Traffic flow forecasting has been regarded as an important issue for traffic planning, design and management. In this study, we propose a multimodal deep learning approach for short-term traffic flow forecasting, which can jointly and adaptively learn the variation regulation of long temporal trend, seasonality, and residual of traffic flow produced by multi-dimensional decomposition. According to the highly nonlinear characteristics of traffic flow, the module of our approach consists of one-dimensional Convolutional Neural Networks (1D CNN) and Bi-directional Long Sort-Term Memory (Bi-LSTM) with the attention mechanism for fusion. The former is to capture the local trend features of residual and the latter is to capture the temporal regulation of trend and seasonality. As the daily, weekly and monthly periodicity of traffic flows, target prediction is related with various irregular sequences. By introducing an attention mechanism, the strongly correlated historical information is connected for data fusion of the final prediction. Compared with the baseline methods ARIMA and LSTM, the experimental results indicate that the proposed multimodal deep learning model is capable of dealing with complex nonlinear urban traffic flow forecasting with satisfying accuracy and effectiveness.

Traffic flow forecasting;Multimodal deep learning;Long Sort-Term Memory;Convolutional Neural Networks;Multi-dimensional decomposition