Multi-Task Learning approach for Fault Location Estimation in a Power Distribution Network

Abstract

Faults in the power distribution system have a great impact on the reliability and quality of the system. However, due to numerous branches and loads in the distribution system, the traditional algorithm shows a high error, and for this reason, research introducing artificial intelligence is becoming active. However, previous machine learning and deep learning methods require a lot of data and are greatly affected by problems such as data loss. In response, a method for faster learning by combining previous learning experiences with current learning has been studied, and Multi-Task Learning (MTL) for learning related tasks at the same time has been studied. In this study, we propose MTL that performs the tasks of fault section classification and fault distance estimation from the three-phase current acquired from the substation and DERs side for minimal input data. In Chapter 2, we select algorithms suitable for classification and regression in fault situations through comparison between conventional algorithm and machine learning methods. The proposed MTL constructs the specific layer of the ANN, extracts features, classifies the fault section, and optimizes it with backpropagation to estimate the fault distance in the specific layer of the XGBoost. This approach is compared in performance with previous machine learning methods that are widely used for fault location estimation by configuring the same data and environment on the IEEE-34 and IEEE-123 node test feeder. On the IEEE-34 node test feeder, compared to single machine learning, both classification and regression performance achieved the highest accuracy, and on the IEEE- ii 123 node test feeder, 100% in classification and relatively good performance in regression were achieved.