Carbon dioxide (CO2) emission forecasting is crucial for efficient carbon reduction management. The majority of carbon emission prediction models are developed based on limited data, which are often collected annually and spatially sparse, and hence face the problems of overfitting and low robustness. Aiming at reliable estimation of CO2 emissions with a small-scale of data, we propose a CO2 prediction framework that incorporates a hybrid predictor with feature selection. The hybrid predictor, formed by a fractional-order Grey multivariate model (FGM) and an ensemble learning model, XGBoost, can capture both linear and nonlinear variations of CO2 emissions, demonstrating strong predictive ability. ReliefF is used for feature selection due to its ability to balance features’ importance and diversity, which helps reduce model overfitting. The forecasting effect of the proposed framework is validated on the county-level CO2 emissions in Shanxi Province, China, from 2012-2022. The results show that the proposed model is superior to other linear and machine learning prediction models and achieves a good forecasting effect, with RMSE, MAE, and R2 values of 1.73, 1.03, and 0.93, respectively. The Likelihood Ratio (LR) test, the soundness test, and the heterogeneity test have confirmed the generalizability and stability of our proposed hybrid model for CO2 emissions predictions, as well as the effectiveness of feature selection. Consequently, the prediction results of Shanxi’s CO2 emissions provide a reliable basis for spatial correlation analysis using Moran’s Index.