基于灰色关联与BP神经网络的台风非台风暴雨洪水分类模拟

Simulation of Storm-Floods during Typhoon and Non-Typhoon Seasons Based on Grey Correlation Analysis and BP Neural Network

为进一步研究BP神经网络模型在我国东南沿海地区不同类型暴雨洪水模拟的适用性,基于1956—2011年东南沿海西溪流域暴雨洪水实测资料,将洪水划分为台风和非台风暴雨洪水两类,选取并统计影响洪峰流量的7个要素,采用灰色关联法,分别分析洪水、台风暴雨洪水、非台风暴雨洪水的洪峰流量与各个要素之间的相关性,应用BP神经网络模型对三种系列洪水进行分类模拟。结果表明:(1)各个要素分别与台风和非台风暴雨洪水的洪峰流量的关联度大小、排序明显不同,不同类型洪水的洪峰流量与影响要素的之间相关程度存在较大的差异;(2)构建的多种BP神经网络模型结果都较为满意,可用于西溪流域洪峰流量的模拟预测,且进行台风与非台风暴雨洪水分类后的模型性能更优;(3)分别选取4个主要影响要素建立的台风与非台风暴雨洪水BP神经网络模型,模拟和预测的精度同样较高,能够有效地预测洪峰流量。

Convective rainstorms and typhoon storms occur frequently in the southeastern coastal area of China. In order to investigate the applicability of BP neural network model for simulating storm-floods in this region,Xixi watershed was selected as a study area. Based on observed data of annual rainstorm flood events occurring from1956 to 2011,seven indices describing the characteristics of storm-floods were calculated（ rainstorm volume,rain intensity,rain duration,rainfall time coefficient of variation,center position of rainstorm peak,rainfall space coefficient of variation,and starting discharge of flood）. First,all storm-floods were divided into two types： the typhoon storm-floods and non-typhoon storm-floods. Then,the correlation between peak discharge and each of the seven indices was estimated respectively with a grey correlation method,regarding three flood groups（ all events,typhoon,non-typhoon） separately. Finally,a variety of BP neural network models were constructed and the performances of models were compared. The results show that the grey correlation degrees of peak discharge and the indices were different among three storm-floods groups,having a higher correlation for the classified floods. All the BP neural network models proposed worked reasonably well for all flood events,but performed better for typhoon storm-floods and non-typhoon storm-floods respectively. The differing correlations and improved model performance suggest that it is necessary to divide the floods into the typhoon and non-typhoon types. Furthermore,by only selecting four indices based on the grey correlation analysis,the constructed BP neural network models are also applicable for the simulation of peak discharges in the study area.