The invention provides a power distribution method for the downlink of a low-orbit satellite. The method comprises the steps of: initializing the low-orbit satellite, and building a Markov decision process; observing the current state St; randomly selecting an action or selecting an optimal action according to the probability of an exploration factor; obtaining a new state St+1 and the instant reward rt of the current state St according to an action at, and storing the (st, at, rt, st+1) tetrad into an experience pool; under the condition that a cycle period is greater than a training number,training a current network; under the condition that the current time slot is integer multiples of the updating frequency of a target network, updating parameters of the target network; adding 1 to the current time slot; repeating the above steps until the current time slot is greater than a time slot counter, setting the current time slot to be 1 and adding 1 to the count of the cycle period; andrepeating the steps until the cycle period is greater than the training network period number. According to the distribution method, a deep reinforcement learning algorithm is adopted to dynamicallydistribute the subcarrier power of the multi-beam low-orbit satellite, so that the capacity of the low-orbit satellite is maximized, and the spectral efficiency is improved.