A spatio-temporal synchronous graph machine learning for prediction and optimization of regional traffic flow

The performance of urban regional traffic forecasting is poor due to its powerful spatio-temporal data representation ability and small sample data processing ability. This paper proposes a regional traffic flow prediction and optimization method using Spatio-temporal Synchronous Graph Machine Learning (SSGML). The proposed model combines the advantages of deep neural network in time series prediction and spatial feature extraction, and can more accurately capture the spatio-temporal variation of traffic flow. Similarly, the model can be quickly and effectively trained and deployed in small sample sizes, and has good predictive performance. In the experimental hyperparameter settings, the optimal experimental effect is achieved when Epoch= 240, model dimension is 32, the number of encoder decoder layers are 5, and the number of attention heads are 16. The experimental results on the PEMS04 and PEMS08 datasets show that the proposed SSGML model outperforms other comparison models in terms of performance metrics such as MAE, MAPE and RMSE. Specifically, compared with the optimal baseline model (DCRNN), the MAE, MAPE and RMSE of SSGML model on PEMS04 dataset are reduced by 2.02%, 7.83% and 7.10%, respectively; MAE, MAPE and RMSE on PEMS08 data set are reduced by 14.92%, 18.31% and 21.02%, respectively.