Abstract: In deep learning short-term traffic flow prediction, there are some problems, such as the weak real-time performance in data processing and the poor prediction performance caused by the insufficient reuse and correction ability of the algorithm for traffic flow data. To address these problems, a kind of short term traffic flow prediction method using deep bidirectional long short-term memory neural network which incorporated a hysteretic extreme learning machine was proposed. Firstly, a weight-sharing mechanism was introduced to optimize the structure of the bidirectional long and short-term memory network model, and the weights and biases were continuously updated during the training process of the model, so as to make full use of the reverse-order data to enhance the reuse and correction ability of the data. Secondly, in order to further improve the real-time performance of the algorithm, the extreme learning machine model was introduced and the hysteretic parameters of the biological neural system were embedded in the activation function of its neurons, which optimized the computational efficiency and enhanced the overall real-time performance of the algorithm. The experimental results show that the proposed method has improved prediction accuracy and algorithm real-time performance, its average absolute error is respectively reduced by 6.82 and 6.47 and its computation speed is respectively improved by 12 and 19 seconds, compared with the classical method CNN-BiLSTM and multivariate ensemble CNN-LSTM, which has good short-term traffic flow prediction capability and real-time performance.