水文时间序列变点分析的可靠性检验

Reliability test for detecting change-point of hydrological time series

针对不符合正态分布的水文时间序列,建立用于时间序列变点分析的贝叶斯数学模型.研究了某水文站1942—2008年共67年年径流最大值时间序列的突变,得到1989年是最大可能变异点.贝叶斯数学模型中仅给出了变点位置的后验概率,但是最大后验概率时间位置处是否确实发生变异,则有待进一步讨论.因此文中认为需要进一步检验1989年前后的变化是否在随机波动范围内或者是确实发生了变异.文中采用再抽样自助法对结果进行深入的可靠性分析,证实1989年前后确实发生了突变,变异程度远远超过了随机波动范围.考虑到时间序列的随机性,建议在今后工作中采用贝叶斯变点分析模型确定变异点时,采用自助再抽样方法对结果进行可靠性分析.

A Bayesian model for analyzing change-points of time series is established to study the abrupt change of the mean value of strong autocorrelated and abnormal distributed hydrological time series.The maximum annual runoff time series from 1942 to 2008 of one hydrometric station is analyzed;the results of the Bayesian model show that 1989 is the most probable change-point location with the largest posterior probability.But whether there exists abrupt change on earth cannot be detected from the Bayesian model result;further reliability test should be done to testify the change around 1989 belongs to chance fluctuation or there exists abrupt change indeed.Resampling method is used to determine the reliability of the estimated change-point.The result shows that there exists abrupt change indeed;the change around 1989 is far more than stochastic fluctuation.Taking the randomness of observed time series into account,reliability test should be done when using Bayesian model to detect the change-point location in future work and bootstrap resampling method is recommended.