EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Explainable Artificial Intelligence for Three-Hour-Ahead Natural Gas Consumption Forecasting with Uncertainty Quantification

Dmytro Tymoshchuk, Grigorii Shymchuk, Iaroslav Lytvynenko, Serhii Lupenko, Mariusz Rząsa

Applied Sciences · 11.11.2026 · 16 (22) · 11175

DOI: 10.3390/app162211175 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Applied Sciences, online erschienen am 11.11.2026. Verfasst von Dmytro Tymoshchuk und 4 weiteren Autor:innen. Die Arbeit nennt 45 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

This study investigates short-term natural gas consumption forecasting in the city of Ternopil and the contribution of different feature groups to forecasting accuracy. Synchronized natural gas consumption and meteorological observations from 2007 to 2019 were used. A total of 203 calendar, meteorological, and derived features were constructed. Six machine learning models, three naive baseline forecasts, and a seasonal ETS model were compared. Hyperparameters were optimized using Optuna, model performance was evaluated using six temporal folds with an expanding training window, and final testing was conducted on data from 2019. XGBoost achieved the lowest mean root mean squared error of 4307.13 ± 280.29 m3 across the cross-validation folds and a root mean squared error of 6517.29 m3 on the test set. No statistically significant difference was found between XGBoost and Random Forest, although both models outperformed the baseline forecasts. Feature ablation analysis showed that adding historical gas consumption features to calendar features reduced the mean root mean squared error by 81.42%. Adding meteorological features to a feature set that already included consumption history further reduced this error by 0.51%. Among the features evaluated for XGBoost, gas consumption 24 h before the target interval had the highest mean absolute SHAP value. The adaptive online conformal method produced prediction intervals that contained the observed consumption in 89.32% of cases, compared with a nominal coverage level of 90%. However, these intervals had the largest mean width among the prediction interval methods considered. The findings are limited to forecasting for a single city and one test year.

Zitieren

APA

Tymoshchuk, D., Shymchuk, G., Lytvynenko, I., Lupenko, S., & Rząsa, M. (2026). Explainable Artificial Intelligence for Three-Hour-Ahead Natural Gas Consumption Forecasting with Uncertainty Quantification. Applied Sciences, 16(22), 11175. https://doi.org/10.3390/app162211175

Vancouver

Tymoshchuk D, Shymchuk G, Lytvynenko I, Lupenko S, Rząsa M. Explainable Artificial Intelligence for Three-Hour-Ahead Natural Gas Consumption Forecasting with Uncertainty Quantification. Applied Sciences. 2026;16(22):11175. doi:10.3390/app162211175

BibTeX
@article{Tymoshchuk2026,
  author = {Tymoshchuk, Dmytro and Shymchuk, Grigorii and Lytvynenko, Iaroslav and Lupenko, Serhii and Rząsa, Mariusz},
  title = {Explainable Artificial Intelligence for Three-Hour-Ahead Natural Gas Consumption Forecasting with Uncertainty Quantification},
  journal = {Applied Sciences},
  year = {2026},
  volume = {16},
  number = {22},
  pages = {11175},
  doi = {10.3390/app162211175}
}
Kostenlos starten →