Keywords

Spatiotemporal predictions / Koopman operator / PM_2.5 / Graph attention network

References

[1] Abd-Elhameed WM. Novel expressions for the derivatives of sixth kind Chebyshev polynomials: Spectral solution of the non-linear one-dimensional Burgers’ equation. Fractal and Fractional. 2021, 5(2): 53.

[2] Alsahli MM, Al-Harbi M. Allocating optimum sites for air quality monitoring stations using GIS suitability analysis. Urban Climate. 2018, 24: 875-886.

[3] Avila AM, Mezić I. Data-driven analysis and forecasting of highway traffic dynamics. Nature Communications. 2020, 1112090.

[4] Azencot, O., Erichson, N. B., Lin, V., Mahoney, M. (2020). Forecasting sequential data using consistent koopman autoencoders. In International Conference on Machine Learning (pp. 475–485). PMLR.

[5] Chen Z, Niu B, Zhao X, Zhang L, Xu N. Model-based adaptive event-triggered control of nonlinear continuous-time systems. Applied Mathematics and Computation. 2021, 408: 126330.

[6] Dumont GA, Fu Y. Non-linear adaptive control via laguerre expansion of volterra kernels. International Journal of Adaptive Control and Signal Processing. 1993, 75367-382.

[7] Han, Y., Hao, W., Vaidya, U. (2020). Deep learning of Koopman representation for control. In 2020 59th IEEE Conference on Decision and Control (CDC) (pp. 1890–1895). IEEE.

[8] Hassan H, Latif MT, Juneng L, Amil N, Khan MF, Fujii Y, Banerjee T. Chemical characterization and sources identification of PM2. 5 in a tropical urban city during non-hazy conditions. Urban Climate. 2021, 39: 100953.

[9] Hu J, Wang Z, Liu S, Gao H. A variance-constrained approach to recursive state estimation for time-varying complex networks with missing measurements. Automatica. 2016, 64155-162.

[10] Koopman BO. Hamiltonian systems and transformation in Hilbert space. Proceedings of the National Academy of Sciences. 1931, 175315-318.

[11] Kvalheim MD, Revzen S. Existence and uniqueness of global Koopman eigenfunctions for stable fixed points and periodic orbits. Physica d, Nonlinear Phenomena. 2021, 425: 132959.

[12] Li M, Jiang L. Deep learning nonlinear multiscale dynamic problems using Koopman operator. Journal of Computational Physics. 2021, 446110660.

[13] Lu H, Wang S, Wu Z, Yao S, Han J, Tang X, Jiang B. Variations of polycyclic aromatic hydrocarbons in ambient air during haze and non-haze episodes in warm seasons in Hangzhou, China. Environmental Science and Pollution Research. 2017, 24135-145.

[14] Lusch B, Kutz JN, Brunton SL. Deep learning for universal linear embeddings of nonlinear dynamics. Nature Communications. 2018, 9(1): 4950.

[15] Maaziz MK, Boucher P, Dumur D. A new control strategy for induction motor based on non‐linear predictive control and feedback linearization. International Journal of Adaptive Control and Signal Processing. 2000, 142‐3313-329.

[16] Nansai K, Tohno S, Chatani S, Kanemoto K, Kagawa S, Kondo Y, Lenzen M. Consumption in the G20 nations causes particulate air pollution resulting in two million premature deaths annually. Nature Communications. 2021, 12(1): 6286.

[17] Pan S, Duraisamy K. Physics-informed probabilistic learning of linear embeddings of nonlinear dynamics with guaranteed stability. SIAM Journal on Applied Dynamical Systems. 2020, 191480-509.

[18] Qi Y, Li Q, Karimian H, Liu D. A hybrid model for spatiotemporal forecasting of PM_2.5 based on graph convolutional neural network and long short-term memory. Science of the Total Environment. 2019, 664: 1-10.

[19] Shi Y, Zhao A, Matsunaga T, Yamaguchi Y, Zang S, Li Z, Gu X, Gu X. Underlying causes of PM_2.5-induced premature mortality and potential health benefits of air pollution control in South and Southeast Asia from 1999 to 2014. Environment International. 2018, 121: 814-823.

[20] Sun, Y., Yu, Y., Cao, Z., Zhang, F., Liang, K. (2025). An Adaptation Koopman Model for Predicting PM2. 5 with Distribution Drift. Atmospheric Environment, 343, 120986.

[21] Wehmeyer, C., Noé, F. (2018). Time-lagged autoencoders: Deep learning of slow collective variables for molecular kinetics. The Journal of chemical physics, 148(24): 241703.

[22] Wen C, Hill DJ. Adaptive linear control of nonlinear systems. IEEE Transactions on Automatic Control. 1990, 35(11): 1253-1257.

[23] Yan X, Zuo C, Li Z, Chen HW, Jiang Y, He B, Shi W. Cooperative simultaneous inversion of satellite-based real-time PM_2.5 and ozone levels using an improved deep learning model with attention mechanism. Environmental Pollution. 2023, 327: 121509.

[24] Yu, B., Yin, H., Zhu, Z. (2017). Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875.

[25] Yu Y, Liu D, Wang B, Zhang F. Analysis of global and key PM2.5 dynamic mode decomposition based on the Koopman method. Atmosphere. 2024, 159. 1091

[26] Yu, Y., Zhou, H., Huang, B., Zhang, F., Wang, B. (2024b). Dynamic mode decomposition and short-time prediction of PM_2.5 using the graph Neural Koopman network. International Journal of Geographical Information Science, 1–24.

[27] Zhang X, Ding F, Yang E. State estimation for bilinear systems through minimizing the covariance matrix of the state estimation errors. International Journal of Adaptive Control and Signal Processing. 2019, 3371157-1173.

[28] Zhang L, Na J, Zhu J, Shi Z, Yang L. Spatiotemporal causal convolutional network for forecasting hourly PM_2.5 concentrations in Beijing, China. Computers & Geosciences. 2021, 155: 104869.

[29] Zhu X, Ding C, Jia L, Feng Y. Koopman operator based model predictive control for trajectory tracking of an omnidirectional mobile manipulator. Measurement and Control. 2022, 559–101067-1077.

Funding

the National Natural Science Foundation of China(42171412)

RIGHTS & PERMISSIONS

The Author(s)