In the context of power system operations, traditional satellite fusion networks often suffer from limited bandwidth, leading to slow data transmission and network congestion during peak usage times. Besides, the bandwidth occupation caused by redundant data far exceeds the system’s capacity and user tolerance. In order to meet the needs of multi-user, simultaneous segment and low latency communication, adaptive allocation is introduced into the satellite fusion networks. The transmission capability of network nodes is properly utilized to ensure the balance of bandwidth occupied by tasks. Based on the above scenarios, this paper proposes a Sequential Double Deep Q-Network model to meet the bandwidth requirements in typical power services of complex scenarios. First of all, to solve the problem of low utilization of experience samples, the success and failure experiences are classified and stored, and the LSTM framework is used to improve the failure history information. Secondly, a hybrid experience playback mechanism is introduced in DDQN to improve the efficiency of network training. Through simulation, we prove that the model is effective and advanced.