基于模块化交通流组合预测模型

Traffic Flow Combination Prediction Model Based on Modularization

短时交通流预测是智能交通系统的核心能力组件之一,为城市交通管理、交通控制和交通引导提供智能决策支撑。针对交通路网交通流呈现的非线性、动态性和时序相关性,提出一种基于模块化的交通流组合预测模型ICEEMDAN-ISSA-BiGRU。采用改进的基于完全自适应噪声集合经验模态分解(Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise,ICEEMDAN)方法对交通流非线性时间序列进行分解,获取本征模态分量(Intrinsic Mode Functions,IMF);利用双向门控循环单元(Bi-directional Gate Recurrent Unit,BiGRU)挖掘交通流量序列中的时序相关性特征;基于动态自适应t分布变异方法改进麻雀搜索算法(Improved Sparrow Search Algorithm,ISSA),实现对BiGRU网络权值参数的迭代寻优,避免了短时预测结果陷入局部最优;基于公开PeMS数据集对短时交通流预测性能进行性能评估与验证。实验结果表明,所提组合模型的短时交通流预测性能优于10个传统模型,改进后的交通流量预测平均绝对误差(Mean Absolute Error,MAE)指标接近10.98,平均绝对值百分误差(Mean Absolute Percentage Error,MAPE)指标接近10.12%,均方根误差(Root Mean Square Error,RMSE)指标接近12.42,且在不同数据集下所提模型具有较好的泛化性能。

Short-term traffic flow prediction is one of the core competence components of intelligent transportation system,which provides intelligent decision support for urban traffic management,traffic control and traffic guidance.In this paper,Iceemdan-Isa-Bigru,a modular combined traffic flow prediction model,is proposed based on the nonlinear,dynamic and temporal correlation of traffic flow in the traffic network.Firstly,an Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise(ICEEMDAN)method based on fully adaptive noise set empirical Mode decomposition was used to decompose the nonlinear time series of traffic flows and obtain the intrinsic mode component.Secondly,bidirectional gated cycle unit is used to explore temporal correlation characteristics of traffic flow sequence.Then,based on the dynamic adaptive distributed variation method,the SSA algorithm is improved to achieve iterative optimization of the weight parameters of Bi-directional Gate Recurrent Unit(BiGRU)network,which avoids short-term prediction results falling into local optimal.Finally,the short-time traffic flow prediction performance is evaluated and verified based on open PeMS data set.Experimental results show that the short-time traffic flow prediction performance of the proposed combined model is better than that of the 10 traditional models.The improved Mean Absolute Error(MAE)index is close to 10.98,the Mean Absolute Percentage Error(MAPE)index is close to 10.12%,and the Root Mean Square Error(PMSE)index is close to 12.42.The proposed model has better generalization performance under different data sets.