Attention-Enhanced GCN-LSTM with Residual Connections for Robust PM2.5 Forecasting

Abstract

Predicting PM2.5 concentration is a critical task for public health and urban planning. While deep learning models such as Long Short-Term Memory (LSTM) and Convolutional Neural Network (CNN) have shown promise in capturing temporal dependencies, they often neglect the complex spatial correlations between monitoring stations. Graph Convolutional Network (GCN) attempts to model these spatial dependencies but often suffers from over-smoothing and an inability to weigh neighboring influences dynamically. This paper proposes an Improved GCN-LSTM architecture that integrates Graph Attention Network (GAT) with Residual Connections to capture spatiotemporal patterns effectively. Using the Beijing Multi-Site Air-Quality Dataset, we compare our proposed model against the baseline LSTM, CNN-LSTM, and standard GCN-LSTM. Experimental results demonstrate that the proposed Improved GCN-LSTM significantly outperforms other architectures, achieving a Mean Squared Error (MSE) of 363.52 and a Mean Absolute Error (MAE) of 10.69. This represents an 89.48% decrease in MSE and a 73.54% decrease in MAE compared to the standard GCN-LSTM.