Method & limits

How the numbers on this site were produced, why these choices were made, and what they can't tell you. The full write-ups, with equations and 37 cited sources, are in the project repository under docs/models/.

Pipeline

1. Collect (plain Python, no AI):
   • NEA daily operational reports (PDF)
   • NEA monthly and annual reports (PDF → tables, OCR where tables are images)
   • the NEA homepage widget
   • ERA5 weather, GloFAS river flow and its forecast
   • India customs trade and the CERC power-exchange report
   • World Bank GDP
   • a sourced register of flood-damaged plants
2. Store in a DuckLake lakehouse (Parquet + DuckDB). Raw files are kept, every fetch is logged, and the store is time-travel queryable.
3. Analyse: seasonality, temperature, river flow, growth, and the agreement between sources (the Analysis page).
4. Model: a rolling-origin backtest, then model selection, then P10–P90 ranges (the Models page).
5. Forecast: daily 1–30 days and a monthly winter outlook, adjusted for flood losses (the Forecast page).
6. Publish: the lake is exported to Parquet, and this static site queries it in your browser with DuckDB-WASM (the SQL explorer).

Why these models

• Simple first, complex only if proven. In international load-forecasting competitions (GEFCom 2012/2014/2017), regression with temperature and calendar terms appears in every winning entry. Our ridge model is that, adapted to Nepal: two temperature regions, Nepali holidays, river flow.
• No deep learning. We have two winters of daily data. A Nepal peak-load study found LSTM beat a 30-day moving average by only 0.05 percentage points. Tree models and even linear models beat deep networks on data this size.
• Honest testing.
  • Every forecast is made only from data available at the time.
  • Future weather comes from the long-term average, not the actual weather.
  • The backtest scores winter error separately.
  • A Diebold–Mariano test guards against "better by luck".
• Ranges, not just numbers. P10–P90 bands come from each model's own recent out-of-sample errors. Their coverage is reported, including where it misses.

Results in one line each

• Daily demand: 2.5% error next day, 5–8% at 30 days (baseline: 3.0% → 12.5%).
• Daily peak: 2.4% next day (a Nepal LSTM study: 2.95%).
• Monthly winter demand: 3–6.5% error per month, forecast 5–10 months past the last published NEA figure.
• Winter 2026/27 import: about 300–350 GWh/month in Jan–Apr after flood losses (about 200 without).

Limits

• One test winter for daily models (Nov 2024 – Jan 2025), because NEA stopped publishing daily reports in January 2025. Treat daily error figures as indicative.
• The flood adjustment uses estimates: plant lists from news reports, and a rule-of-thumb dry-season output of 35%. It could be off by ±30%.
• GloFAS is a global model (~5 km grid), used as a basin-scale signal, not plant-level truth.
• Not modelled: India-side supply limits and prices, and new plant commissioning dates.
• Plant locations on the map are approximate (valley level).

Data credits

• Nepal Electricity Authority: reports and website.
• Open-Meteo (CC BY 4.0): ERA5 reanalysis (ECMWF) and GloFAS river flow (Copernicus).
• Other data: World Bank, India Ministry of Commerce (TradeStat), CERC.
• Flood damage reporting: Kathmandu Post, Spotlight Nepal, Mongabay, Al Jazeera.
• Inspiration: the NDOR scraping idea came from nea-data-warehouse by Aayush Paudel.

This is an educational, non-commercial project.