In recent times, the real-time and online identification and prediction of power transformer malfunctions play a vital part in maintaining the dependability and stability of electrical system operations. Early detection of transformer malfunctions is imperative to prevent unexpected outages and potential blackouts. Not only do such occurrences result in financial losses and physical damage, but they also pose a significant threat to the overall stability and functionality of the electrical grid. Therefore, this paper introduces an innovative approach leveraging long short-term memory techniques, coupled with the features of wavelet transformation, to effectively identify and predict internal defects (IDs) in power transformers amidst inrush currents (ICs) in online applications. A key highlight of this approach is its ability to accurately differentiate IDs from ICs within a quarter cycle of the fundamental frequency, enabling real-time fault prediction—a capability overlooked in previous research endeavors. furthermore, a novel technique is proposed to mitigate the adverse impacts of current transformer saturation on the functionality of differential relays. The efficacy of the proposed methodology and the extracted features for deep learning applications is corroborated through a series of MATLAB simulations conducted on a three-phase transformer (33/11 kV, 31.5 MVA). Ultimately, the simulation results and testing validate the remarkable accuracy of 99.65% achieved by the proposed approaches in real-time fault detection for power transformers. This signifies a significant advancement in enhancing the reliability and operational efficiency of power systems
