Mekong River Flow Prediction Based on Multi-Modal Decomposition and Dynamic Weighting Coordination

本研究针对湄公河流域1924~1987年自然基准态流量序列构建多模态分解与动态赋权协同的预测框架，提出EEMD-Transformer-GRU混合模型，旨在揭示气候自然振荡与人类活动的多尺度耦合作用机制。通过集合经验模态分解(EEMD)将原始流量数据解耦为8个本征模态函数(IMF)和残差项，其中IMF1~IMF2捕捉月–季尺度极端事件，IMF3~IMF5解析ENSO事件与季风相位耦合关系，IMF6~IMF8分离1970年后人类活动主导的流量变异。在此基础上，融合Transformer的全局注意力机制与GRU门控时序记忆，构建跨尺度水文动力学解析模型。实验结果表明，模型在测试集取得R2 = 0.95、RMSE = 2239.8 m3/s、MAPE = 0.326%的优异性能，较次优模型Transformer-GRU分别提升2.2%、4.4%和0.03%，且通过EEMD预处理使输入序列维度降低38.2%。对比实验显示，该模型在i5-4200U处理器上仅需1.2 ms/sample推理耗时与600 MB内存占用，参数量仅8.3 M，较传统Transformer-GRU降低46.8%计算需求。研究证实，IMF6准12年周期与PDO相位同步性达0.89，1978~1982年丰水期预测误差控制在3.2%以内，为区分自然变异与人类干预贡献提供量化依据。该成果构建了兼顾预测精度与计算效率的多尺度水文分析范式，为湄公河流域水资源管理提供了轻量化技术工具，尤其适用于发展中国家基层水文站的实时监测场景。This study develops a hybrid EEMD-Transformer-GRU model to analyze the multi-scale coupling mechanisms between natural climate oscillations and anthropogenic activities in the Mekong River’s natural baseline flow regime (1924~1987). The Ensemble Empirical Mode Decomposition (EEMD) decomposes the original flow data into eight intrinsic mode functions (IMFs) and a residual component, where IMF1~IMF2 (high-frequency components) capture monthly-to-seasonal extreme events, IMF3~IMF5 (interannual components) reveal ENSO-monsoon phase coupling, and IMF6~IMF8 (decadal components) isolate post-1970 human-induced flow variations. By integrating Transformer’s global attention mechanism with GRU’s gated temporal memory, the model achieves superior performance, yielding a test-set R2 of 0.95, RMSE of 2239.8 m3/s, and MAPE of 0.326%, outperforming the suboptimal Transformer-GRU model by 2.2%, 4.4%, and 0.03%, respectively. Computational efficiency is enhanced through EEMD preprocessing, reducing input sequence dimensionality by 38.2%. Benchmarking demonstrates the model’s lightweight deployment capability, requiring only 1.2 ms/sample inference time and 600 MB memory on an i5-4200U processor, with 8.3 M parameters-46.8% fewer computational demands than conventional Transformer-GRU. Key findings include a 0.89 synchronization between IMF6’s quasi-12-year cycle and PDO phases, and a <3.2% prediction error for the 1978~1982 wet period, providing quantitative criteria for distinguishing natural variability from anthropogenic impacts. This framework establishes a multi-scale hydrological analysis paradigm balancing accuracy and efficiency, offering a practical tool for real-time monitoring in resource-constrained basins.