基于预插补阶段和生成对抗网络的空气质量缺失值插补方法研究

打开文本图片集

An Imputation Method for Missing Air Quality Data Based on Pre-ImputationStageandGenerative Adversarial Network

GAO DongdongLIU Haizhong （Lanzhou JiaotongUniversity，Lanzhou 73oo7O， China）

Abstract： [Purposes]To address the issue of missing data in the field of air quality，a missing value imputation method based on a pre-imputation stage and generative adversarial network is proposed. [Methods] First，samples with a low missng rate were selected and preprocessed using traditional imputation methods to obtain an imputed dataset.Then，a clustering algorithm was applied to the imputed dataset to generate pseudo-labels，and a classifier was trained using the imputed dataset and the pseudolabels.Finally，the pre-trained clasifier was used to constrain the generator during the training of the BGGAN model. [Findings] Experimental results demonstrate that the proposed method outperforms other imputation methods under three missing scenarios： Missing Completely at Random （MCAR），Temporal Continuous Missing （TCM），and Spatial Continuous Missing （SCM）. Particularly under a high missing rate of 70% ，compared with the GAIN model， the proposed model reduces the Root Mean Square Error （RMSE）， Mean Absolute Error （MAE），and Mean Absolute Percentage Error （MAPE） by 0.027 6， 0.015 5，and 0.O31 9，respectively，in the MCAR scenario.[Conclusions] This method significantly improves imputation performance across MCAR， TCM，and SCM missing scenarios and can be effectively

applied to missing value imputation in the field of air quality.

Keywords：missing value imputation;generative adversarial network; bidirectional long short term memory network; graph neural network

0 引言

在空气质量监测中，由于传感器故障、通信故障和数据传输错误等原因，往往会出现数据缺失的情况，损害了数据的可解释性，从而对空气质量的准确评估和预测产生一定的偏差[1]。（剩余6545字）