Abstract: Emergency control is an important means of maintaining power system transient security and stability following serious faults. The current popular "human-in- the-loop" offline emergency control decision-making method has some drawbacks, including low efficiency and heavy reliance on expert experience. Therefore, this paper proposes an intelligent emergency generator rejection decision-making method based on knowledge fusion and deep reinforcement learning (DRL). First, a DRL-based emergency generator rejection decision-making framework is built. Then, when the agent deals with multi-generator decisions, the resulting high-dimensional decision space makes the agent training difficult. There are two solutions proposed: decision space compression and the application of a branching dueling Q (BDQ) network. Next, to further improve the exploration efficiency and the decision-making quality of the agent, the knowledge and experience related to emergency generator rejection control are integrated to the agent training. Finally, the simulation results in the 10-machine 39-bus system show that the proposed method can quickly give effective emergency generator rejection decisions in multi-generator decision- making. Applying a BDQ network has better decision performance than decision space compression. The knowledge fusion strategy can guide the agents to reduce ineffective decision- making explorations and improve decision-making performance.