CDTDNet: A neural network for capturing deep temporal dependencies in time series

Abstract

The current research in time series forecasting is still deficient in extracting the time dependencies in depth. For this reason, a novel deep learning framework is proposed in this paper to extract deep temporal dependencies from time series data, and effectively feature-fuse temporal dependencies with other time series features. The Cell State Capture Recurrent Unit is used as a novel recurrent neural network together with Temporal Convolutional Network to capture the deep temporal dependencies of the data. Historical statistical information is constructed to introduce linear correlation variables for the model. Novel temporal attention coordinates the importance of time series time steps. Coupled attention improves the decoder's ability to interpret the encoded information. Finally, the AutoEncoder is employed as a prediction calibrator to improve the accuracy and robustness of the network. Comparisons with baseline methods and state-of-the-art strategies on datasets from four different domains confirm the effectiveness as well as the robustness of the proposed predictive network. In addition, the Cell State Capture Recurrent Unit can be considered a benchmark for time series forecasting instead of being limited to the Long and Short-Term Memory or Gated Recurrent Unit.