基于注意力机制的泊位占有率预测模型研究  

Research on berth occupancy prediction model based on attention mechanism

作　　者：王竹荣[1] 薛伟 牛亚邦 崔颖安[1] 孙钦东[1] 黑新宏[1] WANG Zhurong;XUE Wei;NIU Yabang;CUI Ying’an;SUN Qindong;HEI Xinhong(School of Computer Science and Engineering,Xi’an University of Technology,Xi’an 710048,China)

机构地区：[1]西安理工大学计算机科学与工程学院,陕西西安710048

出　　处：《通信学报》2020年第12期182-192,共11页Journal on Communications

基　　金：国家重点研发计划基金资助项目(No.2018YFB1201500);国家自然科学基金资助项目(No.61773313);陕西省重点研发计划基金资助项目(No.2017ZDXM-GY-098);陕西省教育厅重点实验室基金资助项目(No.17JS100)。

摘　　要：为解决泊位占有率的预测精度随步长增加而下降的问题,提出了一种基于注意力机制的泊位占有率预测模型。通过卷积神经网络获得多变量的时间模式信息作为模型的注意力机制。通过对模型训练、学习特征信息,并对相关性高的序列分配较大的学习权重,来实现解码器输出高度相关的有用特征预测目标序列。应用多个停车场数据集对模型进行测试,测试结果及对比分析表明,所提模型在步长达到36时对泊位占有率的预测数据能较好地估计真实值,预测精度和稳定性相比LSTM均有提高。To solve the problem that the berth occupancy prediction accuracy decreases while the prediction step was increasing,a berth occupancy prediction model based on an attention mechanism was proposed,which was the multivariate time pattern information obtained by convolutional neural networks(CNN).The characteristic information was learned by the model training,and the sequence with higher correlation was assigned a larger learning weight,so that the highly correlated features output from the decoder could be used to predict the target sequence.Data sets of multiple parking lot were adopted to test the model.The test results show that the proposed model can estimate the real value well when the step length of berth occupancy prediction reaches 36.The prediction accuracy and stability of the model are improved compared with long short-term memory(LSTM)model.

关 键 词：时间序列预测 泊位占有率预测 注意力机制 序列到序列模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]