公路工程

高精度的共享单车出行需求预测是精细化单车系统运营的关键，但骑行数据不易获取，且传统的研究也常常忽略交通需求变化的时间动态性和空间相关性。因此提出了一种借助注意力机制的深度时空域适应网络模型解决上述问题，命名为DTSA-GCN。首先，使用堆叠的3D时空图卷积层学习原始序列的数据表示，编码时空依赖，将源域和目标域嵌入到共同的潜在表示空间；其次，借助深度自适应网络(Deep Adaptive Networks, DAN)的思想，使用最大均值差异(Maximum Mean Difference, MMD)作为约束，学习两个域之间的可迁移特征；然后，通过注意力机制计算不同输入特征向量的权重；最后，使用一个全连接层对数据进行输出。通过公开的CitiBike数据集和NYCTaxi数据集的测试验证，结果表明，在60 min时间粒度划分下，所提出的预测模型得到3种均方根误差分别为0.711、0.542和0.046,相较于BP神经网络、长短期记忆神经网络(LSTM)、差分回归移动平均模型(ARIMA),均方根误差平均降低了61.7%,平均绝对误差平均降低了28.7%,平均绝对百分比误差平均降低了17.0%,证明DTSA-GCN模型能够通过有限的骑行数据表现出较好的预测效果，可以用作共享单车系统需求的预测模型。能够克服小样本数据对共享单车需求预测的局限，可为城市共享单车平衡调配提供技术参考。

High-precision forecast of shared bicycle travel demand is the key to refined bicycle system operation, but riding data is not easy to obtain, and traditional research often ignores the temporal dynamics and spatial correlation of traffic demand changes.Therefore, a deep time spatial adaptation network model with attention mechanism is proposed to solve the above problems, named DTSA-GCN.First, the data representation of the original sequence is learned using stacked 3D spatiotemporal convolutional layers, encoding spatiotemporal dependencies, and embedding the source domain and target domain into a common potential representation space.Secondly, using the idea of Deep Adaptive Networks(DAN), Maximum Mean Difference(MMD) is used as a constraint to learn the transferable features between two domains.Then, the weights of different input feature vectors are calculated by the attention mechanism.Finally, the data is output using a fully connected layer.Through the testing and verification of the publicly available CitiBike data set and NYCTaxi data set, the results show that under the 60 min time granularity division, the three root-mean-square errors of the proposed prediction model are 0.711, 0.542 and 0.046, respectively.Compared with BP neural network, LSTM network and differential regression moving average model(ARIMA), the average root-mean-square error decreases by 61.7%, average absolute error decreases by 28.7%, and average absolute percentage error decreases by 17.0%.It is proved that DTSA-GCN model can show better prediction effect through limited riding data, and can be used as a prediction model for shared bicycle system demand.This study can overcome the limitations of small sample data on the demand prediction of shared bicycles, and can provide technical reference for the balanced allocation of shared bicycles in cities.