Time-series Prediction Based on VMD and Stack Recurrent Neural Network

Time-series prediction is a hot research field. How to build an effective model to improve the accuracy of long-term prediction is a difficult issue. In this paper, we propose a stack recurrent neural network with variational modal decomposition (VMD-SRNN) for long-term time-series prediction. First, a time series is decomposed into multidimensional subsequences by using variational modal decomposition to reveal the potential hidden information of the original time series and improve the prediction accuracy of time series. In addition, we build a stack recurrent neural network (SRNN) model to predict subsequences. The hidden layer of SRNN model has two reservoirs and these reservoirs effectively excavate the internal correlation information of subsequences, which enhances the long-term prediction ability. Besides, the links of reservoir neurons are improved into a special sparse connection structure to ensure the generalization ability of SRNN. Finally, we report the experimental results on multi-step prediction on the Lorenz-x time series and the NO2 time series. The results show that the model has a prominent prediction ability in long-term prediction.