The invention discloses a cross-mode traffic flow prediction method based on domain adaptation. The method comprises the steps that data of a source city and data of a target city are acquired and preprocessed; dividing a city into grid regions with equal sizes, and constructing city data tensors with corresponding sizes and structures; performing data enhancement on the data in the grids; pre-training a deep network prediction model based on the source city data; performing multi-layer data distillation by calculating spatial feature correlation of the source city and the target city; similar region matching is carried out by calculating the correlation of the flow characteristics of the source city and the target city; and carrying out shared knowledge migration by optimizing the objective function and training a final prediction model. According to the method, shared knowledge can be learned and migrated from a source city with rich data more fully, and the method still has a good prediction effect for cities with different modes or scarce traffic flow data.
本发明公开了一种基于域适应的跨模式交通流量预测方法,包括:获取源城市和目标城市数据并进行预处理;将城市划分成等大小的网格区域,并构建相应大小结构的城市数据张量;对网格中的数据进行数据增强;基于源城市数据预训练深度网络预测模型;通过计算源城市和目标城市的空间特征相关性进行多层数据蒸馏;通过计算源城市和目标城市的流量特征的相关性进行相似区域匹配;通过优化目标函数进行共享知识迁移并训练最终的预测模型。本发明方法能够更加充分地从数据丰富的源城市学习共享知识并进行迁移,对于模式不同的或交通流量数据稀缺的城市,仍具有良好的预测效果。
Cross-mode traffic flow prediction method based on domain adaptation
基于域适应的跨模式交通流量预测方法
2024-06-11
Patent
Electronic Resource
Chinese