基于趋势提取下多尺度时空残差网络的路段速度预测模型  

Roadway speed prediction model based on multiscale spatiotemporal residual network and detrending

作　　者：朱子俊 宋国华[1] 范鹏飞 黄健畅 ZHU Zi-jun;SONG Guo-hua;FAN Peng-fei;HUANG Jian-chang(School of Traffic and Transportation,Beijing Jiaotong University,Beijing 100044,China)

机构地区：[1]北京交通大学,交通运输学院,北京100044

出　　处：《交通运输工程与信息学报》2023年第3期74-85,共12页Journal of Transportation Engineering and Information

基　　金：中央高校基本科研业务费专项资金资助项目(2022YJS058);国家自然科学基金项目(71871015)。

摘　　要：速度预测模型在速度变异,即短时间、大幅度的速度变化场景中会出现模型失准的现象,为了获得更好的预测精度和对速度变异的鲁棒性,本研究建立了多尺度时空残差网络。时间上,通过提出的去趋势方法将速度分为趋势项和残差项,预测速度是预测残差和对应趋势项的直接加和。空间上,同时考虑了预测路段所在的局部路网和整体路网对预测的影响,构建了基于路段速度相关性的前馈神经网络模块和基于AlexNet的图像识别技术分别提取两种不同尺度的空间特征。预测路段的速度趋势、速度残差、局部路网特征和整体路网特征作为构建的编码器-解码器结构的输入以计算最终的预测速度。本研究提出了误差差异率指标来描述速度变异导致的模型失准程度,基于北京市路网浮动车数据,分析路段速度的日间和日内变化特征,选取具有不同速度变化特性的路段进行数据实验和模型验证。结果表明,模型不仅在整体测试集上具有出色的预测精度,在速度变异场景下具有更显著的性能优势。模型具有更好的鲁棒性,误差差异率相比于对照模型平均下降了约30%。Speed prediction models suffer from accuracy degradation in scenarios with rapid and significant speed variations.Thus,a multiscale spatiotemporal residual network was developed to increase the prediction accuracy and improve the robustness to this scenario.For the temporal features,the speed was divided into the trend term and the residual term via detrending,and the predicted speed was the sum of the predicted residual and the corresponding trend term.For the spatial features,the effects of both the subregional road network and the overall road network on the prediction were considered,and a feed for ward neural network module based on the roadway speed correlation was constructed.An AlexNet-based image recognition technique was used to extract spatial features at two different scales.The speed trends,speed residuals,subregional road network features,and overall road network features were input into the constructed encoder–decoder to predict the speed.The error variance rate was used to measure the model accuracy degradation due to rapid and significant speed variations.According to the floating car data of the road network of Beijing,the speed variability of the day and across days were analyzed,and data experiments and model validation were conducted for roadways among which the speed variability differed.The results indicated that in addition to superior prediction performance to benchmark models for the overall test set,the proposed model had superior performance in a scenario with rapid and significant speed variations.The error variance rate of the proposed model was reduced by approximately 30%on average compared with the benchmark models,implying better robustness to this scenario.

关 键 词：智能交通 短时交通流预测 深度学习 鲁棒性 去趋势 图像识别 

分 类 号：U495[交通运输工程—交通运输规划与管理]