A step-by-step, plain-language and technical guide to what Türkiye Grid Foundation Model is, why it was created, what real EPİAŞ evidence has already been verified, what the project has achieved, and how future researchers can extend it.
It is an open research platform for studying three coupled hourly signals in Türkiye's electricity system: electricity consumption, renewable generation and day-ahead MCP/PTF.
The project does not assume that one large AI model must be better. It tests whether a shared temporal representation creates useful cross-task transfer across +1 h, +6 h and +24 h forecasts, while also measuring uncertainty and the possibility of negative transfer.
The project started from a stronger question than “can AI forecast electricity?”. It asks whether a shared temporal representation can learn useful structure across electricity demand, renewable generation and day-ahead market price, and when that sharing causes negative transfer.
Before opening final evidence, the repository implemented the EPİAŞ client, timestamp alignment, calendar features, leakage-resistant windows, baselines, a compact shared Transformer, uncertainty utilities, tests, CI and a synthetic smoke path. This separates software debugging from scientific inference.
The first empirical request was deliberately bounded to 1–31 January 2024. The runner was designed not to accept arbitrary dates, and the future confirmatory holdout remained undefined. This created an evidence firewall before data inspection.
Three official hourly streams were retrieved: real-time consumption, source-level generation and day-ahead MCP/PTF. Each returned 744 of 744 expected hours, with zero missing expected timestamps, zero duplicates, zero key numeric nulls and zero out-of-scope timestamps.
All three targets aligned into 744 complete hourly rows. Timezone behavior, schema fields, ranges, missingness, descriptive summaries, correlations and SHA-256 provenance were recorded. Raw third-party records remained local and gitignored.
The first generation closure check showed a residual because the schema contained an importExport field that had not been included in the source sum. A corrective audit was run on the exact same local snapshot with zero new EPİAŞ requests.
When importExport was added to the listed source sum, the reported EPİAŞ total closed exactly across all 744 hours: mean absolute residual, P95 absolute residual and maximum absolute residual all became zero to numeric precision.
Two solar observations of -0.01 MWh were found. Their total absolute magnitude was only 0.02 MWh. They were not silently deleted or clipped; the project records them so any later preprocessing rule can be declared before confirmatory testing.
The current achievement is not a claim that a Transformer wins. The achievement is a verified, reproducible empirical data layer with a traceable audit trail, a resolved generation accounting convention, open code and a protected future confirmatory study.
The project can now expand historical data only on the exploratory/training side, audit schema stability across years, benchmark persistence/seasonal/linear baselines, and measure CPU/GPU/runtime and storage costs. Only after that should model families, seeds, metrics, exclusions and the untouched holdout be frozen.
744 complete aligned rows; timezone +03:00; provenance hashes preserved.
total = listed source sum + importExport for all 744 audited hours to numeric precision.
Two sun = -0.01 MWh observations documented, retained and not silently altered.
Authenticated client, bounded acquisition, schema inspection, timestamp alignment and local raw-data policy.
Chronological splits, 168-hour windows, calendar features, baselines, shared model and uncertainty utilities.
Scope lock, aggregate reports, SHA-256 provenance, public Pages, publication roadmap and preregistration gate.
The scientific object is the hypothesis, not the model's reputation. A tie or negative transfer remains informative.
Future researchers can replace models while preserving the data contract, audit logic and evaluation firewall.
A reusable bridge between official electricity-market data and reproducible AI/time-series experimentation.
No final model ranking, no Transformer-superiority claim, no causal claim and no confirmatory forecasting result exists yet. This is deliberate methodological protection, not a missing piece being hidden.
Faramarz Kowsari is an author and researcher based in Istanbul. Focusing on the intersection of technology, education, and personal growth, he has published over 80 digital titles on international platforms. His areas of expertise span Artificial Intelligence, prompt engineering, modern trading strategies (Smart Money Concepts & algorithmic trading), as well as classical literature and mindfulness. In addition to writing, he develops web-based educational tools and creates specialized instructional video content.
Official Website · Google Scholar · GitHub · LinkedIn · ORCID
The current research lifecycle state is now archived on OSF as a public, persistent registration. Registration DOI: 10.17605/OSF.IO/FMCYQ. Associated OSF project: osf.io/tz5pw.
The registration time-stamps the completed January 2024 real-data EPİAŞ audit, corrective generation-accounting audit, reproducibility rules, and the boundary between exploratory development and a later confirmatory study. It is not the final confirmatory preregistration; the confirmatory holdout remains undefined and untouched.