Non-orthogonal multiple access (NOMA) has become a promising candidate aiming to enhance the system performance in the fifth generation of wireless communication. However, most existing studies on the resource allocation of NOMA systems only consider the short-term performance improvement but ignore the problems caused by the imperfect time-varying channels and the dynamic traffic arrivals, which will result in unsuccessful transmission issues. Different from the existing approaches, we propose a long-term cross-layer resource allocation model with dynamic traffic arrivals and limited channel information. With the low-overhead one-bit feedback, the optimal decoding order, user scheduling, and power allocation are analyzed. Specifically, the problem is formulated as a stochastic optimization framework to minimize long-term power consumption. By applying the Lyapunov optimization framework, the problem can be transformed into a joint traffic rate control and mixed-integer programming resource allocation problem. This NP-hard problem is difficult to be solved directly. Therefore, we devise an efficient sub-optimal algorithm with the dynamic penalty factor. In our theoretical analysis, we prove that the delay-power tradeoff can be achieved by tuning a control parameter. The simulation results confirm that our proposed algorithms can efficiently reduce the power consumption compared with the baseline algorithms while satisfying the QoS requirements.