基于数据分解和深度强化学习的交通流预测方法  

Traffic Flow Prediction Method Based on Data Decomposition and Deep Reinforcement Learning

作　　者：刘嘉辉[1] 杜金 LIU Jiahui;DU Jin(School of Computer Science and Technology,Harbin University of Science and Technology,Harbin 150080,China)

机构地区：[1]哈尔滨理工大学计算机科学与技术学院,黑龙江哈尔滨150080

出　　处：《牡丹江师范学院学报（自然科学版）》2023年第3期28-34,共7页Journal of Mudanjiang Normal University：Natural Sciences Edition

基　　金：黑龙江省自然科学基金项目(LH2021F031)。

摘　　要：提出一种基于数据分解和深度强化学习(DRL)的交通流预测框架.为了减轻不规则波动的影响,利用局部加权回归时间序列分解方法将数据分解为趋势分量、季节分量和剩余分量.趋势分量由门控循环单元(GRU)训练,季节分量和剩余分量作为环境状态采用策略梯度算法和强化学习模型学习,根据门控循环单元网络的趋势预测结果,环境状态对预测结果进行及时调整.实验结果表明,本文提出的方法优于其他模型.A traffic flow prediction framework based on data decomposition and deep rein forcement learning(DRL)is proposed.In order to mitigate the impact of irregular fluctuations,the time series decomposition method of local weighted regression is applied to decompose data into trend component,seasonal component and residual component.The trend component is trained by gated recurrent unit(GRU),while the seasonal and residual component are trained as environmental states by the strategy gradient algorithm and reinforcement learning model,and the prediction results are timely adjusted according to the gated recurrent unit network's trend prediction results and environmental state.The experimental results show that the proposed method is superior to other models.

关 键 词：交通流预测 强化学习 时间序列 策略梯度 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]