A robust stacking model for predicting oil and natural gas consumption in China

Abstract

ABSTRACT Accurate prediction of oil and natural gas consumption (ONGC) is crucial for energy security and greenhouse gas emission control. This study uses machine learning to improve forecast accuracy by transforming time series predictions into supervised learning models. A novel stacking learning method, with added cross-validation, enhances model diversity and robustness. The key findings are: (1) The stacking model outperforms base models in predicting China’s ONGC. It achieves R2 scores of 94.44% for oil and 98.33% for natural gas, with corresponding RMSE scores of 0.5325 and 0.2919. (2) When comparing the scores of the models in the validation set using cross-validation, it can be observed that the stacking model exhibits the most consistent performance. (3) Through the diversification of models, the stacking approach enhances robustness and achieves better generalization on new datasets. The study provides fresh insights into model stacking for energy consumption prediction.