Page 1 of 19

Transactions on Machine Learning and Artificial Intelligence - Vol. 10, No. 6

Publication Date: December, 25, 2022

DOI:10.14738/tmlai.106.13645. Nansamba, S., & Harb, H. (2022). Developing a Neural Network Based Fault Prediction Tool for a Solar Power Plant in Uganda.

Transactions on Machine Learning and Artificial Intelligence, 10(6). 52-70.

Services for Science and Education – United Kingdom

Developing a Neural Network Based Fault Prediction Tool for a

Solar Power Plant in Uganda

Salmah Nansamba

Engineering Institute of Technology, Perth, WA 6005, Australia

Hadi Harb

Engineering Institute of Technology, Perth, WA 6005, Australia

ABSTRACT

Solar photovoltaic (PV) systems are one of the fastest growing renewable energy

technologies and plenty of research has been and continues to be carried out in this

domain. Maximization of solar PV power plant production, efficiency and return on

investment can only be achieved by having adequate and effective maintenance

systems in place. Of the various maintenance schemes, predictive maintenance is

popular for its effectiveness and minimization of resource wastage. Maintenance

activities are scheduled based on the real time condition of the system with priority

being given to the system components with the highest likelihood of failure. A good

predictive maintenance system is based on the premise of being able to anticipate

faults before they occur. In this study therefore, a fault prediction tool for a solar

plant in Uganda is proposed. The hybrid tool is developed using both feed forward

and long short term memory neural networks for power prediction, in conjunction

with a mean chart statistical process control tool for final fault prediction. Results

from the study demonstrate that the feed forward and long short term memory

neural network modules of the proposed tool attain mean absolute errors of 4.2%

and 6.9% respectively for power production predictions. The fault prediction

capability of the tool is tested under both normal and abnormal operating

conditions. Results show that the tool satisfactorily discriminates against the fault

and non-fault conditions thereby achieving successful solar PV system fault

prediction.

Keywords: Machine learning; Feedforward Neural Networks; Long Short Term Memory;

Solar fault prediction.

INTRODUCTION

There is an ever growing shift from the use of conventional power sources to renewables such

as wind and solar energy sources. Photovoltaic has an ever replenishing, vast and clean energy

supply and is therefore the most promising renewable energy source [1]. Solar photovoltaic

systems are subject to various environmental constraints that result in faults. These faults can

potentially lower the annual power generated by a PV system by about 18.9% [1]. It is therefore

important that adequate maintenance is done in solar PV systems so as to maximize energy

production and the use of these systems till their end of life thereby ensuring the highest return

on investment. Maintenance strategies are crucial towards maintaining the technical and

financial performance of solar photovoltaic systems. The early detection and prediction of

Page 2 of 19

53

Nansamba, S., & Harb, H. (2022). Developing a Neural Network Based Fault Prediction Tool for a Solar Power Plant in Uganda. Transactions on

Machine Learning and Artificial Intelligence, 10(6). 52-70.

URL: http://dx.doi.org/10.14738/tmlai.106.13645

faults in photovoltaic systems is crucial towards increasing the reliability, efficiency and

lifetime of solar photo voltaic systems [1]. The lack of a predictive component in some

maintenance systems however is a hindrance to the minimization of downtime costs. In Uganda

particularly, grid connected solar PV generation is on the rise, with four solar plants

contributing up to 50MW annually, equivalent to 21.277 GWh [2]. There is still a lot of potential

for growth, it is therefore prudent that the existing solar PV systems are adequately maintained

to ensure efficient performance. The existing solar power plants have Supervisory Control and

Data Acquisition (SCADA) systems that monitor crucial plant parameters such as AC power

output, string current and voltage and module temperature.

Without automatic analysis of these huge volumes of important collected data in order to

predict faults, there remains a limitation in the development of the most effective predictive

maintenance work plans for the solar plants. This is a gap that needs to be filled and as such

forms the basis of this research project. In this paper, a fault prediction system for a solar power

plant is proposed. The system is based on machine learning techniques, namely Feedforward

Neural Networks and Long Short Term Memory.

FAULT DETECTION IN SOLAR PV SYSTEMS

Faults in Solar PV Systems

Faults in solar photovoltaic systems are failures that happen in the solar modules, electrical

conductors and other system components such as the inverters. Generally, the operating

lifetime of photovoltaic panels is set to 25 years by manufacturers. However, decline in power

output performance is usually observed before this time elapses and is attributed to the faults

that occur in the system [3]. During the course of operation, the photovoltaic system is exposed

to unfavourable conditions; solar cells may experience overheating owing to the higher

ambient temperatures, dirt and droppings on solar modules can cause shading that results into

hotspot problems in solar modules. Inverters on the other hand, owing to their major electrical

component composition are prone to sudden total failure [4]. These resulting solar system

faults not only lower the PV system efficiency but also its useful life. A brief discussion of PV

system faults as presented in [5] follows.

Arc faults

Arc faults typically occur when fault current is discharged into air and other dielectrics. This

can be as a result of a break in the electrical conductors which causes discontinuity in the

electrical circuit. It can also be as a result of insulation damage which leads to conductors of

different voltages being in close proximity with each other. Arc faults can potentially destroy

entire PV strings and also cause electrical fires [5].

Line to line faults

Line to line faults occur when there is an accidental connection between two points at different

potentials in a PV string. They can also arise as a result of the creation of an undesired path of

low impedance between two points. This can be due to various causes for example conductor

aging and DC connector damage. Line to line faults are dangerous because they not only destroy

PV strings and the system conductors but are also potential sources of electrical fires [4] [5]

[6] [7].

Page 3 of 19

54

Transactions on Machine Learning and Artificial Intelligence (TMLAI) Vol 10, Issue 6, December - 2022

Services for Science and Education – United Kingdom

Ground faults

Also referred to as zero efficiency faults, ground faults are the commonest fault type in electrical

systems. They occur as a result of breakdown in insulation of the equipment, which causes

leakage currents to flow to the ground [7].

Hot spots

Photovoltaic hotspots occur when there is reverse bias operation of one or more solar cells as

a consequence of solar cell degradation and cell mismatch. The cells then dissipate the power

produced instead of transmitting it and this eventually causes abnormally high cell

temperatures. Consequently, the output power of the entire PV module gradually declines.

Hotspots often result in open circuits and damage of the affected solar cells [5] [8].

Diode faults

Bypass diodes serve to facilitate the flow of current around cracked or shaded solar cells

thereby maintaining optimal power production by the cell. When these diodes are not exposed

to sufficient air flow to enable cooling, they undergo thermal stress which makes them defective

[9].

PV array faults

The faults in the PV panels themselves can be temporary or permanent. Temporary faults occur

as a result of the accumulation of dirt and dust on the panels while some of the permanent

defects’ causes include burnt solar cells, cracking and yellowing of the cells, bubbles and

delamination. Normal production in the event of permanent faults can be restored by replacing

the faulty modules. Temporal faults on the other hand do not always necessitate replacement

of the affected components; for example, cleaning the dirty solar panels can be effective in

restoring optimum power production [5].

Machine Learning

Machine learning is a subset of artificial intelligence approaches that is focused on training

computer systems to gain the ability to learn from data and make meaningful predictions and

classifications. No explicit programming is required for machine learning. The resulting

algorithms have the ability to adapt to new situations in form of new data that is presented to

them and this facilitates the making of more accurate classifications and predictions. Machine

learning is classified into supervised and unsupervised approaches. Under supervised learning,

labelled datasets are used to train the algorithms while unlabelled data sets are used in training

for unsupervised learning. Both traditional machine learning and deep learning techniques find

a lot of application in the photovoltaic arena such as in maximum power point tracking,

photovoltaic design parameter extraction, modelling of photovoltaic modules, solar irradiance,

solar power production forecasting, photovoltaic storage system energy management and

anomaly detection [10]. This study is focused on the application of machine learning to solar

power production and fault prediction.

Predictive Maintenance in Solar PV Systems

Preventive maintenance is about mitigating system failure by intervening to equipment before

they fail. It is hinged on the premise that before a system component actually fails, it bears signs

of deterioration; whether visible or not. Focus is therefore on the ability to perceive these signs

and this is accomplished by monitoring and analysing defined performance parameters in