Electricity Forecasting — The Turkish Grid

Can a pretrained time-series foundation model, with no task-specific training, compete with the official hourly demand plan that EPİAŞ publishes for the Turkish grid? I built this forecasting and analysis application with Universal Software’s energy team to find out.

Streamlit backtest page with April 2026 metrics and overlaid actual, TimesFM, and EPİAŞ demand curves — The backtest view for April 2026 at a 24-hour horizon: TimesFM (*predict*) against observed demand and the EPİAŞ plan over 720 scored hours, loaded from cached forecasts.

What it does

Pulls hourly consumption, the EPİAŞ demand plan, and meter data for all 21 distribution companies from the EPİAŞ Transparency Platform into Parquet.
Runs TimesFM 2.5 zero-shot with a 1,024-hour context window.
Backtests 24-hour and 168-hour horizons from rolling forecast origins, using only information available at each forecast time. Results are cached, so repeat analysis is instant.
Explores hourly meter multipliers by distributor, profile group, and meter-reading type.

The distributor dashboard: hourly meter multipliers for two distribution companies, broken down by profile group and meter-reading type.

Results and limits

Rolling backtests over April–June 2026, with about 1,800 scored hours per horizon, showed a clear trade-off:

24-hour MAPE	TimesFM	EPİAŞ plan
All reported hours	3.47%	3.13%
Excluding public holidays	2.39%	2.92%
Public holidays only	12.67%	4.89%

Daily mean absolute error of TimesFM and the EPİAŞ plan from April to mid-June 2026, with public holidays shaded — Daily MAE across the 24-hour backtest. TimesFM’s largest errors fall on public holidays, above all Eid al-Adha.

On ordinary days the zero-shot model beat the official plan. On holidays the ranking reversed sharply. That gap is explainable: TimesFM sees only demand history, with no holiday calendar or weather, so it cannot anticipate a national holiday, and at 168 hours the depressed demand lingers in its context window afterwards. The next improvement is calendar awareness, not a bigger univariate model.