Unmanned aerial vehicle (UAV) intelligent air combat is a disruptive technology that can change the operational form of future warfare, with deep reinforcement learning serving as a crucial paradigm for achieving intelligent decision-making in air combat scenarios. Despite recent advancements in this area, reinforcement learning models designed based on virtual simulation interactions continue to struggle with low learning efficiency and poor generalization, making it difficult to achieve effective sim-to-real transfer. To enhance the transferability of air combat decision-making models from virtual simulation environments to real-world physical spaces, this paper proposes a close-range air combat maneuvering decision-making model based on the EfficientZero algorithm enhanced with domain randomization. The model efficiently learns decision-making strategies through self-play by leveraging interaction data generated in the process. Additionally, domain randomization techniques are employed to improve the model's robustness. Simulation results demonstrate that the intelligent decision-making model based on the EfficientZero algorithm efficiently utilizes air combat sample data, avoiding the common issue of strategy cycles in self-play. Furthermore, the use of domain randomization significantly improves the generalization performance and effectively enhances the robustness of the reinforcement learning model under reality gap conditions.