Manaz Afzal

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.

5.4%average hourly error; regression with weather 10.4%
4.2%error on each day’s peak hour (simple model 7.7%)
80%of hours inside the stated 80% range

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.

5.4%average hourly error over 7,296 test hours of 2025 (95% CI 5.2% to 5.7%)
4.2%error on each day’s peak hour, against 7.7% for the simple model
80.2%of hours inside the stated 80% range

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.

AEMOFive-minute demandVIC1Weather forecastTemperature andsolar, next dayModelHourly forecastretrained monthlyRanges80% and 95% fromrecent errorsDeskCSV and chartby midnight
Daily workflow. Only data available before midnight is used.

Results

0.0 GW2.0 GW4.0 GW6.0 GW8.0 GW01-2401-2501-2601-2701-2801-2901-30ActualForecastSame hour last weekHeat peak
Shaded: the 80% range. Over the test period 80.2% of hours fell inside it. Data: AEMO NEMweb; Open-Meteo historical forecasts
Same hour last week620 MW off on average13.6%Same hour yesterday508 MW off on average11.0%Linear regression + weather479 MW off on average10.4%Gradient boosting, no weather397 MW off on average8.7%Gradient boosting + temperature forecast only299 MW off on average6.5%Gradient boosting + weather forecast252 MW off on average5.4%
Each step adds information a desk would have at midnight. The final model cuts average error to 5.4% from 10.4% for a regression with weather (5.0 points, 95% CI 4.5 to 5.5).

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.

RegressionFinal<15°C4.6%15–204.9%20–256.1%25–305.9%30–356.2%35°C+6.2%
The gap is widest on hot days, when demand and prices both spike.
0%5%10%15%20%00:0004:0008:0012:0016:0020:00Linear + weatherFinal modelMidday rooftop solar
Error rises around midday, when rooftop solar hides household demand from the grid.
2025-0186%2025-0277%2025-0386%2025-0492%2025-0586%2025-0672%2025-0774%2025-0881%2025-0977%2025-1071%Target 80%
The 80% range held in five of ten months; February, June, July, September and October fell short (lowest 71%), so buffers are reviewed each month.
MethodMAPE (95% CI)MW off, avgBias, MW
Same hour last week13.6% (12.5–14.6)620-17
Same hour yesterday11.0% (10.3–11.8)508-4
Linear regression + weather10.4% (9.8–11.0)479-101
Gradient boosting, no weather8.7% (8.2–9.3)397+8
Gradient boosting + temperature forecast only6.5% (6.2–6.9)299-56
Gradient boosting + weather forecast5.4% (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

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 dayDayErrorForecast top °CObserved top °CHoliday
2025-01-02Thursday14.6%20.820.0No
2025-01-26Sunday14.3%25.823.0No
2025-04-18Friday13.9%30.630.6Yes
2025-01-28Tuesday13.0%22.021.4No
2025-02-05Wednesday12.7%23.422.9No

What a team would do next

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.

Analysis code (Python) · Data download

Financial impact

About 8,760 MWh less forecast error a year on a 20 MW book

Lower forecast error means less energy bought or sold at spot to cover the gap between forecast and actual demand. The value depends on the book’s contracts and spot exposure.

InputValueSource
Average demand of the book20 MW small retailer or a group of large sites
Hourly error: regression with weather10.4%Measured
Hourly error: final model5.4%Measured
Average cost of each MWh of forecast errorA$30 spot imbalance against contract price
Hours a year8,760Public figure

Calculation

  1. Average error removed1.0 MW
  2. Error removed a year8,760 MWh
  3. Annual valueA$262,800

Annual value of lower forecast errorA$262,800

The study measures grid-level demand; a single book is noisier and must be backtested on its own meter data before this figure is used.

Implementation and rollout

Eight weeks from meter data to a forecast the desk trusts each morning

  1. Week 1

    Connect

    Connect the site’s or book’s meter data and the weather feed.

    Analyst, Manaz

  2. Weeks 2 to 3

    Backtest

    Backtest on the last 12 months against the current method.

    Manaz

  3. Weeks 4 to 8

    Run alongside

    Forecast issued by midnight beside the current one; the desk compares each morning.

    Trading or operations desk

  4. Then

    Extend

    Add price forecasts for battery scheduling if the demand gain holds.

    Desk lead, Manaz

Tools

AEMO NEMweb data, Open-Meteo or BoM weather forecasts, Python gradient boosting with conformal ranges, Excel or a BI dashboard for the desk

Team

  • Trading or operations lead owns hedging decisions
  • Analyst reviews the forecast and ranges each morning
  • Manaz pipeline, model, monitoring and monthly range review

The timeline is a plan for a typical engagement; it depends on data access and how many systems are involved.

Risks and mitigations

What could go wrong in production, and the control for each

RiskMitigation
Holiday weeks behave unlike normal daysHoliday flags and wider ranges; four of the five worst days were summer holiday weeks.
Range coverage drifts with the seasonMonthly coverage check and recalibration from the last 60 days.
Data feed fails before the midnight deadlineFallback to the same-hour-yesterday forecast with an alert.
Demand forecast read as a price forecastKept separate; buying decisions also need price forecasts and contract positions.

Key takeaways

Key takeaways

  • Recent demand and forecast temperature carry most of the gain; observed weather in hindsight barely helps (5.43%).
  • The peak hour, which drives the most expensive purchases, is forecast within 4.2%.
  • Raw quantile ranges covered only 58% of hours; conformal ranges fixed that and are checked monthly.
  • Rolling-origin tests over ten months show how the forecast behaves across seasons, not on one lucky split.