Abstract
A novel deep learning approach, using Convolutional Gated Recurrent Unit with Temporal Spatial Neural Net-work (ConvGRU-TSNet), is presented in this paper. It is proposed to solve the mixed pattern challenges faced by multivariate time series prediction. Gated Recurrent Unit (GRU) and Convolutional GRU (ConvGRU) are utilized to extract long-term patterns among different periods such that the short-term patterns in one period can be detected. Additionally, the traditional autoregressive (AR) model is employed to tackle the scale insensitive problem. The experiments are carried out using four datasets with complex mixed patterns. It is observed that the ConvGRU-TSNet presents enhanced performance exceeding several baseline methods. Among the four selected datasets, the Relative Squared Error (RSE) can be reduced by 7.604% at most and the Correlation Coefficient (CORR) can be increased by 1.198% at most. It is demonstrated that the proposed ConvGRU-TSNet offers an alternative perspective of solving the mixed-pattern challenges with improved prediction accuracy. The code and four datasets are available at https://github.com/privateos/ConvGRU-TSNet