GPS数据驱动的车辆跟驰行为建模与分析

为研究道路交通流特性，基于车载高精度GPS跟驰试验数据进行车辆跟驰建模研究，结合深度学习理论和数据驱动方法，构建了基于粒子群优化(particle swarm optimization, PSO)的长短期记忆(long short term memory, LSTM)车辆跟驰模型。首先，清洗和平滑车辆轨迹数据，并对驾驶特征行为参数及相关关系进行研究，如加速度、车头时距以及速度与跟驰距离特性关系等。在此基础上，制定跟驰状态筛选规则；其次，构建考虑时间序列的PSO-LSTM模型，识别跟驰数据样本集，将当前时刻的前车速度、车头间距和上一时刻的车头时距作为模型输入，预测当前时刻的跟驰车速度；接着，选用25辆车跟驰试验的高精度GPS数据验证PSO-LSTM模型性能；最后，为验证该模型的优越性，选用传统机器学习SVR(support vector regression)模型以及深度学习LSTM模型作为对比。结果表明，基于粒子群优化的长短期记忆模型预测精度高达0.993，整体预测效果高于SVR模型和LSTM模型，其中预测误差指标MAPE(mean absolute percentage error)较SVR和LSTM分别降低了60.02%、1.52%。PSO算法进行超参数优化后的PSO-LSTM模型，能更好地模拟车辆的跟驰行为。

Modeling and Analysis of Car following Behavior Based on GPS Data

In order to study the characteristics of road traffic flow, a car-following model based on vehicle-mounted high-precision GPS car-following test data was studied. Combining the deep learning theory and data-driven method, a Long and Short Term Memory (LSTM) car-following model based on Particle Swarm Optimization (PSO) was constructed. Firstly, the vehicle track data was cleaned and smoothed, and the parameters related to driving behavior, such as acceleration, headway, and the relationship between speed and car-following distance, were examined. Car-following status screening rules were formulated based on this analysis; Secondly, a PSO-LSTM model considering time series was constructed to identify the sample set of car-following data. The current time''s preceding car speed, headway, and the preceding car''s headway at the previous time were utilized as inputs for the model to predict the car-following speed at the current time; Subsequently, the performance of the PSO-LSTM model was verified using high-precision GPS data from 25 car-following tests; Finally, traditional machine learning SVR (Support Vector Regression) model and deep learning LSTM model were selected for comparison in order to validate the superiority of the proposed model. The results demonstrate that the Long and Short Term Memory model based on particle swarm optimization achieves a prediction accuracy as high as 0.993, outperforming the SVR and LSTM models overall. Moreover, the prediction error index MAPE is reduced by 60.02% and 1.52% compared to SVR and LSTM, respectively. The PSO-LSTM model, optimized through the PSO algorithm for hyperparameter tuning, provides a better simulation of the car-following behavior of vehicles.