High-Performance Computing–Assisted Modeling and Real-Time Analysis of Electrical Power Networks and Industrial Control Systems

Abstract

This study examined the performance, robustness, and real-time feasibility of high-performance computing–assisted modeling and integrated analytical pipelines for electrical power networks coupled with industrial control systems. A quantitative experimental design was implemented using 360 structured run instances distributed across small, medium, and large benchmark grid tiers, multiple telemetry regimes, and standardized contingency libraries. Dependent outcomes included update latency, sustained throughput, state-estimation error, contingency ranking stability, and anomaly detection delay, while independent variables included compute architecture type, parallelism level, grid size tier, telemetry regime, contingency volume, and disturbance severity. Mixed-effects regression models with clustered random intercepts were used to quantify performance drivers. Results showed that mean update latency across the full sample was 148.2 ms (P95 = 268.0 ms), while sustained throughput averaged 6.8 updates per second. Hybrid CPU–GPU configurations reduced latency by 24.63 ms relative to CPU-only baselines (p < .001) and increased throughput by 1.18 updates per second. Large-grid tiers increased latency by 72.84 ms and reduced throughput by 4.27 updates per second compared with small-grid tiers (p < .001), confirming scale as the dominant workload driver. Convergence rates averaged 96.9%, indicating stable numerical robustness across tiers. Analytical quality improved under richer telemetry, with PMU-only configurations reducing estimation error by 0.008 per unit relative to SCADA-only (p < .001). Contingency ranking stability averaged 0.88, while missed critical contingencies averaged 0.37 per run under standardized runtime budgets. Anomaly detection delay averaged 312.5 ms and increased significantly under higher disturbance severity and larger grid tiers. Overall, the findings demonstrated that real-time cyber-physical feasibility in power–ICS environments was primarily governed by grid scale, telemetry design, and workload composition, while compute acceleration and parallelism provided statistically significant but workload-dependent performance improvements.