Electromagnetic Torque Prediction and Modeling of a Doubly Fed Induction Generator for Wind Energy Conversion Systems Using Machine Learning and Deep Learning Algorithms

According to the 2023 Wind Energy Report published by the Global Energy Council, the total installed power of wind energy conversion systems worldwide is around 1 TW. In addition, in 2024 and the following years, an average annual increase of around 15% on this installed capacity is envisaged. This situation reveals the importance and rapid development of wind energy conversion systems (WECS) in renewable energy systems. Accordingly, during the design, modeling and production of AC generators at different power levels used in wind turbines, new generation design and modeling techniques are used in addition to classical modeling methods, and wind turbine generator R&D is developing rapidly. New design and optimization methods have begun to be used in the modeling and performance analysis of Double Fed Asynchronous Generators (DFIG), which are frequently used in the field for different output powers. Modeling DFIG with classical numerical modeling and FEA-based magnetic simulation programs is a time-consuming operation, especially in transient or dynamic analysis. Depending on the performance of the computer, obtaining a transient field distribution solution may take hours or even days to obtain iteration-based field distribution solutions that use the finite difference method as a reference. Therefore, machine learning and deep learning-based iterative optimization and prediction methods stand out as a powerful alternative. In this study, electromagnetic torque values obtained through FEA-based simulations for three different DFIGs numerically modeled at medium power levels (250 kVA) with different winding materials (copper and aluminum) were used as reference. These torque curves were estimated using deep neural network algorithms based on K Nearest Neighbors (KNN), Support Vector Regression (SVR), Extra Tree (ET), Random Forest (RF), and Long Short-Term Memory (LSTM). Thus, the FEA results were compared with the predictions obtained from these algorithms, and the predictive performance of the algorithms was evaluated. The performances of the aforementioned algorithms in trainings and cross-validations were compared using R 2 , MAE, and RMSE metrics. The LSTM-based deep neural network outperformed the other algorithms for electromagnetic torque estimation. Using this approach, R 2 values of 0.990, 0.976 and 0.994 were obtained for DFIG-1, DFIG-2 and DFIG-3 in cross-validation, respectively.