Research Papers 论文研究 4h ago Updated 1h ago 更新于 1小时前 43

A Machine Learning Surrogate for Component Criticality Ranking in Interdependent Power-Communication Networks 用于互依赖电力-通信网络组件关键性排序的机器学习代理模型

A Gradient Boosting machine learning surrogate is developed to predict cascading failure severity in interdependent power-communication networks, replacing computationally expensive high-fidelity simulations. The model achieves high accuracy with Spearman correlations of 0.849 for severity prediction and 0.853 for component criticality ranking on the IEEE 118-bus system. Performance is limited by the reproducibility of the ground-truth simulator (MIIM), which also caps at approximately 0.85 corr 针对电力-通信网络级联故障评估计算成本高的问题,提出基于梯度提升的机器学习代理模型以替代高保真模拟器。 在IEEE 118节点系统上,该代理模型在预测故障严重性和组件关键性排名方面均达到约0.85的Spearman相关系数。 模型性能受限于底层采样管道的经验上限,且主要优势来源于层间依赖信息而非单纯拓扑中心性。 确立了“代理模型快速初筛+高保真模拟器选择性验证”的两阶段韧性规划工作流。

55
Hot 热度
70
Quality 质量
60
Impact 影响力

Analysis 深度分析

TL;DR

  • A Gradient Boosting machine learning surrogate is developed to predict cascading failure severity in interdependent power-communication networks, replacing computationally expensive high-fidelity simulations.
  • The model achieves high accuracy with Spearman correlations of 0.849 for severity prediction and 0.853 for component criticality ranking on the IEEE 118-bus system.
  • Performance is limited by the reproducibility of the ground-truth simulator (MIIM), which also caps at approximately 0.85 correlation under the same sampling pipeline.
  • Feature ablation reveals that inter-layer dependency information is the primary driver of the surrogate’s predictive advantage over standard topological centrality measures.
  • The study validates a two-stage workflow where the fast ML surrogate ranks candidates for hardening, reserving the slow simulator only for selective verification.

Why It Matters

This research addresses a critical bottleneck in cyber-physical infrastructure resilience planning: the computational prohibitive cost of evaluating large N-k contingency sets. By providing a reliable, fast surrogate model, it enables practitioners to perform comprehensive risk assessments and prioritize infrastructure hardening efforts that would otherwise be impossible due to time constraints.

Technical Details

  • Methodology: Utilizes the Modified Implicative Interdependency Model (MIIM) as the ground-truth simulator for cascade propagation, training a Gradient Boosting classifier/regressor on leakage-free structural features of the network.
  • Performance Metrics: On the IEEE 118-bus test case, the surrogate demonstrates stability across three independent datasets, achieving Spearman rank correlations of 0.849 (severity) and 0.853 (criticality ranking).
  • Baseline Comparison: Standard topological centrality measures serve as a baseline with lower performance (Spearman 0.60–0.69), highlighting the necessity of incorporating specific interdependency features.
  • Empirical Ceiling Analysis: The study notes that the surrogate’s performance aligns with the intrinsic noise/reproducibility limit of the MIIM simulator itself (~0.85), suggesting the model has reached the maximum possible fidelity given the underlying simulation data.
  • Feature Importance: Ablation studies confirm that the inclusion of inter-layer dependency structures is the key factor distinguishing the surrogate’s performance from simple graph-theoretic baselines.

Industry Insight

  • Workflow Optimization: Infrastructure operators should adopt a hybrid assessment strategy, leveraging lightweight ML surrogates for initial broad-scope screening and resourcing-intensive physics-based simulators only for final validation of top-ranked risks.
  • Data Quality Awareness: When deploying surrogate models for critical infrastructure, it is essential to quantify the inherent variability and reproducibility limits of the ground-truth simulation engine, as the surrogate cannot exceed these empirical ceilings.
  • Feature Engineering Priority: For interdependent systems, models must explicitly capture cross-layer dependencies rather than relying solely on intra-layer topology; ignoring these interactions significantly reduces predictive accuracy for cascading failures.

TL;DR

  • 针对电力-通信网络级联故障评估计算成本高的问题,提出基于梯度提升的机器学习代理模型以替代高保真模拟器。
  • 在IEEE 118节点系统上,该代理模型在预测故障严重性和组件关键性排名方面均达到约0.85的Spearman相关系数。
  • 模型性能受限于底层采样管道的经验上限,且主要优势来源于层间依赖信息而非单纯拓扑中心性。
  • 确立了“代理模型快速初筛+高保真模拟器选择性验证”的两阶段韧性规划工作流。

为什么值得看

本文展示了如何利用轻量级机器学习模型加速复杂物理系统的韧性评估,解决了大规模N-k故障集仿真中的算力瓶颈。对于从事关键基础设施保护、电网安全分析及AI for Science的研究人员而言,这种结合领域知识(MIIM模型)与数据驱动方法的混合范式具有重要的工程应用价值。

技术解析

  • 核心方法:使用Modified Implicative Interdependency Model (MIIM)作为地面真值模拟器,训练Gradient Boosting机器学习代理模型,利用无泄漏的结构特征预测故障严重程度及组件关键性。
  • 性能指标:在IEEE 118-bus系统上,代理模型实现了0.849的严重程度预测Spearman相关系数和0.853的关键性排名相关系数,且在三个独立采样数据集上保持稳定。
  • 基准对比与归因:仅基于全网络拓扑中心性的基线模型Spearman系数仅为0.60-0.69;特征消融实验表明,代理模型的优越性主要由层间依赖信息驱动。
  • 性能天花板分析:研究指出MIIM本身在现有采样管道下的复现Spearman系数约为0.85,表明当前代理模型已达到该采样设置下的经验性能上限。

行业启示

  • 计算效率优化:在关键基础设施韧性规划中,可采用ML代理模型进行大规模初步筛选,仅对高风险候选组件保留高保真仿真,从而显著降低计算资源消耗。
  • 特征工程导向:在跨域耦合系统(如网-信耦合)建模中,显式的层间依赖关系比单纯的拓扑结构特征更具预测力,应在特征工程中予以重点考量。
  • 评估局限性认知:需警惕采样偏差对模型性能上限的影响,当代理模型性能逼近底层模拟器的统计一致性极限时,应优先优化采样策略或引入更复杂的物理约束,而非单纯增加模型复杂度。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Research 科学研究 Machine Learning 机器学习 Security 安全