As for the challenges of local motion planning subject to constrained state for quadrotor in evaluating the health status of photovoltaic power generation in complex environment, including fast replanning, limited onboard computational resources, a practical framework for constructing motion trajectory from the bounded states is proposed in this work. The references and nominal inputs are made by the scheme of receding horizon fashion of model predictive control (MPC) by means of gradient-based method in the continuous domain. In order to avoid the complex calculation of the optimal gradient, the ability of approximation of a neural network make it possible for the fast evaluation and online optimization. The simulations with variety types of trajectories are made by means of a quadrotor to demonstrate the good efficiency and real-time of the proposed approach.