向模型预测控制的风电机组深度Koopman 全局线性建模方法

To address the challenges posed by the complex nonlinear dynamics of wind turbines in model predictive control (MPC)-oriented modeling, as well as the limitations of existing nonlinear and local linear modeling methods in terms of model complexity and accuracy, a global linear modeling method for wind turbines is proposed. Based on Koopman operator theory and deep learning techniques, a statespace mapping neural network is designed. A training strategy incorporating a Frobenius norm-based regularization term is developed to enhance the long-term prediction accuracy of the established model. Through data-driven training of the proposed network, a high-dimensional global linear dynamic model of the wind turbine is established. Simulation results demonstrate that the prediction errors for rotor speed and pitch angle are 0. 869% and 0. 026%, respectively, which are significantly lower than those of the three comparative methods. Compared with the local linear dynamic model, the wind farm MPC strategy based on the established high-dimensional global linear dynamic model reduces the rotor speed tracking error and overshoot by 92. 58% and 95. 85%, respectively. The findings provide a theoretical reference for MPC-oriented dynamic modeling of wind turbines.