Multimodal Prediction Of Traffic Flow By Sparse Gaussian Process Mixture Model

At present,the imbalance between traffic demand and road resource supply is becoming increasingly obvious with the urban population growing year by year,traffic congestion has become an important factor restricting social development.Effective traffic flow prediction is the key to traffic control.Traffic flow types include shared bicycle flow,pedestrian flow and short-term car flow.Shared bicycle flow prediction can be used to guide the rational launch of bicycles and improve user’ experience.Pedestrian flow prediction is used to ease crowds and prevent accidents.Short-term car flow prediction is conducive to the efficient utilization of urban road network carrying capacity and the improvement of travel efficiency.However,different types of traffic have their own characteristics,and the traditional prediction model is unable to accurately analyze its change rules,the prediction is inaccurate and time-consuming.Therefore,the in-depth study of traffic flow prediction is of great practical significance.Gaussian process mixture(GPM)model is a combination of several Gaussian process(GP)models.If it is used for traffic flow prediction,it can depict slight changes of traffic flow sequence and realize time-segment multimodal prediction,which can effectively improve the prediction effect.On the basis of GPM,the improved sparse Gaussian process mixture(S-GPM)model reduces the computational complexity,further improves the prediction efficiency.This paper analyses the characteristics and rules of different types of traffic,and applies GPM and S-GPM models to traffic prediction with different characteristics including sharing bicycle flow,pedestrian flow and short-term car flow.The main contents are as follows:(1)multimodal prediction of shared bicycle traffic by multi-core Gaussian process mixture modelGPM is used to predict the shared bicycle traffic.The false nearest neighbor method is introduced to obtain the optimal embedding dimension.Hard-cut algorithm is used to train GPM parameters to enhance the prediction performance of the model.And the multi-core GPM model is constructed for multi-modal prediction by combining multiple kernel functions to fully fit the complex characteristics of shared bicycle traffic.The results show that the multi-core GPM can display the prediction results in different modes,and the root mean square error can be reduced to 0.0875.Compared with single-core GPM and traditional model,multi-core GPM has smaller prediction error than other models.(2)multimodal prediction of pedestrian flow and shared bicycle traffic by Sparse Gaussian process mixture modelFor large data volume traffic series,the training calculation of model is large and time-consuming.Therefore,sparse strategy is introduced into GPM model,that is,selecting representative samples for training,which reduces the computational complexity and builds S-GPM model,and is used for pedestrian and shared bicycle traffic prediction.The results show that the prediction time can be shortened to 10.4535 s and 12.3102 s.Comparing the prediction results of S-GPM with GPM(usually using Loo CV and Variational algorithm),S-GPM has better prediction accuracy and time-consuming than traditional GPM.(3)multimodal prediction of short-term car flow by Sparse Gaussian process mixture modelS-GPM is applied to short-term car flow prediction.Combining with the principle of multi-modal prediction,the predicting results are displayed in different modes,and the confidence interval of the predicting results is given,and the reliability of the predicting results is analyzed.The results show that the root mean square error of S-GPM can be reduced to 0.0476 and the time-consuming is only 8.6516 s.Compared with GPM and traditional model,S-GPM shows superiority in error,time-consuming and anti-noise.Through the analysis of the experimental results,it can be concluded that GPM has wide applicability,and S-GPM used in traffic prediction can improve the prediction accuracy,and shorten the time-consuming.This study provides a more accurate and reliable solution to the problem of traffic flow predicting.