耦合BP神经网络的MIKE11模型预报无资料断面水位

Investigation of water level prediction at ungauged sections by coupled BP neural network and MIKE11 model

为解决无资料断面水位预报问题,提出BP神经网络与MIKE11模型耦合方法。通过小时尺度的洪水场次数据,选取水位、降雨和前期影响因子进行相关性分析;筛选出预见期12 h及24 h兰江流域水位关联度最高的预报因子,以多元线性回归(MLR)和反向传播神经网络(BPNN)模型构建兰溪站不同预见期下的水位预报模型;并利用Mike NAM-HD机理模型推演未设测站断面的水位数据,通过BP神经网络进一步构建兰江流域无资料断面的水位预报模型。结果表明:(1)BPNN模型对长预见期下的水位预报效果优于MLR模型;(2)随着预见期的延长,BPNN模型的纳什效率系数逐渐降低;(3)对比筛选预报因子前后BPNN模型的纳什效率系数,12 h及24 h预见期下分别提升9.0%、34.7%;(4)提出的耦合BP神经网络和MIKE11模型方法可应用于无资料断面的水位预报。

To solve the problem of water level prediction at ungauged sections,the method of coupled BP neural network and MIKE11 model was proposed in this study.First,the water level,precipitation and antecedent influencing factors were collected for correlation analysis according to the hourly flood events data.Second,the input factors with the highest correlation degree of water level in the Lanjiang River Basin in the prediction period of 12 h and 24 h were selected,the water level predictions at Lanxi station were built by MLR and BPNN models for different forecasting lead times.Then,the BPNN-based water level prediction at ungauged sections in Lanjiang basin was constructed through the synthetic water level data generated by MIKE NAM-HD model.The results indicate that:(1)The performance of BPNN is better than MLR model for water level predictions of long lead times;(2)The NSE obtained from BPNN gradually decreases with the increasing of forecasting lead time;(3)After input variable selection,the improvements of BPNN at lead times of 12 h and 24 h are about 9.0%and 34.7%in terms of NSE;(4)The proposed BPNN coupled with MIKE11 model is able to apply to ungauged section water level predictions.