Forecasting PV-Diesel Microgrid Campus Load Using Machine Learning - The University of the Free State QwaQwa Campus Microgrid

Aug 5, 2026

Upload by Violet

Global energy consumption continues to rise, creating increasingly complex demand patterns that require continuous monitoring and effective control to improve management efficiency. This thesis applies machine learning to forecast short-term electricity load for a photovoltaic (PV)-diesel microgrid at the University of the Free State’s QwaQwa Campus, installed to reduce frequent power shortages in South Africa. Multiple ML models, including LSTM, GRU, hybrid CNN, XGB, and Random Forest regressors, are trained and evaluated using historical consumption data. Results show XGB delivers the strongest accuracy and supports better peak-demand planning, reducing diesel generator reliance.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages94

Reading Time3 h 57 min

AccessFree to download

Abstract

Introduction

Background

Problem Statement

Objectives

Theoretical Framework

Electric Power Systems

Fundamentals of Machine Learning

Related Literatures

Case Study: QwaQwa University Campus

QwaQwa University Campus Overview

Understanding the Power Consumption Patterns at QwaQwa

Understanding the Time series Data

Methodology

Data Collection and Preprocessing

ML Model Development

Use of generative AI

Result and Discussion

Model Results

LSTM

Master's Thesis 2024 30 ECTS Faculty of Science and Technology Forecasting PV-Diesel Microgrid Campus Load Using Machine Learning: The University of the Free State QwaQwa Campus Microgrid David Ajaegbu Data Science Abstract Global energy consumption has been increasing non-stop since the first humans made fire. Energy consumption methods have evolved significantly over the centuries, leading to complex consumption patterns. This complexity requires the monitoring and control of this data to improve energy management and eﬀiciency. With the advent of Machine Learning (ML), coupled with the vast amount of available data today, comes the unique opportunity to make sense of consumption data. Energy consumption prediction can be achieved with data-driven models powered by ML. This thesis used ML techniques to forecast the load of a photovoltaic (PV)-diesel-powered microgrid on the QwaQwa Campus at the University of Free State in South Africa. The hybrid PV-diesel system is installed to mitigate frequent power shortages faced in South Africa. As energy management becomes a critical issue in developing countries, this research aims to optimize energy usage by leveraging predictive models. The methodology involved training multiple ML models, including Long Short-Term Memory (LSTM) networks, Gated Recurrent Units, hybrid Convolutional Neural Networks, Extreme Gradient Boosting (XGB), anda Random Forest regressor, using historical consumption data. These models were evaluated based on their performance in accurately predicting short-term electricity consumption. The results demonstrate that the XGB model achieved the best performance with a Mean Absolute Percentage Error of 3 .9%, Mean Absolute Error of 8.169 and Mean Squared Error of 191.878, suggesting that simpler models can be more effective in certain contexts than complex ones like LSTMs. The XGB model’s performance in accurately predicting peak demands is beneficial for optimizing solar PV output and minimizing diesel generator usage. These findings highlight the potential of ML models in effectively reducing reliance on diesel generators, thereby improving the campus’s financial planning and energy eﬀiciency. Future work could involve improving the robustness of predictive models, including the ability to forecast using real-time data by integrating the models into real-time data streams. This thesis provides a robust and adaptable framework for optimizing hybrid microgrids in university settings and beyond, contributing to the broader discourse on sustainable energy management. Acknowledgements First and foremost, I give thanks to God Almighty, who made the completion of this thesis possible and endowed me with His grace and wisdom throughout this journey. My deep and sincere gratitude goes to my parents Mr. and Mrs. C.O. Ajaegbu for their never-ending support and sacrifice in my journey of life academically. I love you both. To my firm but patient supervisor Prof. Leo Rydin, I want to say a big thank you for driving me to work diligently on this thesis and for encouraging me to be more thorough in this thesis. I am especially thankful to Dr. Maritz, Mr. Armand Bester and the entire team at UFS for your invaluable guidance, support and generosity throughout the development of this thesis. Your comprehensive assistance ensured that my research journey was smooth. Your warm welcome during my visits and your willingness to share your expertise madea significant difference in my learning experience. My appreciation also goes Tomi, Aditya, Awo, Daniel and everyone who offered their help at various stages of this thesis. God bless you all. Contents 1 Introduction 1 1.1 Background ................................ 1 1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3 1.3 Objectives ................................. 4 2 Theoretical Framework 5 2.1 Electric Power Systems ......................... 5 2.1.1 PV-dies

FAQs

What microgrid system is used for the forecasting task in this thesis?

The thesis forecasts the load of a photovoltaic (PV)-diesel-powered hybrid microgrid at the University of the Free State’s QwaQwa Campus. The system is used to mitigate frequent power shortages in South Africa.

Which machine learning models are trained for short-term electricity consumption prediction?

The methodology trains several ML models, including LSTM networks, Gated Recurrent Units, hybrid Convolutional Neural Networks, Extreme Gradient Boosting (XGB), and a Random Forest regressor. Models are evaluated for their ability to predict short-term electricity consumption.

Why does the thesis highlight the XGB model’s performance?

The results indicate XGB achieves the best performance, with a Mean Absolute Percentage Error of 3.9% and strong absolute error metrics. This accuracy is described as useful for optimizing solar PV output and minimizing diesel generator usage, improving energy efficiency and planning.

Aug 5, 2026

Upload by Violet