Figures (14)  Tables (3)

    • Figure 1. 

      Benefits of traffic flow prediction.

    • Figure 2. 

      Schematic diagram of spatial location distribution.

    • Figure 3. 

      Relationships between Artificial learning (AL), Machine learning (ML) and Deep learning (DL).

    • Figure 4. 

      Difference between simple neural network and deep learning neural network.

    • Figure 5. 

      Track data is matched to the map.

    • Figure 6. 

      Multi-source data fusion process.

    • Figure 7. 

      Grid structure of traffic network.

    • Figure 8. 

      Graph structure of traffic network.

    • Figure 9. 

      Structure of RNN.

    • Figure 10. 

      Structure of LSTM.

    • Figure 11. 

      Structure of TCN.

    • Figure 12. 

      Structure of Transformer.

    • Figure 13. 

      Structure of CNN.

    • Figure 14. 

      Structure of GCN.

    • Machine learning categoryMain methods
      Supervised learningSupport vector machine(SVM)[9−11]
      K-nearest neighbors(KNN)[12,13]
      Logistic regression[9, 13]
      Linear regression[12,13 ]
      Decision trees[14−19]
      Random forest[20−22]
      Unsupervised learningK-means clustering[9,13 ]
      Principal component analysis[9]
      Latent dirichlet allocation[23]
      Reinforcement learningQ-learning[24,25]
      Monte Carlo tree search[26]

      Table 1. 

      Main methods of machine learning.

    • TypeDetection
      data type      		Characteristic
      Loop
      detection data      		Traffic flow
      Speed
      Occupancy      		High detection accuracy but detection accuracy decreases in traffic congestion
      Geomagnetic detection dataTraffic flow
      Speed
      Occupancy      		Unable to detect stationary and slow-moving vehicles
      Microwave
      detection data      		Traffic flow
      Speed
      Occupancy
      Density
      Queue      		Detection errors may occur when large vehicles obstruct the reflection waves of small vehicles

      Table 2. 

      Fixed detection data.

    • ModelPerformance (PeMS04)Optimization prospect
      MAERMSE
      ARIMA32.1144.59Improve the robustness
      LSTM28.8337.32Multi-LSTM stack/
      Increase Dropout
      GRU28.3240.21Bi-GRU/
      Use semi-supervised training
      Transformer[81]18.9221.28Reduce quadratic complexity
      DCRNN[62]24.7033.60Replacement activation function/
      Add a Residual connection
      STGCN[32]25.1531.45Change convolution kernel size/
      Join Jump connection
      ASTGCN[33]21.8028.05Modified spatial convolution/
      Integrate external factors
      STSGCN[106]21.1924.26Add multi-grained information/
      Use semi-supervised learning
      to increase robustness

      Table 3. 

      Baseline model performance and its prospects.