Abstract: Accurate El Niño-Southern Oscillation (ENSO) predictions are important for mitigating extreme weather and climate events. Although significant progress has been made in deep learning for ENSO prediction, current models still have limitations in capturing complex dynamic spatial and temporal correlations. To address the above problems, the present study uses global SST data provided by the UK Met Office Hadley Centre and proposes an ENSO prediction model based on the Dual Spatial Temporal Network (DualST) architecture, which innovatively designs a Spatiotemporal Graph Convolutional Network (ST-GCN) that learns the connectivity between ENSO spatiotemporal data through a novel edge computing module and converts it to graph data, enabling the graph convolutional network to accurately capture the complex spatiotemporal dynamic characteristics of ENSO. In addition, the advantage of the attention mechanism, which is good at capturing long-range dependencies, is exploited to improve the long-term prediction accuracy of ENSO. Experimental results show that DualST significantly outperforms other state-of-the-art deep learning models in terms of the Nino 3.4 index prediction correlation coefficient and can achieve effective ENSO prediction up to 20 months with moderate increases in MAE and RMSE. Therefore, the DualST model can effectively model the dynamic connectivity between SST fields through innovative modules such as edge calculation, and significantly improves long-term prediction of ENSO.