山区流域GRU神经网络洪水预报模型研究

Study on GRU Neural Network Flood Forecasting Model in Mountain Watershed

摘要: 运用具有更新门和重置门控制的深度学习循环神经网络(GRU)技术，选取1997~2021年期间崇阳溪上游流域29场降雨径流过程，其中21场过程作为训练集，以上游岚谷等6个雨量站逐时雨量和下游控制断面武夷山水文站前期流量为输入，以该断面相应流量为输出，依据RMSE最小方法确定网络隐含层单元数和迭代轮数，在GRU层之后增加全连接层，并对其进行Dropout化处理，构建GRU神经网络预报模型。采用该模型对余下的8场洪水进行测试，并与共轭梯度PRPB神经网络模型结果进行对比。结果表明，GRU模型预测效果更好，其洪水过程预测误差均小于PRPB模型，在洪峰流量预测精度方面总体上略高于PRBP模型，模型的纳什效率系数也比后者高，因此适合用于山区流域的洪水预报。

Abstract: The deep learning recurrent neural network technology with renewal gate and reset gate control (GRU) was used to select 29 rainfall runoff processes in the upper reaches of Chongyangxi River from 1997 to 2021, among which 21 floods were selected as the training set. Hourly precipitation records from six upstream rain gauges—Langu among them—together with the antecedent discharge measured at the downstream Wuyishan station were employed as model inputs. Taking the corresponding flow of this section as the model output, the root-mean-square error minimum criterion was used to analyze the number of hidden layer units and the number of network iteration rounds. At the same time, a full-connection layer was set after the GRU layer and the full-connection layer was processed by Dropout to construct a GRU neural network model for mountain watershed. The model was used to test the remaining 8 floods and compared with the artificial neural network PRPB model. The results show that the GRU model performs better in prediction, and its flood process prediction error is smaller than that of PRPB model, and the accuracy of flood peak flow prediction is slightly higher than that of PRBP model. The Nash efficiency coefficient of the model is also higher than that of the latter, so it is suitable for flood forecasting in mountain basins.