grid.md

Every morning at 09:00 Moldova time, a forecast of how much electricity Moldova will use tomorrow, hour by hour, given as a range rather than one number and checked against the grid operator's own plan. The page is in English, Romanian and Russian.

Live: iliev-blnk.github.io/grid.md

Result

Backtest over one test year (Oct 2025–Sep 2026, 8,693 hours), with the design fixed beforehand (full report):

	typical error (MAPE)	hours inside the 90% range
grid.md	8.6%	88.4%
Moldelectrica's own plan	18.7%	–
the plan with its recent bias removed	15.5%	–
same hour last week	15.0%	89.1% (range 1.8× wider)

• Most of the gap comes from spring and summer. Moldelectrica's load is consumption net of distributed solar, and on sunny afternoons it falls by half; only a forecast that sees tomorrow's sunshine follows it.
• In winter the operator's plan is as good as or slightly better than the model.

Moldova's hidden solar

Moldelectrica's load is net of any solar it does not meter, so sunny days pull midday load down. Measuring how far, month by month, gives the solar the grid operator cannot see. The result for May–September 2026 is ≈668 MW (range 557–814), plus ~184 MW that it does meter. The official figure for rooftop prosumers is 195 MW. The estimate matches CNED's national total almost exactly at the end of 2024 (391 vs 394 MW) and is ~20% low at the end of 2025. The pre-registered first version failed that check; the report shows both versions and why: research/SOLAR.md.

The story

Moldova's grid, 2021–2026, in seven numbers, each with its chart:

1. 97% from one power station before the war.
2. 290 MW at 15:00 on the day of the November 2022 blackout.
3. 80% of the hours of 2023 as a corridor from Romania to Ukraine.
4. Cuciurgan falling to 222 MW after the gas cutoff.
5. The midday dip deepening from 1.05× to 0.50× of the evening.
6. The hidden solar.
7. Tomorrow's forecast.

English, Romanian and Russian, like the main page.

How it works

• Data:
  • Moldelectrica's public 15-minute archive (load, plan, generation, flows), from 2021.
  • Open-Meteo archived weather forecasts for Chișinău, Bălți, Cahul and Ungheni.
  • Quirks found along the way are in research/DATA_FACTS.md.
• No hindsight:
  • The forecast for D+1 is built at 09:00 on D only from data that existed then: load hours that ended by 08:00, and weather forecasts already issued.
  • A test checks the weather rule on every row.
  • The operator's plan is never an input.
• Model:
  • LightGBM quantile regression (7 quantiles) of load minus the same hour last week.
  • Features: calendar, Moldovan holidays, recent load, weather, and a 14-day estimate of how strongly sunshine currently cuts net load.
  • A rolling conformal correction keeps the ranges honest as the system changes.
  • Design choices were made on a separate validation half-year (research/VALIDATION.md) and pinned before the test year was scored.
• Every day:
  • A systemd timer on my machine runs scripts/daily.sh: refresh data, retrain monthly, issue tomorrow's forecast, score finished days, commit and push. It runs locally because Moldelectrica blocks cloud runners.
  • Forecasts in forecasts/ are committed before the day starts and never edited. Scores land in scores/.

Run it

python -m venv .venv && .venv/bin/pip install -e ".[dev]"
.venv/bin/gridmd ingest load --from 2021-12-01     # Moldelectrica → data/load.csv.gz
.venv/bin/gridmd ingest weather --from 2022-01-01  # Open-Meteo → data/weather.csv.gz
.venv/bin/gridmd backtest dataset|tune|test        # research/
.venv/bin/gridmd retrain --force && .venv/bin/gridmd forecast --date 2026-10-12
.venv/bin/pytest -q && pnpm install && pnpm --dir site dev
deploy/install.sh                                   # the daily systemd user timer

Data: Moldelectrica operative data; Open-Meteo (CC BY 4.0). MIT licence.