Swarm aerial robots are required to maintain close proximity to successfully traverse narrow areas in cluttered environments. However, this movement is affected by the downwash effect generated by the other quadrotors in the swarm. This aerodynamic effect is highly nonlinear and hard to model by classic mathematical methods. In addition, the motor speeds of quadrotors are risky to reach the limit when resisting the effect. To solve these problems, we integrate a Neural network Downwash Predictor with Nonlinear Model Predictive Control (NDP-NMPC) to propose a trajectory-tracking approach. The network is trained with spectral normalization to ensure robustness and safety on uncollected cases. The predicted disturbances are then incorporated into the optimization scheme in NMPC, which handles constraints to ensure that the motor speed remains within safe limits. We also design a quadrotor system, identify its parameters, and implement the proposed method onboard. Finally, we conduct an open-loop prediction experiment to verify the safety and effectiveness of the network, and a real-time closed-loop trajectory tracking experiment which demonstrates a 75.37% reduction of tracking error in height under the downwash effect.