Weather normalization: separate degree days and base demand
Energy sales can fall during warmer weather even when the customer base is unchanged. A useful research comparison holds weather constant while preserving the non-weather part of demand. Multiplying all sales by a degree-day ratio usually fails that task when a material base load does not depend on heating weather.
The deliverable is a service-area degree-day sheet and an explicit counterfactual demand calculation. All volumes and coefficients below are fictional. The calculation is a descriptive model, not proof of causal effects, an earnings forecast or a prediction of future weather.
Locate the degree-day definition and data type
NOAA Climate Prediction Center’s Degree Day Introduction defines heating and cooling degree days using a 65°F reference, daily mean temperature and a zero floor. It also explains population weighting and describes the page’s outlook ranges. Locate Introduction and Explanation of tables.
Use the page to establish definitions, not to relabel an outlook as observed weather. Actual-period analysis needs the relevant observed daily series. Reference-weather scenarios need a separately identified historical normal or chosen benchmark, including its period and method. The example’s one thousand HDD is a teaching benchmark, not a NOAA normal claimed for a real utility.
Preserve Fahrenheit or Celsius conventions. A coefficient estimated per Fahrenheit degree-day cannot be combined directly with Celsius degree-days. Record station, base temperature, missing-day treatment, period and weighting method beside every accumulated value.
Rebuild daily values before accumulating
For an illustrative daily mean of 55°F, heating degree days are max(65 − 55, 0) = ten and cooling degree days zero. At 75°F, HDD are zero and CDD ten. Across those two days, HDD total ten and CDD total ten.
Taking the two-day average temperature first gives 65°F and would incorrectly produce zero for both if the zero-floor formulas were then applied to that average. The daily threshold and accumulation matter. Do not net a warm day against a cold day before calculating the separate heating and cooling measures.
The demand model also needs separate roles for the two measures. A heating-sensitive product does not automatically respond to a cooling degree day with the same coefficient. If both matter, specify both terms and the evidence supporting them rather than add HDD and CDD indiscriminately.
Rebuild the main demand case
Assume a fixed reporting period and service population. Demand Q, in gigawatt-hours, is modelled as a base of one hundred plus 0.2 GWh for every Fahrenheit HDD. The coefficient is an explicit assumption, not an empirical estimate for an actual issuer.
At the chosen reference weather of one thousand HDD, model demand is 100 + 0.2 × 1,000 = 300 GWh. At actual weather of eight hundred HDD, expected demand is 100 + 0.2 × 800 = 260 GWh. The modelled weather difference is forty GWh.
Suppose observed sales are 250 GWh. The residual relative to the actual-weather model is −10 GWh. To express those sales at the chosen reference weather while retaining the residual, add the modelled weather difference: 250 + 0.2 × (1,000 − 800) = 290 GWh.
| Quantity | GWh | Interpretation |
|---|---|---|
| Reference-weather model demand | 300 | Base plus assumed heating response |
| Actual-weather model demand | 260 | Same base and coefficient, different HDD |
| Observed sales | 250 | Fictional measured input |
| Reference-weather adjusted sales | 290 | Observed sales plus forty weather units |
| Remaining difference to reference model | −10 | Residual retained, cause unproven |
The remaining ten-GWh gap is not automatically customer loss, efficiency improvement or a price effect. It can reflect omitted factors, measurement differences or an unsuitable coefficient. Treat it as a question for further evidence rather than a completed causal attribution.
Test the whole-volume ratio shortcut
Multiplying all 250 GWh by 1,000 ÷ 800 gives 312.5 GWh. That scales the base load as though it were entirely weather-sensitive and exceeds the additive adjustment by 22.5 GWh. It answers a different model, one without the stated fixed base.
The shortcut can only be justified if the appropriate evidence supports its assumptions. Do not choose it because it restores a larger adjusted sales figure. Show the base, weather-sensitive term and residual separately so the result can be reproduced and challenged.
The base is fixed only for the example’s matched period and population. Different billing durations or customer counts require their own treatment. A monthly base cannot simply be carried unchanged into a period with a different number of service days while claiming all other conditions are constant.
Test service-area weighting
For a separate two-area case, suppose Area A records six hundred HDD and B 1,200. Customer-exposure weights are 75% A and 25% B, giving a company-weighted value of 750 HDD. A different population weighting of 25% A and 75% B gives 1,050.
The difference is three hundred HDD, worth sixty GWh under the assumed 0.2 coefficient. Neither weighting is universally correct: it depends on the population being represented. NOAA’s regional population weighting does not automatically match a utility’s customers or the part of its business being analysed.
Obtain the service territory and exposure weights relevant to the product. Customer counts, delivered volumes or other weights answer different questions and must be labelled. If weights change between periods, preserve a constant-weight comparison or explicitly attribute the composition change instead of calling it weather.
Examine parameter stability and missing data
An unchanged linear coefficient is a modelling assumption. Changes in heating equipment, customer behaviour, efficiency, prices or industrial activity can make an old relationship unsuitable. These are hypotheses to test with company and demand data, not conclusions supplied by the degree-day definition.
If one station has missing days, do not treat them as zero HDD. Use a documented replacement method or report the incomplete accumulation and a conditional range. If station weights or service geography are unavailable, report the published regional measure as context rather than pretend it is the company’s measured exposure.
Separate fitted values from actual sales. A model can fit an earlier period well and still leave meaningful error later. Record the estimation period, coefficient version and any validation evidence when an empirical model replaces the invented coefficient. More decimal places cannot remove uncertainty about the specification.
Use the weather verification sheet
| Input | Obtain | Check |
|---|---|---|
| Weather | Daily observed temperatures or documented degree days | Base, units, missing days and dates |
| Territory | Company service map and weights | Customer exposure distinct from population average |
| Demand | Matched delivered-energy or sales series | Same product, period and service population |
| Model | Base, coefficients and estimation record | Assumed versus fitted and stable-period scope |
| Reference | Named benchmark and source vintage | Chosen scenario not relabelled official normal |
| Adjustment | Reproducible arithmetic and residual | Base not scaled with all weather-sensitive demand |
When reports conflict, first compare observed versus forecast weather, Fahrenheit versus Celsius, accumulation dates, weights and reference periods. Do not average incompatible HDD totals or select the weather series that makes adjusted growth strongest.
Update after corrected observations, revised demand, changed territory, a new reference baseline or evidence of parameter drift. Preserve original and revised adjustments so the effect of weather data can be separated from model changes. The conclusion should state the counterfactual weather, explicit coefficient and remaining residual, without presenting a weather-normalised series as proof of underlying demand causation.
Test coefficient choice and regional response
For a sensitivity that changes only the assumed HDD coefficient, 0.1 gives adjusted sales of 270 GWh and 0.3 gives 310, compared with 290 at 0.2. These are conditional results, not a confidence interval. The associated model predictions and residuals also change, so the original negative-ten residual cannot be retained as though the specification were unchanged.
In the separate two-area example, assume A’s coefficient is 0.1 and B’s 0.3 with the stated 75%/25% weights. The weighted heating term is 75% × 0.1 × 600 + 25% × 0.3 × 1,200 = 135 GWh. Using a common 0.2 times the weighted 750 HDD gives 150. The fifteen-GWh difference arises from the different response assumptions. A weather-weighting series alone cannot establish the appropriate demand coefficient.
FAQ
Can all sales be multiplied by the HDD ratio?
Only if that model is supported. It fails the stated example with a separate base load.
Are NOAA regional population weights automatically a company’s weights?
No. The service population and weighting purpose need to match.