Carbon Emission Prediction: A Classification and Evaluation of Traditional and Data-Driven Methods

Accurate carbon emission forecasts are extremely important for climate policy formulation and carbon emission assessments. This paper divides prediction technologies into three categories: traditional statistical models, machine learning and deep learning models, and hybrid frameworks. Existing models such as STIRPAT, ARIMA and GM(1,1) are highly explanatory but difficult to deal with nonlinear mechanisms or data analysis. On the contrary, data-driven methods represented by Long Short-Term Memory (LSTM) have high prediction accuracy but usually lack transparency. The hybrid approach combines grey neural networks with the Economic Complexity Index (ECI) and Green Complexity Index (GCI) to improve prediction accuracy and overall performance and transcend the constraints of a single country's context. Interpretability tools such as SHAP and LIME are increasingly used to clarify model logical transparency and support strategy development. In short, integrating interpretability, multi-source data fusion and policy-oriented scenario simulation are of great significance to improving the scientificity and practicality of carbon emission prediction tools.