On Deep Machine Learning Based Harmonic Localization Using High resolution Merging Units

Harmonic distortion in electrical power systems poses significant challenges, leading to power losses and potential damage to transmission and distribution equipment. Traditional methods for detecting and identifying harmonic sources often lack the precision required for effective power quality management. This study introduces an advanced approach that combines high-resolution data acquisition with sophisticated computational techniques to enhance the detection and identification of harmonic sources. Utilizing high-performance merging units (MUs) from Ecophi, capable of sampling at 28.8kHz, we collected detailed voltage and current data from three feeders over a one-month period. We applied Fast Fourier Transform (FFT) preprocessing to segment and transform the time-domain signals into frequency-domain data, focusing on harmonic content while excluding the fundamental frequency. This processed data served as input to a three-layer attention-based Long Short-Term Memory (LSTM) network designed to model temporal dependencies and prioritize significant features for accurate classification. The model was trained on 20 days of data and tested on the remaining 10 days, achieving an average classification accuracy of 95% in identifying the origin of harmonic disturbances across the feeders. Notably, the high accuracy was maintained even with unbalanced data distributions, demonstrating the model’s robustness. This integrated approach offers a practical and effective solution for real-time monitoring and analysis of power quality issues, with significant implications for industry practitioners aiming to manage harmonic distortions and enhance the reliability of electrical power systems