Energy Flow

An exploratory model of US energy, not a forecast. Pick a region and year to see how energy moves from fuels through generation into end uses. Past the historical boundary the chart switches to projection mode — open Build a scenario below to pull levers (build caps, EV adoption, retirements, carbon price, …) and watch the picture shift.

How the model works every source & equation Why it's hard to plug in queues, wires & who blocks what Where it gets built siting explorer — sister project ↗
What this is for

The point is what-if, not what-will-be. Move a lever, render the scenario, and the consequences flow through generation, end-use sectors, the emissions panel, and the impact-delta readout side by side. Read the numbers as directional: "this lever, moved this way, shifts the picture this much" — not "2050 will look like this."

Treat it as a sandbox. Build a scenario, ask whether the change you made should have moved the chart that way, then iterate. Saved baskets keep favourite scenarios around to reload and compare. Prefer plain language? The assistant turns "make solar cheaper, phase out coal by 2035" into scenario levers for you.

Build a scenario

Catalog

Editor

Basket

1 Quad equals 293 TWh of electricity — or about four days of everything the US runs on 8.3 billion gallons of gasoline ≈ 470 billion hamburgers — burned for their calories, ~540 kcal each For scale: the whole US runs on about 93 Quads a year — so one Quad is close to four days of total US energy.
Methodology & data sources

Historical artifacts are derived from the EIA SEDS dataset (State Energy Data System) via the project's fetch_eia pipeline. Values use the EIA's captured-energy methodology (October 2023 onward), where noncombustible renewables (solar, wind, hydro, geothermal) count at the heat-rate constant 3,412 Btu/kWh — replacing the previous fossil-fuel-equivalency convention that inflated those sources by a ~3× factor. As a result, renewable shares in this chart will read smaller than older "Estimated" LLNL charts published before the methodology change; they match LLNL's current Actuals series exactly. See EIA SEDS change log for the methodology details.

Projection-mode artifacts (years past the historical boundary) come from the v2model engine — a cost-driven annual loop with Wright's-law learning curves, per-source max-build caps, CES electrification elasticity, and the empirical-correction variables (retirement slippage, capacity-factor degradation, construction-delay drag). Per-state research overrides apply for state regions; the "Build a scenario" panel exposes the full lever surface for what-if exploration.

Lever defaults sit in research/**/*.md with quality tiers L1 (primary EIA / NREL / IRENA) through L4 (transcribed unverified). When a projection consumes any L4 value the meta line surfaces a "⚠ less-verified source" callout. See docs/concepts.md for the v2 engine spec and DECISIONS.md for the full design history.

Explore the model architecture → for a guided tour of the engine's generation sources, submodels, feedback loops, levers, and decision history. Research findings → for the longer-form things this project noticed along the way (literature gaps, empirical surprises, honest negative results).