Fast Implementation of Model Predictive Control for PMSG Wind Turbine Generator Based on Deep Learning

Due to the randomness of wind energy, this paper applies Model Predictive Control (MPC) to control the wind turbine generator to improve power tracking performance and disturbance rejection at the maximum power point, thereby enhancing the efficiency and stability of wind turbines. Since MPC requires solving optimization problems with constraints online, a supervised learning-based approximation method for wind turbine generator MPC is proposed to meet real-time requirements, aiming to maintain system performance while improving computational efficiency. The proposed method uses a deep feedforward neural network to model the input-output relationship of the MPC. An unsupervised learning technique is employed to obtain optimal control inputs, whose accuracy matches that of the training data, and the objectives and constraints of the learning problem are aligned with those of the MPC. The proposed method also incorporates input-tight constraints in the model training, ensuring that the network outputs under all feasible states satisfy the system’s input constraints. Finally, the proposed fast implementation method is applied to a Permanent Magnet Synchronous Generator (PMSG) wind turbine. Numerical simulation results show that the proposed optimization-control implementation method offers high computational efficiency and superior performance.