Multi‐step traffic speed prediction model with auxiliary features on urban road networks and its understanding

Abstract

Multi-step prediction of long-term traffic speed is an important part of the intelligent transportation system. Traffic speed is affected by temporal features, spatial features, and various environmental features. The prediction of traffic speed considering the above features is a big challenge. This study proposed a multi-step prediction model named embedding graph convolutional long short-term memory network (EGC-LSTM) for urban road network traffic speed prediction which can deal with spatial–temporal correlation and auxiliary features at the same time. Firstly, a graph convolutional network (GCN) for capturing directed graph properties is proposed. Based on the GCN, the LSTM and sequence to sequence model are further applied to realise multi-step prediction considering the spatial–temporal correlation of the traffic network. To improve the performance of the model and obtain the importance of each step in the historical data, the attention mechanism is introduced. Then, one-hot encoding is applied to the category-type auxiliary features. Considering that the dimension becomes larger after the features are one-hot encoded, the dimensions are reduced using embedding. The experiment results prove that the proposed model's performance is better than other models, and the model is interpreted in detail.