基于CNN与多尺度特征融合的城市交通流预测模型

An Urban Traffic Flow Prediction Model Based on CNN and Multi-scale Feature Fusion

随着出租车和网约车的日益普及,GPS数据生成大量的时空视频流数据,对城市交通流预测提供坚实的数据价值。传统城市流量预测方法存在精度低,目标区域受周围区域影响等问题。卷积神经网络在交通流预测上表现出色,但仍存在目标区域受全局信息的干扰、低层网络的特征表征能力弱及高层下采样损失过多特征等问题。该文提出一种基于卷积神经网络(Convolutional Neural Network,CNN)与多尺度融合机制的交通流预测模型MS-RSCNet(Multi-scale Residual Self Checking Network)。该模型采用了一种残差自校验网络(Residual Self Checking Network,RSCNet)结构,并引入融合多尺度特征的双向门控循环单元设计方案。通过公开数据集对交通流预测性能进行测试验证,相较于ST-ResNet、ARIMA、STAR等模型,MS-RSCNet模型具有更优的交通流预测性能。

With the increasing popularity of taxis and online car hailing,GPS data generates a large number of spatio-temporal video stream data,which provides solid data value for urban traffic flow prediction.The traditional urban flow prediction method has some problems,such as low accuracy and target area affected by surrounding area.Convolutional neural network performs well in traffic flow prediction,but there are still some problems,such as the interference of global information in the target area,the weak feature representation ability of low-level network and too much loss of high-level down sampling.We propose a traffic flow prediction model MS-RSCNet(multi-scale residual self checking network)based on convolutional neural network(CNN)and multi-scale fusion mechanism.The model adopts a residual self checking network(RSCNet)structure and introduces a design scheme of two-way gating cycle unit integrating multi-scale features.The traffic flow prediction performance is tested and verified through the public data set.Compared with St RESNET,ARIMA,STAR and other models,MS-RSCNet model has better traffic flow prediction performance.