基于神经网络与半分布式水文模型相结合的缺资料区径流估计模型——以莺落峡流域为例

Estimation of streamflow in ungauged basins using a combined model of black-box model and semi-distributed model:case study in the Yingluoxia watershed

针对黑河上游莺落峡流域海拔较高区域降水缺测问题,基于半分布式水文模型TOPMODEL,利用神经网络(ANN)融合海拔较高区模拟降水与流域稀疏站点观测作为降水输入计算产汇流,并利用粒子群优化算法进行参数全局优化以同时率定TOPMODEL与ANN参数,从而构建基于ANN与半分布式水文模型TOPMODEL相结合的径流估计模型P-NN-TOP.利用1990—1995年莺落峡水文控制站观测日径流进行模型进行参数率定,利用1996—2000年的相应观测进行模拟验证,并借助所发展模型P-NN-TOP对2001—2010年的日径流进行估计与分析.结果表明:在模型率定和验证期,新发展的P-NN-TOP模型较其他模型表现出明显的优势,在观测站点稀疏情况下能较好地融合站点观测和模式模拟信息;并在兼具输入数据可获取性的同时提高了径流模拟的精度.此外,结果还显示2001年为模拟与观测日径流相关性的年突变点并从2004年开始明显降低,与上游水电站的开发使用在时间上基本一致,表明所发展的P-NNTOP模型能较为合理描述2001—2010年出山流量日变化,在探索径流估计新方法的同时也为类似流域的洪涝预警和水资源调配提供一定参考.

To estimate streamflow in ungauged basins,an integrated scheme of artificial neural network （ANN）and TOPMODEL is proposed.ANN is used to fuse rainfall data from different observation stations at lower altitude areas with a regional climate model for a higher region.Rainfall input to TOPMODEL was directly replaced by output of ANN model. Generation and routing of runoff were conformed to the TOPMODEL.Particle Swarm Optimization was adopted for global parameter calibration for the whole scheme. The Yingluoxia watershed was used to validate performance of the new scheme in ungauged catchments. Observed discharges from 19901995 and 19962000 were used for model calibration and validation, respectively.Natural discharges from 20012010 were then estimated and analyzed. Compared with TOPMODEL,R-NN-TOP and CLM4.5-RTM,the proposed scheme can better tradeoff the predictive accuracy and data availability in sparse rainfall sites so that its performance as indicated by several evaluation indexes was more satisfactory,especially during peak flood periods.Moreover,reconstructed discharge hydrograph from 20012010 showed reasonable increase from 20012009 and a slight decrease later.Hence integrated scheme of black-box model and semi-distributed model may be promising in similar watersheds,to provide reference for flood warning and water resources management.