An online power allocation algorithm based on deep reinforcement learning in multibeam satellite systems

Abstract

Dynamic power allocation (DPA) is the key technique to improve the system throughput by matching the offered capacity with that required among distributed beams in multibeam satellite systems. Existing power allocation studies tend to adopt the metaheuristic optimization algorithms such as the genetic algorithm. The achieved DPA cannot adapt to the dynamic environments due to the varying traffic demands and the channel conditions. To solve this problem, an online algorithm named deep reinforcement learning‐based dynamic power allocation (DRL‐DPA) algorithm is proposed in this paper. The key idea of the proposed DRL‐DPA lies in the online power allocation decision making other than the offline way of the traditional metaheuristic methods. Simulation results show that the proposed DRL‐DPA algorithm can improve the system performance in terms of system throughput and power consumption in multibeam satellite systems.