Traffic speed prediction: spatiotemporal convolution network based on long-term, short-term and spatial features

With the more and more in-depth research on intelligent transportation, many scholars have proposed their models for accurate prediction of traffic. In this paper, we analyze the advantages and disadvantages of the existing models and propose our own model. In our model, the temporal and spatial factors are taken into account. Gate Recurrent Unit (GRU) and Gated Linear Units (GLU) are used to learn the short-term temporal features of traffic data, and Graph Convolutional Network (GCN) is used to learn the spatial features of traffic data. In order to fully learn short-term feature changes, a multi time step perception layer is proposed. A new network GCGRU is proposed to learn the long-term features of traffic data. As the sensor will be affected by urban canyon, weather, and other factors, there will be missing value and noise in the collected data. We created a short-term trend based missing value filling up algorithm to fill in missing values and use Singular Spectrum Analysis (SSA) algorithm to eliminate noise of training data set. In order to reduce the process of adjusting parameters manually in the model training process, we propose k-block search method based on fuzzy extreme points. Finally, the model is compared with the existing traffic forecasting models, and the analysis results show that our model has advantages in many indicators.