Research Of Short-term Traffic Flow Forecasting Model Based On Ensemble LSTM

In order to solve traffic congestion,many cities have begun to deploy and develop Intelligent Transport System(ITS).Short-term traffic flow forecasting,as an important basic technology of ITS,has always been a research focus and hotspot in the field of ITS.This thesis utilizes vehicle inspection data from radio frequency identification(RFID),and implements long short-term memory neural network(LSTM)to study short-term traffic flow prediction.The main work of this thesis is as follows:Firstly,the existing methods of short-term traffic flow prediction are summarized in detail.Basic parameters and characteristics of short-term traffic flow are analyzed,and related technologies of LSTM are expounded which lay a foundation for the subsequent model construction and implementation.Secondly,a single LSTM model is built to predict short term traffic.Considering that traffic flow has complex historical dependence,we select LSTM and implements it on open source machine learning library.The data set comes directly from real RFID vehicle inspection.After trained with the training set,the model predicts the traffic flow of a certain intersection for one day,and its accuracy is verified.Next,the thesis builds an ensemble LSTM model on the basis of previous work on single LSTM model,in order to improve its generalization ability and prediction accuracy.Additional,several influencing factors such as holidays and weather condition are taken into consideration this time.Different data subsets are constructed for them by data extraction;then different training sets and different initial weight intervals are set for training to obtain multiple differentiated LSTM neural network models.Then,using Bagging technique,an ensemble LSTM neural network model is established.The weight coefficients of each model are obtained through training,and the final prediction result is obtained by weighted averaging the predicted values from each single LSTM model.Finally,the trained ensemble LSTM model is tested its performance compared with other models such as Back Propagation(BP)model,Autoregressive Integrated Moving Average(ARIMA)model and single LSTM model.Tests showed that the ensemble model achieved the lowest mean absolute percentage error of 8.86%,and root mean square error of 7.18,and the highest average accuracy of 91.14%.In addition,the predictive stability of the ensemble model is also the best when increasing the amount of data.It showed that the ensemble LSTM model has higher prediction accuracy and better generalization ability.