基于图卷积和注意力机制的高速公路交通流预测

Expressway Traffic Flow Forecast Based on Graph Convolution and Attention Mechanism

为了精确取得高速公路各断面交通流参数预测值,为交通管理控制提供依据,基于注意力机制的Seq2Seq模型提出了一种考虑时空分布的高速公路短时交通流多步预测方法。将预测断面和可能影响上下游路段的历史和当前时间序列数据作为输入,建立了基于“注意力机制-双向门控循环单元-序列对序列”和图卷积神经网络的高速公路交通流多步预测模型。采用图卷积提取了道路上下游对预测路段影响因素的空间特征,采用双向门控循环单元模型提取了时间临近性、日周期性、周周期性的时序特征。将提取的空间和时间特征加权融合值输入到门控循环单元模型,采用注意力机制获取了各特征值的权重,计算得到交通流参数多步预测值。以均方误差最小化为目标,采用Adam优化器梯度下降法进行迭代训练取得并更新了模型参数,以京港澳高速高碑店—定州南段为例进行试验分析。结果表明:预测精度随步数的增加而减少,但减少趋势缓慢,且3步内的精度均在合理范围内;预测精度随时间粒度的增加而减小,5~10 min时最佳;自由流状态指标值小于拥堵交通状态下指标值;与不考虑空间特征方法、图卷积注意力模型方法、Seq2Seq模型方法相比,本交通流多步预测方法预测精度最高,进一步证明了本预测模型更好地映射了内部的非线性关系,能够满足实时交通控制需要。

In order to obtain more accurately forecasted data of traffic flow for each section of expressway and then provide the basis for expressway management and control,a multi-step forecast method of short-time traffic flow considering time and space distribution is proposed based on attention mechanism Sequence to Sequence(Seq2Seq)network.Using the historical and real-time time series data of the forecast section,upstream and downstream sections that may affect the predicted section as inputs,a multi-step forecast model of traffic flow on expressway based on“attention mechanism,Bi-Gated Recurrent Unit(Bi-GRU)and Seq2Seq”and Graph Convolution Network(GCN)is established.The space features including the influence of upstream and downstream on the forecast section are extracted by GCN model and the time series features including time proximity,daily periodicity,and weekly periodicity are extracted by Bi-GRU model.The multi-step forecast values of traffic flow parameters can be computed by using GRU model,where the weighted fusion value of the above extracted features is used as the input of the model,and the weight of each feature value is obtained by attention mechanism.The model parameters can be obtained and updated by using Adam optimizer gradient descent method for iterative training,where the goal is to minimize mean square error.The proposed method is validated through the simulation analysis of Gaobeidian-Dingzhounan segment of Beijing-Hong Kong-Macao expressway.The result shows that(1)the forecast accuracy decreases with the increase of the number of steps,but the decreasing trend is small,and the accuracy is within a reasonable range within 3 steps by using the proposed method;(2)the forecast accuracy decreases with the increase of time granularity,and the best time is 5-10 min;(3)each indicator value in free flow state is smaller than the one in congested state;(4)compared with the method without considering spatial features,the method of using the attention model of graph convolution and the method of using the Seq2Seq model,the proposed method has the highest forecast accuracy,which shows that the internal nonlinear relationship can be better mapped and the demand of real-time traffic control can be met.