Non-orthogonal multiple access (NOMA) has been considered as a key candidate technology in 5G networks. To improve the system efficiency of 5G systems, the design and principles of Multiple-input-multiple-output (MIMO) can be adopted into NOMA. MIMO helps to improve the performance gain without the requirement of additional resources. In this paper, power allocation (PA) problem was investigated for downlink MIMO-NOMA systems by utilizing non-deep learning and deep learning approaches. In order to improve the energy efficiency of such MIMO-NOMA system, multi-dimensional power allocation problem was formulated which is a mixed integer non-linear programming problem. With this insight, the problem is first dealt with non-deep learning approach and it is solved by the proposed PASR (Power allocation based on shifting additional resources) scheme. This scheme is investigated with two different assumptions: fixed (F-PASR) and variable inter-cluster power (V-PASR). Finally, a machine learning framework was proposed to deal with power allocation problem. An energy efficient deep neural network model (EE-DNN) was proposed for power allocation to reduce the complexity and latency of the aforementioned problem. Extensive simulation results are provided to confirm the effectiveness of the proposed energy efficient schemes. It was clearly demonstrated that proposed F-PASR and V-PASR algorithms outperform conventional OMA (Orthogonal multiple access) scheme. It was also validated that deep learning based framework performs slightly poorer than non-deep learning schemes but with less computation time.