Machine learning for accelerating the discovery of high-performance low-cost solar cells

Bhatti, Satyam (2023) Machine learning for accelerating the discovery of high-performance low-cost solar cells. MPhil(R) thesis, University of Glasgow.

Abstract

Solar energy has the potential to enhance the operation of electronic devices profoundly and is the solution to the most important challenge facing humanity today. Such devices primarily rely on rechargeable batteries to satisfy their energy needs. However, since photovoltaic (PV) technology is a mature and reliable method for converting the Sun’s vast energy into electricity, innovation in developing new materials and solar cell architectures is becoming more important to increase the penetration of PV technologies in wearable and IoT applications. Moreover, artificial intelligence (AI) is touted to be a game changer in energy harvesting. The thesis aims to optimize solar cell performance using various computational methods, from solar irradiance and solar architecture to cost analysis of the PV system. The thesis explores the PV cell architectures that can be used for optimized cost/efficiency trade-offs. In addition, machine learning (ML) algorithms are incorporated to develop reconfigurable PV cells based on switchable complementary metal-oxide-semiconductor (CMOS) addressable switches, such that the output power can be optimized for different light patterns and shading.

The first part of the thesis presents a critical literature review of a range of ML techniques applied for estimating solar irradiance, followed by a review on accurately predicting the levelized cost of electricity (LCOE) and return on investment (ROI) of a PV system and lastly, presents a systematic review (SR) on the discovery of solar cells. Furthermore, the literature review consists of a thorough systematic review that reveals that ML techniques can speed up the discovery of new solar cell materials and architectures. The review covers a broad range of ML techniques that focus on producing low-cost solar cells. Additionally, a new classification method is introduced based on data synthesis, ML algorithms, optimization, and fabrication process. The review finds that Gaussian Process Regression (GPR) ML technique with Bayesian Optimization (BO) is the most promising method for designing low-cost organic solar cell architecture. Therefore, the first part of the thesis critically evaluates the existing ML techniques and guides researchers in discovering solar cells using ML techniques. The literature review also discusses the recent research work done for predicting solar irradiance and evaluating the LCOE and ROI of the PV system using various time-series forecasting techniques under ML algorithms.

Secondly, the thesis proposes an ML algorithm for accurately predicting solar irradiance using the wireless sensor network (WSN) relying on batteries that need constant replacement and are hazardous waste. Therefore, WSNs with solar energy harvesters that scavenge energy from the Sun are proposed as an alternative solution. Consequently, the ML algorithms that enable WSN nodes to accurately predict the amount of solar irradiance are presented so that the node can intelligently manage its energy. The nodes use the panel’s energy to power its internal electronic components, such as the processor and transmitter, and charge its battery. Accordingly, this helps the node access an exact amount of solar irradiance predictions to plan its energy utilization more efficiently, thereby adjusting the operation schedule depending on the expected solar energy availability. The ML models were based on historical weather datasets from California, USA, and Delhi, India, from 2010 to 2020. In addition, the process of data pre-processing, followed by feature engineering, identification of outliers, and grid search to determine the most optimized ML model, is evaluated. Compared with the linear regression (LR) model, the support vector regression (SVR) model showed accurate solar irradiance forecasting. Moreover, from the predicted output calculated results, it was also found that the models with time duration of 1 year and 1 month have much better forecasting results than 10 years and 1 week, with both root square mean error (RMSE) and mean absolute error (MAE) less than 7% for California, USA.

Consecutively, the third part of the thesis evaluates the parameter LCOE using demographic variables. Moreover, LCOE facilitates economic decisions and quantitative comparisons between energy generation technologies. Previous methods for calculating the LCOE were based on fixed singular input values that do not capture the uncertainty associated with determining the financial feasibility of a PV project. Instead, a dynamic model that considers important demographic, energy, and policy data that include interest rates, inflation rates, and energy yield is proposed. All these parameters will undoubtedly vary during a PV system’s lifetime and help determine a more accurate LCOE value. Furthermore, comparisons between different ML algorithms revealed that the ARIMA model gave an accuracy of 93.8% for predicting the consumer price of electricity. Moreover, the proposed model with two case studies from the United States and the Philippines is evaluated in detail. Results from these case studies revealed that LCOE values for the State of California could be almost 30% different (5.03 ¢/kWh for singular values in comparison to 7.09¢/kWh using our ML model), which can distort the risk or economic feasibility of a PV power plant. Additionally, the ML model predicts the ROI of a grid-connected PV plant in the Philippines to be 5.37 years instead of 4.23 years which gives a clear indication to the client for making an accurate estimation for the cost analysis of a PV plant.

• Item Type: Thesis (MPhil(R))
• Qualification Level: Masters
• Subjects: T Technology > TA Engineering (General). Civil engineering (General); T Technology > TK Electrical engineering. Electronics Nuclear engineering
• Colleges/Schools: College of Science and Engineering > School of Engineering > Electronics and Nanoscale Engineering
• Supervisor's Name: Hussain, Dr. Sajjad
• Date of Award: 2023
• Depositing User: Theses Team
• Unique ID: glathesis:2023-83618
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 31 May 2023 09:23
• Last Modified: 31 May 2023 09:27
• Thesis DOI: 10.5525/gla.thesis.83618
• URI: https://theses.gla.ac.uk/id/eprint/83618

Related URLs

• Article DOI