基于TFPW-BS-Pettitt法的水文序列多点均值跳跃变异识别

TFPW-BS-Pettitt method for detection of multiple change-points in the mean of hydrological series

为解决水文变异造成的数据"污染"及水文变异识别中存在的异法异解、掩蔽效应、自相关性影响、变异信息交叉等问题,提出将去趋势预置白(TFPW)、二元分割法(BS)与Pettitt检验法相结合,构建TFPW-BSPettitt法进行多点均值跳跃变异识别,并应用水文频率法验证变异点的准确性。结果表明TFPW-BS-Pettitt法缓解了序列的自相关性影响,消除了掩蔽效应,可有效实施多点识别。数据长度对Pettitt法的影响客观存在,且当变点外数据长度小于5时,易出现漏点现象,需做边界处理。研究还发现对均值跳跃和方差变异序列,可依托单纯均值变异段数据构造出近天然数据,缓解实际频率计算中的数据缺乏问题。

Hydrological variation is occurring in many areas, which often causes the ‘pollution＇ of hydrological data and even makes the traditional hydrological frequency calculation in a difficult position. Although a few approaches to this issue have been proposed, some problems, such as different solutions obtained by different methods, masking effect, impact of autocorrelation, and overlapped variation information, are still barriers to developing hydrological design theories in the changing environment. In this paper, a TFPW-BS-Pettitt method is presented for the issue, which combines trend-free pre-whiting（TFPW）, binary segmentation（BS） and Pettitt test to deal with the autocorrelation and masking effects occurring in change-point detection of streamflow data series. And hydrological frequency analysis is used to verify its accuracy in detection of change-point locations by comparing the data distributions before and after the change-points. Results indicate that this method can alleviate the autocorrelation effect on change-point detection through TFPW processing and eliminate the masking problem by BS algorithm, which makes multiple change-points detected effectively. Disadvantage in this new model is that the influence of data length on Pettitt test still remains. Especially when the length of a data sequence beyond the marginalized change-point is less than 5, this point is easily missed and thus a treatment of the edge effect should be done before detection. In addition, it is also found that when only change-points in mean exist in the sequence, an equal near-natural streamflow series can be obtained through calculation according with the data series containing the changes. And it may be used in the design to alleviate the impact of data lack in practical frequency analysis. When change-points in mean and variance both exist, such a series can be obtained only for the piece of data before the first change-point in the variance.