Title: A recurrent neural network based on attention mechanism to predict the trend of univariate time series
Authors: Yunxin Liu
Addresses: School of Computer Science and Technology, China University of Mining and Technology, Xuzhou, 221116, China; Mine Digitization Engineering Research Center of Ministry of Education of the People's Republic of China, Xuzhou 221116, China
Abstract: For the time series with high acquisition frequency and high noise, it is difficult to establish the prediction model directly. If we simply take their average values, we will lose a lot of trend information. Therefore, we studied how to accurately obtain the trend information of the time series and establish its accurate prediction model, and proposed a prediction model based on K-means clustering. The first step of the model is to obtain the trend information of the original time series based on the K-means clustering idea, and the second step is to use the gated recurrent unit based on the attention mechanism to establish a prediction model for the trend information. Experiments on three dataset show that the proposed K-means clustering method can effectively reduce noise interference and accurately obtain trend information. Comparative experiments on different prediction models show that our proposed prediction model has the best prediction accuracy.
Keywords: time series; change trend prediction; K-means clustering; attention mechanism; gated recurrent unit.
International Journal of Collaborative Intelligence, 2020 Vol.2 No.2, pp.108 - 124
Received: 05 Dec 2019
Accepted: 21 Dec 2019
Published online: 08 Dec 2020 *