Research On Traffic Flow Prediction Method Based On Multi-time Granularity Fusion And Deep Learning

In recent years,in order to optimize the scheduling of limited highway resources,timely and accurate traffic flow prediction has become a very important technical area in intelligent transportation systems,which plays a vital role in relieving congestion on highways and urban roads.Currently,for traffic flow prediction sequences,the degree of feature extraction for both periodicity and spatio-temporal correlation can have a direct impact on the prediction results at the next moment,further affecting the accuracy and timeliness of the model.Therefore,how to extract feature information of traffic flow sequences in an integrated way is a critical problem that needs to be solved urgently in deep learning-based traffic flow prediction models.To address this problem,the following work is done in this paper.(1)A multi-temporal granularity fusion algorithm is proposed based on the complex spatiotemporal dependence and dynamics of the traffic flow data itself.The original input sequence is divided into three time-grained sequences of near-periodic,daily-periodic and weeklyperiodic according to different periodicity.And the prediction performance of each experimental model is influenced by different percentages of temporal granularity to obtain the optimal temporal granularity percentage parameter.Finally,the divided time granularity series are reorganized according to the optimal parameters to constitute a new time series data,so that the data not only has the original historical time window,but also has multiple time attributes.The experimental results show that the input time series data obtained by this method can retain the main time dependence of road traffic flow and remove the interference of redundant information,while the overall prediction performance of the experimental model can be improved.(2)Based on the structural principle analysis of Deep Long Short-Term Memory(DLSTM)and Autoencoder(AE),a Deep Long Short-Term Memory neural network model(DLSTM-AE)with Autoencoder as the framework is constructed.In addition to the improved LSTM model,an improved Rectified Adaptive Moment Estimation(RAdam)algorithm is introduced considering the overfitting and gradient disappearance problems of the deep network,which combines the advantages of Adaptive Momentum Estimation(Adam)and first-order momentum stochastic gradient descent(SGD with Momentum,SGDM)algorithms to accelerate the convergence of the model parameters during the training process.Simulation results on the Pe MS dataset show that the proposed model not only has significant advantages in prediction results,but also possesses better curve-fitting ability in terms of traffic flow periodicity.(3)With the advantage that multi-temporal granularity fusion algorithms can help traffic flow models further improve their prediction performance,this paper designs a DLSTM-AE traffic flow prediction model(M-DLSTM-AE)based on multi-temporal granularity fusion.In the fusion mechanism,the preliminary pre-processing of the input sequence by the multitemporal granularity fusion algorithm in advance reduces the complexity of time-dependent encoding of sequence information by the DLSTM module,while accelerating the feature extraction capability of the DLSTM model for sequence information at another level.Through simulation experiments with different real traffic datasets,the prediction performance of the MDLSTM-AE model proposed in this paper is better than the other models in the experiments.Meanwhile,the training time of M-DLSTM-AE model is reduced by nearly 48.6% compared with DLSTM-AE model under the best performance condition,which verifies that the MDLSTM-AE model not only has high prediction accuracy but also has good timeliness.