高速公路突发事件恢复重建期交通量预测
Traffic flow prediction of expressway traffic emergency recovery and reconstruction period

为了提升高速公路突发事件应急管理能力，为突发事件恢复与重建施工及路网拥堵状况预判提供依据，提出一种高速公路突发事件恢复重建期路网交通量预测方法，利用扩散卷积和序列到序列学习框架模型，结合预定采样技术捕捉时间序列时空相关性，以高速公路联网收费数据为基础，建立路网交通量分配模型，实现对交通量时间精度下的分配；利用扩散卷积递归神经网络（DCRNN）构建高速公路路网交通量预测模型，运用扩散卷积运算来捕捉交通量的空间相关性，并使用预定采样编码器？步饴肫鹘峁褂行Ы饩鼋煌？量的时间依赖问题，模型将交通的空间性建模为有向图上的扩散过程，而不是传统的网格划分；并选取自回归滑动平均模型（ARIMA）模型和机器学习BP神经网络模型对模型的准确性及有效性进行了验证。研究结果表明：将河北省石家庄高速公路路网发生突发事件后15 d交通量数据做训练集，后7 d数据进行验证，迭代60次时，测算每15 min间隔内的路网交通量的模型精度达到0.95。提出的模型预测的平均绝对误差（MAE）和平均绝对百分比误差（MAPE）更低，能够有效弥补单一化神经网络预测模型仅能做出时序性预测的不足，可显著提高预测结果的精确性和实用性。
In order to improve the capacity of emergency management of expressway emergencies, provides references and help for emergency recovery and reconstruction or construction for prediction congestion. A method for predicting the traffic volume of an expressway road network during the recovery and reconstruction period after an expressway emergency was proposed, based on previous research results. Based on the diffusion convolution and a sequence？？to？？sequence learning framework, the proposed method captured the space？？time correlation of a time series using a sampling technique. With this method, the distribution of the traffic volume time？？accuracy was first realized by establishing a road network traffic volume distribution model based on highway network charging data, and the expressway road network traffic volume predictive model was then constructed using a diffusion convolution recurrent neural network (DCRNN). The DCRNN model was used to capture the spatial correlation of the traffic volume, and effectively solved the time？？dependent problem of the traffic volume by using the predetermined sampling encoder？？decoder structure. The advantage of the model lies in modeling the spatial property of traffic as a directed graph diffusion process rather than as a traditional grid division, which clearly describes the randomness of the traffic dynamics. To verify the accuracy and validity of the model further, the widely used ARIMA model and the machine learning BP neural network model were selected to calculate and compare the same instance data. The results show that the first 15 days of data in the traffic volume after an emergency of a highway network in Hebei Province Shijiazhuang is used as the training set, and the data after 7 days is verified. When the number of iterations reached 60, the accuracy of the model that calculate the traffic volume per 15 minutes interval reached 0.95. The forecasting method has a lower mean absolute error (MAE) and mean absolute percentage error (MAPE), which can effectively compensate for the defects of the unitized