@Article{ee.2026.077001,
AUTHOR = {Xinyang Zhao, Fei Xie, Jiefeng Duan, Dongran Song, M. Talaat, Mustafa Serdar Genc},
TITLE = {Comparative Study on High-Precision Artificial Neural Network Fitting Methods for NMPC under MPPT Operating Conditions of Wind Turbines},
JOURNAL = {Energy Engineering},
VOLUME = {},
YEAR = {},
NUMBER = {},
PAGES = {{pages}},
URL = {http://www.techscience.com/energy/online/detail/27237},
ISSN = {1546-0118},
ABSTRACT = {Up to now, there remain difficulties in modeling large-scale engineering equipment with complex nonlinear characteristics, especially for the large-scale wind turbines. Based on the powerful learning capability of neural networks, this study adopts an approximation method with fast response characteristics to fit the output results of nonlinear model predictive control (NMPC). First, based on the dynamic characteristics of maximum wind energy extraction (MWEE) of wind turbines, a constrained NMPC rolling optimization prediction model is established, and a dataset of torque control laws with prediction horizons of 5, 6, and 7 steps respectively is generated. Second, the fitting performance of two feedforward networks, namely the artificial neural network (ANN) and feedforward neural network (FFNN), is compared and analyzed, and the positive correlation between the prediction horizon and fitting accuracy is clarified. Finally, three types of temporal neural networks, including the one-dimensional convolutional neural network (1D-CNN), long short-term memory (LSTM) network, and temporal convolutional network (TCN), are introduced, and a comprehensive evaluation of fitting accuracy, convergence characteristics and robustness is conducted based on five quantitative metrics including mean squared error (MSE) and root mean squared error (RMSE). The results show that: feedforward networks can achieve effective approximation of the nonlinear control law of NMPC; temporal neural networks can significantly break through the performance limitations of static networks, and the coefficient of determination (R<sup>2</sup>) of all three temporal networks is greater than 0.99, which fully meets the accuracy requirements of industrial control. This study verifies the feasibility of replacing the online solution of NMPC with a data-driven neural network fitting method, clarifies the differentiated engineering application scenarios of different network architectures, and provides a theoretical basis and technical solution for the high-performance real-time control of wind turbines under maximum power point tracking (MPPT) operating conditions.},
DOI = {10.32604/ee.2026.077001}
}