Wargames require participants to make real-time reasoning and decisions based on complex battlefield environment changes, facing severe challenges brought by large-scale decision spaces and rapidly changing battlefield situations. In recent years, many reinforcement learning algorithms have been continuously applied to wargames to simulate confrontational situations. However, existing methods have not yet provided satisfactory solutions for wargames and have certain limitations. In this context, we propose a hierarchical decision reinforcement learning algorithm based on decision frequency and internal reward mechanism. The algorithm divides decisions into three layers, decomposes complex tasks to reduce decision complexity, and compresses the action space to accelerate convergence. In addition, the decision frequency and internal reward mechanism are introduced to improve decision stability and guide learning. To verify the performance of the algorithm, experiments are designed for wargame scenarios. The experimental results show that the hierarchical structure enables the agent to effectively learn strategies and accelerate convergence.