|
| A condition assessment method for relay protection devices based on edge intelligence and knowledge distillation |
| DOI：10.19783/j.cnki.pspc.260256 |
| Key Words:protection device edge intelligence deep learning knowledge distillation condition assessment |
| Author Name | Affiliation | | YANG Jiarui | Hebei Key Laboratory of Distributed Energy Storage and Microgrid (North China Electric Power University), Baoding 071003, China | | DAI Zhihui | Hebei Key Laboratory of Distributed Energy Storage and Microgrid (North China Electric Power University), Baoding 071003, China | | ZHANG Hongjia | Hebei Key Laboratory of Distributed Energy Storage and Microgrid (North China Electric Power University), Baoding 071003, China | | LUAN Junhao | Hebei Key Laboratory of Distributed Energy Storage and Microgrid (North China Electric Power University), Baoding 071003, China | | DAI Jiabao | Hebei Key Laboratory of Distributed Energy Storage and Microgrid (North China Electric Power University), Baoding 071003, China | | WANG Xue | Hebei Key Laboratory of Distributed Energy Storage and Microgrid (North China Electric Power University), Baoding 071003, China |
|
| Hits: 29 |
| Download times: 7 |
| Abstract:To address the issues of incomplete data caused by inconsistent sampling intervals among condition assessment indicators of protection devices, as well as the difficulty of deploying deep learning models on edge devices with limited computational resources, a condition assessment method of relay protection devices based on edge intelligence and knowledge distillation is proposed. First, a relay protection edge intelligent architecture composed of cloud platform, edge computing layer, and physical entity layer is constructed to provide support for distributed training and on-site deployment of the model. Second, an assessment indicator set for protection devices is constructed, and the Raindrop graph neural network model is employed to analyze and learn indicators with irregular sampling intervals, thereby establishing a directed graph that characterizes the relationships among the indicator data. Finally, the Raindrop model is selected as the teacher model and the TabNet model as the student model to construct a knowledge distillation framework. The soft-label knowledge distilled by the teacher model is used to guide the learning model training, enabling lightweight deployment of the evaluation method on edge devices. Case study results show that the accuracies of the proposed method on the training set and the test set are 98.98% and 98.15%, respectively. Moreover, it has low deployment complexity and computational requirements, enabling accurate local condition analysis and assessment of the operational status of relay protection devices on edge devices. |
| View Full Text View/Add Comment Download reader |
|
|
|