GeoRiskNet: A Spatio-temporal Graph Deep Learning Framework for AQI Forecasting with Regulatory Risk Interpretation

Accurate Air Quality Index (AQI) forecasting is essential for environmental monitoring, public health
protection and regulatory planning in rapidly urbanizing regions. However, reliable station-level AQI prediction
remains challenging due to nonlinear temporal air pollution dynamics, complex meteorological influences and crossregional diffusion among geographically distributed monitoring stations. Conventional statistical models primarily
focus on temporal patterns and often fail to capture structured spatial interdependencies and latent vulnerability
characteristics inherent in urban air pollution systems. To address these limitations, this study proposes GeoRiskNet,
a unified spatio-temporal graph deep learning framework for next-day AQI forecasting with integrated regulatory risk
interpretation. The model formulates AQI prediction as a supervised regression problem using multivariate hourly
pollutant and meteorological observations gathered from fourteen Central Pollution Control Board (CPCB) monitoring
stations in the Chennai metropolitan region between 2018 and 2023. The architecture integrates a Convolutional Neural
Network-Long Short-Term Memory (CNN-LSTM) network with temporal attention to capture short- and long-term
pollution dynamics influencing AQI, a nonlinear autoencoder for compact latent vulnerability representation and a
two-layer Graph Convolutional Network (GCN) to model spatial diffusion using geodesic distance-based adjacency.
The learned representations are fused within an end-to-end regression framework to predict continuous next-day AQI
values, which are subsequently mapped to CPCB-defined severity categories for regulatory interpretation. The
proposed GeoRiskNet achieves strong predictive performance, with Root Mean Squared Error (RMSE) of 0.087, Mean
Absolute Error (MAE) of 0.066 and coefficient of determination (R2
) of 0.953 on normalized AQI values. When
mapped to the original AQI scale (0-500), this corresponds to approximately 43.5 AQI units (RMSE) and 33 AQI units
(MAE), enabling practical interpretation for environmental monitoring and regulatory applications. Compared to
ARIMA (RMSE = 0.124), the proposed framework reduces prediction error by 29.8%. Regulatory severity
classification achieves 91.4% overall accuracy with a macro F1-score of 0.906, confirming balanced and interpretable
urban air pollution forecasting.