Research And Implementation Of Vehicle Trajectory Prediction Algorithm Based On Neural Network

With the continuous development of mobile internet technology,global positioning technology,the scale of trajectory data is continuously expanding.Based on the trajectory data,location-based services such as mobile location recommendation,trip planning,urban traffic congestion mitigation and road planning are also prospering.In any case,vehicle location prediction is the core and foundation of the above services.In response to this demand,this paper proposes a vehicle trajectory prediction method based on neural network by utilizing the GPS log data of urban vehicles.Through three steps including trajectory data preprocessing,sparse data completion based on K-Modes,LSTM prediction model base on GA's optimization,trajectory data quality improvement and vehicle position prediction are achieved,which provides stable and reliable underlying support for other location-based services.In this paper,after reviewing the status quo of the current vehicle trajectory prediction research,research is conducted on sparse trajectory data completion,time-series vehicle trajectory prediction,and LSTM initial parameter optimization,etc.Finally,by using the KModes clustering algorithm,the LSTM vehicle trajectory prediction model optimized by GA algorithm is finally determined.The main work of this paper includes the following aspects:(1)Data preprocessing and sparse data completion.The quality of trajectory data has a great influence on the accuracy of the final vehicle position prediction.However,existing research tends to focus on the trajectory prediction method,few people evaluate and improve the quality of historical data.In this paper,after studying the characteristics of the trajectory itself,duplicate trajectory points are filtered,and a set of vehicle trajectories is formed by slicing the original trajectory.Then,K-Modes clustering algorithm is conducted to achieve trajectory data clustering.Based on the results,the sparse trajectory data is complemented without the aid of third-party road network data,and the quality of vehicle trajectory data is further improved.(2)The LSTM model optimized by GA algorithm.For the vehicle trajectory,which has obvious temporal characteristics,this paper utilizes LSTM model to build the vehicle position prediction model.In the process of model building,the randomness of LSTM initial parameters leads to the poor convergence.The GA algorithm is used to search and optimize the initial parameters of the LSTM model globally,which can quickly shorten the convergence speed and improve the performance.(3)Establish a vehicle position prediction model and conduct experiments.Based on the initial set of parameters obtained from GA optimization experiments,a vehicle position prediction model is established and experiments are organized.Firstly,the experiment compares the number of total trajectory points and the average number of trajectory points in a single trajectory,which shows that sparse data complementing effectively improves the quality of trajectory data.Then,the optimization effect of GA optimization algorithm is given.It shows that the LSTM model optimized by the GA algorithm has better convergence performance.Finally,we compare the accuracy of position prediction before and after data completion.For the obvious error points,the error calibration is introduced to correct false predictions,which eventually increases the accuracy to 82.9%.Experimental results show that the vehicle trajectory prediction algorithm based on neural network presented in this paper is accurate,efficient and feasible.