水文变异对深度学习模型训练性能的影响研究

Research on the Influence of Hydrological Variation to the Training Performance of Deep Learning Model

受气候变化和人类活动的双重影响,黄河流域水文情势发生剧烈演变,水文序列颠覆传统的“一致性”假设。为使深度学习模型更好适应于变异水文序列预测,采用Mann-Kendall检验方法确定径流时间序列变异点,对基于LSTM模型的k-fold交叉验证方法进行改进,提出考虑变异点的k-fold交叉验证方法,分析水文变异对深度学习模型训练的影响。选取黄河支流渭河华县站和黑河祁连山站为研究对象,对比分析不同变异程度的中长期水文序列预测结果。结果表明:渭河华县站的径流时间序列变异发生在1986年和2008年,黑河祁连山站径流时间序列变异发生在1988年;LSTM模型在中长期径流预测中,模型性能主要受时间序列长度的影响,其次受水文变化趋势及变异的影响。

Due to the dual impact of climate change and human activities,the hydrological regime in the Yellow River Basin has undergone drastic evolution,and the hydrological sequence has overturned the traditional assumption of"consistency".In order to better adapt the deep learning model to the prediction of variable hydrological series,the Mann⁃Kendall test method was used to determine the variation points of the runoff time series.The k⁃fold cross validation method based on the LSTM model was improved,and a k⁃fold cross validation method con⁃sidering the variation points was proposed to analyze the impact of hydrological variation on the training of the deep learning model.Selecting Huaxian Station on the Weihe River and Qilianshan Station on the Heihe River as the research objects,we compared and analyzed the mid to long term hydrological series prediction results with different degrees of variation.The results show that the variation of runoff time series at Huaxian Station on the Weihe River occurred in 1986 and 2008,while the variation of runoff time series at Qilian Mountain Station on the Heihe River occurred in 1988.The performance of the LSTM model in medium to long⁃term runoff prediction is mainly influenced by the length of the time series,followed by the trend and variation of hydrological changes.