Home

Welcome to the Talk2PowerSystem wiki!

Talk2PowerSystem (a.k.a. CIMon) is a Statnett × Graphwise research project that "democratises power system analytics" by enabling Natural Language Querying over electrical CIM Knowledge Graphs through SPARQL generation, diagrams, time-series federation, and BIM-CIM integration.

Go to cim.ontotext.com/chat for a live demo of the chatbot!

Table of Contents Project Data and Schema Datasets CIM Schema and Ontology Stack External / Non-Semantic Data Talk2PowerSystem Chatbot Overview KGQA Corpus and Code Generation Diagrams Quality Evaluation LLM Experiments and Fine-Tuning BIM-CIM Integration Demos Dissemination Blog Posts

Project

• Project Description (Short): mission, goals, expected benefits, TRL, timeline
• Related Repositories: a list of github repositories related to the project
• ModelOps SchemaOps MLOps: lifecycle of instance data vs schema vs ML models
• Wiki Editing: conventions for this wiki (MediaWiki vs Markdown, semantic line breaks, images)

Data and Schema

Datasets

• Datasets: overview — Nordic44, Telemark-120, Svedala, ENTSO-E, Assets / CIM4Enterprise
  • Dataset Telemark-120: detailed description of the Telemark-120 distribution dataset

CIM Schema and Ontology Stack

• CIM Primer: introduction to CIM containment, voltage levels, substations, lines
• CIM Semantic data: — ontology stack curation, LLM-friendly subsetting, the cimr: inferred extension, the cimd: diagrams ontology, unit / currency / vendor extensions, cims:pragmatics LLM hints, and the instance-data loading pipeline
• Areas and Zones and Bidding Zone: geographical and market areas
• Named Graphs: named-graph layout used in GraphDB
• Inference: cim.pie rules and cimr: shortcut predicates / property chains
• Connection Questions: worked examples of cimr:connectedThroughPart connectivity
• Data Checks: validation queries against the loaded data

External / Non-Semantic Data

• Time Series: federated integration with Cognite Data Fusion via the mRID ↔ external_id contract

Talk2PowerSystem Chatbot

Overview

• Talk2PowerSystem Chatbot: agent design, tools (SPARQL, Autocomplete, N-shot, Display Graphics, Cognite, Now), history vs memory
• Architecture: high-level architecture and Azure infrastructure
• Components-Healthcheck-Corrective Actions: operational monitoring and recovery
• Chatbot Backend Application and Chatbot UI Application: backend & UI implementation details

KGQA Corpus and Code Generation

• KGQA Corpus Generation: how the Q&A corpus is built
  • Manual Templates for Questions: question templates used for generation
• Code Generation: generating code (e.g. pandas / PowerPanda) from CIM

Diagrams

• Diagrams: overview of diagram support in the chatbot
  • Electrical Diagrams: PowSyBl SVG diagrams (SLD, NAD)
  • Visual Graph: embedded GraphDB Visual Graphs

Quality Evaluation

• Quality Evaluation: evaluation methodology
  • LLM and KGQA Evaluation Tools: frameworks and metrics
  • Evaluation Results: measured results across configurations
  • Evaluation Slow Query Performance: analysis and mitigation of slow queries

LLM Experiments and Fine-Tuning

• Experiments: ad-hoc and systematic LLM experiments
• SOTA on Local LLMs: state of the art for locally deployable LLMs
• Fine‐Tuning Experiments: fine-tuning experiments and findings
• Local LLM fine‐tuning Blueprint: blueprint for fine-tuning local LLMs

BIM-CIM Integration

BIM-CIM addresses BIM–CIM integration around Substation data:

• RSE IFC CIM Work
• ICDD as CIM Container
• BIM data for CIM components
• BIM CIM Semantic Interoperability
• BIM CIM Integration Requirements and Technical Design

Demos

Question sets that we tried to work out with the bot:

• Demo0: questions in the original proposal
• Demo1: 5 questions before kickoff
• Demo2: targeting 15 questions for milestone M2
• Demo3: targeting further 20 questions for milestone M2 or later
• Demo4 Telemark 120: queries unique to classes found in Telemark-120 but not in Nordic-44
• Demo5 CimSparql: complex real-world queries used in production
• Demo6 Diagrams: diagram competency questions
• Demo7 Timeseries: time-series competency questions

Dissemination

• Dissemination: timeline of dissemination, communication, exploitation, sustainability
• Training Materials

Blog Posts

• Blog Ontology Simplification for LLM
• Blog Using Semantic Reasoning to Help LLM with SPARQL Generation in Electrical CIM
• Blog Mapping Electrical Resources with GeoSPARQL
• Blog Using Double Optional to Select Preferred Value in SPARQL
• Blog Trying ChatGPT for Electrical CIM Data Validation
• Blog GW Chatbot knows about CIMon