基于VMD-TPE-LSTM模型的月径流预测方法研究

Research on Monthly Runoff Forecasting Method Based on VMD-TPE-LSTM Model

摘要: 科学、准确、可靠的径流预测对防汛抗旱、水资源高效利用、水利设施综合效益的发挥至关重要。受到气候变化和人类活动的影响，径流过程易呈现出高度的非线性、非平稳性特征，给径流预测带来了更大的挑战。本文提出了一种基于变分模态分解(VMD)、剪枝优化算法(TPE)、长短期记忆网络(LSTM)等方法结合的月径流预测模型(VMD-TPE-LSTM)，采用受钱塘江上游新安江水库调控影响的富春江水库1969~2022年月径流序列对VMD-TPE-LSTM模型进行了训练、验证与测试，月径流预测结果表明：VMD-TPE-LSTM模型的纳什效率系数达到了0.91，能够对峰值流量实现较好的预测，模型具有良好的泛化性能；进一步开展了对照实验，揭示了各因素在组合径流预测模型中对模型预测性能影响程度排序为：预处理技术 > 基准模型 > 模型参数。因此，耦合预处理技术和参数优化算法的径流预测方法能有效解决气候和人类活动影响的径流非平稳性问题，从而提高月径流预测精度和能力。

Abstract: Scientific, accurate, and reliable hydrological forecasting is crucial for flood control, drought resistance, efficient water resource utilization, and comprehensive water conservancy facility benefits. The runoff process is influenced by climate change and human activities, showing a high degree of non-linearity and non-smoothness, posing a greater challenge to runoff forecasting. In this paper, a monthly runoff forecasting model (VMD-TPE-LSTM) is proposed by coupling variational mode decomposition (VMD) with tree-structured parzen estimator algorithm (TPE) and long short-term memory network (LSTM). The model is trained, validated, and tested using the monthly runoff process of the Fuchun River Reservoir influenced by the upstream reservoir regulation from 1969 to 2022. The monthly runoff forecasting results show that the Nash efficiency coefficient of the VMD-TPE-LSTM model is 0.91 and achieves accurate forecast of peak flow, which has good generalization performance. Further controlled experiments reveal that the factors in the combined runoff forecasting model influence the model forecasting performance in the following order: pre-processing technique > baseline model > model parameters. It can be seen that the coupled pre-processing technique and parameter optimization algorithm can effectively solve the runoff non-smoothness problem influenced by climate change and human activities, thus improving the accuracy and capability of monthly runoff forecasting.