Newly built microgrids lack historical operation data, and conventional data-driven methods are difficult to accurately predict renewable power output, which in turn affects the accuracy of scheduling plans. In this paper, a microgrid optimal scheduling method for newly built microgrids with small sample data scenarios is proposed. Firstly, an improved network structure named DANN-LSTM integrating domain adversarial neural network (DANN) and long-short-term memory network (LSTM) is designed, and the domain adversarial idea and gradient inversion mechanism are introduced into the migration learning to improve the generalization ability of the model, reduce the difference in the domain distribution of the data, and use the rich operating data of power stations with similar output characteristics to predict the output of the target station, which overcomes the problem of poor accuracy of the output prediction under the conditions of small samples. The problem of low accuracy of output prediction under small sample conditions is overcome. Further, the optimal scheduling model is transformed into a Markov decision process, which is solved using the double-delay deep deterministic policy gradient algorithm (TD3). Finally, the effectiveness of the proposed method is verified with an example of an improved CIGRE 14- node microgrid.