Energy forecasting with foundation models
Predict grid load, renewable generation, and peak demand using zero-shot time series models. Foundation models handle the non-stationarity, weather dependence, and seasonality of energy data without per-asset training pipelines.
{
"model": "amazon/chronos-bolt-base",
"inputs": [{
"item_id": "grid-zone-NE-load-MW",
"target": [14280, 13950, 14520, 15100, 16340, 17200, 18050, 17600, 16800, 15400, 14100, 13800],
"start": "2026-04-01T00:00:00Z"
}],
"parameters": {
"prediction_length": 48,
"freq": "h",
"quantiles": [0.1, 0.5, 0.9]
}
}48-hour ahead grid load forecast with prediction intervals for dispatch planning.
Non-stationarity
Foundation models adapt to shifting load patterns without retraining
Speed
No per-asset model fitting — send history, get forecasts immediately
Uncertainty
Quantile forecasts for reserve margin and capacity planning
Why foundation models suit energy forecasting
Energy time series are non-stationary, weather-dependent, and seasonally complex. Foundation models handle these properties because they have seen millions of diverse patterns during pre-training.
No per-asset training pipelines
Classical energy forecasting requires a fitted model per meter, zone, or generation asset. Foundation models forecast any energy series zero-shot — add new assets or zones without retraining.
Robust to non-stationarity
Grid loads shift with EV adoption, behind-the-meter solar, and demand response programs. Foundation models generalize across distribution shifts because they were pre-trained on diverse, non-stationary series.
Prediction intervals for grid operations
Quantile forecasts give you the uncertainty bounds that dispatch, reserve scheduling, and capacity planning require. No separate conformal prediction step needed.
View quantile docsGetting started with energy forecasting
Most energy teams start with a single load zone or generation asset, validate against held-out data, then scale to their full asset fleet.
- 1
Pick one load zone or asset
Start with an hourly load series or a solar/wind generation series that has at least a few weeks of history. Send it to the API and compare the forecast against your existing model or actuals.
- 2
Evaluate across models
Try 2-3 foundation models on the same series. Compare MASE, CRPS, and prediction interval coverage. Energy series often favor models with strong handling of multiple seasonalities.
- 3
Scale to your fleet
Use the batch endpoint to forecast all zones, meters, or generation assets in a single request. Integrate results into your EMS, SCADA, or trading platform.
Frequently Asked Questions
Forecast your first load zone
Send an hourly load series to the API and get back point forecasts with prediction intervals. No training pipeline required.