数据驱动与物理机制耦合的菜子湖流域水位预报模型研究

Research on Water Level Prediction Model of Caizi Lake Basin Based on Coupling of Data Driven and Physical Mechanism

菜子湖作为引江济淮工程的关键调蓄湖泊之一,湖泊水位的变化同时受到自然降水和工程调蓄的影响。为了准确模拟及预报菜子湖水位,构建了菜子湖流域四水源新安江模型和菜子湖水位预报的神经网络(LSTM)模型。在此基础上,采用外部耦合的方法,构建了四水源新安江-LSTM耦合模型,进一步将物理机制模型模拟的入湖流量作为补充因子,驱动神经网络(LSTM)模型模拟菜子湖水位,从而实现两种不同模型在湖泊水位预报中的耦合应用。结果得出:直接模拟水位的洪水误差小于0.1 m,耦合模拟水位的洪水误差小于0.02 m,后者相较前者,水位误差精度提升了0.08 m。直接模拟水位验证期的洪水误差在0.02 m之内,纳什系数R^(2)分别为0.89、0.75及0.88,均方根误差RMSE分别为0.034、0.027及0.015;耦合模拟水位验证期的洪水误差在0.015 m之内,纳什系数R^(2)分别为0.91、0.82及0.88,均方根误差RMSE分别为0.019、0.021及0.008。研究结果表明,与单驱动因子得出的结果相比,双驱动因子得出的结果更有效地提高了水位的模拟精度。同时在考虑对应降雨的洪水过程中,数据驱动和物理机制相结合的方法与直接预测水位误差相对比,有效地提高了场次洪水水位预报的精度,得到更精确的模拟结果。为引江济淮工程的调水提供了重要的参考依据,也为相似调水工程的洪水水位预报提供一定的参考。

As one of the key storage lakes in The Yangtze-to-Huai Water Diversion Project,the change of water level in Caizi Lake is affect-ed by both natural precipitation and engineering storage.In order to accurately simulate and forecast the water level of Caizi Lake,this paper constructs a model of Xinʹan River,the four water sources in Caizi Lake basin,and a neural network(LSTM)model of Caizi Lake water level forecasting.On this basis,an external coupling method is used to construct the four water sources Xin′an River-LSTM coupling model,and the inlet flow simulated by the physical mechanism model is further used as a complementary factor to drive the neural network(LSTM)mod-el to simulate the level of Caizi Lake,so as to realize the coupling application of the two different models in the lake water level forecasting.The results show that the flood error of the directly simulated water level is less than 0.1 m,and the flood error of the coupled simulated water level is less than 0.02 m,and the accuracy of the water level error of the latter is improved by 0.08 m compared with that of the former.The flood error of the directly simulated water level during the validation period is within 0.02 m,and the Nash coefficients,R^(2),are 0.89,0.75,and 0.88,respectively,and the root-mean-square errors,RMSE,are 0.034,0.027,and 0.015,respectively;the flood error of the coupled simulated water level validation period is within 0.015 m,the Nash coefficient R^(2) is 0.91,0.82 and 0.88,and the root mean square error RMSE is 0.019,0.021 and 0.008,respectively.The results of the study show that compared with the results derived from a single driving fac-tor,the results derived from the dual driving factor are more effective in improving the water level simulation accuracy.Meanwhile,in consid-ering the flooding process corresponding to rainfall,the combination of data-driven and physical mechanisms effectively improves the accura-cy of the field flood water level prediction and obtains more accurate simulation results compared with the direct prediction of water level er-ror.It provides an important reference basis for the water transfer of The Yangtze-to-Huai Water Diversion Project,and also provides some references for the flood level prediction of similar water transfer projects.