Abstract
The rapid growth of demand of electrical energy and depletion of fossil opened door for renewable energy. The expeditious broadening of Photovoltaic (PV) energy has attracted the private and government precinct world-wide, due to the reduction of costs and being cleaner source of energy. However, most of the maximum power point tracking (MPPT) controllers are inefficient under rapidly changing environmental conditions. In addition, under partial shading conditions (PSC) MPPT controllers fail to track global maximum power point (GMPP). Therefore, it is essential to propose an MPPT controller that will be able to locate GMPP using historical weather data. The work is undertaken to investigate the feasibility of using machine learning (ML) based MPPT techniques to harness maximum power on a PV system under PSC. If successful, the ML based MPPT algorithms could lead to a reduction in power losses in a PV system. In this dissertation, certain contributions to the field of PV systems and ML based were made by introducing two online and eleven artificial intelligence (AI) MPPT techniques, by presenting four experiments under different weather conditions. The first contribution is concerned with an MPPT system that harvests maximum power under PSC, using Johannesburg real-time weather data. The system consists of an MPPT controller cascaded with a PID controller, to reduce errors of the MPPT algorithms, to improve the system’s performance. First, the system evaluates and compares the online [Perturb & Observe (P&O) and Incremental Conductance (INC)], to determine the most powerful MPPT algorithm. Secondly, the system validates the performance of the eleven AI MPPT methods [Fuzzy Logic Control (FLC) and Recurrent Neural Network (RNN), Support Vector Machines (SVM), the Weighted K-nearest neighbour (WK-NN), a Gaussian process regression (GPR), Decision Tree (DT), Multivariate linear regression (MLR), Linear discriminant analysis (LDA), Naïve Bayes classifier (NBC), Bagged Tree (BT) and Boosted Tree (BoT)] under PSC. The ML based techniques are evaluated using four types of error [root mean squared error (RMSE), mean absolute error (MAE), Mean squared error (MSE) and R-squared (𝑅2)]. For the first experiment, online methods are empirically compared in the form of four case studies conducted under various weather conditions. The results showed that INC performed significantly better than P&O under PSC. INC overperformed P&O in all case studies in terms of power extraction. Nevertheless, P&O has less settling time around maximum power point...
M.Phil. (Electrical and Electronic Engineering)