新型耦合数据驱动模型在降雨径流模拟中的应用研究

Application of A New Coupled Data-driven Model in Rainfall-Runoff Simulation

为解决传统数据驱动模型的不足,使其能实现降雨径流过程高精度连续模拟,提出新型耦合数据驱动模型——PEK,即:基于偏互信息的输入变量选择、基于新型集成神经网络的出流量预测和基于K最近邻模型的出流量误差预测。PEK模型具有以下特点:(1)提出了基于分离式选择策略和滑窗累积雨量的模型候选输入向量,并与基于偏互信息的输入变量选择方法联合使用,提高了输入信息的充分性和无冗余性,对建立精度高、泛化能力强的高质量模型意义重大;(2)提出了新型集成神经网络——EBPNN及其率定方法。联合使用NSGA-II多目标优化算法和早停止Levenberg-Marquardt算法,通过一次优化过程同时确定全局最优个体网络个数、各个体网络拓扑结构和网络参数。个体网络权重由基于AIC信息准则的权重优选方法确定。EBPNN在模拟精度和网络复杂度间取得了良好折衷,精度高、泛化能力强、率定结果客观;(3)PEK模型能够进行多步外推预报,实现了非实时校正模式下的高精度连续模拟,增长了预见期;(4)PEK模型不需要进行流域状态变量的计算,仅需初始出流量就可进行出流量的连续模拟。在呈村流域应用PEK和CLS两个数据驱动模型进行次洪降雨径流模拟及精度比较。结果表明PEK模型使用简便,模拟精度高于CLS模型,实现了多步外推的高精度连续模拟,增长了数据驱动模型的预见期。

For the purpose of overcoming the disadvantages of the traditional data-driven model and implementing the high accura- cy rainfall-runoff simulation by data-driven model, a new coupled data-driven model named PEK has been proposed in this paper. The PEK model was developed by coupling partial mutual information based on the input variable selection, novel ensemble back- propagation neural network based on discharge forecasting and K-nearest neighbor based on discharge error forecasting. The PEK model has the following characteristics： （1） The separate IVS strategy and the sliding window accumulative rainfall based on model candidate input vector. These two methods combined with the PMI-based IVS approach ensure the sufficiency and parsimony of the input information and is very important to the construction of the high quality data-driven model; （2） The novel ensemble back- propagation neural network and the methodology of calibration were proposed in this paper. The global optimal number of compo- nent networks, topology and network parameters were obtained simultaneously by using the NSGA-II multi-objective optimization al- gorithm and the early stopping LM algorithm. The combination weights of the component networks were obtained by the AIC-based component network weights assignment approach. The EBPNN model can make a good compromise between simulation accuracy and network complexity; （3） The PEK model implemented multi-step forecasting and high accuracy simulation under the non-updating mode. The forecast period was also increased; （4） The PEK model doesn＇t need to compute the catchment state variables and im- plements continuous simulation only by using the initial discharge. In this paper, the PEK and CIS models were applied in hourly rainfall-runoff simulation in the Chengcun watershed and the results were compared. The simulation results indicate that the PEK model is easy to use, is better than CLS in simulation accuracy, can realize multi-step high accuracy simulation and increase the forecasting period of the data-driven model.