Adaptive vector oriented control of doubly fed induction generator wind turbines using M5P model tree for robust dynamic performance Authors: Koudri Benyoucef, Moualidia Abdelhafidh, Boudana Djamel, Mujammal Ahmed Hasan Mujammal, Khechafi Sofiane, Zemirline Nassim, Mohit Bajaj, Oleksandr Rubanenko Published: 2025 ABSTRACT AbstractWind energy systems using doubly fed induction generators (DFIGs) rely on vector-oriented control (VOC) to achieve decoupled regulation of active and reactive power. Conventional VOC strategies typically employ proportional-integral (PI) controllers for rotor current control; however, these controllers struggle under nonlinear and variable operating conditions, such as wind speed fluctuations, leading to degraded dynamic performance and power quality. To overcome these limitations, this paper proposes an intelligent VOC approach based on the M5-Pruned model tree (M5P) algorithm, complemented by a comparative study with fuzzy logic controllers (FLC). The M5P-based controller introduces a data-driven mechanism that adapts to system dynamics, ensuring improved accuracy and robustness compared to traditional PI and FLC methods. A detailed MATLAB/Simulink simulation of a grid-connected wind turbine with DFIG evaluates the controllers under varying wind profiles. Results demonstrate that the proposed M5P strategy significantly reduces overshoot and settling time, enhances power regulation, and improves overall system stability. These findings highlight the advantages of machine learning-based control in addressing the shortcomings of conventional VOC techniques for renewable energy applications.Keywords: Vector oriented control (VOC), Doubly fed induction generator (DFIG), Wind energy conversion system (WECS), Fuzzy logic controller (FLC), M5P algorithmSubject terms: Energy science and technology, Engineering Introduction The global transition toward sustainable energy has intensified the integration of renewable sources into modern power systems, with wind energy emerging as one of the most viable alternatives to fossil fuels1. Among the various technologies employed in wind energy conversion systems (WECS), the doubly fed induction generator (DFIG) has gained prominence due to its high efficiency, variable-speed operation, and capability for independent control of active and reactive power2,3. This control flexibility is typically achieved through rotor-side converters that implement vector-oriented control (VOC) strategies, which decouple the torque and flux components for improved dynamic performance4. However, conventional VOC schemes rely heavily on proportional–integral (PI) controllers, whose performance deteriorates under nonlinear operating conditions and parameter uncertainties, thereby motivating the development of more adaptive and intelligent control approaches5. Extensive research has been conducted to enhance the performance of DFIG-based wind energy systems through advanced control strategies6,7. Conventional PI controllers, despite their simplicity and ease of implementation, exhibit poor adaptability under nonlinear dynamics, parameter variations, and external disturbances such as wind speed8,39. Additionally, subsynchronous resonance (SSR) phenomena in series-compensated transmission lines connected to WECS have been shown to significantly influence system stability, with damping characteristics strongly affected by controller gains, wind speed, and compensation levels9. The motivation for this work stems from the increasing penetration of wind energy into power grids, which demands highly adaptive and robust control strategies. Conventional PI-based VOC fails to meet these requirements under fluctuating wind and grid conditions, creating a need for intelligent, data-driven solutions that can ensure stability, efficiency, and power quality in real-time applications. To overcome these limitations, intelligent control techniques have been introduced, including Fuzzy Logic Controller. (FLC)10,37 and artificial neural networks (ANN)11,12, which offer improved robustness and adaptability. FLCs, in particular, have demonstrated superior performance in handling system nonlinearities without requiring an accurate mathematical model13,14. However, their rule-based structure can become complex and computationally demanding for large-scale systems. Recent advancements in machine learning have opened new opportunities for data-driven control approaches, such as model tree algorithms, which combine decision tree structures with regression models to achieve high accuracy and adaptability in dynamic environments15,40. Despite the widespread adoption of VOC in DFIG-based wind energy systems, its conventional implementation using PI controllers faces several challenges16,17. Moreover, their fixed gain structure limits adaptability, leading to degraded performance during nonlinear operating scenarios and grid disturbances18. These limitations often result in slower dynamic response, increased overshoot, and reduced robustness, particularly when the system operates under variable wind speeds or experiences sudden load changes. Consequently, there is a growing need for advanced control strategies that can overcome these drawbacks by providing improved adaptability, robustness, and accuracy without significantly increasing computational complexity19. Similarly, autonomous microgrids integrating wind and diesel generation require advanced frequency regulation schemes, where intelligent controllers such as fuzzy logic-based sliding mode control have demonstrated superior robustness under load perturbations and dynamic conditions20. To address the limitations of conventional PI-based VOC, intelligent control techniques have emerged as promising alternatives due to their ability to handle nonlinearities and adapt to varying operating conditions21. FLC has been widely adopted because they do not require an accurate mathematical model and can effectively manage uncertainties in wind energy systems. However, their performance depends on the design of membership functions and rule sets, which can become complex for high-dimensional systems22. Recent advances in intelligent and optimization-based control for power systems further illustrate the trend towards hybrid and data-driven methods. For example, Harris Hawks Optimization (HHO)23 has been applied to tune fractional-order fuzzy- PID controllers for frequency regulation in multi-microgrid environments, showing improved robustness under load variations and renewables penetration. Similarly, HHO-guided model predictive control has been proposed for stochastic wind-integrated multi-source power systems to handle uncertainty in generation and demand24. While these optimization-driven schemes offer strong performance, they typically rely on online/iterative optimization or relatively complex controller structures. In contrast, the proposed M5P model-tree approach aims for a different trade-off: it learns compact, locally linear models from simulation data to produce an interpretable and computationally light controller that can generalize across operating conditions without requiring continuous optimization during runtime. M5P combines decision tree structures with linear regression models, enabling accurate prediction and control with reduced computational complexity10,25. This capability makes M5P particularly suitable for real-time applications in renewable energy systems, where dynamic adaptability and robustness are critical26. While ANN-based controllers have been explored27,28, they often require large datasets, complex tuning, and significant computational resources, making them less suitable for real-time applications. In contrast, FLC offers simplicity and interpretability, whereas the M5P model tree algorithm provides an adaptive, data-driven approach with a favorable balance of accuracy, efficiency, and transparency. By constructing locally linear regression models organized in a tree structure, M5P enables fast learning, smooth decision boundaries, and interpretable control logic, making it well-suited for MATLAB/Simulink implementation and potential real-time applications in renewable energy systems. These characteristics position FLC and M5P as ideal techniques for comparison in this study. Building on these considerations, a smart VOC strategy for DFIG-based wind energy systems is proposed by replacing the conventional PI controllers used for rotor current regulation and with the M5P model tree algorithm. This modification introduces a data-driven and adaptive mechanism into the traditional VOC framework, significantly improving its ability to handle nonlinearities and parameter variations. The integration of M5P not only enhances the dynamic performance of the system but also adds an intelligent dimension to the control process, enabling superior adaptability under fluctuating wind and grid conditions. The main contributions of this work are as follows: Propose an intelligent VOC strategy for DFIG using the M5P model tree algorithm. Propose an intelligent VOC strategy for DFIG using the M5P model tree algorithm. Perform a comparative analysis of PI, FLC, and M5P controllers under realistic wind profiles. Perform a comparative analysis of PI, FLC, and M5P controllers under realistic wind profiles. Validate the proposed approach through MATLAB/Simulink simulations, demonstrating improved dynamic response, power regulation, and robustness. Validate the proposed approach through MATLAB/Simulink simulations, demonstrating improved dynamic response, power regulation, and robustness. Highlight the applicability of machine learning techniques for real-time renewable energy control. Highlight the applicability of machine learning techniques for real-time renewable energy control. The results demonstrate that the M5P-based controller significantly outperforms conventional methods, offering improved accuracy, faster response, and enhanced robustness, thereby paving the way for intelligent control integration in renewable energy systems. The main body of this article starts with the Introduction, setting the context and objectives of the work. This is followed by sections dedicated to the Modeling of System (or theoretical framework), which include subsections for specific models or components. After that, the Control Strategy Proposition (or methodology) is presented, often with subsections for different approaches or techniques. The paper then moves to the Discussion, where results are analyzed and interpreted, and finally to the Summary and Conclusions, which encapsulate the findings and implications. Modeling of system Wind turbine modeling The wind turbine converts the kinetic energy of moving air into mechanical power, which is then transmitted to the generator through a gearbox29,33,34. The aerodynamic power extracted from the wind is expressed as in Eq. (1): where ρ is the air density, A = πR2 is the rotor swept area with blade radius R, vw is the wind speed, β is the blade pitch angle, and Cp (λ,β) is the power coefficient. The latter captures aerodynamic efficiency and depends on the tip-speed ratio30. To emulate realistic operating conditions, the wind speed is modeled using a composite sinusoidal function as expressed in Eq. (2): This formulation combines multiple harmonics to represent both short-term gusts and long-term variations, which are typical in real wind conditions. The resulting profile varies between approximately 2.5 m/s and 16 m/s, covering the operational range of most wind turbines. By using this dynamic wind model, the simulation captures realistic fluctuations rather than assuming constant wind speed, allowing for a more accurate evaluation of controller performance under variable conditions. Figure 1 illustrates the generated wind speed profile. Wind speed profile. with the turbine (low-speed shaft) angular velocity. Mechanical torque: The mechanical torque at the turbine shaft follows from (1), is expressed as in (4). It is subject to aerodynamic and structural constraints (e.g., pitch limits, speed limits). Mechanical torque: The mechanical torque at the turbine shaft follows from (1), is expressed as in (4). It is subject to aerodynamic and structural constraints (e.g., pitch limits, speed limits). Gearbox relations: A gearbox with ratio G > 0 maps low-speed shaft quantities to the generator (high-speed) shaft as in (5): Gearbox relations: A gearbox with ratio G > 0 maps low-speed shaft quantities to the generator (high-speed) shaft as in (5): where anddenote the generator-side mechanical speed and torque, respectively. Shaft dynamics (generator side): Grouping the turbine and generator inertia to the generator side yields the one-mass torsional model expressed as in Eq. (6). Shaft dynamics (generator side): Grouping the turbine and generator inertia to the generator side yields the one-mass torsional model expressed as in Eq. (6). where Te is the electromagnetic torque of the generator, and. with , and the turbine and generator inertias and viscous friction coefficients, respectively. For studies requiring shaft elasticity, (6) can be extended to a two-mass model with shaft stiffness and damping. Power coefficient model: The function is typically provided by manufacturers as a map or fitted surface. A widely used analytical approximation is expressed in Eq. (8). Power coefficient model: The function is typically provided by manufacturers as a map or fitted surface. A widely used analytical approximation is expressed in Eq. (8). with the auxiliary quantity expressed as: and empirical coefficients c1,. . ., c6 chosen to match the specific turbine (typical starting values used in the literature are c1 = 0.5176, c2 = 116, c3 = 0.4, c4 = 5, c5 = 21, c6 = 0.0068; these should be tuned for your rotor). Figure 2 shows the main components of the system. Wind turbine to generator (DFIG) mechanical pathway: aerodynamics, gearbox, and shaft dynamics. Modeling of doubly fed induction generator (DFIG) DFIG is widely used in wind energy systems due to its ability to operate at variable speeds and control active/reactive power35,36,38. It consists of a wound rotor induction machine with the stator directly connected to the grid and the rotor connected via slip rings to a back-to-back converter31. Assumptions • Balanced three-phase system. Sinusoidal steady-state operation. Sinusoidal steady-state operation. dq-axis reference frame aligned with stator flux or grid voltage. dq-axis reference frame aligned with stator flux or grid voltage. Neglecting core losses and magnetic saturation. Neglecting core losses and magnetic saturation. Electrical Equations in dq Reference Frame: Stator: Stator Equations. Stator Equations. The d and q components of the stator voltage are expressed, respectively, in Eqs. (10) and (11). Rotor Equations. Rotor Equations. The d and q components of the rotor voltage are expressed, respectively, in Eqs. (12) and (13). Electromagnetic Equations: The d and q components of the stator electromagnetic flux are expressed in Eqs. (14) and (15), respectively, and the rotor stator electromagnetic flux d and q components are expressed, respectively, in Eqs. (16) and (17). Electromagnetic Torque: The Electromagnetic Torque is expressed as in Eq. (18), and the mechanical dynamics is expressed as in Eq. (19). Modeling of rotor side vector oriented control strategy In this work, the VOC strategy is applied exclusively to the Rotor Side Converter (RSC) of the DFIG. The control objective is to regulate the stator active and reactive power by controlling the rotor currents in a synchronous dq reference frame aligned with the stator flux or grid voltage. Rotor Current Components: The rotor currents are decomposed into two orthogonal components: Rotor Current Components: The rotor currents are decomposed into two orthogonal components: controls the stator active power. controls the stator reactive power. Stator Power Expressions The stator active and reactive powers in the dq frame are given by: Stator Power Expressions The stator active and reactive powers in the dq frame are given by: These powers are regulated using PI controllers that generate reference rotor currents and . The rotor-side converter then tracks these references using current control loops. Decoupled Control. By aligning the reference frame with the stator voltage vector, the VOC strategy decouples the control of active and reactive power. This simplifies the control structure and enhances dynamic performance. Control Objectives. Maintain desired stator active power output. Maintain desired stator active power output. Regulate stator reactive power to support grid voltage. Regulate stator reactive power to support grid voltage. Ensure stable operation under varying wind speeds. Ensure stable operation under varying wind speeds. This control strategy enables efficient and flexible operation of the DFIG in wind energy applications, allowing for grid support and optimal power extraction. Proposed control strategies To enhance the dynamic performance and robustness of the DFIG control system, this section presents a comparative study of three control strategies applied to the RSC. The baseline approach is the conventional VOC, which decouples the control of active and reactive power using PI controllers. In the second strategy, the conventional PI controllers are replaced by an independent FLC to improve adaptability under nonlinear and uncertain operating conditions. The third and most significant strategy introduces the M5P decision tree algorithm, replacing both PI controllers with a data-driven model that enables intelligent and adaptive control based on real-time operating data. Conventional vector oriented control The conventional VOC strategy is implemented in a synchronous dq reference frame aligned with the stator voltage vector. Figure 3 represents the embedded controllers PI, FLC, and proposed M5P with the voltage compensator in the system. VOC strategy with voltage compensator. This alignment allows the decoupling of the stator active and reactive power control through the rotor current components. The control structure consists of two PI controllers: The controller regulates the stator reactive power. The controller regulates the stator reactive power. The controller regulates the stator active power. The controller regulates the stator active power. The stator power expressions in the dq frame are given by: The PI controllers generate reference rotor currents and which are tracked by the rotor-side converter through current control loops. The PI controller gains were tuned empirically to optimize transient response and steady-state accuracy. The final parameters were Kp = 0.0115, Ki = 0.5521 for the Idr loop and Kp = 0.01, Ki = 10 for the Iqr loop. This strategy provides a reliable and well-established method for controlling DFIGs under normal operating conditions. Figure 4 indicates the PI controllers with the rotor voltage compensation. Structure of and PI controllers of VOC strategy. Rotor voltage compensation in VOC From the figure, the rotor voltage compensation in the VOC strategy is based on the decoupling terms that cancel cross-coupling effects in the dq-axis model. The expressions are expressed in Eq. (24) Where: compensation voltages for d and q axes, compensation voltages for d and q axes, ωs: synchronous speed, ωs: synchronous speed, ωr: rotor speed, ωr: rotor speed, Lr: rotor inductance, Lr: rotor inductance, idr, iqr: rotor current components. idr, iqr: rotor current components. These terms are added to the PI controller outputs to decouple the d and q axes, ensuring independent control of active and reactive power. In Field-Oriented Control (FOC), the machine equations are expressed in the synchronous rotating dq frame aligned with the rotor flux (or PM flux for PMSM). The d and q electrical dynamics are expressed as in Eqs. (25) and (26), respectively: Where: vd, vq are the dq-axis stator voltages, vd, vq are the dq-axis stator voltages, id, iq are the dq-axis currents, id, iq are the dq-axis currents, Rs is the stator resistance, Rs is the stator resistance, Ld, Lq are the dq-axis inductances, Ld, Lq are the dq-axis inductances, λf is the permanent magnet flux linkage, λf is the permanent magnet flux linkage, is the electrical angular speed, with p the number of pole pairs. is the electrical angular speed, with p the number of pole pairs. The electromagnetic torque is expressed in Eq. (27): For a surface-mounted PMSM (Ld ≈ Lq), Eq. (27) simplifies to (28): The mechanical dynamics is expressed as in (29): where J is the rotor inertia, B the viscous friction coefficient, and TL the load torque. The electrical and mechanical speeds are related by: FOC Principle: By aligning the d-axis with the rotor flux, torque control is achieved by regulating iq, while id is set to zero (or a negative value for reluctance torque in IPMSM). Vector oriented control based on FLC In conventional VOC, the inner current control loops for the d-axis and q-axis components (id and iq) are typically implemented using PI controllers. While PI controllers offer simplicity and satisfactory performance under nominal conditions, their effectiveness can degrade in the presence of parameter variations, nonlinearities, and external disturbances32. Figure 5 represents the VOC controller based on FLCs, while Fig. 6 shows the membership function types of the FLC algorithm. To enhance robustness and adaptability, this subsection proposes replacing the two PI controllers with FLCs. The FLC approach leverages linguistic rules and membership functions to handle system uncertainties without requiring an accurate mathematical model. By tuning the control action based on fuzzy inference, the proposed method aims to improve dynamic response, reduce overshoot, and maintain stability under varying operating conditions. The membership functions for the input error, change of error, and output control signal used in the FLC-based VOC scheme are illustrated in Fig. 6. Vector Oriented Control based on FLC. Membership Functions of inputs and output of FLC. The mathematical representation of a leaf model defines appropriate input variables (such as current error and its derivative), constructs a rule base, and implements a defuzzification strategy to generate the control signals for the d- and q-axis voltage references. The overall structure of the VOC remains unchanged, ensuring compatibility with standard field-oriented control schemes while introducing intelligent control at the current regulation level. Two input variables (current error and its derivative) and one output variable (voltage control signal) were used, each with seven triangular membership functions. To reduce computational complexity, the FLC inputs and outputs were normalized using gain factors (eIdr = 0.05, ∆eIdr = 0.01, output = 1.2; eIqr = 0.025, ∆eIdr = 0.00001, output = 0.1). These adjustments ensured effective control performance while minimizing the computational burden of the fuzzy inference process. Table 1 illustrates the FLC table for inputs and output membership functions. The error and its derivative are calculated according to Eqs. (31) and (32), respectively. Table 1Two-input fuzzy rule base: error e vs. Change of error ∆e.e\∆eNBNMNSZEPSPMPB NB NBNBNMNMNSZEZE NM NBNMNMNSZEZEPS NS NMNMNSZEZEPSPM ZE NMNSZEZEPSPMPM PS NSZEZEPSPMPMPB PM ZEZEPSPMPMPBPB PB ZEPSPMPMPBPBPB Two-input fuzzy rule base: error e vs. Change of error ∆e. The error and its derivative are calculated according to Eqs. (31) and (32), respectively. VOC based on the proposed M5P decision tree algorithm VOC is a widely used technique for controlling AC machines due to its ability to decouple torque and flux components. Traditionally, VOC employs PI controllers to regulate the direct-axis (Id) and quadrature-axis (Iq) currents. However, PI controllers exhibit limitations under parameter variations, nonlinearities, and dynamic operating conditions, which can degrade performance. To address these issues, this work proposes replacing the conventional PI controllers with an M5P decision tree-based model, offering adaptive, data-driven control without requiring precise system modeling. From a computational perspective, the M5P-based controller introduces a moderate increase in processing requirements compared to the PI approach but remains substantially more efficient than the FLC structure. After training, the control action is derived from simple linear equations at the leaf nodes, ensuring rapid computation with minimal memory usage. This balance between accuracy and efficiency suggests that M5P can be feasibly implemented for real-time control applications using standard embedded processors. The structure of the proposed M5P-based VOC controller, including the decision-tree based adaptive current regulation, is presented in Fig. 7. Vector Oriented Controller based on M5P. Modeling of the proposed M5P algorithm The M5P algorithm is a model tree approach that combines decision tree partitioning with linear regression models at the leaves. Its main steps are: Tree Construction: The input space (e.g., reference currents, speed, and system states) is recursively partitioned based on attributes that minimize the standard deviation of the target variable. Tree Construction: The input space (e.g., reference currents, speed, and system states) is recursively partitioned based on attributes that minimize the standard deviation of the target variable. Linear Models at Leaves: Instead of assigning a constant value at each leaf, M5P fits a multivariate linear regression model, improving prediction accuracy and smoothness. Linear Models at Leaves: Instead of assigning a constant value at each leaf, M5P fits a multivariate linear regression model, improving prediction accuracy and smoothness. Pruning and Smoothing: The tree is pruned to avoid overfitting, and smoothing is applied to ensure continuity between adjacent regions. Pruning and Smoothing: The tree is pruned to avoid overfitting, and smoothing is applied to ensure continuity between adjacent regions. In this study, the M5P model tree was trained using three key input features: current error ,change of error , and control voltages . These variables were chosen because they directly influence the dynamic performance of the DFIG and provide the necessary information for accurate prediction and control of rotor currents. By incorporating both instantaneous error and its rate of change along with machine speed, the model can adapt to nonlinearities and varying operating conditions, ensuring robust and precise current regulation. The training dataset was generated from FLC-based simulations, where the input signals were error and change of error, and the outputs were the control actions produced by the FLC. This ensures that the M5P model learns the relationship between system states and optimal control signals, enabling adaptive and robust current regulation. The mathematical representation of a leaf model is expressed as in Eq. (33): where xi are input features (e.g., speed error, torque demand), and a.i. are regression coefficients. Importance of Replacing PI Controllers. Adaptive Behavior: Unlike PI controllers, which require manual tuning and are sensitive to parameter changes, M5P adapts to nonlinearities and varying operating conditions. Adaptive Behavior: Unlike PI controllers, which require manual tuning and are sensitive to parameter changes, M5P adapts to nonlinearities and varying operating conditions. Data-Driven Control: M5P leverages historical and real-time data, reducing dependency on exact motor parameters as shown in Fig. 8. Data-Driven Control: M5P leverages historical and real-time data, reducing dependency on exact motor parameters as shown in Fig. 8. Improved Dynamic Response: By learning complex relationships between inputs and outputs, M5P can provide faster torque and flux regulation. Improved Dynamic Response: By learning complex relationships between inputs and outputs, M5P can provide faster torque and flux regulation. Reduced Overshoot and Steady-State Error: The regression models at leaves allow fine-grained control compared to fixed PI gains. Reduced Overshoot and Steady-State Error: The regression models at leaves allow fine-grained control compared to fixed PI gains. Learning process of M5P Decision tree. Performance Evaluation Metrics. The performance of the proposed model is evaluated using the following error metrics: Mean Absolute Error (MAE): Root Mean Squared Error (RMSE): Relative Absolute Error (RAE): Root Relative Squared Error (RRSE): where: is the actual value, is the predicted value, is the mean of actual values, and n is the number of observations. M5P model tree generation The training dataset for the proposed M5P controller was generated from MATLAB/Simulink simulations of the FLC-based control system, which provided a larger and more representative dataset than the PI-based configuration. Data were sampled at s, resulting in 300,001 instances. After preprocessing, the dataset was trained using the WEKA platform to build the M5P model tree, which was then integrated into MATLAB/Simulink using a customized implementation to evaluate its control performance. Figure 9 illustrates the learned M5P model-tree structure, showing the hierarchical rule partitions and corresponding linear regression leaves used for current control generation. M5P decision tree algorithm generated by Weka. Figure 10a, b, c and d illustrate data classifications of inputs, output, and desired predictive output. It can be seen that the data of the output is the same data of the predictive output. Data classifications of inputs, output, and desired predictive output. Simulation findings Simulation results are essential to validate the proposed control strategy under realistic operating conditions. They demonstrate the dynamic performance of the system, robustness, and ability to handle nonlinearities. By comparing the M5P-based VOC with PI and fuzzy controllers, the results confirm significant improvements in response speed and power quality. Performance of system under wind variation Figure 11 shows the turbine speed. The M5P controller achieves the closest tracking to the reference compared to PI and FLC, supported by its lowest error metrics: MAE = 0.0081 and RMSE = 0.2025, which indicate minimal absolute and squared deviations from the reference, resulting in smoother waveforms. Additionally, RAE = 0.0871% and RRSE = 0.9206% confirm that the proposed method significantly reduces relative error compared to PI and FLC. Together, these metrics validate the observed reduction in oscillations and harmonic distortion. Turbine speed. Figure 12 illustrates the rotor direct current response under wind variation, highlighting the comparative performance of the three control strategies. The conventional PI controller exhibits significant overshoot and persistent oscillations during transient states, which can compromise system stability and increase current ripple. The FLC mitigates these oscillations and improves reference tracking; however, it still suffers from slower convergence and a noticeable steady-state error. In contrast, the M5P-based controller demonstrates superior dynamic behavior, achieving rapid convergence with minimal overshoot and negligible steady-state error. This performance is supported by its error metrics: MAE and RMSE indicate very small absolute and squared deviations from the reference, which translates into closer tracking and smoother waveforms, as can be seen from the zoomed-in section in Fig. 12. Additionally, RAE and RRSE confirm that these improvements are significant relative to PI and FLC controllers, validating the reduction in oscillations and harmonic distortion. Together, these metrics provide strong evidence of the proposed approach’s enhanced accuracy and robustness under variable operating conditions. Rotor direct current wind variation. Figure 13 shows the electromagnetic torque profile reveals the limitations of the PI controller, which suffers from pronounced oscillations and overshoot during transient periods, leading to reduced system stability. The FLC improves the dynamic response by reducing oscillations and smoothing the torque curve, although minor fluctuations persist under disturbances. Conversely, the M5P controller exhibits the most stable and accurate performance, characterized by rapid settling, minimal overshoot, and strong damping of oscillations. These improvements confirm the M5P approach as the most robust and reliable method for torque regulation in VOC systems. Electromagnetic torque under wind variation. Figure 14 represents the active power response, demonstrates the superiority of the M5P controller in tracking the reference trajectory. While the PI controller exhibits strong oscillations and instability, and the FLC provides moderate improvements with reduced overshoot and faster convergence, the M5P controller achieves near-perfect tracking with minimal oscillations and enhanced robustness against disturbances. This superior performance underscores the effectiveness of the M5P approach in ensuring accurate power regulation under dynamic operating conditions. Active power under wind variation. Figure 15 shows the reactive power regulation and highlights the weaknesses of the PI controller, which shows significant oscillations and transient spikes, particularly during disturbance periods. The FLC improves performance by reducing oscillations and providing smoother tracking, but deviations remain under dynamic conditions. The proposed M5P controller, however, delivers exceptional performance, almost perfectly following the reference with negligible overshoot and steady-state error. This demonstrates its ability to handle nonlinearities and uncertainties effectively, confirming its superiority in VOC applications. Reactive power under wind variation. Performance of system under Q variation Across all Qs variation scenarios, the PI controller consistently demonstrates poor performance with strong oscillations and instability, particularly during transient events. The FLC improves stability and reduces oscillations but still exhibits small deviations under dynamic changes. The M5P controller, however, consistently achieves smooth tracking of the reference with minimal overshoot and fast settling. These results reinforce the effectiveness of the M5P method in suppressing transients and ensuring robust and stable control of reactive power and associated variables. Figure 16 Shows The rotor direct current Idr response under Qs variation, demonstrating the limitations of the PI controller, which exhibits strong oscillations and instability during transient events, resulting in significant deviations from the reference trajectory. Although the FLC improves stability and reduces oscillations, it still shows small deviations under dynamic conditions. In contrast, the M5P-based controller achieves outstanding performance, providing smooth and accurate tracking with minimal overshoot and rapid settling. The zoomed-in sections confirm its ability to suppress transients effectively, highlighting its robustness and superior control capability compared to the traditional and fuzzy approaches. Rotoric direct current Idrunder Qs variation. Figure 17 presents the electromagnetic torque profile under Qs variation, further illustrating the comparative performance of the three controllers. The PI controller suffers from pronounced oscillations and overshoot during transients, compromising system stability. The FLC reduces oscillations and improves damping, but residual fluctuations remain under disturbances. The M5P controller, however, delivers the most stable and accurate torque response, characterized by rapid convergence, minimal overshoot, and strong damping of oscillations. These results confirm the M5P approach as the most reliable method for torque regulation under reactive power variations. Electromagnetic torque under Qs variation. Figure 18 shows the active power response under Qs variation, highlighting the weaknesses of the PI controller, which exhibits significant oscillations and instability during dynamic changes. The FLC improves performance by reducing oscillations and providing smoother tracking, yet it does not fully eliminate deviations. The M5P controller demonstrates superior performance, achieving near-perfect tracking with negligible overshoot and fast settling time. This confirms its effectiveness in ensuring accurate power regulation under varying reactive power conditions. Active power under Qs variation. Figure 19 illustrates the reactive power regulation under Qs variation, showing that the PI controller struggles with large oscillations and transient spikes, particularly during disturbance periods, indicating poor robustness. The FLC improves stability and reduces oscillations but still exhibits minor deviations. The M5P controller, however, achieves exceptional performance, almost perfectly following the reference with negligible overshoot and steady-state error. This demonstrates its capability to handle nonlinearities and uncertainties effectively, reinforcing its superiority in VOC applications. Reactive power under Qs variation. Figure 20 presents the quadrature current Idr response under Qs variation, validating the advantages of the M5P controller. The PI controller exhibits large oscillations and instability during transients, while the FLC significantly improves the response by minimizing overshoot and oscillations. Nevertheless, the M5P controller outperforms both, maintaining excellent stability and almost perfect tracking with negligible fluctuations. This confirms the robustness and precision of the M5P approach in managing current control under reactive power variations. Rotor quadrature current Iqr under Qs variation. Conclusion This study introduced an advanced and smart VOC strategy for DFIG-based wind energy systems, leveraging the M5P model tree algorithm to overcome the significant limitations of conventional PI and FLC controllers. By integrating a data-driven approach into the VOC framework, the proposed method demonstrated superior adaptability and robustness under nonlinear and variable operating conditions. The proposed M5P algorithm used FLC data in the training process and achieved an excellent performance for both currents Idr and Iqr with a Correlation Coefficient: 1.0000 that main a perfect learning, the proposed M5P decision tree algorithm distinguishes of the other machine learnings with the rapidity and clarity where it provide with a seen program in form of tree acceptable to the modifications that allow to incorporate it in the implementation platforms as MATLAB/Simulink. Comparative simulations confirmed that the M5P-based controller significantly improved dynamic response, reduced overshoot, and improved power quality compared to traditional PI and fuzzy logic controllers. These findings underscore the potential of machine learning techniques to revolutionize renewable energy control systems, paving the way for smarter, more resilient wind energy integration into modern power grids. Despite the promising results obtained in simulation, the M5P-based controller presents some limitations that merit further exploration. Its performance depends on the quality and representativeness of the training data, and its robustness under unmodeled dynamics or measurement noise remains to be assessed experimentally. Future work will focus on implementing the proposed control strategy on real-time platforms and extending the approach toward hybrid or adaptive machine learning frameworks to enhance generalization and robustness. Acknowledgements The authors thank contributors and colleagues for support during simulation and manuscript preparation. Author contributions Benyoucef Koudri: Formal analysis, investigation, methodology, resources, software, design, and Writing – original draft.Moualidia Abdelhafidh and Boudana Djamel: Data curation, Funding acquisition, Methodology, Project administration, Supervision, Validation, Visualization, Writing – review and editing.Mujammal Ahmed Hasan MUJAMMAL, Khachafi Sofiane, and Zemirline Nassim: Software, Visualization, Methodology, Writing - Review and Editing.Mohit Bajaj and Oleksandr Rubanenko: Funding, project administration, resources, and evaluation. Funding This work received no external funding. Data availability The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. Code availability MATLAB/Simulink models and WEKA model files used to produce the training dataset and simulations are available from the corresponding author on reasonable request. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Mohit Bajaj, Email: mohitbajaj.ee@geu.ac.in. Oleksandr Rubanenko, Email: rubanenko.o.y@vntu.edu.ua. References Associated Data Data Availability Statement The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request. MATLAB/Simulink models and WEKA model files used to produce the training dataset and simulations are available from the corresponding author on reasonable request.