Hybrid machine learning and deep learning framework for multi-stage fault diagnosis in modern power systems

Modern power system especially smart grids, integrating with different renewable energy resources, distributed generations, communication networks, and control systems, operate in islanded or grid-connected modes becoming more complex with time. This introduces a new challenge in faults detection during transition of modes which leads to changing behavior and characteristics of faults. This paper proposed a AI-based multi-stage fault diagnosis framework for modern power system by presenting hybrid random forest and convolutional neural network (HRF–CNN) Model. Unlike previous approaches relay on single model for faults detection, classification and localization (FDCL). Proposed hybrid model is not used only for detection of faults also determine Islanded or Grid-Connected mode of operation in smart grid. In RF model, which consists of two output layers, one for detection of faults to determine grid status (Normal or Fault), as well as a second one classify the operational mode of the smart grid (Islanded or Grid-Connected). If a fault occurs as recognized by the RF, then moves to two distinct CNN models separately. The first classifies a classification of faulty bus, while the second use for the classification of faults from among a set of 11 predefined short circuit faults. We employed Modified IEEE 14-Bus system integrated with renewable energy sources on MATLAB/Simulink to generate varied fault datasets as well as to carry out extensive simulations under different fault scenarios. The outcome validates that the integrated system accurately detects grid status, determines smart grid mode, detects the location of faults, as well as classifies types of faults in grid connected as well as in islanded modes. The proposed framework achieves 98% accuracy of fault detection, 95% accuracy of mode identification, 80% accuracy of fault bus localization, and 87% accuracy of fault classification, outperforming conventional machine learning and deep learning models such as SVM, KNN, and RNN. The results demonstrate that the HRF–CNN framework provides a scalable and efficient solution for fault diagnosis in complex smart grid environments.