Virtual Pipeline Netback

Should a producer with stranded gas truck it as CNG or LNG, or keep flaring it? This project answers that every trading day for 18 real, satellite-detected flares in the Permian and Eagle Ford. It uses only public data: EIA prices, World Bank/NOAA VIIRS flare detections, EIA basin boundaries and OSRM road routing. A Postgres star schema, idempotent ingestion with data-quality checks, a netback model written in readable SQL, a sensitivity analysis and a dashboard with assumption controls sit on top.

▶ Live dashboard: kaavinb.github.io/virtual-pipeline-netback: no install, works on a phone. Drag the sliders to ask "what if?"; every view is a shareable link.

Findings (prices as of 2026-10-20: Henry Hub $3.10/MMBtu, Gulf Coast diesel $5.90/gal)

1. Selling trucked gas into a pipeline loses money at every site: −$2.60 to −$5.16/MMBtu. Henry Hub would need to be $5.70–$8.26 to break even. Permian sites fare worst, because their only nearby hub, Waha, trades about $2 below Henry Hub. Since 2010, hub sales beat flaring on most Eagle Ford days only once: in 2022, when Henry Hub averaged $6.45.
2. Selling the same gas as a diesel substitute to an in-basin customer (rig, frac fleet, off-grid power) nets about +$20/MMBtu at every site. That conclusion holds even though the customer discount is a guess: flaring only wins if the customer demands more than ~89% off diesel parity. The real constraint is finding demand, not price.
3. CNG beats LNG at every site. On a road chassis an LNG ISO tank hits the 80,000 lb limit at about 69% full, so LNG's energy-per-load advantage shrinks to 1.8×. Small-scale liquefaction then costs about $7/MMBtu, so LNG only becomes cheaper than CNG beyond 641 mi (fuel sales) or 895 mi (hub sales). Every route here is under 600 mi.
4. No single input flips any decision. 25 drivers × 18 sites × 2 markets were swung across their sourced or data-driven ranges (08).

The bottom-up cost model reproduces the World Bank's independent full-chain estimates within 3–9% (06).

Why this matters to a virtual-pipeline company

A virtual pipeline (CNG/LNG in ISO containers on trucks) earns its keep where pipelines don't reach. The questions a commercial or BI team asks every day are the ones this model answers:

• Which sites and delivery points net the most per MMBtu?
• How far can we haul before it stops paying?
• When does LNG beat CNG?
• How exposed is each answer to gas, basis and diesel prices?

The pipeline-versus-fuel result is the strategic point: the value sits in delivered-fuel markets, not pipeline injection, and distance only matters at the margin.

How it works

flowchart LR
    subgraph Sources["Public sources"]
        EIA["EIA API v2<br/>Henry Hub daily · diesel weekly"]
        WB["World Bank / NOAA VIIRS<br/>flare volumes by location"]
        GIS["EIA basin boundaries"]
        OSRM["OSRM routing"]
        Y["assumptions.yaml<br/>38 sourced / labelled inputs"]
    end
    subgraph Ingest["Python ingestion (idempotent, logged, retried)"]
        I1["eia.py"] --- I2["sites.py"] --- I3["routes.py"] --- I4["assumptions.py"]
    end
    subgraph PG["PostgreSQL 17"]
        CORE[("core<br/>star schema")]
        OPS[("ops<br/>ingest runs · DQ results")]
        MART["mart (SQL views)<br/>v_netback_daily · v_site_latest<br/>scenarios · tornado · curves"]
    end
    EIA --> I1 --> CORE
    WB --> I2
    GIS --> I2 --> CORE
    OSRM --> I3 --> CORE
    Y --> I4 --> CORE
    CORE --> MART
    CORE --> DQ["dq.py: 14 checks"] --> OPS
    MART --> MB["Metabase dashboard<br/>(live, with assumption controls)"]
    MART --> CSV["CSV extracts → Tableau Public"]

Netback = delivered price − (treatment + compression or liquefaction) − trucking − (decompression or regas) − fees, all in $/MMBtu at the wellsite. Trucking is built bottom-up:

trucking $/MMBtu = fixed + 2 × (road miles + last mile) × (diesel $/gal ÷ mpg + non-fuel $/mile) ÷ MMBtu per load

MMBtu per load is computed: the smaller of what the container holds and what is legal at 80,000 lb. Delivered price is Henry Hub + basis for hubs, or diesel parity × (1 − discount) for a fuel customer. Diesel parity is taken on a lower-heating-value basis: engines burn on LHV, gas is sold on HHV, and skipping the correction overstates the gas value by 3.3%. The full walk-through is in 01 — value chain and netback.

All business logic is SQL (sql/migrations). Python only moves data in and out. What-if analysis never copies the formulas: mart.site_latest_scenario('{"fuel_discount_to_diesel": 0.6}') sets a transaction-local override that mart.v_assumption honours, and then the same views recompute. The sensitivity tornado and the dashboard's controls both run through it.

Run it

Requirements: Docker, uv, make, and a free EIA API key.

cp .env.example .env        # paste EIA_API_KEY, choose a Metabase admin password
make setup                  # Python 3.12 deps via uv
make all                    # Postgres → migrations → EIA → sites → assumptions → routes → DQ → sensitivity → CSV extracts (~90 s)
make dashboard              # Metabase on http://localhost:3000, dashboard built from code
make test                   # 70 tests

make help lists every target. make reset && make all rebuilds from an empty database deterministically: the same 18 sites, and identical numbers for the same source data.

여기까지가 README 앞부분입니다. 나머지는 원문에서 읽을 수 있습니다. README 전문 보기 ↗