Artificial Intelligence based Sensor Data Analytics Framework for Remote Electricity Network Condition Monitoring

Abstract

Rural electrification demands the use of inexpensive technologies such as single wire earth return (SWER) networks. There is a steadily growing energy demand from remote consumers, and the capacity of existing lines may become inadequate soon. Besides, the existing SWER networks are very inefficient and experience poor voltage regulation. Furthermore, high-impedance arcing faults (HIF) from SWER lines can ignite bushfires such as the catastrophic 2009 Black Saturday bushfires in Victoria (Australia). Replacing SWER lines by cables as recommended by the Victorian Royal Commission comes at an astronomical cost and service providers are not able to comply with. As a solution, reliable remote electricity networks can be established through splitting the existing system into microgrids, and existing SWER lines can be utilised to interconnect those microgrids. The development of such reliable networks with better energy demand management will rely on having an integrated network-wide condition monitoring system. As the first contribution of this thesis, a distributed online monitoring platform is developed that incorporates power quality monitoring, real-time HIF identification and transient classification in SWER network. Characteristic features are extracted from the current and voltage signals, and Artificial Intelligence (AI) based classification techniques are developed to classify the faults and transients. The proposed approach demonstrates higher HIF detection accuracy (98.67%) and reduced detection latency (115.2 ms). Secondly, to facilitate electricity demand management, a remote consumer load identification methodology is developed to detect the load type from its turn-on transients. An edge computing-based architecture is proposed to facilitate the high-frequency analysis for load identification with minimum data transmission. Computationally efficient load identification methodologies are developed to enable their real-time deployment on resource constrained devices. The proposed approach is evaluated in real-time, and it achieves an average accuracy of 98% in identifying different loads. Finally, a deep neural network-based energy disaggregation framework is developed to separate the load specific energy usage from an aggregated signal. A generative approach is applied to model energy usage patterns. The proposed framework is evaluated using a real-world data set. It improves the signal aggregate error by 44% and mean aggregate error by 19% in comparison with other state-of-the-art techniques.