Adaptive Mixed Variation Tuning For Neural Network Traffic Flow Prediction

With the development of science and technology in our country to promote social progress,the scale of urbanization has expanded,the level of residents’ travel has improved,and the number of private vehicles on the road has increased.In addition,the road environment is also plagued by external factors such as weather and holidays,resulting in more complex and changeable road traffic,has strong nonlinear,time-varying and random characteristics,and the road is more prone to traffic jams and traffic accidents.To solve this problem,intelligent transportation has become a research hotspot of many scholars.Its excellent intelligence induction ability and road real-time analysis ability can provide car owners with better driving routes and reduce road driving time.Traffic flow prediction is a key research direction in intelligent transportation.Real-time and accurate traffic flow prediction can loosen roads and improve the efficiency of road vehicles.The swarm intelligence optimization algorithm optimizes the neural network model to approximate various complex functions with arbitrary precision,and has achieved good applications in various nonlinear forecasting problems including traffic flow forecasting.In order to improve the accuracy of traffic flow prediction,this paper proposes to use two neural networks as the traffic flow prediction model,and uses two different operators to improve the sparrow search algorithm to improve the local optimization ability.The neural network parameters are optimized,and two traffic flow prediction models with combined optimization are built.Finally,the proposed new model is used to conduct experiments on domestic and foreign data sets,respectively,to verify the effectiveness of the model.The main work and innovative achievements of this paper include the following three aspects:（1）In view of the fact that the original traffic flow data has stray noise,which will affect the actual effect of traffic flow prediction,it is proposed to use the Gaussian smoothing filter method to process the original traffic flow data,remove the road noise signal,and obtain a new time reconstruction sequence as the experimental data,which can effectively improve the prediction accuracy of the model.（2）A chaos-optimized adaptive neural network traffic flow prediction method is proposed.The Back Propagation Neural Network（BPNN）with nonlinear characteristics can be arbitrarily close to various nonlinear problems by setting the number of network layers.Aiming at the characteristics of poor global optimization ability of sparrow algorithm,Tent chaotic operator is used to disturb the initial position of sparrow algorithm,increase the diversity of the population,and increase the traversal uniformity and randomness of the search,so as to optimize the global optimization.The experimental results show that the proposed Tent chaotic sparrow algorithm combined with the BP neural network traffic flow prediction model improves the average prediction accuracy by0.52%-4.01% compared with the commonly used swarm algorithm optimization of the BP neural network method.The chaotic adaptive search optimization of the predictive model is helpful to improve the predictive ability of the model.（3）An extreme learning machine（ELM）traffic flow prediction method is proposed that integrates adaptive hybrid mutation tuning.In the iterative process of the sparrow algorithm,the adaptive t distribution and the optimal Gaussian tuning variation are performed on the flock respectively,which increases the population variability and improves the local optimization ability before the flock converges.Then,the algorithm after mutation optimization is combined with ELM to optimize the weight of ELM and establish a new adaptive traffic flow prediction model.Using three sets of data sets of three stations on a highway near an airport in the UK for prediction,The results show that the proposed ELM traffic flow prediction model with adaptive mixed variation tuning can obtain smaller generalization error and better prediction accuracy with less time consumption.