適応的信号分解と時空間特徴抽出を統合した動的降雨流出予測のための解釈可能なディープラーニング

Accurate runoff prediction is essential for water resource management. However, runoff data often exhibit nonlinear and non-stationary characteristics, influenced by complex spatial and temporal factors. To address these challenges and improve prediction accuracy, this study proposes a rainfall-runoff prediction model that integrates adaptive signal decomposition with spatiotemporal feature extraction. First, the adaptive particle swarm optimization variational mode decomposition algorithm is applied for multi-scale decomposition and denoising of runoff time series, extracting multi-scale features from complex signals. Next, spatiotemporal attention mechanisms are employed to dynamically explore the temporal dependencies and spatial driving relationships of rainfall-runoff data, enhancing the precision and specificity of feature extraction. Finally, gated recurrent units are employed to effectively capture the dynamic relationships between rainfall and runoff. Moreover, the visualization of attention weights provides insights into the model's decision-making process, enhancing interpretability. To comprehensively evaluate the performance of the proposed model, multi-step prediction tasks were designed. Ablation experiments and comparative analyses with other models were conducted to validate the contribution of each module to overall performance. Additionally, an uncertainty analysis was performed to quantitatively evaluate the impact of individual components and their interactions on runoff prediction. The results demonstrate that the proposed model consistently outperforms traditional baseline models and other combined models across different prediction horizons. Notably, the model shows significant advantages in short-term predictions (NSE = 0.9977, RMSE = 1.1793). The data decomposition module has the strongest main effect, contributing up to 62.1 %. Meanwhile, the interaction between the data decomposition module and the spatiotemporal feature extraction module plays a crucial role, with average contributions of 31.74 %, 32.04 %, 28.47 %, and 32.98 % to RMSE, NSE, MAE, and R, respectively. These findings highlight the effectiveness of the proposed approach, as this study not only enhances the accuracy of runoff prediction but also provides new insights into modeling and forecasting complex nonlinear spatiotemporal data.