面向灾难性遗忘问题的交通流量预测持续学习技术研究
Study on the Continuous Learning Technology for Traffic Flow Prediction in the Face of Catastrophic Forgetting Problem
金涛斌 1陈仲岚 1刘苑 1郭朕宇 1祝瑞1
作者信息
- 1. 广州大湾区轨道交通产业投资集团有限公司,广东广州 510450
- 折叠
摘要
交通流量预测是建设智能交通管理系统的核心,目前基于图深度神经网络的交通流量预测算法已取得显著成效.然而传统的深度学习模型需要事先获得大量样本,使用离线训练得到的固定模型对未来发生的情况进行预测,当遇到未知情况时模型面临失效的风险.针对交通流量预测中流式数据对模型更新带来的挑战,以减轻模型灾难性遗忘、同时增强模型对新模式的学习能力为目标,研究并设计了一种基于最小化损失保持和图结构聚合保持的交通流量预测持续学习模型.该模型通过分别评估各参数对损失函数以及对图拓扑结构的贡献度提取重要参数,在面临新任务时通过对不同重要性的参数赋予不同权重系数、设计惩罚项更新损失函数,使得模型能够自适应地学习新模式,同时记住旧任务.在国际标准交通流量预测数据集上的测试表明,文章提出的方法有效缓解了流式时序序列预测的灾难性遗忘现象,并且对新任务具有良好的预测精度.
Abstract
Traffic flow prediction is at the core of building an intelligent traffic management system. At present,the traffic flow prediction algorithm based on a graph-deep neural network has achieved significant results. However,the conventional deep learning model needs to obtain a large number of samples and uses a fixed model obtained from off-line training to predict future circumstances,so the model has the risk of failure when encountering unknown circumstances. In face of the challenge brought to model updating by stream data in traffic flow prediction,the article aims to alleviate the catastrophic forgetting of the model and improve the ability to learn new patterns of the model. It studies and designs a continuous learning model for traffic flow prediction based on minimized loss preservation and graph structure aggregation preservation. This model extracts important parameters by separately evaluating the contribution of each parameter to the loss function and the graph topology structure. When facing new tasks,it assigns different weight coefficients to parameters of different importance and updates the loss function by designing penalty terms so that the model can adaptively learn new patterns and remember old tasks at the same time. The test on the international standard traffic flow prediction dataset shows that the proposed method effectively alleviates the catastrophic forgetting phenomenon in flow time series prediction and has good prediction accuracy for new tasks.
关键词
城市轨道交通/交通流量预测/持续学习/灾难性遗忘/损失函数Key words
urban rail transit/traffic flow prediction/continuous learning/catastrophic forgetting/loss function引用本文复制引用
出版年
2024
NETL
NSTL
维普
万方数据