Türkiye Grid Foundation Model

A Reproducible Data Audit and Open Research Infrastructure for Multi-Task Electricity Forecasting

Abstract

Electricity demand, renewable generation, and day-ahead market price are not independent signals: they arise within the same physical and market system and may therefore share useful temporal structure. That observation, however, does not imply that a shared multi-task model will necessarily improve forecasting. The longer-term question of the Türkiye Grid Foundation Model project is whether shared temporal representations can produce positive transfer across electricity consumption, renewable generation, and day-ahead Market Clearing Price (MCP; PTF in Turkish terminology), and under which regimes such sharing may instead produce negative transfer. This preprint addresses an earlier methodological requirement: whether the empirical data layer is sufficiently well defined, complete, aligned, and provenance-preserved to support a later model-comparison study.

Before the first empirical acquisition, the audit interval was committed as a scope lock from 2024-01-01T00:00:00+03:00 through 2024-01-31T23:00:00+03:00, corresponding to 744 expected hourly timestamps. Three official EPİAŞ Transparency Platform datasets were examined: real-time consumption, source-level real-time generation, and day-ahead MCP/PTF. Each dataset contained 744 raw rows and 744 unique hourly timestamps, with zero duplicates, zero missing expected hours, zero key numeric nulls, and zero observations outside the locked interval. Timestamp alignment yielded 744/744 complete three-target rows.

The initial audit also identified a generation-accounting residual when the reported total was compared with the listed generation-source sum. A corrective audit made no new EPİAŞ request and reused the exact locked January 2024 generation snapshot. When the importExport field was added to the listed source sum, the reported generation total reconciled to numerical precision across all 744 rows. Two solar observations of -0.01 MWh were documented and preserved without deletion or clipping.

The contribution of this preprint is deliberately bounded: a reproducible empirical data audit, an explicit evidence firewall between exploratory and future confirmatory work, and an open research infrastructure for subsequent multi-task forecasting and uncertainty-calibration experiments. No claim of forecasting superiority, state-of-the-art performance, or confirmatory model advantage is made. The confirmatory holdout remains undefined and untouched.