LLaTSA: Large Language Model-Aligned General-Purpose Transient Stability Analysis
| Source: arXiv AI
Tags: LLM-applications, power-grid, transient-stability, mixture-of-experts, time-series
LLaTSA adapts a pretrained sparse mixture-of-experts LLM to predict power grid transient stability trajectories across heterogeneous systems, encoding operating conditions and disturbance attributes as structured text prefixes to enable generalization without full system-specific retraining.
Details
Data-driven transient stability analysis (TSA) predicts whether a power grid remains stable after a fault, but most models require retraining from scratch when network configurations change. LLaTSA builds on Uni-TSA's general-purpose approach by more carefully aligning numerical trajectory data with LLM representations. It encodes operating conditions, disturbance attributes, and state-variable identity as structured textual prefixes, aligns normalized temporal patches to a TSA-related vocabulary, then processes them through a sparse decoder-only mixture-of-experts backbone. A state-variable coupling module captures coordinated post-fault dynamics that channel-independent models miss, while teacher forcing and rollout-based training support long-horizon prediction. Case studies across multiple test systems demonstrate accurate trajectory prediction and reliable stability discrimination on unseen scenarios. The method's key practical value is avoiding full retraining when grid configurations change — reducing engineering overhead for utilities managing evolving generation mixes. No open code release or deployment timeline is mentioned; real-world integration would require extensive safety validation beyond academic benchmarks.