Ask the Grid
PreviewFoundation model · time-series + geospatialBase model

Geotemporal Foundation Model

GTFM

A time-series foundation model that understands where the grid is, not just when.

Our base model — a time-series foundation model pretrained on years of ISO history jointly with the grid's geography, so a forecast at one location carries what's happening around it.

How it works

A transformer-based time-series foundation model conditioned on geospatial structure. Each series is patch-tokenized rather than read one timestep at a time, and a node's tokens carry positional embeddings drawn from its real coordinates and network topology alongside the usual temporal position — so attention runs jointly across time and across space, rather than treating each node as an isolated series scored independently. Output is a probability distribution per horizon, not a single point estimate. Pretraining is self-supervised masked-forecasting over multi-year, multi-ISO history, so the model learns the shape of grid dynamics generally before any task-specific fine-tuning narrows it. It is the backbone the other three models are built from, either by fine-tuning it directly or by post-training on top of its learned representations.

Inputs

  • Nodal and system time series — prices, load, generation, and fuel mix
  • Grid geography — node coordinates and network topology / connectivity
  • Weather fields — forecast and observed
  • Calendar and market context

Outputs

  • Probabilistic forecasts across horizons, each with location and as-of
  • Learned spatiotemporal embeddings for downstream models
  • Nowcasts of current grid state

Three ways to consume it

product · inference · MCP
01

In the product

The Geotemporal Foundation Model powers Ask the Grid directly — the live map, the radar, and the grid agent read it in place, so the same model you'd call is the one already reasoning over the grid on screen.

02

Developer inference service

A direct, typed inference API: request forecasts or embeddings by node and horizon and get structured results back. This is a plain inference endpoint — not MCP — for pipelines, notebooks, and services that want the model output without an agent in the loop.

See the developer surface
03

MCP

The same model outputs, exposed as tools your own agent can call. Point a compatible client at the Ask the Grid MCP server and the model becomes a tool in your agent's own loop.

Build with MCP

Intended use

  • Nodal and system price, load, and renewable forecasting
  • Transfer to new nodes and markets with little local history
  • Embeddings that feed the downstream Spike, Event, and Reschedule models

Limitations

  • Forthcoming — not generally available; capabilities described are forward-looking
  • Forecasts are probabilistic, not guarantees
  • Coverage rolls out per ISO as data and evaluation land

Consume it in the product, over the inference API, or through MCP.

Every model reaches the same three surfaces. Start where you work and the model output comes back with its source and as-of attached.