Measure residual exposure interval by interval
A flat hedge can match annual energy while leaving substantial peak exposure. Forecast errors and price can be correlated: a hot evening can increase both demand and price. Keep volume, shape and locational basis distinct.
The lesson
7 chaptersBrowse the chapters
01Define what is being forecast
The Monday morning question
Your manager asks for tomorrow's energy exposure. A regional demand forecast is available, but the portfolio contains industrial meters, rooftop solar and a new data centre that is commissioning in stages. Start by identifying the quantity the decision actually needs. A regional MW figure cannot stand in for the portfolio's metered MWh.
| Field | Example for this desk exercise |
|---|---|
| Boundary | Net imports at the portfolio's customer meters |
| Resolution | Average MW in six consecutive 30-minute intervals |
| Issue time | Before the first interval; freeze both forecast versions |
| Horizon | The following three hours |
| Decision | Compare forecast quality and the resulting energy-volume mismatch |
| Exclusions | Network charges, losses, contract premiums, credit and taxes |
Check your understanding
Rooftop solar falls while customers use the same amount of electricity. Can net grid imports rise?
Show reasoning
Yes—more of the same consumption may need to come from the grid.
First check the meter boundary and onsite generation. A change in grid imports is not necessarily a change in the customer's total electricity use.
02Build a fair forecast comparison
Keep tomorrow out of today's inputs
Freeze the forecast issue time and preserve the data available then. Realised weather, revised meter readings and a commissioning date confirmed afterwards can make a historical reconstruction look deceptively accurate. They were not necessarily available when the decision had to be made.
From historical data to an honest test
- 01
Fit
Estimate the model using earlier observations.
- 02
Validate
Choose settings using a later validation period; compare with a simple baseline.
- 03
Test
Freeze the choice and evaluate on an untouched later period.
| Record | Why it matters |
|---|---|
| Issue time and target interval | Prevents using a late forecast against an earlier decision |
| Input versions and missing values | Makes revisions and data gaps visible |
| Benchmark | Tests whether added complexity improves on a simple method |
| Held-out dates and operating conditions | Shows whether peaks, outages and seasonal changes are represented |
Check your understanding
Can actual tomorrow temperatures be used when scoring a day-ahead model?
Show reasoning
Not as if they were known at the day-ahead issue time.
Use archived weather forecasts available then, or label the exercise as a hindsight diagnostic. The distinction changes what the test demonstrates.
03Read the profile, then calculate the error
Two forecasts, one observed load profile
| Interval | Actual MW | A MW | B MW | A$/MWh |
|---|---|---|---|---|
| 1 | 100 | 100 | 100 | 50 |
| 2 | 110 | 120 | 110 | 60 |
| 3 | 120 | 100 | 120 | 70 |
| 4 | 130 | 120 | 130 | 100 |
| 5 | 140 | 160 | 140 | 300 |
| 6 | 160 | 180 | 120 | 500 |
State the sign convention
eᵢ = forecastᵢ − actualᵢ; bias = Σ eᵢ / n; MAE = Σ |eᵢ| / n
All intervals here have equal duration and the same MW boundary. Positive bias means overforecasting. Some references define error with the opposite sign.
Work through forecast A
Subtract actual from forecast
Errors: 0, +10, −20, −10, +20, +20 MW
Do not discard the sign when checking systematic over- or underprediction.
Separate bias and magnitude
Bias = 20 / 6 = +3.33 MW; MAE = 80 / 6 = 13.33 MW
A small net bias can conceal sizeable positive and negative errors.
Result: Forecast B has a lower MAE of 6.67 MW, but its last-interval error is −40 MW.
Check your understanding
Do errors of +20 MW and −20 MW cancel for MAE?
Show reasoning
No. Their mean absolute error is 20 MW, although their bias is zero.
Bias measures net direction. MAE measures average error size. Neither alone describes when the errors occur.
04A better average can hide a worse peak
Give larger errors greater weight
RMSE = √(Σ eᵢ² / n)
The square root returns the metric to MW. Compare on the same target series, units and held-out intervals.
| Measure | Forecast A | Forecast B |
|---|---|---|
| Bias (MW) | +3.33 | −6.67 |
| MAE (MW) | 13.33 | 6.67 |
| RMSE (MW) | 15.28 | 16.33 |
| Last-interval shortfall (MW) | No shortfall; 20 MW over | 40 MW under |
Choose the test to match the decision
B wins on MAE, while A wins on RMSE in this small sample. Neither is automatically the operational choice. A capacity decision also needs peak shortfall, timing, forecast uncertainty and feasible corrective actions. Report which criterion matters and why it was selected before inspecting the result.
Check your understanding
Why is B's RMSE larger despite its lower MAE?
Show reasoning
The single 40 MW miss is squared before averaging.
For B, RMSE is √(1,600/6) = 16.33 MW. For A it is √(1,400/6) = 15.28 MW. Check both the aggregate and the individual intervals.
05Translate forecast error into energy exposure
A deliberately simplified commercial assumption
Assume a buyer has fixed-price financial coverage equal to the forecast volume in each interval, matched to the same spot reference as its load. The residual below is only the uncovered spot component. It excludes the fixed-price payment and all other costs. A forecast by itself is not a hedge or an executed purchase.
Residual spot component
Residual A$ = Σ (actual MW − covered MW) × 0.5 h × spot A$/MWh
Coverage is assumed to equal the forecast solely for this exercise. Positive is a residual cost; negative is a credit. Real contracts require their own volume, price, loss and settlement conventions.
Follow forecast B's peak miss
Reconcile the energy
Actual: 380 MWh; coverage: 360 MWh
Add interval MW multiplied by 0.5 hours. Do not add MW and label the result MWh.
Price the shortfall where it occurs
(160 − 120) MW × 0.5 h × A$500/MWh = A$10,000
The entire mismatch is in the last interval. Using an unweighted average price would conceal its timing.
Stress the price, not the arithmetic
At A$1,000/MWh: 20 MWh × A$1,000/MWh = A$20,000
This is an invented sensitivity, not a market-price forecast or a probability claim.
Result: B has the better MAE but leaves a costly peak shortfall under the stated coverage assumption.
| Quantity | A | B |
|---|---|---|
| Forecast / covered energy | 390 MWh | 360 MWh |
| Actual energy | 380 MWh | 380 MWh |
| Residual spot component | −A$7,100 | A$10,000 |
Check your understanding
Would a financial hedge remove a physical 40 MW connection shortfall?
Show reasoning
No. Financial coverage changes cash flows, not the site's physical capacity.
Keep the commercial and engineering workstreams separate. A data-centre ramp needs feasible connection capacity and operating arrangements as well as a procurement plan.
06Separate renewable resource, availability and export
A wind forecast is not an export promise
For this original example, a 50 MW wind farm's weather-to-power model predicts 30 MW before outages. Assume 80% of the turbines are available, with equal forecast output per turbine. The available-resource forecast is then 24 MW. An assumed site export ceiling of 15 MW limits delivered power further. Keep each reduction visible rather than labelling every shortfall ‘forecast error’.
Three distinct quantities
- 01
Resource forecast
30 MW from the weather-to-power model, before outages.
- 02
Available plant
30 MW × 80% = 24 MW under the equal-output assumption.
- 03
Export ceiling
min(24 MW, 15 MW) = 15 MW delivered at the defined site boundary.
Convert the export gap into energy
(24 − 15) MW × 0.5 h = 4.5 MWh
This is the energy gap caused by the assumed ceiling relative to available-resource output, not the loss from outages or the error in the original weather forecast.
Check your understanding
How much is exported in the half-hour, and why is it below the 15 MWh resource potential?
Show reasoning
7.5 MWh is exported: 3 MWh is removed by assumed availability and another 4.5 MWh by the ceiling.
The energy bridge is 15 − 3 − 4.5 = 7.5 MWh. Attribute these separately before judging the weather model. Different turbine output or a changing ceiling would require interval-level modelling.
07Write the recommendation your manager can check
Your assignment
Prepare one page comparing A and B. Include the six-row worksheet, an error chart, MAE, bias, RMSE, energy reconciliation and the residual spot component. Recommend what to test next rather than declaring a production winner from three hours of invented data.
| Deliverable | What a reviewer should be able to reproduce |
|---|---|
| Scope | Meter boundary, horizon, issue time and data versions |
| Calculations | 13.33 vs 6.67 MW MAE; 15.28 vs 16.33 MW RMSE |
| Energy and cost | 380 MWh actual; −A$7,100 vs A$10,000 residual under the stated assumption |
| Stress test | B's A$20,000 residual if the final price doubles |
| Recommendation | A longer test, a simple baseline, peak diagnostics and an accountable review date |
Connect the rest of the course
- 01
Units and data
Topic 1: MW to MWh. Topic 31: interval alignment and revisions.
- 02
Contracts
Topic 21: swap cash flows. Topic 22: PPA shape and basis.
- 03
Decision
Topic 35: coherent stresses. Topic 42: commissioning and data-centre load ramps.
Check your understanding
What is missing from ‘B is better because MAE is lower’?
Show reasoning
The decision criterion, peak consequences, evaluation scope and uncertainty.
A defensible brief explains what the metric captures, what it omits and which additional evidence is needed before committing capital or changing operations.
Watch the explanation
Key terms in plain English
- MWh
- Megawatt-hour: a quantity of energy. A steady 1 MW for one hour delivers 1 MWh.
- Hedge
- An arrangement that changes exposure to uncertain prices or other variables. It reduces specified risks, not necessarily every risk in a business.