The invention provides an interpretable space-time analysis method for traffic jam prediction. Key features causing a jam event and deep relation between roads are extracted from interpretation. A traditional data mining method often explores the correlation relation between traffic spatio-temporal data from the perspective of statistics, and the deep relation and key factors of traffic congestion are difficult to fully reveal. Therefore, the STGCN-based spatio-temporal interpretation generation model is provided, the characteristic that the neural network is good at discovering hidden features is utilized, and key input features concerned by the neural network are extracted by using an interpretability technology of deep learning. The model generates a mask by using a disturbance-based interpretation method, and generates gradient mapping of the mask by using a gradient-based interpretation method; and in view of the problems of coarse airspace mask granularity and poor pertinence, a step-by-step mask method is provided to reduce the interpretation granularity. Thus, effective extraction of hidden information is increased, and more accurate and comprehensive congestion key information is obtained.
本发明提供一种面向交通拥堵预测的可解释性时空分析方法,从解释中提取得到引发拥堵事件的关键特征和道路间的深层联系。传统的数据挖掘方法往往从统计学角度探索交通时空数据间的相关性联系,难以充分揭示交通拥堵的深层联系和关键因素。因此本发明提出基于STGCN的时空解释生成模型,利用神经网络善于发现隐藏特征的特点,使用深度学习的可解释性技术提取神经网络关注的关键输入特征。模型使用基于扰动的解释方法生成掩膜(mask),使用基于梯度的解释方法生成掩膜的梯度映射;又鉴于空域掩膜粒度粗、针对性差的问题,提出分步掩膜方法降低解释粒度。如此,增加了对隐藏信息的有效提取,从而获得了更加准确全面的拥堵关键信息。
Interpretable space-time analysis method for traffic jam prediction
面向交通拥堵预测的可解释性时空分析方法
2022-12-20
Patent
Elektronische Ressource
Chinesisch
Interpretable traffic flow prediction stacking method
Europäisches Patentamt | 2024
|Traffic flow prediction method considering space-time similarity
Europäisches Patentamt | 2023
|