A Hybrid End-to-End Neural Network for Accurate Global Natural Gas Trade Forecast

Accurate forecast of global natural gas trade is crucial for a country to promptly optimize and adjust its energy strategies for energy security. Recently, a variety of machine learning-based approaches have been adopted to address this issue. However, global natural gas trade is influenced by various complex (e.g., reserves of natural gas, energy imports and exports, economy, population, etc.) and uncertain factors (e.g., climate change, international turmoil, and global pandemics). Existing approaches either partially or loosely explore these factors by nonendto-end learning, resulting in suboptimal feature extraction and temporal integration for forecasting. To fill this gap, this article proposes a novel hybrid end-to-end neural network (HENN) with three components: graph attentional network (GAT), temporal transformer (TT), and multilayer perceptron (MLP). First, GAT learns the latent features of annual trade network among various countries year by year. Second, TT integrates these annual latent features with temporal information to forecast the trade network’s representations for the upcoming year. Third, MLP employs the forecasted representations to forecast the trade network of next year. The three components are trained by end-to-end learning, ensuring the feature extraction and the integration of temporal information are both comprehensive and optimal for forecasting. Extensive experiments with global natural gas trade data from 2000 to 2023 show that the proposed HENN significantly outperforms advanced models in forecasting the global natural gas trade.