Abstract:

This paper proposes a machine learning method to predict curbside parking occupancy in dynamic parking policies. We apply the convolutional long short term memory neural network (ConvLSTM) to learn the temporal and spatial features of the data simultaneously. Based on the 4.92 million transaction records of parking meters in San Francisco, we train a policy model that incorporates the information of dynamic pricing and parking limits, and a non- policy model without other information. The results show that both the policy and non- policy model can predict the curbside parking occupancy, and the policy model has better performance in training efficiency and prediction accuracy. Meanwhile, the errors of the non-policy model increase with different parking policies, whereas the policy model can always obtain high prediction accuracy.

Key words: intelligent transportation, curbside parking occupancy prediction, deep learning, long short term memory, convolutional neural networks

摘要：

城市停车已逐步实现信息化和动态化管理，本文对动态管理模式下大范围路侧泊位占有率预测方法进行研究.在收集美国旧金山492万条停车交易数据的基础上，利用可同时提取数据空间关联和时序趋势特征的卷积长短时记忆神经网络(Convolutional LSTM Network，ConvLSTM)，分别构建考虑停车费率和时限动态变化的有政策模型，和没有动态管理信息输入的无政策模型.结果显示，有政策模型的训练效率和预测精度会显著提升.在政策平稳阶段，两种模型均能够有效预测泊位占有率；在政策发生变化时段，无政策模型的预测误差出现激增，但有政策模型的预测误差依然保持平稳，表明本文提出的方法能够很好地应对动态管理模式下停车需求的变化.

关键词: 智能交通, 停车泊位占有率预测, 深度学习, 长短期记忆单元, 卷积神经网络