Study On Short-Term Traffic Flow Prediction Considering Dynamic Spatio-Temporal Features In Urban Road Network

Due to the lack of complete directly measured data,which contains the all vehicle identity information of the whole city.On the one hand,existing studies cannot adequately mine the dynamic spatial-temporal correlations feature of traffic flow in urban road networks.On the other hand,most researchers have not considered the vehicle type of the traffic flow.Electronic registration identification(ERI),which is an emerging technology for uniquely identifying a vehicle,can help collect the travel records of all vehicles.This inspires us to employ ERI big data for traffic flow prediction.To solve the problem of the dynamic spatial-temporal correlations feature mining.We propose a dynamic spatial-temporal feature optimization method with ERI big data based on a gradient–boosted regression tree(DSTO-GBRT).We analyze the dynamic spatial-temporal correlations to extract feature and optimize the original training data.Finally,we utilize gradient descent method to fit the features and future traffic flow.In the experiment,compared with ST-GBRT,ARIMA,DSTO-BPNN,DSTO-SVM,DSO-GBRT and DTO-GBRT,the results demonstrate that DSTO-GBRT can provide timely and adaptive prediction even in rush hour,when traffic conditions change rapidly.And we propose a hybrid method named FT-ST-LSTM,which treat traffic flow as two components: periodicity and volatility.The period traffic flow is treated as a function of time and period behaviors are modeled with period sine and cosine follow the idea of Fourier Transform.The volatility is determined by the surrounding environment of the current location,so the spatial-temporal correlations are extracted as input features of long-short term memory network(LSTM).In the experiment,we compared FT-ST-LSTM with other FT model(FT-ST-SVR,FT-ST-GBRT)and non-FT model(ARIMA,ST-LSTM,ST-SVR,ST-GBRT)to verify the FT-ST-LSTM.