Prediction Study Based on TCN-BiLSTM-SA Time Series Model

Abstract

To enhance the accuracy of time series prediction, this study proposes a hybrid network model called TCN-BiLSTM-SA, which combines Temporal Convolutional Network (TCN), Bidirectional Long Short-term Memory (BiLSTM), and Self Attention (SA). The TCN is employed to learn sequence features, while the BiLSTM model captures preceding and succeeding states to extract more information for prediction. The self-attention mechanism calculates weights for each time step's output, effectively utilizing the cell memory information of BiLSTM to capture global features and improve prediction accuracy. Experimental results on the Beijing PM 2.5 dataset demonstrate that the TCN-BiLSTM-SA network outperforms the BiLSTM model in terms of RMSE, MAE, and MAPE error rates, while also exhibiting greater stability. This model holds promising potential for various time series prediction applications. It has broad application prospects in time series prediction.