Energy Intelligence Platform

See demand sooner. Decide with confidence.

GridPulse forecasts hourly electricity demand for the 51 balancing authorities that run the US grid, and scores every forecast against the operator's own.

51 balancing authorities · ~100% of lower-48 demand Retrained daily · gated on the real serve path, not just a holdout Scored hourly against live EIA data 4.35% median per-BA error on data the models never saw (served ensemble, 168 h recursive holdout)
The live GridPulse Overview for Texas (ERCOT): current demand, 7-day KPIs, actuals with the forward forecast segment, and real holdout metrics

The live dashboard — real EIA demand, weather-aware forecasts, honest holdout metrics.

Why it exists

Four questions, answered from one place

Demand data, weather models and forecast accuracy normally live in separate places, and reconciling them is manual. GridPulse puts them on one screen.

Forecast with context

Hourly demand against the weather driving it, with the recent history the forecast was built from on the same axis.

Know what the number is worth

Prediction intervals, the model’s recent live error and its drift grade are shown beside the forecast itself.

See risk before it lands

Severe-weather alerts, demand anomalies and degraded model grades, tracked per balancing authority.

Test a what-if

Move temperature, wind or solar and see how demand responds. Every answer comes from a real forecast run for that region, computed in advance.

The platform

What each view is for

Each view opens on one question. Picking a role — grid operations, renewables, trading or data science — changes which view loads first and which numbers it leads with.

Overview What changed, and what matters now? Mission-control summary, key KPIs, and a role-aware insight card.
US Grid Which balancing authorities are stressed? 51-BA cards, map, and polygon views with utilization and capacity context.
Forecast What will happen? Forward predictions from Prophet, SARIMAX, XGBoost, and the ensemble — with confidence bands.
Risk Where is operating risk rising? Severe-weather signals, anomalies, stress indicators, degraded conditions.
Models How trustworthy is the forecast? Per-model holdout metrics, live drift, residuals, SHAP feature importance.

Built in the open

Where the numbers come from

Accuracy is measured the way production actually forecasts. Data-quality exclusions are named on the page that uses them, with the count of dropped hours. The architecture decisions, including the ones that turned out wrong, are written down.

The decision record

Every architecture decision — including the ones that turned out to be wrong — is written down, from the ensemble weighting to why a freshly retrained model must replay sanely through the real serve path before it may serve.

ADR trail in PRD.md →

Accuracy, measured honestly

Holdout backtests use the recursive multi-step protocol production serves — each prediction builds on the model’s own previous ones, roughly 2× harder than the one-step-ahead (“teacher-forced”) numbers dashboards usually quote. All 51 BAs, published. Head-to-head against each grid operator’s own day-ahead forecast, GridPulse runs within a fraction of a point of the operators’ median — and holds steady where their forecasts break down. Wins and losses published, per operator.

The blend GridPulse serves is a hedge, not a free win, and the number says so: it trails the best single model at the median — 4.35% against 3.69% — and beats that model on 21 of 51 authorities. What it buys is error decorrelation, because which model wins for a given authority does not hold still between retrains. Serving the blend is that trade, made deliberately and published per authority rather than smoothed away.

Live right now: within 0.22 points of the operators’ own median error across 44 operators (30-day live window, ~24h lead) — 3.99% against their 3.76%, with a fraction of their best-to-worst spread (ours 4.5×, theirs 18.6×).

See the head-to-head benchmark → What the ensemble costs → Backtest results →

Data provenance, disclosed

EIA publishes provisional readings that can revise heavily before settling. GridPulse measures each feed's revision behavior, classifies it live, and discloses exclusions in plain language on the dashboard.

The anchor-conditioning study →

Failure analysis, published

When a forecast collapsed off provably clean inputs, the diagnosis — an ablation study across 67 model vintages — shipped to the repo, and its fix became a permanent acceptance gate.

The forecast-dive diagnosis →