An Attention-Based Spatio-Temporal Graph Convolutional Network for Gas Concentration Prediction

Abstract

Spatio-temporal modeling of data is the key to predict gas concentration, which is mainly reflected in the modeling and prediction of spatio-temporal dependence in time and space dimensions. However, current studies mainly focus on the temporal aspect of gas concentration, often ignoring the influence of spatial distribution, which makes the prediction accuracy low. Therefore, we propose an Attention-based Spatio-temporal Graph Convolutional Network (AGT-GCN) for gas concentration prediction. This network comprises several key components: Graph Convolutional Network (GCN), Gated Recurrent Unit (GRU), spatio-temporal attention mechanism and residual convolution. Initially, we plot gas concentration distributions by selecting gas concentrations over different time periods to understand the spatio-temporal characteristics of the gas. Subsequently, AGT-GCN is employed to accomplish three main tasks: 1) Capturing the dynamic spatio-temporal correlations within the data using a spatio-temporal attention mechanism; 2) Utilizing GCN and GRU to capture the spatial dependencies of gas concentration and to model the temporal features respectively; 3) Employing residual convolution to maintain consistency in the number of input and output channels, thus preventing issues such as gradient vanishing or explosion during model training. Finally, we compare the prediction performance of AGT-GCN with Long Short-Term Memory (LSTM), GCN, GRU, Temporal Graph Convolutional Network (T-GCN) and Temporal Pattern Attention-Long Short-Term Memory (TPA-LSTM). Experimental results demonstrate that the proposed model can effectively capture the spatio-temporal correlations of gas concentration and outperforms the aforementioned baseline models in prediction accuracy.

Keywords: Spatial distribution, Attention mechanism, Spatio-temporal correlation, Gas distribution map, Graph Convolutional Network