Development of a surrogate model using machine learning for current-voltage behavior of dual optical element concentrated photovoltaic system

In Today, the escalating global energy demand and the detrimental environmental impacts of fossil fuels have necessitated a shift towards renewable energy sources. Among these technologies, Concentrated Photovoltaic (CPV) systems possess the potential to offer significantly higher energy conversion efficiencies compared to standard photovoltaic systems by focusing sunlight onto small-area, high-efficiency solar cells using optical elements. However, the multiplicity of optical and geometric parameters affecting the performance of CPV systems, coupled with the complex, non-linear relationships among these parameters, complicates design processes and escalates experimental costs.The primary objective of this thesis is to develop a machine learning-based, multi-layered surrogate model to predict the complex electrical behavior of dual optical element (Fresnel lens) CPV systems and to optimize system performance. The model architecture, developed using experimental data, consists of the integration of an XGBoost (Extreme Gradient Boosting) based Multi-Output Regression layer, optimized via the Optuna framework, and a Neural Network (NN) layer. The developed model utilizes the system's optical parameters (concentration ratios C1, C2 and focal ratios f1, f2 of the primary and secondary lenses) and environmental conditions (irradiance intensity G) as inputs, producing maximum power (Pmax), maximum current (Imax), and maximum voltage (Vmax) values as outputs with high accuracy. A unique contribution of this study is the estimation of the critical "Inter-Lens Distance" (LD) that ensures the system's maximum performance. Through the Neural Network-based approach, a virtual sweep is conducted across a broad parameter space, enabling the identification of optimum LD and optical configuration combinations for varying irradiance levels. This methodology offers the capability to test the system in a "virtual laboratory" environment, thereby minimizing time and cost losses associated with physical prototype manufacturing. The validity and predictive performance of the model were systematically cross-validated against the Akbaba ve Alattawi (1995) I-V curve model, a widely accepted analytical method in the literature. Comparative analyses revealed that the I-V curves generated by the machine learning model exhibited a high correlation with the reference model, particularly in the Open Circuit Voltage (Voc) and Short Circuit Current (Isc) regions. The obtained low normalized Root Mean Square Error (nRMSE) values and minimal deviation rates demonstrate that the model produces reliable results consistent with physical behavior. Consequently, this study presents a data-driven and generalizable decision-support tool that accelerates engineering decisions in the design and optimization processes of CPV systems.