资料匮乏地区径流降尺度模型构建及预测

The Implication of Stream Flow Downscaling Method in Ungauged Regions and Scenario Prospective

基于贝叶斯神经网络 ,构建了资料匮乏地区的径流降尺度模型 ,模拟了叶尔羌河卡群站月平均径流 ,与BP神经网络的结果进行了对比 ,验证了BNN的优越性 ,并结合CMIP5三种气候模式GFDL_ESM2G ,GFDL_ESM2M 及MIROC5的RCP 4 .5 ,RCP 6 .0 ,RCP 8 .5三种情景 ,对未来3个时段（2020年代 ,2050年代 ,2080年代）卡群站月平均径流进行了预测 ,并定量计算了预测的不确定性区间 ,研究表明 ：贝叶斯神经网络降尺度模型可以较好地捕捉叶尔羌河的径流特征 ,即相关系数达到0 .9以上 ,效率系数达到0 .8 ,且模拟效果比ANN较优;未来情景下 ,叶尔羌河流域受气温升高影响 ,3个时段年径流均呈现增加的趋势 ,增加幅度分别为75% ～92% ,83% ～110% ,88% ～127% ,其中RC P8 .5情景下的径流增加幅度比其他情景较明显 ;不同月份径流存在不同程度的增加趋势 ,其中5 -8月份变化趋势相对较明显.

Based on the newly presented stream flow downscaling method on the basic of Bayesian Neural Networks （BNNs） ,month-ly stream flow of Yarkant River was simulated and proved to be outperforming the results from conventional BP artificial neural net-works （ANNs） .The future monthly stream flows （2020s ,2050s ,2080s） in response to climate change on Kaqun hydro-station ,as well as the uncertainty interval ,were projected with consideration of three emission scenarios （RCP4 .5 ,RCP6 .0 ,RCP8 .5） provid-ed by three different global climate models （GFDL_ESM2G ,GFDL_ESM2M ,MIROC5） .Results indicate that the stream flow downscaling method has better performance in capturing the inner-annual and inter-annual stream flow changes ,with the correlation coefficient over 0 .9 and the efficiency coefficient reaching 0 .8 .Affected by rising temperature ,the monthly stream flow in Kaqun Station are expected to increase under all the future scenarios .,especially in RCP8 .5 ,and May-Aug monthly stream flow shows greater upward trend than Sep-Apr monthly stream flow .