To address the poor timeliness and simple prediction functions of stock forecasting models, we propose a model abbreviated as BiLSTM-SA-TCN, which combines Bi-directional Long Short-Term Memory (Bi-LSTM) neural network, Self-Attention (SA) and Temporal Convolution Network (TCN).The learning unit and prediction unit in the proposed model can effectively learn important stock data, capture long-term dependency information, and output the predicted next day close price.The experimental results indicate that the BiLSTM-SA-TCN model has more stable prediction results on multiple data sets and has higher modle generalization ability.Furthormore, incomparative experiment, the BiLSTM-SA-TCN model achieves the lowest root mean square error, the lowest mean absolute error, and the best fitting degree of R² on the majority of datasets.