This paper proposes a method for early dynamic reserves prediction of carbonate gas wells based on machine learning algorithms. To achieve accurate prediction of dynamic reserves, fundamental parameters of operational carbonate gas wells in the Sichuan Basin were utilized to construct average single-well dynamic reserve prediction models using Least Squares Support Vector Machine (LSSVM), Artificial Neural Network (ANN), and Light Gradient Boosting Machine (LightGBM). The results indicate that the LightGBM algorithm performs excellently in average single-well dynamic reserves prediction, with R2 values of 0.894 and 0.867 on the training and test sets, respectively, significantly outperforming the other two algorithms. LSSVM and ANN models also exhibit high prediction accuracy but are slightly inferior to LightGBM in terms of stability and generalization ability. This study provides crucial technical support and theoretical basis for the early development and rational deployment of carbonate gas wells, demonstrating the extensive application prospects of machine learning in oil and gas reserve prediction.