Prediction of gas concentration evolution with evolutionary attention-based temporal graph convolutional network

Abstract

Accurate prediction of gas concentration is of great importance in many safe-based systems and applications. However, prediction accuracy of gas concentration is limited by not only the temporal evolution of gas concentration but also the spatial characteristics of gas dispersion. To capture the spatial and temporal dependences simultaneously, an evolutionary attention-based temporal graph convolutional network (EAT-GCN) is proposed, which has three outstanding features: (1) graph convolutional network (GCN) is used to capture spatial dependence by learning topological structures of a gas sensor network; (2) gated recurrent unit (GRU) is adopted to retain temporal dependence by learning dynamic changes of gas concentration, and (3) evolutionary attention is introduced to improve the ability of GRU to pay different degrees of attention to the sub-window features within multiple time steps. Finally, kernel extrapolation distribution mapping algorithm is employed to visualize the predicted results of gas concentration and update the gas distribution map. Compared with CNN, GCN, GRU, T-GCN, A3T-GCN and EA-GRU models, the proposed EAT-GCN model improves the prediction accuracy by 13.46%, 124.21%, 33.92%, 23.39%, 46.63%, and 23.97%, respectively. Experiments demonstrate that the designed model captures spatiotemporal correlation from gas concentration data and achieves better prediction accuracy than state-of-the-art baseline methods.