In recent years, with the increasing impact of climate change, achieving the "dual carbon" goal has become increasingly urgent. Fossil fuel consumption is one of the main sources of carbon emissions and is closely related to economic development. Therefore, it is necessary to decouple economic growth from fossil fuel consumption as soon as possible. However, the small sample defect in fossil energy consumption panel data has caused great difficulties for machine learning algorithms. To address this defect, this study uses TimeGAN for data augmentation and compares the performance of six regression analysis methods (XGBoost, CatBoost, LGBM, KNN, linear regression, decision tree) in predicting fossil energy consumption. In addition, feature importance and SHAP are used simultaneously to explore the main driving factors of fossil energy consumption. The research results show that the decoupling index of most provinces fluctuated relatively little from 2011 to 2018. Compared to other machine learning algorithms, XGBoost and CatBoost perform better in predicting fossil energy consumption. Finally, based on two interpretable analysis methods, we found that population has the most significant impact on fossil energy consumption, followed by GDP, and urbanization rate is the least important. The research results provide important information for formulating energy sustainable development strategies and further discuss energy-saving solutions based on data methods.