Evaluating deep learning data imputation for subway indoor air quality: Accuracy, efficiency, and implications for downstream tasks

Missing data is a pervasive issue in environmental and building systems and poses significant challenges for indoor air quality (IAQ) management in subway stations, where sensor readings directly guide ventilation control, exposure assessment, and risk modeling. This paper evaluates a range of imputation techniques for subway IAQ time-series, comparing their reconstruction accuracy, computational efficiency, and impact on downstream tasks. Twelve models are assessed across twelve distinct missing data scenarios that vary in pattern and severity. The methods include nine advanced deep learning approaches covering sequence and attention models, probabilistic generative frameworks, graph aware techniques and spectral or convolutional architectures and three naive statistical methods. Experiments use a newly collected dataset from Yeongtong station in the Seoul metro system giving real data insights into IAQ imputation challenges. To our knowledge this is the first comprehensive evaluation linking algorithmic performance to both ventilation control implications and forecasting outcomes in a multi-zone IAQ monitoring context. This study demonstrates the importance of aligning imputation strategies with specific missingness characteristics and forecasting requirements by means of reconstruction accuracy, which improved downstream IAQ task performances of forecasting and ventilation control. All implementation details are publicly accessible.¹