CIM ModelOps vs SchemaOps vs MLOps – A Power System Digital Twin Perspective

ModelOps

In our work with Talk2PowerSystem, we use the term ModelOps to describe the lifecycle of changes to the electrical CIM instance data: the operational model of the power system represented in the Knowledge Graph. For example:

• A new transmission line is built and energized.
• That line must now be reflected in the operational model used for simulations, market calculations, and planning studies.

This is about managing the evolving state of the grid to maintain an up-to-date digital twin for system studies.

SchemaOps

Parallel to that, we have SchemaOps, focused on changes to the schema (CIM/CGMES ontology) itself:

• Extending CIM to support new equipment types or concepts.
• Aligning the schema with other standards or adding new details driven by regulatory changes or market design updates.

Both processes are critical — but distinct — in the lifecycle of power system models.

MLOps

Interestingly, when looking at the MLOps community (https://ml-ops.org/), we see many similar challenges:

• Managing model versions
• Ensuring data consistency
• Validating operational correctness
• Integrating reasoning or rule-based logic
• Monitoring "model drift", i.e. long-term degradation of performance due to slow-changing instance data, enumerations and ontologies

The project includes a similar-named task, described as follows:

T2.3 ML Operations This task deals with the technical aspects of ML and LLM deployment.

• Deploy open (locally deployable) LLMs. This is the foundation of reaching TRL 7 since Statnett data cannot leave the operational environment. But we can use more powerful LLMs to generate training datasets for less powerful LLMs.
• Deployment (inference) architectures (GPU vs CPU e.g. Llama.cpp or Ollama) and optimizations (e.g. quantization). Experimentation with various LLM models and sizes (variants), taking into account required features and tradeoffs (e.g. between memory/speed and NLQ performance)
• LLM fine tuning and adaptation (e.g. Low Rank Adaptation, LoRA). This increases LLM performance without requiring a full retraining.
• ML ops: ML server configuration (we expect that Statnett will provision the servers). LLM deployments. Monitor server performance. Selection and optimization of LLM size and parameters. Training and retraining.

Conclusion

There is a lot we can leverage from MLOps best practices when defining our own ModelOps and SchemaOps: especially as we start adding reasoning rules and automated analyses into the Knowledge Graph.

The key differences:

• CIM ModelOps: Managing the state of the power system, not a machine learning / statistical model.
• SchemaOps: Maintaining semantic integrity as the underlying CIM schema evolves.
• MLOps: Managing machine learning models that adapt, learn, and drift over time.

Yet, in the end: they all face the same core DevOps challenges:

• Lifecycle management
• Validation
• Production deployment

This perspective strengthens the bridge between traditional power system modeling and modern DevOps-driven approaches.