Energy · Forecasting
Tomorrow’s demand, hour by hour, with a range a desk can plan around
A day-ahead forecast for Victoria’s grid, tested at 10 monthly origins over 7,296 hours of 2025.
Executive summary
A day-ahead forecast for Victoria’s grid halves average hourly error against a regression with weather, 5.4% against 10.4%, and states a range that held 80% of the time.
Situation
Retailers, spot-exposed sites and battery operators commit a day ahead in a market where prices went negative in 21% of five-minute intervals and reached $17,500/MWh. A closer forecast with a stated error changes how much is hedged.
Approach
Gradient boosting on AEMO demand and Open-Meteo forecast weather, retrained at 10 monthly origins and tested a day ahead on each following month, with split-conformal ranges.
Result
Error falls 5.0 points (95% CI 4.5 to 5.5) against regression with weather. Holidays remain harder (7.0%), and range coverage fell short in five of ten months, so buffers are reviewed monthly.
The business problem
Energy retailers and large buyers in the National Electricity Market set hedges and spot positions from a forecast of tomorrow’s demand. AEMO publishes its own forecasts and reports their accuracy each year, and published day-ahead models reach low single-digit percentage errors, so a forecast is only useful if it is at least that good and comes with a range a trading desk can plan around.
I built a day-ahead hourly forecast for Victoria with an 80% range and tested it at monthly origins through 2025 against a linear regression on the same inputs.
- Team
- One or two analysts in a trading or operations team
- Today
- Same-day-last-week rules or a spreadsheet regression
- Systems
- AEMO NEMweb data, weather API, Excel or a BI tool
- Deadline
- Forecast issued by midnight for the next trading day
What was built
A daily job that pulls AEMO five-minute demand and the Open-Meteo weather forecast, builds hourly features (recent demand, forecast temperature and solar radiation, calendar and Victorian holidays) and issues 24 hourly forecasts with calibrated 80% and 95% ranges for the next day.
Results
The daily peak hour, which drives the most expensive purchases, was forecast within 4.2% on average against 7.7% for the simple model, across 304 days.
| Method | MAPE (95% CI) | MW off, avg | Bias, MW |
|---|---|---|---|
| Same hour last week | 13.6% (12.5–14.6) | 620 | -17 |
| Same hour yesterday | 11.0% (10.3–11.8) | 508 | -4 |
| Linear regression + weather | 10.4% (9.8–11.0) | 479 | -101 |
| Gradient boosting, no weather | 8.7% (8.2–9.3) | 397 | +8 |
| Gradient boosting + temperature forecast only | 6.5% (6.2–6.9) | 299 | -56 |
| Gradient boosting + weather forecast | 5.4% (5.2–5.7) | 252 | -50 |
| Gradient boosting + observed weather (hindsight) | 5.4% (5.2–5.7) | 252 | -38 |
Data and method
- Demand
- AEMO NEMweb, VIC1, 298,080 five-minute readings in 34 monthly files, 2023-01-01 to 2025-11-01; no duplicates, gaps or non-positive values
- Weather
- Open-Meteo historical forecasts for Melbourne; forecast against observed temperature differs by 1.1°C on average
- Test
- Rolling origin: retrained at 10 monthly origins, each forecasting the next month a day ahead; 7,296 test hours
- Ranges
- Split conformal: widths set from relative errors on the last 60 days before each origin
Using observed rather than forecast weather barely helps (5.43%), so the weather forecast is not the bottleneck. Raw quantile ranges covered only 58% of hours and were replaced with conformal ranges.
Assumptions and limitations
- Grid-level demand for one region; a single site or retailer book is noisier and needs its own data
- Holidays are harder: 7.0% error on 10 holiday days against 5.3% otherwise
- The forecast does not predict price; buying decisions also depend on price forecasts and contracts
- Range coverage drifts with the season and is re-checked monthly
Where it misses
Four of the five worst days fall in the summer holiday weeks (2 January, the Australia Day weekend and early February) and one is Good Friday. Demand in those weeks behaves unlike a normal weekday, and two years of history hold few examples.
| Worst day | Day | Error | Forecast top °C | Observed top °C | Holiday |
|---|---|---|---|---|---|
| 2025-01-02 | Thursday | 14.6% | 20.8 | 20.0 | No |
| 2025-01-26 | Sunday | 14.3% | 25.8 | 23.0 | No |
| 2025-04-18 | Friday | 13.9% | 30.6 | 30.6 | Yes |
| 2025-01-28 | Tuesday | 13.0% | 22.0 | 21.4 | No |
| 2025-02-05 | Wednesday | 12.7% | 23.4 | 22.9 | No |
What a team would do next
- Week 1: connect the site’s or book’s meter data and the weather feed
- Weeks 2 to 3: backtest on the last 12 months against the current method
- Weeks 4 to 8: run alongside the current forecast; the desk compares each morning
- Then: add price forecasts for battery scheduling if the demand gain holds
Technical depth
Gradient boosting on lagged demand (24 and 168 hours), forecast temperature, solar radiation, hour, weekday and holiday. The biggest single input is demand at the same hour yesterday, then a week ago, then forecast temperature.