Time Series Segmentation of Flood Flow Based on Bi-LG-LSTM Neural Network

Abstract

With fast development of big data technology, how to mine useful information from a large amount of time series data is an important task in hydrology. Such process of information mining should be built on the basis of pre-processing, including data cleaning, aligning, segmentation and so on. In this paper, we mainly focus on an efficient algorithm to accurately segment time sequences in hydrology domain. Although traditional methods of time series segmentation can segment time series easily, there are three drawbacks in traditional methods; First, traditional segmentation methods pay more attention to the evaluation of the compression ratio and the fitting rate, but ignore the accuracy of segmentation points; Secondly, traditional segmentation methods do not take the information of time series data itself into account, such as it is easy to divide a whole time series into two time series segments, which leads to losing semantic information in time series analysis. Finally, in these traditional evaluations of time series segmentation, the number of segments and the position of the segmentation points are not comprehensively taken into account. In this paper, we propose a time series segmentation method based on Bi-LG-LSTM neural network, which is modified on the base of the LSTM/Bi-LSTM neural network. Through the supervised learning method on flood flow data sets, we can extract a large number of effective segments of the time series. Meanwhile, we propose a novel evaluation method, which utilizes the dynamic DTW distance method to evaluate the results of the time series segmentation method. The evaluation method takes the position of segmentation points and the number of time series segments into consideration. Experiments on the flood flow time series show that our proposed method has an obvious effect.