High-efficiency chaotic time series prediction based on time convolution neural network

Abstract The prediction of chaotic time series is important for both science and technology. In recent years, this type of prediction has improved significantly with the development of deep learning. Here, we propose a temporal convolutional network (TCN) model for the prediction of chaotic time series. Our TCN model offers highly stable training, high parallelism, and flexible perception field. Comparative experiments with the classic long short-term memory (LSTM) network and hybrid (CNN-LSTM) neural network show that the TCN model can reduce the training time by a factor of more than two. Furthermore, the network can focus on more important information because of the attention mechanism. By embedding the convolutional block attention module (CBAM), which combines the spatial and channel attention mechanisms, we obtain a new model, TCN-CBAM. This model is comprehensively better than the LSTM, CNN-LSTM, and TCN models in the prediction of classical systems (Chen system, Lorenz system, and sunspots). In terms of prediction accuracy, the TCN-CBAM model obtains better results for the four main evaluation indicators: root mean square error, mean absolute error, coefficient of determination, and Spearman's correlation coefficient, with a maximum increase of 41.4%. The TCN-CBAM has also the shortest training times among the previous classic four models.