一种无延迟的水文时间序列预测方法

A Delay-free Hydrological Time Series Forecasting Method

为解决水文时间序列预测中,序列连续值间存在着高度的自相关性,而引起峰值点的预测时间落后于真实时间的问题,提出一种基于CSVMD-LSTM-ELM的无延迟预测方法。首先,将变分模态分解算法(variational mode decomposition, VMD)与布谷鸟搜索算法(cuckoo search algorithm, CS)相结合,其中,VMD用于削弱时间序列间的相关性,CS用于全局搜索VMD参数的最优解,并重点关注预测延迟的问题,为此,定义了一种新的适应度函数;其次,为分解得到的子序列建立了长短期记忆神经网络(long and short-term memory neural network, LSTM)和极限学习机(extreme learning machine, ELM)两种网络结构,分别讨论了单一网络和组合网络预测效果的优劣;最后,在秦淮河流域数据集上进行实验验证,与原有的LSTM和VMD-LSTM-ELM方法进行比较。结果表明,所提方法相较于其他方法,预测的峰值时间延迟更小,预测误差更低。可见,所提方法能够解决预测的延迟问题。

In order to solve the problem that there is a high degree of autocorrelation between consecutive values in the hydrological time series prediction, which causes the prediction time of the peak point to delay behind the real time, a delay-free prediction method based on CSVMD-LSTM-ELM was proposed. Firstly, the variational mode decomposition(VMD) algorithm was combined with the cuckoo search algorithm(CS). Among them, VMD was used to weaken the correlation between time series, and CS was used to search the optimal solution of VMD parameters globally, focusing on the problem of predicting delay. For this purpose, a new fitness function was defined. Secondly, two network structures were established for the sub-sequences obtained by decomposition, including long and short-term memory neural network(LSTM) and extreme learning machine(ELM),and the pros and cons of prediction effect of the single network and the combined network were discussed separately. Finally, it was verified by experiments on the Qinhuai River Basin data set, and compared with the original LSTM and VMD-LSTM-ELM methods. The results show that, compared with other methods, the proposed method has shorter peak time delay and lower prediction error. It can be seen that the proposed method can solve the prediction delay problem.