统计与数据科学学院知识库

  随着社会经济的不断发展，汽车走进了千家万户，汽车保有量得到了迅速的提升。汽车的普及表明社会经济的变革丰富了人们的生活，但由此带来的社会性问题也不容小觑。汽车保有量的迅速增长使得我国各大城市中汽车与停车位的配比严重失衡，同时，由于车位信息在供给端和需求端的不对称，很多停车场的空置率高达50%以上，这两种看似矛盾的现象造成了各大城市中广泛存在的交通拥挤问题，停车难题正困扰着每一个司机，给道路资源管理带来了巨大的压力。

  车位数量预测信息对于提高车位利用率、缓解交通拥堵具有重要意义，它可以平衡停车场车位占用数量的分布。然而目前的车位预测算法大多仅关注数据的时间相关性，对于空间相关性的研究相对较少。为了深入挖掘空间信息在车位数量预测中的重要作用，进一步提高车位预测算法的精度，本文提出了一种多视角图注意力门控单元网络（Multi-View Graph Attention Gating Unit Network，MGA-GRU）模型，它对门控循环单元神经网络（Gate Recurrent Unit，GRU）进行了改造，在原先只能用于提取时间信息的两个门控单元的基础上增加了两个图注意力卷积门控单元（GA），用于提取空间信息。图注意力卷积门控单元使得MGA-GRU模型能够同步的提取与处理时间特征和空间特征，减少了时空信息之间的相互干扰，其中所采用的注意力机制可以根据不同节点序列间的注意力系数来对邻居节点的信息进行聚合，更加科学合理。此外，MGA-GRU模型在构造关系图时，融合了多种空间信息构造多视角拓扑图，使空间特征更加全面与准确，进一步的提高了模型的预测效果。

  为了验证模型效果，我们在现实数据集兰州停车场数据集上进行了实验，证明了MGA-GRU模型不论是在短期还是长期车位预测效果上都优于大多数的现有模型，预测结果在多个指标上达到了最优。最后，本文还进行了实用性探索，为了能够将模型进行推广并获得更广泛的使用，本文以MGA-GRU模型为基础开发了车位预测应用，挖掘模型的实用价值。

  With the continuous development of social economy, automobiles have entered thousands of households, and the number of automobiles has been rapidly increased. The popularity of automobiles shows that social and economic changes have enriched people's lives, but the resulting social problems should not be underestimated. The rapid growth of car ownership has caused a serious imbalance in the ratio of cars and parking spaces in major cities in my country. At the same time, due to the asymmetry of parking space information on the supply side and the demand side, the vacancy rate of many parking lots is as high as 50%. Seemingly contradictory phenomena have caused widespread traffic congestion problems in major cities, and parking problems are plaguing every driver, bringing enormous pressure to road resource management.

   The prediction information of the number of parking spaces is of great significance for improving the utilization rate of parking spaces and alleviating traffic congestion. It can balance the distribution of the number of parking spaces occupied by the parking lot. However, most of the current parking space prediction algorithms only focus on the temporal correlation of data, and there are relatively few studies on spatial correlation. In order to deeply explore the important role of spatial information in the prediction of the number of parking spaces and further improve the accuracy of the parking space prediction algorithm, this paper proposes a Multi-View Graph Attention Gating Unit Network (MGA-GRU) Model, which transforms the Gate Recurrent Unit (GRU) neural network, and adds a graph attention convolution gate unit (GA) for extracting spatial information. The graph attention convolution gating unit enables the MGA-GRU model to simultaneously extract and process temporal features and spatial features, reducing the mutual interference between spatio-temporal information. The attention mechanism adopted can be based on the attention between different node sequences. It is more scientific and reasonable to aggregate the information of neighbor nodes by using the force coefficient. In addition, when the MGA-GRU model constructs the relationship graph, it integrates a variety of spatial information to construct a multi-view topological graph, which makes the spatial features more comprehensive and accurate, and further improves the prediction effect of the model.

   In order to verify the effect of the model, we conducted experiments on the real data set Lanzhou parking lot data set, which proved that the MGA-GRU model is superior to most of the existing models in both short-term and long-term parking space prediction, and the prediction results are in many reached the optimum in terms of indicators. Finally, this paper also conducts a practical exploration. In order to promote the model and obtain wider use, this paper develops a parking space prediction application based on the MGA-GRU model to tap the practical value of the model.