Attention-Based Spatio-Temporal Graph Neural Network for Multi-Pollutant Urban Air Quality Prediction

In modern times, air quality prediction become an essential component of smart city planning and environmental observation because of increasing urban pollution levels. However, existing Gated Recurrent Unit (GRU)-based spatiotemporal approaches face challenges such as restricted spatial adaptability, lack of attention mechanisms, and high computational cost. To address these challenges, an attention-based spatiotemporal graph convolutional network (ASTGCN) is proposed for accurate multi-pollutant air quality forecasting. Initially, multivariate time-series data were collected from Global Urban Air Quality Index Dataset. Furthermore, linear interpolation was used for missing value imputation, and min-max normalization with sliding window segmentation was used to prepare temporally aligned inputs. Then, a city-level graph is constructed using geographical proximity, where each node represents a city, and edge encodes spatial relations. Then, data fed into ASTGCN model, where graph convolution layers extract spatial features and temporal Convolutional Neural Network (CNN) layers will identify crucial time steps. Furthermore, a random search was used to tune hyperparameters, which enabled model to achieve improved generalization while reducing training cost. Finally, experimental results demonstrated that ASTGCN achieved an improved Mean Absolute Error (MAE) of 13.1%, Root Mean Squared Error (RMSE) of 19.5, and R² of 95%.