How it works

Ask the grid a question. Get back an answer you can check.

This site turns ~500 billion data points of official GB electricity-grid data into something you can interrogate in plain English — and every answer comes with the data and the method, so you can verify it rather than take it on trust.

1 What this is

Behind the site sits the GDA data lake — 701 datasets, ~500 billion data points, ~200 GB of parquet: official feeds from NESO, National Grid, the DNOs, Elexon, GridWatch, the Carbon Intensity API, weather and more, plus derived datasets and a full power-flow solve of the GB grid for every second of the last 15+ years. Everything here is built from public data only.

There are four ways to use it:

Ask the data (GDALLM) — a plain-English question answered from the lake, packaged with its data and validation.
The library — every question anyone has already asked, anonymised, free to download.
Solves — download the second-by-second solved grid, one day at a time.
Mount the lake — read-only access to the whole parquet tree for your own analysis.

2 What happens when you ask

Your question→ Grounded in the catalogue→ Read-only query, sandboxed→ Packaged ZIP

You type a question like "how much did NESO spend on balancing during the World Cup final half-time, and what was the demand swing?". Behind the scenes the assistant is handed the machine-readable catalogue of all 701 datasets — their columns, partitions and time ranges — and uses it to write a read-only SQL query (DuckDB) over exactly the datasets that answer your question.

That query runs in a locked-down sandbox: no network, read-only access to the data, a row cap and a memory limit. It cannot change anything. When it finishes, the answer, the data it produced, the exact query and the steps to reproduce it are bundled into a single ZIP for you to download.

3 What you get back

Every completed query produces a downloadable ZIP containing:

summary

A short, plain-English narrative of what the query did and what it found.

result.csv / .parquet

The actual data the query returned — open it in Excel, pandas, DuckDB, anything.

query.sql

The exact SQL that was run. Nothing is hidden; you can read and re-run it.

validation_instructions.md

Step-by-step notes on which datasets were used and how to reproduce the answer yourself.

4 How to validate it

The whole point is that you don't have to trust the answer — you can check it. Because you get the exact SQL and the list of datasets used, you can:

Read query.sql and satisfy yourself it asks what you meant.
Mount the lake read-only and re-run the same query against the same parquet — you should get the same numbers.
Cross-check against the raw source (NESO / Elexon / etc.) named in validation_instructions.md.

Same data in, same query, same answer out. That reproducibility is the product.

5 Asking good questions

Be specific about time. A date, a window, or a named event ("18:00–20:00 on 9 July 2018") beats "recently".
Name the metric and the units you want — MW, £, GVA·s, Hz, tCO₂.
Say how you want it broken down — per settlement period, per fuel, per region, per interconnector.
One question at a time gives the cleanest, most checkable result.

"What was the lowest system frequency in 2018, when did it happen, and what was the RoCoF?"

"Total BSUoS-relevant balancing spend per month in 2023, and the demand each month."

"During the 2022 heatwave week, how much did behind-the-meter solar reduce measured demand at midday?"

6 The honest caveats

This is a beta tool. Answers are computed directly from source data, but the assistant can misread a question or pick the wrong dataset — that's exactly why every answer ships with the query and the means to validate it. Always check the data before you rely on it. It works with official public data only, and it is not personalised advice of any kind.

Running a query costs real compute and model budget, so it's registration-gated with a small daily limit per account. Browsing the library, downloading existing answers and downloading solves are all free and need no login. Need a higher limit for legitimate work? Email mark@twinscrollgridbalancer.co.uk.

Ask the data → Browse the library