A Prediction Method For Ozone Concentration Based on GCN-BiLSTM-Attention

As one of the major air pollutants, ground-level ozone has significant impacts on human health and the environment. To effectively address air quality issues and improve urban living conditions, predicting ozone concentration has become an urgent task. Existing neural network prediction models increasingly consider the correlation of time and space, known as spatiotemporal prediction. These models take into account both temporal and spatial features, often outperforming simple sequence models such as Recurrent Neural Networks (RNN). Ozone predictions typically exhibit spatiotemporal correlations, with ozone detection values in different areas influencing each other, especially in urban settings. Therefore, this paper incorporates these spatiotemporal features into the modeling process and proposes a GCN-BiLSTM-Attention spatiotemporal neural network for ozone concentration prediction. To integrate spatiotemporal features simultaneously, BiLSTM is first used to compress features along the time dimension of each node. Then, GCN is employed to transfer and fuse information between nodes, ultimately obtaining the embedded representation of the nodes in the graph. The features of all nodes are then pooled to obtain a comprehensive representation of the input spatiotemporal data, which is finally fed into a fully connected layer to predict ozone values for the next 24 hours. Experimental results show that this model achieves the best prediction accuracy, with an average MAE value of 10.7021 and a goodness of fit of 0.91. Compared to baseline models such as GRU and LSTM, the performance is significantly improved.