Comparative study on monthly natural gas vehicle fuel consumption and industrial consumption using multi-hybrid forecast models

• Published in: Energy (Oxford) 2023-01, Vol.263, p.125826, Article 125826
• Main Author: Pala, Zeydin
• Format: Article
• Language: English
• Subjects: automobiles; comparative study; data collection; energy; energy use and consumption; Long-term forecasting; memory; Multi-hybrid models; natural gas; prediction; Time series; time series analysis; USA natural gas industrial consumption; USA natural gas vehicle fuel consumption
• ISSN: 0360-5442
• DOI: 10.1016/j.energy.2022.125826

Accurate natural gas consumption forecasting plays a significant role in production, supply, and dispatching. Therefore, in this study, a new multi-hybrid model methodology is proposed that combines both statistical and deep learning models to obtain better prediction results beyond individual models or constrained hybrid models in linear and non-linear modeling. Here, long-term natural gas consumption future forecast analyzes were performed for the USA natural gas vehicle fuel (NG-VFC) dataset from January 1997 to October 2021 and for the USA natural gas industrial consumption (NG-IC) dataset between January 2001 and October 2021. The values obtained as a result of the analyzes using multi-hybrid models based on statistical and deep learning models were evaluated with popular metric values such as mean absolute percentage error and mean absolute scaled error within reference measures. In all analyzes using NG-VFC and NG-IC time series, the best MAPE values were obtained as 5.40% and 3.19% for afnt (equi-weighted) and af (CV-weighted) multi-hybrid models, respectively. While the first of the equi-weighted and CV-weighted approaches featured here required less computation time, the latter required more computation time. In terms of prediction accuracy, the suggested multi-hybrid model outperforms most existing state-of-the-art approaches without sacrificing time or memory complexity. •A multi-hybrid model is proposed for linear and non-linear datasets.•All combinations of six different models in the proposed hybrid model is used.•The relationship between model collaboration and forecasting performance is revealed.•The proposed methodology overcomes shortcomings of deep learning in forecasting tasks.•Compared to alternatives, the proposed model's forecast accuracy is better than many ones.