<p>Accurate global natural gas trade forecast is crucial for a country to optimize and adjust its energy strategy to ensure energy security. Since various complex and uncertain factors influence global natural gas trade, traditional forecasting methods based on expert knowledge and statistical methods make it difficult to achieve accurate forecasts. Recently, approaches based on machine learning (ML) have been adopted to model such complex and uncertain factors based on historical data associated with global natural gas trade. However, global natural gas trade has three special characteristics: (1) network characteristic, i.e., the trade network relationships between various countries, (2) dynamic characteristic, i.e., the trade relationships dynamically change over time, and (3) complex-factor characteristic, i.e., the trade relationships are influenced by many complex factors. Unfortunately, none of the existing ML models can comprehensively capture these three characteristics for forecasting. To fill this gap, this paper proposes a novel hybrid ML model by integrating three ML models of Graph Attention Networks (GAT), Gated Recurrent Unit (GRU), and Transformer, which is termed GAT-GT. The main idea of GAT-GT is threefold: (1) GAT is employed to learn the latent features of the annual trade network among various countries year by year, (2) the learned time-sequential latent features are input into GRU to model the dynamic trade relationships, and (3) Transformer is exploited to combine various complex factors with the learned latent features for forecasting. As such, GAT-GT enjoys all the merits of GAT, GRU, and Transformer, making it able to comprehensively capture these three characteristics of the global natural gas trade. In the experiments, the global natural gas trade data from 2000 to 2023 were collected to benchmark our empirical studies. The results demonstrate that our GAT-GT model significantly outperforms eight advanced ML competitors in forecasting the global natural gas trade.</p>