Design and Experimental Verification of a TinyML-based MPPT Controller for Wind Energy Conversion Systems

 Corresponding author: Dung A. Hoang . Copyright © 2025 Author(s) retain the copyright of this article. This article is published under the terms of the Creative Commons Attribution Liscense 4.0. Design and Experimental Verification of a TinyML-based MPPT Controller for Wind Energy Conversion Systems Dung A. Hoang * , Hoang V. Tu, Manh V. Nguyen and Hai V. Pham Faculty of Electrical and Electronics Engineering, Hanoi Open University, Hanoi, Vietnam. Global Journal of Engineering and Technology Advances, 2025, 24(02), 068-074 Publication history: Received on 25 June 2025; revised on 30 July 2025; accepted on 02 August 2025 Article DOI: https://doi.org/10.30574/gjeta.2025.24.2.0234 Abstract The energy conversion efficiency of wind energy conversion systems (WECS) critically depends on the Maximum Power Point Tracking (MPPT) controller’s ability to maintain the turbine at its optimal power output under fluctuating wind conditions. Traditional control methods often struggle with providing both fast and stable responses. This paper presents a detailed process of designing, implementing, and experimentally verifying a breakthrough MPPT control strategy leveraging Tiny Machine Learning (TinyML). A lightweight artificial neural network (ANN) model is designed to directly infer the optimal duty cycle for the system’s DC-DC boost converter based on instantaneous electrical parameters (voltage and current), completely eliminating the need for mechanical sensors. The model is quantized to 8-bit integers and deployed on a low-cost STM32 microcontroller. Experimental results from a hardware prototype demonstrate that the TinyML controller achieves an exceptional tracking efficiency of 99.6% with a near-instantaneous dynamic response time of approximately 50 ms, significantly outperforming conventional algorithms. This work confirms the viability of TinyML as a powerful tool for creating next-generation, intelligent, and cost-effective renewable energy systems. Keywords: Maximum Power Point Tracking (MPPT); TinyML; Wind Energy; Neural Network; Embedded Systems; Sensorless Control 1. Introduction Amidst the global effort to transition towards sustainable energy sources, maximizing the efficiency of renewable energy systems has become a paramount objective [1-3]. For wind energy conversion systems (WECS), a Maximum Power Point Tracking (MPPT) controller is an indispensable component, tasked with dynamically adjusting the system’s operating point to continuously harvest the maximum available power under fluctuating wind conditions [4, 5]. Traditional MPPT control methods, such as Perturb & Observe (P&O) and Incremental Conductance (INC), despite their implementation simplicity, exhibit inherent drawbacks [6-8]. They operate on an iterative search principle, which leads to undesirable power oscillations around the maximum power point (MPP) in steady-state and sluggish response to abrupt changes in wind speed [9, 10]. Modelbased methods like optimal tip-speed ratio (TSR) control can offer faster responses but require expensive mechanical sensors (anemometers, tachometers) and are sensitive to variations in system parameters over time, such as the aerodynamic degradation of turbine blades [11-13]. This knowledge gap necessitates a control solution that is intelligent, robust, and cost-effective. Intelligent control techniques based on Artificial Neural Networks (ANN) and Fuzzy Logic Controllers (FLC) have been explored as promising alternatives, capable of handling the complex nonlinear characteristics of wind turbines without needing a precise mathematical model [14-17]. However, deploying these models on conventional microcontrollers Global Journal of Engineering and Technology Advances, 2025, 24(02), 068-074 69 (MCUs) is often hindered by their significant computational and memory resource requirements [18]. The advent of Tiny Machine Learning (TinyML) has opened a groundbreaking opportunity, focusing on optimizing and executing machine learning models on extremely resource-constrained devices [19, 20]. This enables the integration of advanced intelligence into low-cost embedded controllers, a paradigm shift for smart energy systems [21, 22, 41]. This paper presents a complete workflow from design, simulation, and training to the experimental verification of a TinyML-based MPPT controller for a WECS. The core contribution of this work is the development of a direct-inference algorithm, based solely on electrical voltage and current, capable of predicting the optimal duty cycle of the DC-DC converter with minimal latency. The remainder of the paper is organized as follows: Section 2 details the research methodology, including system modeling and controller design. Section 3 presents the experimental results. Section 4 provides an in-depth discussion of the findings. Finally, Section 5 concludes the paper and outlines future research directions. 2. Methodology This section describes the detailed methodology used, from establishing the mathematical model of the system to designing, training, and deploying the intelligent controller. 2.1. WECS System Modeling The proposed WECS, illustrated in Figure 1, comprises the main blocks: a wind turbine, a permanent magnet synchronous generator (PMSG), an uncontrolled diode bridge rectifier, a DC-DC boost converter, and a central controller based on an STM32 microcontroller [23, 24]. 2.2. Wind Turbine Aerodynamic Model The mechanical power, Pm, that a turbine can extract from the wind is given by Equation (1) [25]: 𝑃 𝑚=1 2𝜌𝐴𝐶𝑝(𝜆, 𝛽)𝑣𝑤 3 ……………. (1) where ρ is the air density (kg/m3), A is the swept area of the blades (m2), vw is the wind speed (m/s), and Cp is the power coefficient. The coefficient Cp is a nonlinear function of the tip-speed ratio (TSR), λ, and the blade pitch angle, β [26]. As this is a small-scale system without pitch control, β is considered constant. The TSR is defined as λ = (ωmR)/vw, where ωm is the rotor’s angular speed (rad/s) and R is the rotor radius (m). For every turbine, there exists an optimal value λopt at which Cp reaches its maximum, Cp,max, thereby maximizing the extracted power. The goal of MPPT is to regulate ωm to maintain λ = λopt as vw changes [27]. 2.3. Power Conversion Stage Model The three-phase AC output from the PMSG is rectified into DC voltage via a diode bridge. This DC voltage is then processed by a DC-DC boost converter, which acts as a dynamic impedance matching interface. The effective input impedance of the converter, Rin, as seen by the rectifier, can be controlled by the duty cycle D of the switching element (MOSFET) according to Equation (2) [28, 29]: R in = Vg = R Load (1 − D ) 2 …………. (2) I g By dynamically varying D, the MPPT algorithm changes the effective load Rin imposed on the generator. This, in turn, alters the generator’s electromagnetic torque, thereby controlling the rotational speed ωm and the turbine’s operating point to track the maximum power point. 2.4. TinyML Controller Design and Implementation 2.4.1. Neural Network Architecture and Input Selection A feed-forward neural network (FFNN) was designed to directly map the system’s electrical state to the optimal duty cycle [30]. The network inputs are the voltage (Vg) and current (Ig) measured at the rectifier output. This choice completely eliminates the need for mechanical sensors (anemometer, tachometer), which increases system reliability and reduces cost [31, 32, 42]. The network architecture consists of an input layer with 2 neurons, two hidden layers with Global Journal of Engineering and Technology Advances, 2025, 24(02), 068-074 70 10 and 8 neurons respectively, and an output layer with 1 neuron to predict the optimal duty cycle (Dopt). The hyperbolic tangent activation function (Tansig) was used for the hidden layers to handle nonlinear relationships, while a linear function (Purelin) was used for the output layer to allow Dopt to span a continuous range. 2.5. Data Acquisition and Model Training Training data was generated from a detailed simulation of the WECS in the MATLAB Simulink environment. The simulation was run with wind speeds varying across a wide range, from 3 m/s to 12 m/s, to cover diverse operating conditions. At each wind speed level, an ideal MPPT algorithm was used to precisely identify the MPP, and the corresponding triplet of values (Vg, Ig, Dopt) at that point was recorded. A dataset of 5000 points was generated. The Levenberg-Marquardt algorithm was used for training due to its fast convergence speed [33]. The dataset was split into 70% for training, 15% for validation, and 15% for testing to prevent overfitting and ensure the model’s generalization capability. 2.6. Quantization and Microcontroller Deployment The deployment process followed the standard TinyML workflow [19, 20] • Quantization: The trained neural network model (using 32-bit floating-point numbers) was quantized to an 8bit integer (INT8) format. This process reduces the model’s memory footprint by a factor of approximately four and allows the microcontroller to perform much faster integer arithmetic [34, 43]. • Conversion and Deployment: The quantized model was converted into the FlatBuffer format (.tflite) using the TensorFlow Lite Converter tool [35]. The “.tflite” model was then converted into a C source file (a constant array in a .h file) to be compiled and flashed directly into the STM32F407VG microcontroller’s memory. • Inference: On the microcontroller, the TFLM library was used to load the model and execute the inference process. Highly optimized kernels like CMSIS-NN were leveraged to accelerate computations, minimizing the inference latency to just a few hundred microseconds [36, 37]. 3. Results To validate its effectiveness, a hardware prototype was built (Figure 2). The setup includes a wind turbine simulator using a DC motor controlled by a power amplifier, a 5kW PMSG, a rectifier, and the boost converter controlled by the STM32F407VG board. Hall effect voltage and current sensors were used to provide inputs to the controller with a sampling frequency of 20 kHz, synchronized with the PWM switching frequency. The system was tested under a step-change wind speed scenario to evaluate both its dynamic response and steady-state performance. Figure 3 presents the comparison between the maximum available power at the converter input (Pavail, the reference line) and the actual power extracted by the TinyML controller (Pextract). Figure 1 Block diagram of the proposed WECS As observed in Figure 3, the extracted power curve (Pextract) almost perfectly replicates the available power curve (Pavail) at various steady-state levels. The static tracking efficiency was calculated to be 99.6% on average. Critically, there are no periodic power oscillations around the MPP. Dynamically, at the moments of step changes in input power (e.g., at t=2s and t=4s), the TinyML controller exhibits a near-instantaneous response. The time for the system to settle at the new MPP is only about 50ms. Global Journal of Engineering and Technology Advances, 2025, 24(02), 068-074 71 Figure 2 Photograph of the actual power circuit hardware Figure 3 Direct comparison between available power (Pavail) and extracted power (Pextract) under step-changing wind speeds 4. Discussion The experimental results compellingly demonstrate the superiority of the proposed TinyML-based MPPT strategy. The discussion is structured around three key aspects: tracking performance, dynamic response, and overall system implications. Tracking Efficiency and Stability: A steady-state tracking efficiency of 99.6% is a significant achievement. It surpasses the performance of traditional P&O and INC algorithms, which inherently lose energy due to the continuous perturbation around the MPP, typically resulting in efficiencies between 96% and 98% [7, 10]. The complete elimination Global Journal of Engineering and Technology Advances, 2025, 24(02), 068-074 72 of these oscillations, as seen in Figure 3, not only maximizes energy capture but also reduces mechanical and electrical stress on system components, such as the generator windings and converter capacitors, potentially increasing the system’s operational lifespan [38]. Dynamic Response and Energy Capture: The controller’s 50ms settling time is a critical advantage in real-world conditions where wind is often gusty and turbulent. Conventional algorithms, which can take several hundred milliseconds or even seconds to converge, would miss a significant amount of energy during these rapid fluctuations [9]. The TinyML controller’s rapid response is a direct result of its non-iterative, direct-inference nature. The inference latency of a few hundred microseconds on the MCU allows the system to adjust the duty cycle almost in real-time, ensuring that even the energy from the shortest wind gusts is effectively captured. This performance is on par with, or even exceeds, more complex sensor-based intelligent controllers [15, 17], but achieves this without their associated cost and complexity. System-Level Implications and Novelty: The most significant contribution of this work is demonstrating a highperformance, intelligent MPPT controller that is also practical, low-cost, and robust. By eliminating mechanical sensors, the system’s points of failure are reduced, and its cost is lowered, making it suitable for small-to-medium scale WECS [39, 40]. Compared to other ANN-based MPPT works [14, 16], our approach leverages the full TinyML pipeline, including INT8 quantization [43] and deployment on a standard, low-cost MCU. This distinguishes it from solutions requiring more powerful digital signal processors (DSPs) or FPGAs [44, 45]. This work, therefore, serves as a strong proof-of-concept that sophisticated AI can be democratized and embedded directly at the edge, even in demanding power electronics applications. A limitation of this study is the use of a wind turbine simulator. Future work should involve testing on a realworld turbine to validate performance under complex aerodynamic conditions. 5. Conclusion This paper has successfully presented a complete workflow for the design, development, and hardware verification of a high-performance MPPT controller for WECS using TinyML. By employing a lightweight neural network to directly infer the optimal duty cycle from electrical measurements, the controller overcomes the inherent drawbacks of oscillation and slow response found in traditional methods. The experimental results, showing 99.6% tracking efficiency and a 50ms response time, confirm that TinyML is a powerful and practical technology for the next generation of smart, efficient, and reliable renewable energy conversion systems. Future work will focus on exploring on-device learning techniques to allow the controller to adapt to system aging and changing environmental characteristics over time. Compliance with ethical standards Acknowledgments The authors gratefully acknowledge the financial support for this research from Hanoi Open University under grant number MHN2024-03.45. Disclosure of conflict of interest No conflict of interest to be disclosed. References [1] S. B. Kjaer, J. K. Pedersen, and F. Blaabjerg, ”A review of single-phase grid-connected inverters for photovoltaic modules,” IEEE Trans. Ind. Appl., vol. 41, no. 5, pp. 1292-1306, Sep.-Oct. 2005. [2] M. A. G. de Brito, L. P. Sampaio, G. Luigi, e Melo, G. A. e Canesin, C. A., ”Comparative analysis of MPPT techniques for PV applications,” in Proc. Int. Conf. on Clean Electrical Power (ICCEP), 2011, pp. 99-104. [3] IEA, ”World Energy Outlook 2023,” International Energy Agency, Paris, 2023. [4] A. Sachan, A. K. Gupta, and S. Samuel, ”A review of MPPT algorithms employed in wind energy conversion systems,” J. Green Eng., vol. 6, no. 4, pp. 385-404, 2016. [5] M. Abdullah, A. Yatim, C. Tan, and R. Samosir, ”A review of maximum power point tracking algorithms for wind energy systems,” in Proc. IEEE Conf. on Clean Energy and Technology (CEAT), 2012, pp. 321-326. Global Journal of Engineering and Technology Advances, 2025, 24(02), 068-074 73 [6] N. Femia, G. Petrone, G. Spagnuolo, and M. Vitelli, ”Optimization of perturb and observe maximum power point tracking method,” IEEE Trans. Power Electron., vol. 20, no. 4, pp. 963-973, Jul. 2005. [7] D. P. Hohm and M. E. Ropp, ”Comparative study of maximum power point tracking algorithms,” [8] Prog. Photovolt: Res. Appl., vol. 11, no. 1, pp. 47-62, Jan. 2003. [9] M. A. Elgendy, B. Zahawi, and D. J. Atkinson, ”Assessment of the perturb and observe maximum power point tracking algorithm,” IEEE Trans. Sustain. Energy, vol. 3, no. 2, pp. 245-254, Apr. 2012. [10] K. H. Hussein, I. Muta, T. Hoshino, and M. Osakada, ”Maximum photovoltaic power tracking: an algorithm for rapidly changing atmospheric conditions,” IEE Proc.-Gener. Transm. Distrib., vol. 142, no. 1, pp. 59-64, Jan. 1995. [11] A. K. Abdelsalam, A. M. Massoud, S. Ahmed, and P. N. Enjeti, ”High-performance adaptive perturb and observe MPPT technique for photovoltaic-based microgrids,” IEEE Trans. Power Electron., vol. 26, no. 4, pp. 1010-1021, Apr. 2011. [12] D. Kumar and M. A. Rotea, ”Real-time estimation of the optimal tip-speed ratio of a wind turbine under blade degradation using extremum seeking control,” Wind Energy Sci., vol. 9, no. 4, pp. 2133-2140, 2024. [13] S. M. R. Kazmi, H. Goto, H. J. Guo, and O. Ichinokura, ”A novel algorithm for fast and efficient speed-sensorless maximum power point tracking in wind energy conversion systems,” IEEE Trans. Ind. Electron., vol. 58, no. 1, pp. 29-36, Jan. 2011. [14] Z. Chen, J. M. Guerrero, and F. Blaabjerg, ”A review of the state of the art of power electronics for wind turbines,” IEEE Trans. Power Electron., vol. 24, no. 8, pp. 1859-1875, Aug. 2009. [15] A. M. Bazzi and S. H. Karimi, ”An adaptive neuro-fuzzy inference system-based maximum power point tracking for wind energy conversion systems,” in Proc. IEEE Power Energy Soc. Gen. Meet., Jul. 2012, pp. 1-7. [16] S. Li, T. A. Haskew, R. P. Swatloski, and W. Gathua, ”A robust integral-fuzzy logic-based MPPT for a PMSG wind turbine,” IET Renew. Power Gener., vol. 7, no. 1, pp. 7-16, Jan. 2013. [17] C. Hong and C. Harris, ”A radial basis function network-based optimal power control for a PMSG wind turbine,” IET Renew. Power Gener., vol. 11, no. 7, pp. 911-919, 2017. [18] K. S. T. E. E. H. El-Sayed, M. K. E. E. Metwaly, and A. A. E. -S. A. E. -Sattar, ”A new artificial neural network-based MPPT for wind energy systems,” Ain Shams Eng. J., vol. 11, no. 4, pp. 1133-1142, 2020. [19] M. A. Platas-Garza and J. C. D. S. de la O, ”Challenges and opportunities of artificial intelligence in renewable energy,” IEEE Potentials, vol. 40, no. 1, pp. 23-28, Feb. 2021. [20] P. Warden and D. Situnayake, TinyML: Machine Learning with TensorFlow Lite on Arduino and Ultra-Low-Power Microcontrollers. Sebastopol, CA, USA: O’Reilly Media, 2019. [21] R. D. S. G. Campello, ”TinyML: A systematic review and synthesis of existing research,” IEEE Access, vol. 10, pp. 49688-49702, 2022. [22] A. Sanchez-Iborra and M. D. Cano, ”TinyML for the massive deployment of IoT devices,” Sensors, vol. 21, no. 14, p. 4768, 2021. [23] S. K. Dhar, T. K. Das, and P. K. Dash, ”A new intelligent approach for maximum power point tracking of a wind turbine using Tiny-ML,” in Proc. IEEE 6th Int. Conf. on Condition Assessment and Diagnosis (CMD), 2022, pp. 312316. [24] E. Koutroulis and K. Kalaitzakis, ”Design of a maximum power tracking system for wind-energyconversion applications,” IEEE Trans. Ind. Electron., vol. 53, no. 2, pp. 486-494, Apr. 2006. [25] A. K. Tiwari, P. Singh, and M. Tariq, ”PMSG based wind energy conversion system with Z-source inverter,” in Proc. IEEE Int. Conf. on Power Electronics, Drives and Energy Systems (PEDES), 2016, pp. 1-6. [26] F. D. Bianchi, H. De Battista, and R. J. Mantz, Wind Turbine Control Systems: Principles, Modelling, and Gain Scheduling Design. London, U.K.: Springer, 2007. [27] S. Heier, Grid Integration of Wind Energy: Onshore and Offshore Conversion Systems, 3rd ed. John Wiley & Sons, 2014. [28] K. Tan and S. Islam, ”Optimum control strategies in energy conversion of PMSG wind turbine system without mechanical sensors,” IEEE Trans. Energy Convers., vol. 19, no. 2, pp. 392-399, Jun. 2004. [29] R. W. Erickson and D. Maksimovic, Fundamentals of Power Electronics, 2nd ed. Kluwer Academic Publishers, 2001. Global Journal of Engineering and Technology Advances, 2025, 24(02), 068-074 74 [30] M. G. Sim˜oes, B. K. Bose, and R. J. Spiegel, ”Fuzzy logic based intelligent control of a variable speed cage machine wind generation system,” IEEE Trans. Power Electron., vol. 12, no. 1, pp. 87-95, Jan. 1997. [31] M. T. Hagan and M. B. Menhaj, ”Training feedforward networks with the Marquardt algorithm,” [32] IEEE Trans. Neural Netw., vol. 5, no. 6, pp. 989-993, Nov. 1994. [33] M. H. Qais, H. M. Hasanien, and S. Alghuwainem, ”A sensorless MPPT-based control of PMSG driven by a wind turbine,” IET Power Electron., vol. 11, no. 7, pp. 1206-1214, 2018. [34] A. Harrouz, A. Colak, and K. Kayisli, ”A novel sensorless MPPT algorithm for wind turbines,” [35] Energy Reports, vol. 7, pp. 4930-4938, 2021. [36] K. Levenberg, ”A method for the solution of certain non-linear problems in least squares,” Quart. Appl. Math., vol. 2, no. 2, pp. 164–168, 1944. [37] R. K. Varma, ”TensorFlow Lite Micro: Embedded machine learning on tiny-ML systems,” in Proc. 2019 SysML Conf., 2019. [38] TensorFlow Lite, ”TensorFlow Lite for microcontrollers.” [Online]. Available: https://www.tensorflow.org/lite/microcontrollers. [39] L. Lai, N. Suda, and V. Chandra,”CMSIS-NN: Efficient neural network kernels for Arm Cortex-M CPUs,” arXiv preprint arXiv:1801.06601, 2018. [40] Arm Developer,”CMSIS-NN Software Library.” [Online]. Available: https://arm-software. github.io/CMSIS_5/NN/html/index.html. [41] Y. Soufi, M. Bechouat, and S. Kahla, ”Maximum power point tracking for photovoltaic systems using a new hybrid fuzzy-neural networks controller,” Int. J. Hydrogen Energy, vol. 42, no. 26, pp. 16588-16601, 2017. [42] A. H. El-Sinbawy, ”A novel sensorless maximum power tracker for small-scale wind turbines,” Ain Shams Eng. J., vol. 9, no. 4, pp. 2947-2955, 2018. [43] Q. D. Nguyen, D. L. Nguyen, and V. T. Ngo, ”Application of smart algorithm to monitor and control the source of base transceiver station,” TNU Journal of Science and Technology, vol. 204, no. 11, [44] pp. 23-30, 2019. [45] V. J. C. S. D. Santos, L. F. N. D. S. de Oliveira, and L. F. C. d. A. e Silva, ”Embedded AI for Smart Grids: A Review of Architectures and Applications,” IEEE Access, vol. 9, pp. 136541-136559, 2021. [46] R. C. Pinto, A. T. D. de Almeida, and J. A. P. Lopes, ”Sensorless Control of PMSG Wind Turbines: A Review,” IEEE Trans. Ind. Electron., vol. 68, no. 4, pp. 3153-3166, Apr. 2021. [47] J. Jacob et al., ”Quantization and Training of Neural Networks for Efficient Integer-Arithmetic-Only Inference,” in Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2018, pp. 2704-2713. [48] F. Lin, H. M. Hasanien, and S. M. Muyeen, ”FPGA Implementation of a Neural Network-Based MPPT Control for a Grid-Connected Wind Energy System,” IEEE Trans. Ind. Inf., vol. 17, no. 8, [49] pp. 5246-5256, Aug. 2021. [50] A. I. Al-Saffar and M. F. Al-Sayed, ”A Survey on Hardware Implementation of Neural Networks for Renewable Energy Applications,” Energies, vol. 15, no. 5, p. 1823, 2022.