Search

DC microgrids (DCMGs) are gaining popularity mainly in shipboards and isolated power systems for their modest control and higher efficiency. However, the newer conception and less literature are forcing the operators to take up tough challenges for smoother operation of the overall system. Protection areas of DCMGs are still immature because of the inexperience of handling DC faults and less availability of DC circuit breakers (DCCBs) in outsized range. In this work, the fault detection and classification issue in low-voltage-DC-microgrid (LVDCMG) has been solved with a bilayered machine learning scheme based on weighted K-nearest neighbor (WKNN) and decision tree (DT). WKNN accurately detects the fault in the line, and DT classifies the fault as PG or PP for further corrective measures. The developed technique has been implemented in an LVDCMG with PV and battery energy sources. Voltage and current measurements from the network have been used as training samples to the classifier models. It has been noticed that the proposed bilayered protection scheme is giving 100% detection accuracy. The outcome of the proposed structure seems to be promising and has been compared with some of the existing machine learning (ML)-based fault detection techniques for microgrids (MGs). [ABSTRACT FROM AUTHOR]

The operation of photovoltaic (PV) systems, like any other system, in fault free environment ensures maximum performance. Hence, accurate and timely fault identification of PV system demands greater importance. Present investigation proposes a stacking ensemble-based PV array fault diagnosis method, which is integrated with automatic feature engineering and selection technique, handling of imbalanced dataset for unbiased classification. The proposed ensemble utilizes decision tree (DT), random forest (RF), extra trees (EXT), extreme gradient boosting machine (XGBoost) as base learners, and light gradient boosting machine (LightGBM) as both base and meta-learner. To validate the proposed technique, a test system of 4.8kWp capacity has been built using the MATLAB/Simulink environment incorporating one-year real-time irradiance and module temperature data. Irradiance, module temperature, voltage, current, and power are collected as primary raw data which are then concocted using the autofeat python library for automatic feature engineering and selection. Subsequently, the proposed fault diagnosis strategy is built in python 3.8.5 using the engineered and class label balanced dataset, prepared using the synthetic minority over-sampling technique (SMOTE). Results demonstrate promising performance of the proposed ensemble technique with 97.04% accuracy in classifying the different faults in the PV array, which accounts for approximately 1–3% improvement over other individual machine learning (ML) models. [ABSTRACT FROM AUTHOR]