Design and Implementation of a TinyML-based Predictive Hybrid MPPT System for Photovoltaic Application
Abstract
Solar photovoltaic systems are important for making renewable energy, but their power output changes all the time because of changes in sunlight and temperature. To get the most out of these systems, operation at the Maximum Power Point is required (MPP). Common MPPT techniques like Perturb and Observe (P&O) and Incremental Conductance (INC) have problems like slow tracking, ongoing oscillations at steady state, and lower accuracy, especially when the weather changes quickly or there is partial shading. This study delineates the creation and simulation of a predictive hybrid MPPT system for photovoltaic applications, utilizing Tiny Machine Learning (TinyML) for power. The method uses a small neural network that has been trained on PV data like voltage, current, temperature, and irradiance to accurately guess the best duty cycle for a DC–DC converter. The model is turned into TensorFlow Lite Micro format after training and put on an ESP32 microcontroller. This makes it possible to make predictions in real time with little energy use. An ESP32, an INA219 current sensor, a PWM-controlled DC–DC converter, and a Li-ion battery module were all used to make a working prototype. To test performance, both simulated and real-world experiments were done. The results show that the suggested TinyML-based MPPT method finds the maximum power point faster, has almost no steady-state fluctuations, and increases efficiency by about 8–12% compared to traditional methods, especially when the amount of sunlight changes quickly.