基于Brown-Forsythe检验的水文序列变异点识别

当前在水文序列变异点识别中常采用的几种统计方法都对数据有较多假设,当假设不满足时,识别结果通常并不理想。本文根据统计学方差分析的原理,建立了基于Brown-For-sythe检验的水文序列变异点识别方法,并采用该方法对新疆开都河大山口站近50年年平均径流序列进行了变异点识别。研究结果表明,该识别方法继承了Brown-Forsythe检验的优点,对数据不做过多假设,且易于进行多变异点识别,在一定程度上具有比当前所用统计方法更优越的性能。

Brown-Forsythe based method for detectingchange points in hydrological time series

The dramatic change of hydrological cycle and natural condition and the increasing intensity of human activity usually result in the step trend of streamflow. Change point detection in streamflow series is an important way to understand and diagnose the step trend. Up to now, many methods have been developed to detect the change point, among which the one derived from basic statistics theory is most widely used for its simplicity, but is also limited for its much assumption for data at the same time. In terms of theory of analysis of variance, a detection method based on Brown-Forsythe test is proposed in this paper, which retains the virtue of simplicity and loose the limitation for data. The method based on Brown-Forsythe test is used in change point detection of Kaidu streamflow of Xinjiang which indicates that 1973 and 1986 are two change points. However, other traditional methods take 1990 as the second change point. Which is true? On the one hand, Kaidu river is mainly influenced by climate, especially precipitation and some researches have indicated that the characteri stics of climate and hydrology of Xinjiang changed acutely in 1973 and 1987. On the other hand, taking 1986 as a change point will make not only the level diff erence between separated series but also the step between change point and next year more significant than in 1990. So it is reasonable to make 1973 and 1986 as the real change points and the method based on Brown-Forsythe test is suitable for hydrological time series. The normality and randomness violation of Kaidu streamflow data is probably the main reason for traditional method to get false results. In the future, the research about change point detection will still be focused on improving the sturdiness of methods for skew data.