Knowledge graph-based multi-hazard risk identification and intelligent reasoning framework for distribution networks

Abstract

With the increasing frequency of extreme weather events and the continuous expansion of renewable energy integration, the uncertainty in the operating environment of distribution grids has significantly increased. Multi-hazard events (such as typhoons, heavy rainfall, and lightning) pose severe challenges to the safe operation of power grids. Traditional power grid risk assessment methods typically rely on statistical models or empirical rules, making it difficult to effectively characterize risk propagation relationships within complex grid structures. At the same time, these methods have notable shortcomings in terms of multi-source data fusion and dynamic risk modeling. To address these issues, this paper proposes a multi-hazard risk identification and inference method for distribution grids based on multi-layer knowledge graphs. First, we construct a multi-layer risk knowledge graph comprising spatial, temporal, and causal layers to uniformly model the power grid topology, the evolution of operational states, and fault propagation relationships, thereby achieving a unified representation of multi-source, heterogeneous information. Second, we employ graph neural network models to learn the representations of nodes in the knowledge graph. By leveraging neighborhood information propagation and structural feature extraction, we capture the latent relationships between power grid nodes, thereby enabling accurate identification of risk propagation patterns. Building on this, we utilize the results of graph representation learning to classify and reason about distribution network nodes, achieving the identification and early warning of high-risk equipment nodes. To validate the effectiveness of the proposed method, this paper constructs multi-hazard disaster scenarios, including typhoons, heavy rain, and lightning, on an IEEE 123-node distribution system and evaluates model performance using metrics such as accuracy, recall, and F1 score. Experimental results demonstrate that, compared to traditional machine learning methods, rule-based systems, and standard graph neural network models, the proposed method achieves superior performance in risk identification tasks, with an F1 score reaching 0.89.