GE-STDGN: a novel spatio-temporal weather prediction model based on graph evolution

摘要

Many crucial tasks in weather prediction require large-scale and long-term spatio-temporal predictions. However, these tasks usually face three challenges: high feature redundancy, dependence of long-term prediction, and complexity in spatial relations of geographical location. To overcome these challenges, the Graph Evolution-based Spatio-Temporal Dense Graph Neural (GE-STDGN) network is proposed in spatio-temporal series prediction1. In this paper, a graph structure learning and optimization method based on an Evolutionary Multi-objective Optimization (EMO) algorithm, called Graph Evolution (GE), is adopted to improve the model’s ability in analyzing complex node correlations. Then, to avoid the over-fitting caused by Graph Neural Network (GNN) and reduce the constraint of graph structure, a GNN based on dense connection, which is named DenseGNN, is applied to message passing. Finally, by combining the multi-head attention mechanism GRU model with DenseGNN, GE-STDGN can handle complex spatio-temporal series prediction tasks. Experimental results on a public real-world weather dataset demonstrate that our model steadily outperforms many state-of-the-art models.