With the increasing penetration of renewable energy, thermal power units frequently operate in a wide load range, resulting in significant variations in the parameters of their Automatic Generation Control (AGC) models. Under traditional grid AGC models and frequency regulation strategies, Frequency regulation capability of the system is difficult to assess and improve accurately. As a result, this paper proposes a Neural Network based Model Predictive Control (NN-MPC) strategy for grid AGC frequency regulation, considering the flexible operation of thermal power units. First, a wide-load frequency regulation model for thermal power units in flexible operation is established. And a Back Propagation Neural Network (BPNN) is used to fit and reconstruct the grid AGC model. Next, the dynamic frequency response characteristics of the grid are predicted based on the BPNN frequency regulation model, which is then used as the predictive model in the model predictive control (MPC) framework. An optimal frequency performance objective function is designed, and the Differential Evolution Algorithm is employed to solve the optimization problem and implement rolling optimization. Finally, simulation results in MATLAB/Simulink demonstrate that the proposed NN-MPC strategy effectively mitigates the mismatch problem of traditional MPC models caused by the wide-load operation of thermal power units. And the proposed method can significantly reduce maximum frequency deviation of the system.