High-Accuracy Transfer Learning-Based Power Quality Disturbance Recognition Under A Noisy Environment

The smart and sustainable development of power grids has brought about numerous power quality disturbances (PQD) that threaten the efficiency and reliability of the smart grids. Detecting and classifying these disturbances is crucial to taking appropriate corrective action to mitigate possible impacts. While machine learning (ML) algorithms are at the center of attention for most PQD detection problems, their performance is limited due to insufficient samples. Therefore, this paper proposes a new model based on continuous wavelet transform (CWT) and transfer learning to overcome the data challenge. In this regard, raw voltage signals are First transformed to 2D contour images using the CWT. Then, converted images are used to fine-tune a pre-trained model on the Imagenet dataset through transfer learning. The simulation results over 29 groups of PQD under different noise levels show its effectiveness over other ML models. Finally, our code is publicly available in a git repository for enthusiastic readers.

I. Introduction

With the ever-increasing usage of eco-friendly energy sources such as wind and solar, electric vehicles (EVs), and energy storage in power systems, the importance of power quality (PQ) cannot be denied. The prevalence of power electronic equipment in such devices, which brings various Power quality disturbances (PQDs), drives the power system toward new issues in PQ mitigation. Generally, These disturbances can lead to non-stationary waveforms that prompt various issues, including delicate power-equipment failure, damage, and accidentally activated safety switches [1]. All of these can ultimately harm the economics of the power grid. PQDs can be categorized into single and compound disturbances. Single disturbances, as mentioned in [2], can be distinguished into amplitude-based disturbance (voltage-swell, sag, and interruption); transient-based disturbances with high-frequency (impulsive oscillatory and transient); steady-state-based disturbances with low-frequency (voltage-fluctuation) and steady-state-based disturbances with high-frequency (harmonic). Compound disturbances arise when the abovementioned single disturbances occur simultaneously. Detecting and classifying these disturbances is crucial to minimize damage and generate appropriate control actions for rectification. By identifying non-stationarities in waveforms, such as amplitude variation, spectral variations, and signal continuity, methods can be developed to detect and classify PQDs effectively.