NovaPower

Inspiration

As power systems engineers, understanding complex models is always a challenge due to the sheer number of interacting components and the size of interconnected systems. We built the this project as an intelligent solution to this problem.

What it does

Our system makes it possible to "talk" to the grid, answering complex questions about system stability, critical contingencies, loading, and generation state. It acts as an intelligent assistant that bridges the gap between raw simulation data and actionable engineering insights, transforming grid interaction from manual modeling to intelligent, automated oversight.

How we built it

We built this project using a hybrid architecture that combines pandapower's rigorous simulation engine with cutting-edge generative AI:

• Core Simulation Engine: We use Python and pandapower to handle the mathematical modeling of the electrical grid, allowing for real-time load flow and contingency analysis.
• The "Brain" (Multi-Agent Architecture): We integrated Amazon Nova models (via AWS Bedrock) to power the natural language understanding. Rather than a monolithic model, we implemented a "Map-Reduce" style architecture: local RegionAgents analyze specific grid clusters, while a central Orchestrator synthesizes global decisions. Utilizing the Bedrock Converse API, these agents parse user queries like "Simulate an outage on Bus 5 and tell me if the grid is stable" and convert them into executable functions.
• Backend: A FastAPI server creates the bridge between the frontend requests, the Amazon Nova agents, and the Python simulation logic.
• Frontend: The interactive user interface, which includes visualization and chat, was built with React and Vite.

Challenges we ran into

• Connecting the agents with pandapower: Bridging the gap between unstructured natural language and a strict mathematical simulation engine was difficult. We had to design robust prompt engineering, specialized tools (like our Scenario Builder), and strict function-calling mechanisms to ensure user queries triggered the accurate load flow calculations without error.
• Hallucination Prevention: Ensuring the Amazon Nova agents relied strictly on the hard math and pandapower simulation results, rather than hallucinating plausible-sounding but physically incorrect power values.

Accomplishments that we're proud of

• The Interface: Creating a chat interface that feels like consulting a senior engineering team—a decentralized group of agents that can instantly run the math, analyze regional data, and report back the synthesized results.

What we learned

• Hybrid Systems: The importance of combining the reasoning and delegation capabilities of GenAI with the verifiable "accuracy" of traditional physics-based solvers to create a reliable engineering tool.

What's next

• Visual Topology Editing: Allowing users to add or remove buses and lines simply by dragging and dropping elements in the UI.
• Refining Agentic Chat: Enhancing the regional agents to handle even more complex queries, such as multi-step cascading contingency analysis and automated, grid-wide optimization scenarios.

Built With

• fastapi
• javascript
• node.js
• nova
• python
• react