Transformer OLTC Operation Monitoring Framework Through Acoustic Signal Processing and Convolutional Neural Networks

The increasing penetration of renewable energy sources (RESs) has a direct impact on the efficiency and reliability of the power grid. In this context, maintaining voltage stability and power quality has become more important than ever. Transformers are crucial components in the grid and are generally equipped with on-load tap changer (OLTC) mechanisms. With the OLTC, transformers can automatically adapt to different voltage levels depending on the dynamics of the grid. Thereby, they assist in the penetration of RESs and improve system operation efficiency. However, the continuous grid fluctuations require increased activity from OLTCs, which, in turn, contributes to elevated stress on equipment and accelerated aging. Existing OLTC monitoring solutions are generally invasive and time-consuming. Accordingly, this research work presents a new cost-effective OLTC monitoring approach based on audio signals. The approach is inspired by self-paced learning (SPL) strategies. Initially, different monitoring solutions are evaluated and tested with respect to accuracy, computational cost, and robustness to noise. After the best architecture is selected, the audio signal processing approach is further refined through signal processing and convolutional neural networks (CNNs). The proposed approach is benchmarked against automated machine learning (AutoML), which shows a competitive performance, but with increased computational cost. The approach has been validated using real data from a transformer in operation.