Accurately and real-time traffic flow prediction is the foundation for intelligent transportation systems (ITS). Since traffic flow time series contains both linear and nonlinear patterns, both theoretical and empirical findings have indicated that a combination of different models outperforms individual models. We propose a novel hybrid autoregressive integrated moving average (ARIMA) and long short-term memory (LSTM) neural network model (ARIMA-LSTM) for multi-step short-term traffic flow prediction. Firstly, we use the ARIMA model to extract linear parts and residuals are regarded as non-linear parts. Then, we formulate a novel neural network contains LSTM layers to capture the temporal feature of data, a concatenation layer to integrate the observed data, non-linearity components, and current linear components and a multi-output layer for multi-step prediction. Finally, the neural network is optimized on the global. To test the performance of proposed model, we use the freeway traffic volume data in Minnesota, U.S. to experiment and employed individual ARIMA, LSTM models and the hybrid ARIMA-ANN model for comparison. The results show that the performance of proposed hybrid ARIMA-LSTM model is improved by 30.59%, 23.97% and 25.62% than individual ARIMA, LSTM, and hybrid ARIMA-ANN model under mean absolute percent error (MAPE) criterion, respectively. In addition, the proposed model is most robust for multi-step prediction. The test results indicate the proposed hybrid ARIMA-LSTM model is a reliable model for multi-step short-term traffic flow prediction.

CICTP