Integrating Transformer Encoders with Graph Attention Networks for River Discharge Prediction

Deep learning models are gaining traction in hydrological forecasting, yet existing approaches often fail to adequately address the complex spatiotemporal nature of river discharge data, especially when it comes to jointly modeling long-term temporal dependencies and spatial interactions over river networks. In order to overcome these limitations, this paper proposes a novel hybrid architecture called the Transformer-Graph Attention Network (Transformer-GAT). The proposed model utilizes a Transformer encoder with positional encoding and multi-head self-attention in order to extract long-range temporal patterns from discharge sequences. Subsequently, a Graph Attention Network (GAT) layer further processes these temporal embeddings to model spatial dependencies through adaptive attention weighting over the connections in river networks. The integrated representations are then utilized for multi-step discharge prediction at target gauge stations. Comprehensive evaluation across multiple flood events demonstrates the predictive advantage provided by the model: superior performance metrics include Nash-Sutcliffe Efficiency (NSE) of 0.8447, Root Mean Square Error (RMSE) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.2448 \mathrm{m}^{3} / \mathrm{s}$</tex>, and Mean Absolute Error (MAE) of <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$0.1596 \mathrm{m}^{3} / \mathrm{s}$</tex>. These results considerably surpass those obtained by established baseline models, hence further validating the effectiveness of the Transformer-GAT framework in addressing the complex space-time challenges of river discharge prediction

Перевод пока недоступен