融合自注意力机制对机场低能见度预测研究  

Research on Low Visibility Predicting of Airports by Integrating Self-attention Mechanism

作　　者：袁敏[1] 李忠堃 YUAN Min;LI Zhongkun(College of Air Traffic Management,Civil Aviation Flight University of China,Guanghan 618307)

机构地区：[1]中国民用航空飞行学院空中交通管理学院,广汉618307

出　　处：《舰船电子工程》2025年第2期45-51,共7页Ship Electronic Engineering

摘　　要：针对现有的机场低能见度预测耗时长、费力多且精度低的问题,论文提出融合自注意力机制的神经网络预测模型,以2003年-2023年4月近20年重庆江北国际机场的逐时报文数据分区间作为输入,实现对未来不同步长的低能见度预测。所构建的模型在传统神经网络RNN、LSTM和GRU框架下分别融合自注意力机制(Self-attention)和ProbSparse(概率稀疏)自注意力机制,提取输入数据序列的信息并对影响低能见度的因子之间复杂非线性关系建模。最后,选用平均绝对误差(MAE)、均方根误差(RMSE)和对称平均绝对百分比误差(SMAPE)作为上述九个模型的评价指标。结果表明:在低能见度两个区间内,融合自注意力机制后的预测模型MAE、RMSE、SMAPE分别降低7.3%、1.86%、4.6%以上,为机场再发生低能见度天气时,可根据要输出步长的不同来选择相应的网络模型,实现较为准确的预测。In response to the problems of time-consuming,labor-intensive,and low accuracy in predicting low visibility at existing airports,this paper proposes a neural network prediction model that integrates self-attention mechanism.The hourly message data partitions of Chongqing Jiangbei International Airport from April 2003 to April 2023 for the past 20 years are used as inputs to achieve low visibility prediction for different step sizes in the future.Under the framework of the traditional neural network RNN,LSTM and GRU,the constructed model integrates the self-attention mechanism(self-attention)and ProbSparse(Probability Sparse)self-attention mechanism respectively,extracts the information of the input data sequence and models the complex nonlinear relationship between factors affecting low visibility.Finally,the average absolute error(MAE),the root mean square error(RMSE)and symmetric mean absolute percentage error(SMAPE)are used as evaluation indicators for the nine models mentioned above.The results show that in the two low visibility intervals,the prediction models MAE,RMSE,and SMAPE fused with self-attention mechanism have decreased by more than 7.3%,1.86%,and 4.6%,respectively.When low visibility weather occurs again at the airport,corresponding network models can be selected according to the different output step sizes to achieve more accurate predictions.

关 键 词：低能见度预测 自注意力机制 概率稀疏自注意力机制 循环神经网络 机场安全运行 

分 类 号：TP393[自动化与计算机技术—计算机应用技术] F562.3[自动化与计算机技术—计算机科学与技术]