Extend CPU-based data generation to GPU: efficient load/topology perturbation

[kibaekkim]

Extend the current CPU-based data generation code #2 to utilize GPU acceleration. This is motivated by the need for efficient perturbation of load and topology (such as N-1 contingencies) on GPU hardware.

Scope:

• Efficiently support load and topology perturbations at scale using GPU.
• The work should allow exhaustive studies such as N-1, while benefiting from GPU parallelism.

Approach Exploration:

1. Investigate using MadNLP + JuMP + GPU for direct problem formulation and solving.
2. Explore ExaModelsPower.jl as an alternative implementation and benchmarking tool.
3. Compare the capabilities, performance, and ease of integration for both approaches.

Tasks:

• Analyze the feasibility and performance of MadNLP + JuMP + GPU solution for the target use case
• Analyze how ExaModelsPower.jl can be used for GPU-accelerated power grid studies
• Develop or integrate GPU kernels for load/topology perturbation
• Document design decisions, code changes, integration steps, and benchmarking results
• Provide examples demonstrating large-scale N-1 studies on GPU

Motivation:
Accelerating data generation and study for contingency analysis using GPUs will enable near real-time computation and support large-scale, high-fidelity grid modeling applications.