【摘要】 针对交通流数据高维非线性和时空依赖性复杂，本文构建了基于特征蒸馏的变分贝叶斯编码器交通流预测模型.对每段时间序列对应的时间窗口特征，构建了基于多模态时间槽和空间槽的交通流特征提取模型.以时空槽特征提取模型作为特征知识蒸馏架构的输入.通过知识蒸馏结构提取的时空特征结晶体，利用教师模型指导学生模型的学习过程，从而提高学生模型的泛化能力.变分贝叶斯编码器对交通流时空特征结晶编码获取交通流数据的隐变量，根据隐变量的生成采样，利用解码器将其解码重构成新的预测值.实验结果表明，本文提出的模型预测性能显著提升，且中长期预测中鲁棒性更优.更多还原

【Abstract】 To improve the accuracy of traffic flow prediction and to solve the problems of high-dimensional nonlinearity and spatio-temporal dependence of traffic flow, a combined feature distillation and variational Bayes encoders traffic flow forecasting model(ST-DVBE) is proposed. First, to extract the time window characteristics corresponding to each time series, the multi-modal time slots and spatial slots are constructed. Second, with spatio-temporal slot feature extraction model as the input of feature knowledge distillation architecture, and space-time feature crystallization extracted by knowledge distillation structure, the learning process of student model is guided by teacher model, so as to improve the generalization ability of student model. Finally, the variational Bayesian encoder is employed to capture the latent variables of traffic flow data by encoding the crystallization of spatiotemporal features. Utilizing the generated latent variables, the decoder reconstructs them into new predicted values. Experimental results demonstrate a significant enhancement in predictive performance with the proposed model, especially with better robustness in mid-and long-term forecasting.更多还原