EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Specification sensitivity in urban air-quality forecasting and simulated control

Sai Charan Addanki, Sunkara Harshitha, Ramit Debnath

Machine Learning: Earth · 01.12.2026 · 2 (2) · 025026

DOI: 10.1088/3049-4753/aeb6f2 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Machine Learning: Earth, online erschienen am 06.11.2026. Verfasst von Sai Charan Addanki und 2 weiteren Autor:innen. Die Arbeit nennt 57 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Model comparisons in urban air-quality forecasting can be sensitive to choices in data preparation and evaluation. We examine this sensitivity using a neural continuous-time model and nine baselines on 59 628 h of composite air quality index observations from Delhi, Mumbai and Kolkata. A separate traffic–pollution simulator examines reward-weight sensitivity in proximal policy optimisation (PPO). Five models achieve the lowest mean forecasting error across the six city–horizon combinations. Withholding the latest observation through a one-step lag-index shift increases the paired C-Mod/LSTM RMSE by 47.9%–76.2% at one hour, compared with a 1.6%–12.8% range among neural architectures. At 24 h, the corresponding lag penalties are 0.9%–4.6%, and architecture and training-seed variability also contribute substantially. In the control experiment, PPO recovers little intervention benefit at the lowest tested benefit–cost weighting, despite a beneficial constant action being available. Recovery improves at higher-benefit settings under all three tested dispersion closures; the transition location depends on the closure and training configuration. Synthetic meteorological and traffic templates restrict the forecasting comparison to the evaluated information set, while assumed simulator coefficients limit the control findings to simulation. The results support explicit reporting of lag availability, scoring rows, random-seed variability and reward weights, with conclusions qualified by forecast horizon and experimental setting.

Zitieren

APA

Addanki, S. C., Harshitha, S., & Debnath, R. (2026). Specification sensitivity in urban air-quality forecasting and simulated control. Machine Learning: Earth, 2(2), 025026. https://doi.org/10.1088/3049-4753/aeb6f2

Vancouver

Addanki SC, Harshitha S, Debnath R. Specification sensitivity in urban air-quality forecasting and simulated control. Machine Learning: Earth. 2026;2(2):025026. doi:10.1088/3049-4753/aeb6f2

BibTeX
@article{Addanki2026,
  author = {Addanki, Sai Charan and Harshitha, Sunkara and Debnath, Ramit},
  title = {Specification sensitivity in urban air-quality forecasting and simulated control},
  journal = {Machine Learning: Earth},
  year = {2026},
  volume = {2},
  number = {2},
  pages = {025026},
  doi = {10.1088/3049-4753/aeb6f2}
}
Kostenlos starten →