COLLIDE — AI-Powered BTM Data Center Site Selection

An AI siting platform that jointly scores candidate sites across land viability, gas supply reliability, and BTM power economics — surfacing a ranked set of locations with quantified risk, cost estimates, and sensitivity analysis.

Built for the ASU Energy Hackathon 2026.

---

What it does

Hyperscale AI data centers need 50–500 MW of power 24/7. Grid interconnection queues now stretch 3–7 years. The alternative — behind-the-meter (BTM) natural gas generation — requires evaluating three interlocking constraints simultaneously:

• Sub-A: Land viability — zoning, fiber proximity, water access, parcel size, flood zone
• Sub-B: Gas supply reliability — pipeline failure probability, hub distance, curtailment risk
• Sub-C: Power economics — BTM spread (LMP minus generation cost), spread durability

COLLIDE scores every candidate site across all three dimensions in seconds, streams a composite scorecard with 20-year NPV estimates, and lets an AI analyst answer what-if questions in natural language.

---

Tech stack

Layer	Technology
Frontend	React 18, Vite, Leaflet, Recharts
Backend	Python FastAPI, async SSE
AI orchestration	LangGraph + LangChain
LLM	Anthropic Claude Haiku + Sonnet
ML models	Random Forest (land, power durability), Gaussian KDE (gas), GMM (regime)
Live data	GridStatus (ERCOT LMP), EIA Open Data, CAISO OASIS
Web enrichment	Tavily API
Data pipeline	DuckDB, Parquet, Pandera, 10 public APIs
Deployment	Vercel (experimentalServices: Vite + FastAPI)

---

ML models

Land viability — Random Forest Classifier

• 300 trees, max depth 8, trained on parcel attribute data
• Top features: highway access (28.5%), substation proximity (23.7%), pipeline distance (19.9%), water distance (14.6%), fiber proximity (8.1%)
• SHAP explainer included for per-site feature attribution

Gas reliability — GPU Gaussian KDE

• Bandwidth 0.5, trained on PHMSA pipeline incident coordinates (severity-weighted)
• Scoring: incident density 40% + interstate pipeline proximity 35% + Waha distance 25%
• Falls back to PHMSA density grid when PyTorch unavailable

Power durability — Random Forest Classifier

• 200 trees, max depth 6
• Top features: LMP level (68.4%), market regime encoding (30.9%), Waha price (0.8%)

Market regime — Gaussian Mixture Model

• 3 components, full covariance, trained on ERCOT 5-feature historical data
• Labels: normal (cluster 0), wind_curtailment (cluster 1), stress_scarcity (cluster 2)
• Training mean: LMP $78/MWh, wind penetration 31%, demand 54.9 GW

Composite score — TOPSIS

• Weights: land 30%, gas 35%, power 35% (user-adjustable)
• 20-year NPV via Monte Carlo (10,000 scenarios, 8% WACC, 100 MW plant)

---

Features

• Interactive map — click any coordinate for a full scorecard; right-click for context menu
• Scorecard panel — land/gas/power breakdown, SHAP explanations, NPV P10/P50/P90, streaming AI narrative
• Compare mode — up to 5 sites side by side
• Grid optimizer — draw a bounding box, stream top-N candidates
• AI Analyst — LangGraph agent with 5 intents: stress test, compare, timing, explain, configure
• Live market data — Waha gas price, ERCOT + CAISO LMP, regime classification (updated every 5 min)
• 72-hour LMP forecast — Moirai model, P10/P50/P90 confidence bands
• Heat layers — composite / gas / LMP score overlays on the map

---

Data sources

Source	Dataset	Cadence
EIA-930	BA demand, net generation	15 min
EIA Open Data	Waha + Henry Hub gas prices	Daily
CAISO OASIS	SP15, NP15, Palo Verde LMP	5 min
NOAA NWS	Phoenix weather forecast	Hourly
BLM GeoBOB	Federal land ownership (AZ, NM, TX)	Static
FCC HIFLD	Dark fiber and FTTP availability	Static
USGS NHD	Water body proximity	Static
FEMA	100-year flood zone boundaries	Periodic
EIA Gas	Pipeline routes ERCOT + WECC	Periodic
PHMSA	Gas pipeline incident records	Periodic

---

Project structure

collide/
├── src/                    # React frontend
│   ├── components/         # 28 UI components
│   ├── hooks/              # 13 custom hooks (useEvaluate, useAgent, ...)
│   ├── docs/               # /docs/* documentation pages
│   └── data/               # Static site definitions
├── backend/
│   ├── scoring/            # Sub-A, Sub-B, Sub-C + TOPSIS + cost
│   ├── agent/              # LangGraph agent (graph.py, tools.py)
│   ├── features/           # Spatial feature extraction
│   ├── pipeline/           # FastAPI routes + background jobs
│   └── data/               # Site definitions + live cache
├── data/models/            # Trained pkl files (RF, GMM, KDE, SHAP)
├── ingestion/              # Data pipeline (see ingestion/README.md)
└── docs/                   # Problem statement + specs + plans

---

Getting started

Frontend

npm install
npm run dev

Backend

python -m venv .venv
.venv/Scripts/activate        # Windows
pip install -r requirements.txt
npm run dev:api

Set environment variables in .env:

ANTHROPIC_API_KEY=sk-ant-...
TAVILY_API_KEY=tvly-...
GRIDSTATUS_API_KEY=...
EIA_API_KEY=...

Data pipeline

See ingestion/README.md for full setup. To run all sources:

cd ingestion
python -m ingestion.run

---

API reference

Method	Path	Purpose
GET	/api/sites	All candidate sites with scores
GET	/api/market	Live gas prices, LMP, BA demand
POST	/api/evaluate	Score a coordinate (SSE stream)
POST	/api/optimize	Grid search (SSE stream)
POST	/api/agent	AI Analyst query (SSE stream)
GET	/api/forecast	72 h LMP forecast P10/P50/P90
GET	/api/regime	Current market regime
WS	/ws/lmp/stream	Live ERCOT LMP WebSocket

Full schema reference: /docs/schema

---

Documentation

Live docs are available at /docs in the app:

• Overview — what COLLIDE does and quick start
• Architecture — system layers, agent graph, data pipeline
• How It Works — ML models, weights, scoring pipeline
• Features — full feature guide
• Data — data sources and pipeline details
• Schema — API contracts, SSE events, data models

---

Links

• GitHub
• Ingestion pipeline docs