AI-based remaining useful life prediction for transmission systems: Integrating operating conditions with TimeGAN and CNN-LSTM networks

Abstract

The remaining useful life (RUL) prediction is key for ensuring the stability of transmission power systems. However, there is no sufficient actual life-cycle, and no mature physics-of-failure model of the power transmission facilities, which make it difficult to predict RUL. In this paper, we propose an AI-based transmission line RUL prediction method which incorporates the measured operating conditions of each line. The proposed method sets the basic linear asset unit as one cable segment and joint boxes on both sides. A feature extraction and piecewise-based RUL model was designed using asset data from 1,458 actual transmission lines accumulated by measuring unit over a period of 44 years. Consequently, the RULs which depend on load operating conditions of target assets can be successfully predicted using CNN-LSTM. In addition, a data augmentation algorithm based on time-series generative adversarial networks was developed to address the issue of imbalanced failure data and further improve the accuracy of RUL prediction. The performance of the proposed RUL estimation method is further verified using real-world data. The proposed method shows an improvement in fault-healthy classification accuracy by 35.72%, 21.43%, and 7.14% compared to existing feature extraction methods, including deep neural networks (DNN), convolutional neural networks (CNN), and autoencoder (AE), respectively. Additionally, when compared to representative deep learning models for RUL estimation, it achieves the best performance with RMSE and Score of 0.074 and 0.066, respectively. Moreover, the proposed method is capable of accurately estimating RUL even for equipment in the early failure period, where the actual operating time is short.