MS2RSHAT-Net: A hybrid Shoebill optimized deep learning framework for accurate electrical fault prediction

Electrical faults arise from insulation failure, aging, and disturbances, causing instability, traditional and deep learning methods show limitations in performance. To overcome these boundaries, this paper presents a novel hybrid Multi-Scale Sparse Recurrent Neural Network (RNN) Stack Hierarchical Attention Transformer Network (MS 2 RSHAT-Net) for precise electrical fault detection and classification. The steps involved data preprocessing using missing value imputation and Exploratory Data Analysis (EDA) using distribution charts. For efficiently represent the data, feature engineering techniques applied Min-Max scaling and feature extraction are applied. For capturing both local and global temporal trends in fault behavior, a hybrid MS 2 RSHAT-Net incorporates a multi-scale sparse recurrent stack for learning fine-grained transient patterns and long-term degradation drifts, while a hierarchical attention transformer for adaptively re-weights temporal contexts across scales, ensuring computational efficiency for embedded or edge deployment and emphasizing diagnostically salient epochs. Additionally, the Shoebill Optimization Algorithm (SOA) is employed to optimize the hyperparameters. The proposed model is executed by Python and evaluated using Electrical Fault Detection and Classification dataset. Experimental results show proposed technique achieves high accuracy of 99.61 %, precision of 99.58 %, recall of 99.53 %, and F1-score of 99.56 %, outperforming existing techniques.