Condition monitoring and fault detection for electrical power systems using signal processing and machine learning techniques

Abstract

Electrical power systems are essential for sustaining daily activities, economic growth, and societal advancement. However, as the demand for electricity increases, these systems become more complex and prone to faults and disturbances. Faults are unexpected deviations from standard operating conditions that can disrupt operations, incur significant maintenance costs, and lead to system failures if not addressed promptly. Transmission lines, responsible for over 85% of faults in power systems, are particularly vulnerable. Common faults in transmission lines include open circuit and short circuit faults. Open circuit faults disrupt power flow and cause voltage fluctuations due to mechanical stress, environmental factors, ageing infrastructure, or operational errors. Short circuit faults create low-impedance pathways that result in current surges, causing severe consequences such as equipment damage, system instability, power outages, and safety hazards like fires and explosions. The evolution of maintenance strategies in electrical power systems has shifted from reactive to more proactive methodologies, with condition monitoring systems playing a crucial role. These systems continuously monitor the performance and status of electrical equipment to detect early signs of deterioration, faults, or failures. Effective condition monitoring involves collecting real-time data through various sensors and devices, which is then processed using advanced fault detection and classification algorithms. Ensuring effective fault detection and classification is essential for minimising downtime and ensuring the reliability and safety of power systems. Advanced technologies and continuous monitoring play a vital role in mitigating the impact of faults and maintaining the overall health of electrical power infrastructure. These are particularly important for developing countries like Djibouti, which relies heavily on Ethiopia for electricity due to its lack of self-energy production despite having high energy potential. With a population of one million and an electrification rate of 55%, Djibouti aims to achieve 100% electrification by 2035. However, due to an outdated maintenance framework, Djibouti's power system, managed by the Electricity of Djibouti (EDD), faces significant reliability and efficiency challenges. The system suffers from frequent operational disruptions and unscheduled downtimes, with approximately 85% of the total Energy not Distributed (END) due to unplanned outages. Current reactive maintenance practices result in high costs, extended outages, and substantial economic losses. Implementing advanced condition monitoring strategies using modern technologies like wavelet transforms and machine learning can significantly enhance the reliability and efficiency of Djibouti's electrical power systems. These strategies enable proactive identification and resolution of potential faults, reducing downtimes and improving system resilience. This thesis explores and validates the application of these advanced technologies in Djibouti's context, establishing robust fault detection and classification models. This study developed three models to address these challenges: two for fault detection and one for fault classification. The models use the data collected by the simulation conducted on the Djibouti power grid to evaluate various fault scenarios using MATLAB/SIMULINK. The simulation involved modelling short-circuit conditions, specifically three-phase faults, under different settings to observe their effects on the power system. Fault types, including single line-to-ground, double-line-to-ground, and line-to-line, were classified and simulated to assess their impact on the system's voltage and current stability. The first fault detection model employed in the Djibouti power grid utilises the Short-Time Fourier Transform (STFT) for non-stationary signal analysis. This method is instrumental in providing a sensitive and real-time assessment of fault characteristics. By focusing on various types of faults, such as single-line-to-ground (SLG), double-line-to-ground (DLG), and three-phase faults, STFT helps distinguish the specific impacts of each fault type on the power system's reliability and efficiency. The simulation results from the STFT analysis reveal that three-phase and DLG faults display specific high-frequency components upon fault clearance, highlighting their significant transient nature. Conversely, SLG faults exhibit a broad frequency band with lower amplitude, indicating a less distinct transient behaviour. However, the fixed-size window of the STFT poses limitations in capturing the full spectrum of SLG fault characteristics, suggesting that more refined techniques, such as wavelet transforms, may be necessary to improve fault detection accuracy and enhance the system's diagnostic capabilities. The second fault detection model addresses the challenge of detecting minor, unseen defects in transmission lines. It employs a novel fault detection methodology utilising a hybrid wavelet transform approach. This methodology, intended for the Djibouti power grid, combines Stationary Wavelet Transform (SWT) and Continuous Wavelet Transform (CWT) to enhance the detection of abnormal voltage signals caused by short-circuit faults and transient phenomena. The process begins with the decomposition of signals into detail and approximation coefficients using SWT. Shannon's Information Criterion (SIC) determines the optimal decomposition level to represent signal features and prevent overfitting effectively. The signals are then reconstructed using an Algebraic Summation Operation (ASO), which amplifies minor defects, making them more visible for the subsequent application of CWT. The Continuous Wavelet Transform (CWT) revealed previously undetectable frequency components, specifically the 12th, 13th, 14th, and 16th components. The effectiveness of this approach is validated through simulations that use artificial signals designed to mimic specific harmonic disturbances known to occur in power systems. The simulation evaluates various fault scenarios, revealing that the hybrid method can detect and analyse fault types successfully. This comprehensive approach allows for precise fault detection and characterisation, which is crucial for maintaining the stability and reliability of the power system. The fault classification in electrical power systems highlights the advantages of machine learning techniques over traditional methods, particularly for the Djibouti power system. It introduces three machine learning classifiers: Decision Trees (DT), Long Short-Term Memory (LSTM) networks, and Support Vector Machines (SVM). The methodology involves data pre-processing using oversampling, feature extraction via Discrete Wavelet Transform (DWT), and evaluation through k-fold cross-validation. The findings show that SVM, particularly with a polynomial kernel, achieves the highest accuracy and precision among the classifiers. Though less accurate, DT provides high interpretability and can improve with parameter tuning. LSTM performs well with sequential data, showing excellent specificity, though its overall effectiveness is slightly less than that of SVM. Each classifier's performance is analysed using confusion matrices, revealing their strengths and weaknesses in handling different fault types. The chapter concludes that integrating advanced machine learning techniques significantly enhances fault diagnosis and system reliability, advocating for a tailored choice of classifier based on the specific needs of the fault classification task. The importance of condition monitoring, fault detection, and classification in maintaining the stability and efficiency of power systems is underscored. Continuous monitoring allows for early fault detection and timely maintenance, preventing unplanned outages and extending equipment lifespan. Integrating advanced technologies such as unmanned aerial vehicles (UAVs), IoT devices, and machine learning algorithms enhances the effectiveness of fault management strategies. In conclusion, implementing advanced condition monitoring, fault detection techniques, and sophisticated fault classification models can significantly improve the reliability and efficiency of power systems in Djibouti. Future work should explore integrating these classifiers into a hybrid model to enhance fault classification accuracy and reliability further. Practical applications, such as intelligent data collection and decision-making robots, can be developed to ensure a more stable, efficient, and robust power infrastructure.