Abstract:
The power distribution network plays a crucial role in delivering electrical power to consumers. However, the lack of feedback mechanisms in legacy power grids prevalent in the Eastern and Southern Africa (ESA) region limits the ability of utility service firms to monitor and respond to abnormal conditions known as electrical faults. This often leads to prolonged blackouts for consumers.
This thesis explores the use of the Internet of Things (IoT) and machine learning to enhance remote fault monitoring on the electrical power distribution grid. The research adopts a design-science approach and investigates both incipient and sudden faults. For incipient faults, the focus is on oil-immersed transformers, with Dissolved Gas Analysis (DGA) data obtained from Kenya Power, a power utility firm in Kenya. The collected data, comprising 2912 records, is subjected to exploratory analysis, confirming its suitability for machine learning training. A machine learning model named KosaNET is developed, trained, and tested, demonstrating superior performance in detecting incipient faults compared to other algorithms, particularly for multinomial data.
To address sudden faults, a LoRa-based platform is developed and deployed. Current sensors are integrated into the distribution grid, and a LoRa-enabled microcontroller transmits data regularly via a gateway. Experimental results reveal that the LoRaenabled platform successfully triggers an alert at the network monitoring centre within approximately 100 msec of a fault occurrence.
A machine-learning multinomial classification model was also offloaded to the edge of the network for condition monitoring of power transformer units. The model was built using the edge impulse service. From the results, an accuracy of 99.9% was attained, which shows that the edge-deployed machine learning model DGA can perform well even when deployed on the periphery of the network.
The original contributions of this study are manifold. Firstly, a machine learning model is developed to accurately interpret DGA data and detect incipient faults. Secondly, a
xiv
LoRa-enabled platform is designed, analysed, and tested for fault monitoring. Thirdly, an intuitive duty cycle is implemented to maximise battery lifespan. Fourthly, data processing and storage functions are offloaded to the edge of the network rather than relying solely on cloud-based solutions. Lastly, future research directions are proposed for further advancements in fault monitoring on the power distribution grid.
Overall, this research demonstrates the potential of IoT and machine learning to enhance fault monitoring and detection, paving the way for more efficient and proactive maintenance of the power distribution grid.