EUVIMEDEUVIMED entdecken
Crossref

Data-dependent Performance of Machine Learning Calibration for Low-Cost Indoor Particle Sensors

Seongjun Park, Kwangbok Jeong, Donghyun Rim, Jaehong Lee, Jaewook Lee

Environmental Pollution · 2027-01 · 412 · 129522

DOI: 10.1016/j.envpol.2026.129522 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Environmental Pollution, 2027-01. Verfasst von Seongjun Park und 4 weiteren Autor:innen. Beteiligte Einrichtung: Department of Architectural Engineering, Dankook University, Yongin-si, Gyeonggi-do, 16890, Republic of Korea. Electronic address: psj@dankook.ac.kr. Die Arbeit nennt 64 Literaturangaben. Gefördert von: Korea Ministry of Education.

AbstractSchlagwörterBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Low-cost particle sensors are increasingly used for high-resolution indoor air quality monitoring; however, their reliability is affected by source-dependent variability and measurement uncertainty. Although machine learning (ML) models have been widely applied for sensor calibration, their advantages over conventional linear approaches remain inconsistent, particularly in indoor environments. This study investigates how training-data volume and source heterogeneity influence the relative performance of ML-based and linear calibration for indoor particle monitoring. Controlled experiments were conducted in a full-scale residential testbed under three representative particle-source conditions. Linear regression (LR) and three ML models were evaluated under intra-source, cross-source, and multi-source training scenarios. Incorporating particle number concentrations improves calibration performance, increasing R2 by 0.05-0.10 and reducing normalized errors by up to 25%. Among the nonlinear models, LightGBM shows the strongest overall performance and benefits more from increased training-data volume. When training-data volume is held constant, increasing source heterogeneity generally reduces the normalized errors and prediction bias of LightGBM, although it does not improve R2 or enable LightGBM to outperform LR. With further increases in training-data volume under multi-source conditions, LightGBM provides greater benefits for selected error- and bias-related metrics, whereas LR retains higher R2. These findings demonstrate that the relative benefit of ML-based calibration depends on both training-data characteristics and the performance metric considered, rather than reflecting universal superiority over linear calibration.

Schlagwörter

Stichwörter der Autor:innen

Indoor Air QualityLight Gradient Boosting Machine (LightGBM)Linear RegressionParticle EmissionPurpleAir

Autor:innen und Institutionen

  • Department of Architectural Engineering, Dankook University, Yongin-si, Gyeonggi-do, 16890, Republic of Korea. Electronic address: psj@dankook.ac.kr.
  • Deep Learning Architecture Research Center, Department of Architectural Engineering, Sejong University, 209, Neungdong-ro, Gwangjin-gu, Seoul, 05006, Republic of Korea. Electronic address: kbjeong7@sejong.ac.kr.
  • Department of Architectural Engineering, Pennsylvania State University, University Park, PA, USA. Electronic address: dxr51@psu.edu.
  • Deep Learning Architecture Research Center, Department of Architectural Engineering, Sejong University, 209, Neungdong-ro, Gwangjin-gu, Seoul, 05006, Republic of Korea. Electronic address: jhlee@sejong.ac.kr.
  • Deep Learning Architecture Research Center, Department of Architectural Engineering, Sejong University, 209, Neungdong-ro, Gwangjin-gu, Seoul, 05006, Republic of Korea. Electronic address: jaewook@sejong.ac.kr.

Förderung

  • Korea Ministry of Education

Zitieren

APA

Park, S., Jeong, K., Rim, D., Lee, J., & Lee, J. (2027). Data-dependent Performance of Machine Learning Calibration for Low-Cost Indoor Particle Sensors. Environmental Pollution, 412, 129522. https://doi.org/10.1016/j.envpol.2026.129522

Vancouver

Park S, Jeong K, Rim D, Lee J, Lee J. Data-dependent Performance of Machine Learning Calibration for Low-Cost Indoor Particle Sensors. Environmental Pollution. 2027;412:129522. doi:10.1016/j.envpol.2026.129522

BibTeX
@article{Park2027,
  author = {Park, Seongjun and Jeong, Kwangbok and Rim, Donghyun and Lee, Jaehong and Lee, Jaewook},
  title = {Data-dependent Performance of Machine Learning Calibration for Low-Cost Indoor Particle Sensors},
  journal = {Environmental Pollution},
  year = {2027},
  volume = {412},
  pages = {129522},
  doi = {10.1016/j.envpol.2026.129522}
}
Kostenlos starten →