NOMA (Non-Orthogonal Multiple Access), as one of the candidate technologies of 5G, can improve the spectrum efficiency and system capacity, and has attracted wide attention. The essence of NOMA is multi-user overlay transmission in power domain, and multiple users will schedule on the same sub-band. The fundamental principle and key technologies of NOMA technology, such as power multiplexing superposition coding technology and SIC (Successful Interference Cancellation) technology, are described. The BER performance of NOMA communication system based on power domain in downlink is studied, and the appropriate power allocation ratio range is obtained. From the overall point of view of the system, users are grouped with the goal of maximizing the channel gain difference of all matched users, and then the problem of maximizing the weighted sum rate is transformed into the problem of minimizing the weighted sum mean square error by using the deep learning algorithm. By updating each optimization parameter iteratively in turn, it converges to the local optimum of the objective function. This method can improve the weighting and rate performance of the system with fewer iterations.