Detection Framework for Few-Shot Voltage Disturbances in Power Systems Using Evaluation Learning

Data-driven detection models for voltage disturbance in power systems have performed exceptionally over the years. However, they require a large number of samples for training, and the monitoring systems that are operated in substations cannot easily meet this requirement. Therefore, few-shot voltage disturbance detection is urgently needed. In this study, a new detection framework and its episodic training mechanism are proposed to address the above need. The proposed framework is comprised of a short-time Fourier transform (STFT), a feature extractor, and a feature evaluator. STFT converts a voltage disturbance into a time-frequency matrix to enrich its information. Then, the feature extractor learns features from the time-frequency matrix. Finally, the feature evaluator identifies the voltage disturbance by evaluating the similarities between its feature and the features of a support set. An episodic training mechanism is proposed to enable the framework to learn how to evaluate similarities between the voltage disturbance to be detected and the voltage disturbance in the support set. Experiments on two measured voltage disturbance datasets demonstrate that the framework can detect a voltage disturbance type on the basis of a few samples of that type in the support set. The proposed framework provides a competitive solution for the detection of few-shot voltage disturbances and other similar electricity data that are difficult to collect.