时空相关的道路网络短时交通流预测模型

A spatial-temporal dependent short-term traffic flow prediction model for road networks

为有效解决复杂路网短时交通流预测问题中涉及的时空特征挖掘问题,提出一种基于改进长短时记忆神经网络(Improved Long Short-Term Memory, ILSTM)的交通流预测模型.首先,通过改进的遗传算法对长短时记忆神经网络(Long Short-Term Memory, LSTM)模型初始参数进行优化获得最优参数组合,解决LSTM初始参数设置对输出结果影响较大的问题.其次,针对复杂路网多路段交通流预测中遇到的空间特征提取问题,通过挖掘相关路段对目标路段交通流预测的影响程度,重新构建LSTM模型的损失函数,采用路网中相关路段对目标路段的影响系数,以损失函数输出值最小为终止条件,构建ILSTM模型.最后,选择加州公路局交通数据进行模型验证实验,采用遗传算法优化LSTM模型(Genetic Algorithm-LSTM, GA-LSTM)和单纯LSTM模型,以及皮尔森相关系数与LSTM组合模型(Pearson Correlation Coefficient-LSTM,PCC-LSTM),对工作日和周末数据的多次实验结果进行对比分析.实验结果表明:ILSTM模型能够充分考虑复杂路网交通流的时间和空间特征,预测平均误差约为1.16%,在收敛效率和预测精度方面均优于其他模型.

To effectively address the spatio-temporal feature mining problem in short-term traffic flow prediction for complex road networks,a traffic flow prediction model based on the Improved Long Short-Term Memory(ILSTM)is proposed.Firstly,an improved genetic algorithm optimizes the initial parameters of the LSTM model,obtaining the optimal parameter combination and reducing the impact of initial parameter settings on output results.Secondly,to tackle the spatial feature extraction problem encountered in predicting multi-segment traffic flow in complex road networks,an ILSTM model is constructed by evaluating the influence of relevant road segments on the target road segment.This involves reconstructing the loss function of the LSTM model using the influence coefficients of relevant road segments in the network,and terminating the optimization when the loss function output reaches its minimum value.Finally,model validation experiments are conducted using traffic data from the California road network.The performance of the ILSTM model is compared against the Genetic Algorithm-LSTM(GA-LSTM)model,the standard LSTM model,and the Pearson Correlation Coefficient-LSTM(PCC-LSTM)model through multiple experiments with weekday and weekend data.The results demonstrate that the ILSTM model effectively captures the temporal and spatial characteristics of complex road network traffic flow,with an average prediction error of approximately 1.16%.The ILSTM model outperforms other models in terms of both convergence efficiency and prediction accuracy.