With the expansion of power grids and the increasing penetration of renewable energy

frequency stability in power systems has become a key research focus. Emergency load shedding is an effective control strategy to mitigate frequency instability; however

the traditional offline-setting and online-matching approach is increasingly inadequate for complex and evolving fault scenarios. To address this issue

a deep reinforcement learning-based method is proposed to achieve online decision-making for emergency load shedding. First

a decision modeling approach based on Markov decision process (MDP) is developed

and a branch dueling Q-network (BDQ) is introduced to handle the high-dimensional load shedding decision space. Furthermore

to overcome the high computational and time costs in traditional reinforcement learning training

a centralized training strategy driven by a pre-generated experience pool is adopted by decoupling sample collection from model training

thereby enabling efficient agent training. Finally

simulation results on a 10-machine 39-bus system demonstrate that the proposed algorithm improves decision effectiveness by 4.94% compared to traditional reinforcement learning methods

while reducing training time to only 8.82% of that of conventional approaches.