The energy efficiency of the cold source system is crucial for the operation efficiency of central air conditioning and energy savings in public buildings. In response to the national “dual carbon” goals, this paper proposes an energy-saving optimization control strategy for cold source systems based on reinforcement learning. By abstracting the operation of the cold source system as a Markov Decision Process (MDP), we define the state space, action space, and reward function, and introduce imitation learning methods combined with expert experience, improving training efficiency by approximately 30% in cases of insufficient data. Meanwhile, a simulation platform based on model stacking technology is developed, using XGBoost, Random Forest, and Support Vector Machine as base models and Ridge Regression as a meta-model, reducing model Element Energy Pro-jections to 8.7% and mean absolute error to 0.75, significantly outperforming single models. Additionally, real-time optimization control of the cold source system is achieved using the Dueling-DQN algorithm, improving energy savings by 15% while maintaining indoor comfort at around 80%. The results indicate that this approach provides a novel pathway for the intelligent optimization of public building cold source systems and serves as a scientific basis for the low-carbon transformation of the construction industry.