• A new dynamic architecture enables real-time modeling of power system equipment interdependencies. • The method uses efficient fine-tuning, updating only 0.1% of parameters to slash computational costs. • It cut fault warning time by 96.4% and reduced GPU memory usage by 78%. The digital transformation of power systems is hindered by inadequate modeling of dynamic equipment interdependencies. Conventional knowledge graphs, limited by static structures, cannot capture temporal state evolution and suffer from the computational cost of full-parameter fine-tuning, causing update latencies that exceed practical maintenance windows. This study introduces a dynamic knowledge graph architecture using Low-Rank Adaptation (LoRA). Instead of adjusting trillions of parameters, our method freezes over 90% of a pre-trained model's parameters and fine-tunes only 0.1% via low-rank matrices. This enables real-time modeling of nonlinear coupling relationships. Key innovations include a dynamic tensor decomposition module that maps multivariate time-series data into a time-variant low-rank space. A conditional edge update mechanism also autonomously triggers subgraph reconfiguration when local parameter variations exceed set thresholds. During the testing process, it is found that the framework achieves a 9-minute advance fault warning (compared to the baseline method), with a 96.4% improvement in warning capability. At the same time, the model update inference latency is reduced to 9.3ms, and GPU memory usage is reduced by 78%. This provides a viable pathway for developing power system digital twins with causal reasoning.
Huina Guo (Fri,) studied this question.