Abstract: With the vigorous implementation of China’s carbon peaking and carbon neutrality goals, renewable energy has gradually come into the public spotlight. The emergence of microgrids is conducive to the rapid absorption of renewable energy, making microgrid dispatching and optimization a research focal point. Traditional methods for low-carbon economic dispatching of microgrids encounter various challenges such as high data dimensionality, complex modeling, limited flexibility, and extended operation times. To address these challenges, a low-carbon economic optimization dispatching strategy for microgrids is proposed based on an improved twin delayed deep deterministic (TD3) policy gradient algorithm. Firstly, a statistical analysis on the strategy results of model predictive control (MPC) is conducted. The obtained operation rules of energy storage batteries are utilized to refine the self-learning rules of the TD3 algorithm, thereby effectively reducing the model’s exploration area. Secondly, the dual buffer technique is introduced to address the issue of sample imbalance, achieving a dual enhancement in efficiency and robustness. Finally, the effectiveness of the proposed algorithm model is verified through the analysis of a case study of a demonstration microgrid.