TATCN: Time Series Prediction Model Based on Time Attention Mechanism and TCN

Prediction is an important research task of time series data analysis. As a powerful tool to solve the problem of time series prediction, Temporal Convolutional Networks (TCN) shows good performance in the prediction task. However, TCN model lacks the consideration of the influence of different historical segments on the prediction value, which limits the prediction accuracy of the model to a certain extent. Therefore, this paper combines the attention mechanism with the data characteristics of time series, proposes a Time Attention mechanism (TA), and integrates it into the TCN model framework to build a prediction model (called TATCN). In TATCN, the TCN output vector of each layer is convoluted, and the sigmoid function is used to generate the weight coefficient, and then the weight coefficient is multiplied by the original output vector to form a new output vector. The new output vector and the input vector of the current layer are added by residual connection as the final output vector of the current layer and input to the next layer network. The experimental results on EEG data and Yanbian electricity fees data show that the Time Attention mechanism in this paper can effectively represent the importance of different historical data to the current prediction. The proposed TATCN model has a significant improvement in the prediction accuracy compared with TCN model, and is also better than RNN prediction models such as LSTM and GRU.