基于时空关联度加权的LSTM短时交通速度预测

A Weighted Long Short-Term Network Model for Short-Term Traffic Speed Prediction Based on Spatio-Temporal Correlation

提出一种基于时空关联度加权的长短期记忆网络(Long Short-Term Memory,LSTM)短时交通速度预测模型。该模型结合综合动态时间规整(Summation Dynamic Time Warping,SDTW)和拓扑邻接关系设计了一种路段速度序列之间时空关联程度的度量方法,然后基于该度量值对路段速度历史观测值进行加权,进而使用LSTM从加权观测序列中提取路段速度的时空变化特征,实现对短时交通速度的预测。实验表明,交通速度预测模型预测结果相比传统的ARIMA模型、SVR模型以及LSTM模型均有提升,实现了更高精度的速度预测。

This paper proposes a weighted long short-term network model for short-term traffic speed prediction based on spatio-temporal correlation.By integrating SDTW(Summation Dynamic Time Warping)with and topological adjacency,we design a method to quantify the spatio-temporal correlation between road speeds.A weight is assigned to each road based on historical values of road speeds and its spatio-tmeprpoal correlation value.The spatio-temporal variance of road speeds is investigated based on weighted observation values using LSTM model.This approach is used to Wuhan city.Results show that the weighted LSTM model has a higher speed prediction accuracy comparing to ARIMA model,SVR model and conventional LSTM model.