An improved deep learning algorithm in enabling load data classification for power system

Abstract

Load behaviors significantly impact the planning, dispatching, and operation of the modern power systems. Load classification has been proved as one of the most effective ways of analyzing the load behaviors. However, due to the issues of data collection, transmission, and storage in current power systems, data missing problems frequently occur, which prevents the load classification tasks from precisely identifying the load classes. Simultaneously, because of the diversities of the load categories, different loads contribute various amounts of data, which causes the class imbalance issue. The traditional load data classification algorithms lack the ability to solve the aforementioned issues, which may deteriorate the load classification accuracy. Therefore, this study proposed an improved deep learning algorithm based on the load classification approach in terms of raising the classification performances with solving the data missing and class imbalance issues. First, the LATC (low-rank autoregressive tensor completion) algorithm is used to solve the data missing issue to improve the quality of the training dataset. A Borderline-SMOTE algorithm is further adopted to improve the class distribution in the training dataset to improve the training performances of biGRU (bidirectional gated recurrent unit). Afterward, to improve the classification accuracy in the classification task, the biGRU algorithm, combined with the attention mechanism, is used as the underlying infrastructure. The experimental results show the effectiveness of the proposed approach.