ORIGINAL RESEARCH
A Novel Framework for Air Quality Forecasting
Using Graph Convolutional Network-Based Time
Series Decomposition
More details
Hide details
1
School of Electromechanical Engineering, Hainan Vocational University of Science and Technology
2
Department of Electronic Engineering, Yeungnam University, Gyeongsan-si 38541, Republic of Korea
3
School of Geography, Nanjing Normal University; Nanjing 210023, China
4
Department of Quantitative Analysis, College of Business Administration, King Saud University, Riyadh, Saudi Arabia
5
Department of Electrical Engineering, College of Engineering, King Saud University, P.O. Box 800, Riyadh 11421,
Saudi Arabia
6
Department of Computer Science, Al Ain University, UAE
Submission date: 2024-01-14
Final revision date: 2024-03-20
Acceptance date: 2024-04-18
Online publication date: 2024-10-25
Corresponding author
Chan-Su Lee
Department of Electronic Engineering, Yeungnam University, Gyeongsan-si 38541, Republic of Korea
Muhammad Tahir Naseem
Department of Electronic Engineering, Yeungnam University, Gyeongsan-si 38541, Republic of Korea
KEYWORDS
TOPICS
ABSTRACT
This study provides an empirical investigation into the effectiveness of several deep learning models in
forecasting ambient concentrations of particulate matter with a diameter of less than 2.5 micrometers (PM2.5),
nitrogen dioxide (NO2), and sulfur dioxide (SO2). These pollutants are critical due to their adverse impacts on
human health and the environment. We evaluated four distinct models: Graph Convolutional Network (GCN),
Empirical Mode Decomposition combined with GCN (EMD+GCN), Ensemble Empirical Mode Decomposition
with Gated Recurrent Unit and GCN, and GCN with an attention mechanism (GCN_ATT). Through rigorous
computational experiments, the models were assessed against multiple statistical metrics including Mean Absolute
Error (MAE), Mean Square Error (MSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of
Determination (R2). The EEMD+GRU+GCN model consistently outperformed the others across all pollutants,
demonstrating the lowest MAE, indicating its strong predictive accuracy. Similarly, it maintained the smallest
MSE, suggesting it was particularly adept at reducing the influence of larger errors in predictions. Moreover, it
achieved the lowest MAPE across the datasets, confirming its robustness in percentage terms relative to the scale
of the actual values, a critical indicator of practical applicability for air quality forecasting. The GCN model,
while foundational, showed significant limitations, especially in the prediction of NO2 and SO2, as evidenced
by its negative R2 values, indicating a poor fit that was outperformed by simple average models. The GCN_
ATT model did not show the expected improvement that the attention mechanism might promise, suggesting
that additional fine-tuning or structural model changes are required. In conclusion, the integration of ensemble
empirical mode decomposition techniques with advanced neural network architectures such as GRUs and GCNs
provides a compelling approach to air quality forecasting. The proposed model’s ability to capture complex
spatiotemporal dependencies in environmental data makes it a promising tool for environmental monitoring and
policy-making, offering significant benefits for public health and ecological protection.