Master's thesis and Internship : Analysis and evaluation of maximum power point tracking algorithms based on machine learning techniques

Abstract

Maximum Power Point Tracking (MPPT) and Global Maximum Power Point Tracking (GMPPT) consist of techniques that are employed to keep a photovoltaic (PV) array operating at its optimal point. This report first investigates the implementation of two Machine Learning (ML) algorithms, namely Support Vector Regression (SVR) and K-Nearest Neighbors (KNN), under homogeneous irradiance conditions. The performance of these two methods is analysed and compared with that of an improved Perturb and Observed (P&O) process. Several scenarios involving varying solar irradiance and temperature conditions were simulated to evaluate the tracking efficiency of the different methods. The obtained numerical results show that ML methods provide a significant advantage when a temperature measurement is included as input to the model, achieving tracking efficiencies higher than 99%, even in the presence of measurement noise. The case of partial shading is also investigated, highlighting the limitations of conventional MPPT techniques and the need for adapted strategies, such as a Q-Learning approach, to identify the global maximum power point. The results indicate that a key challenge lies in resetting the Q-table, which stores previously encountered states, in order to cope with dynamically evolving outdoor conditions. Finally, the concept of Flexible Power Point Tracking is briefly discussed as a potential extension of MPPT strategies.