The accurate prediction of real-time traffic flow is indispensable to intelligent transport systems. However, the short-term prediction remains a thorny issue, due to the complexity and stochasticity of the traffic flow. To solve the problem, a combined prediction method for short-term traffic flow based on the autoregressive integral moving average (ARIMA) model and long short-term memory (LSTM) neural network was proposed. The method could make short-term predictions of future traffic flow based on historical traffic data. Firstly, the linear regression feature of the traffic data was captured using the rolling regression ARIMA model; then, backpropagation was used to train the LSTM network to capture the non-linear features of the traffic data; and finally, based on the dynamic weighting of sliding window combined the predicted effects of these two techniques. Using MAE, MSE RMSE and MAPE as evaluation indicators, the prediction performance of the combined method proposed was evaluated on three real highway data sets, and compared with the three comparative baselines of ARIMA and LSTM two single methods and equal weight combination. The experimental results show that the dynamic weighted combination model proposed has better prediction effect, which proves the versatility of this method.

实时交通流量的准确预测对于智能交通系统是必不可少的。然而，由于交通流的复杂性和随机性，短期预测仍然是一个棘手的问题。针对这一问题，提出了一种基于自回归积分移动平均模型（ARIMA）和长短期记忆（LSTM）神经网络的短期交通流量组合预测方法。该方法可以基于历史交通数据对未来的交通流量进行短期预测。首先，使用滚动回归ARIMA模型捕获交通数据的线性回归特征。然后，使用反向传播训练LSTM网络以捕获交通数据的非线性特征。最后，基于滑动窗口的动态加权结合了这两种技术的预测效果。以MAE，MSE RMSE和MAPE为评估指标，在三个真实的高速公路数据集上评估了所提出的组合方法的预测性能，并与ARIMA和LSTM的三个比较基准，两种单一方法和相等权重组合进行了比较。实验结果表明，提出的动态加权组合模型具有较好的预测效果，证明了该方法的通用性。