TRAISIM

Mon, 01 Jan 2024 00:00:00 +0000

Training Simulator for Power System Operators (TRAISIM)

Funding Agency: CRESYM
Start Date: January 2025
Partners: Cyprus University of Technology (CUT), Réseau de Transport d’Électricité (RTE), Collaborative Research for Energy System Modelling (CRESYM)
Website: sps-lab.org/project/traisim
Code: github.com/SPS-L/

Motivation

Modern power grids are complex cyber-physical infrastructures integrating renewables, distributed resources, and advanced protection. Transmission System Operators (TSOs) must make real-time decisions amid intricate model interactions and emerging AI-support tools. Existing operator training platforms rely on static or simplified modules, reduced-order models, or pre-computed scenario databases — most closed-source and vendor-locked — limiting transparency and excluding smaller TSOs and academia.

Core challenge: Simulate a 6,000-bus network faster than real-time on commodity x86 hardware, within the 1–2 s SCADA refresh constraint, with full model-level fidelity (DAE solvers, protection logic, dynamic loads).

TRAISIM’s answer: An open-source training platform built on Dynawo — algorithmic transparency, no license costs, deployable in existing control centres.

Platform Architecture

TRAISIM integrates six tightly coupled subsystems that deliver a realistic, real-time control-room experience:

Subsystem	Role
Game Master AI	Injects faults & constructs training scenarios
HMI	Control-room user interface
Orchestrator	Time synchronisation & SCADA information flow
Automata	Protection finite-state machines
Physical Simulator	Dynawo / DynaWaltz DAE solver (KINSOL Newton + KLU sparse LU, adaptive timestep)
Trainer	Supervises sessions and injects disturbances

The physical simulator targets the full RTE French Transmission Network (63 kV – 400 kV):

7,075 branches · 335 synchronous generators · 3,039 dynamic loads · 839 load tap changers
Three model fidelity levels: Model 1 (80 k variables, ✓ real-time), Model 2 (210 k, △ marginal), Model 3 (320 k, × requires optimisation)

Benchmarking Results — Year 1 (Phase 1, Jan–Dec 2025)

Using 14 contingency scenarios (Loss of Busbar Section, Loss of Line, Loss of Generator) on the RTE 6,000-bus grid, worst-case solving times versus the 1 s budget:

Model	Configuration	T_sol	Status
M1	LBS Default	0.70 s	✓ Real-time
M2	LBS Default	1.84 s	✗ Exceeds budget
M2	LBS Tuned	0.89 s	✓ Real-time
M3	LBS Default	2.87 s	✗ Exceeds budget
M3	LBS Tuned	1.40 s	✗ Exceeds budget

Key finding: KLU Analyze (symbolic Jacobian factorization) re-analyzes the full sparsity structure on every topology change — consuming 32.5% of solver time and running 2.3× longer in Model 3 vs. Model 2. Tuning fnormtolAlgJ + msbsetAlgJ achieves a 2× speedup for Model 2, but Model 3 (320 k variables) remains infeasible without structural algorithmic changes.

📄 PSCC 2026: “Towards an Open-Source Real-Time Operator-Training Platform: Analysis of Computational Efficiency” — accepted, Limassol, June 8–12, 2026.

Extension 2026 — Work Packages

WP1 · Solver Optimization

Profile and reduce KLU Analyze call frequency; skip re-analysis when discrete events preserve matrix sparsity
Incremental Jacobian structure updates: warm-start BTF permutations & cache sparsity patterns for known post-event topologies
Thread-level parallelism for Euler Evaluation-J & KLU Factor via OpenMP
Target: ~30% overall speedup

WP2 · Adaptive Model Selection (AMS)

AI-driven fidelity control using GATv2 Graph Attention Network with multi-task prediction heads → per-component activity scores
Dynamic fidelity assignment: full detail where needed, Ward-equivalent simplification elsewhere
Trained on ~12,000 dynamic simulations across multiple operating points and N-1 contingencies
Early results: 75.7% continuous variable reduction, 99.35% power prediction accuracy, 99.71% voltage prediction accuracy

WP3 · Orchestrator Co-Design

Slack-time API between orchestrator and simulator: signal available headroom δt = T_sync − T_sol
Dead-time offloading: pre-compute & cache KLU Analyze for anticipated topological changes
Adaptive solver aggressiveness: dynamically tighten/relax tolerances based on real-time headroom budget

WP4 · Validation & Dissemination

Recommendations report with implementation roadmap for upstream Dynawo integration
Open-source release of solver patches & AMS module on SPS-L GitHub
Three peer-reviewed papers: benchmarking (PSCC 2026), solver optimisation (POSYDYS), AMS methodology (journal)

Adaptive Model Selection (AMS)

The AMS module predicts component-level dynamic activity before each training session, selecting only the fidelity needed for the active scenario. A GATv2 graph neural network learns which network components participate actively in disturbance propagation, enabling hybrid models that retain full detail where needed and simplify elsewhere.

Components with activity score > ε retain full-order dynamic models; those below threshold switch to simplified equivalents. Three KPI groups are independently normalised: generator S-power, frequency, and voltage.

Early AMS Results:

Metric	Reduction / Accuracy
Continuous variables	−75.7% (74.5–76.4%)
Discrete variables	−69.2% (67.5–69.7%)
Root functions	−37.4% (37.1–37.6%)
Power prediction accuracy	99.35%
Voltage prediction accuracy	99.71%

Project Timeline

Phase	Period	Milestone
Phase 1 – Task 1	Jan–Apr 2025	Benchmarking methodology; 1 s step validated; headroom metric δt established; 14 scenarios catalogued
Phase 1 – Task 2	May–Aug 2025	Code-level profiling; KLU Analyze identified as primary hotspot (32.5%); 2× speedup via tuning
Phase 1 – Task 3	Sep–Dec 2025	3-model fidelity comparison; PSCC 2026 accepted; Technical Reports 1 & 2 delivered
Phase 2 – WP2 Baseline	Jan–Mar 2026	12 k simulations; GATv2 trained; 75% variable reduction; frequency KPI split identified
Phase 2 – WP2 Refinement	Apr–Jun 2026	Split prediction into 3 models; retrain frequency model; add high-RES operating points
Phase 2 – WP1/WP3/WP4	Apr–Sep 2026	Solver optimisation roadmap; PSCC 2026 presentation; POSYDYS paper; AMS journal planning
AMS Consolidation	Jul–Sep 2026	Full AMS prototype; accuracy–performance trade-off validated; journal paper drafted

Synergy with SPS-Lab Research

TRAISIM is tightly connected to other SPS-Lab projects:

IBM (Interpolation-Based Method): Efficient handling of digital controller-induced discontinuities, enabling accurate simulation of hybrid systems with frequent switching events.
Modeling & Simulation: Scalable, robust simulation tools for large, complex power systems — parallelisation strategies and advanced numerical methods.
TwinEU (Pilot 8, Task 5.5): The core digital twin infrastructure developed within this European project forms the foundation of TRAISIM’s training platform.

Contact: info@sps-lab.org

AI/Machine Learning | Sustainable Power Systems Lab