Energy Baseline Adjustment Worksheet

Energy BaselineAdjustment Worksheet

Compare compressed-air energy at equivalent production and service conditions—not merely one electricity total against another.

1. Engineering answer: adjust the comparison, not the evidence

A compressed-air energy baseline describes how the existing system consumed energy before an improvement, under recorded operating conditions. An adjusted baseline estimates how much that same pre-improvement system would have consumed while serving the reporting period’s legitimate production and service requirements. The difference between that estimate and reporting-period energy is a candidate saving—not an automatic proof of causation or a guaranteed annual benefit. [U2]

The central question is: “What would the old system have used to provide the service actually required during this period?” A plant can consume less electricity because production fell, even while efficiency deteriorates. It can consume more electricity while becoming more efficient because production increased.

The worksheet therefore produces a baseline definition, accepted input dataset, adjustment model, diagnostic record, signed adjustment ledger and reporting-period comparison. It distinguishes measured energy, predicted energy, documented exclusions and unresolved gaps. Its result must remain reproducible without access to the person who originally prepared it.

2. What this page owns—and what it does not

A healthy vibration baseline and a pre-project energy baseline are related records, but they are not interchangeable. [U2–U5]

The worksheet does not select compressor capacity, establish safe pressure limits, certify air purity, set electricity tariffs or calculate an organisation’s carbon inventory. It does not convert a model fit into an independently verified saving.

Nor should a baseline be chosen to maximise apparent benefit. Preserve the actual pre-project operating history, including relevant waste that the project intends to remove. Document abnormal events separately. An old machine’s inefficient but representative operation can be a valid project baseline even though it would not be a suitable “healthy condition” reference.

3. Define the service, periods and comparison question

Record the intervention: leak repair, pressure reduction, sequencing changes, dryer optimisation, distribution improvement, equipment replacement or a defined combination. State when implementation began, when commissioning ended and when stable reporting began. Do not train a pre-project model on commissioning or post-improvement data without explicitly changing the analysis design.

Use a baseline period that covers the operating states relevant to the intended claim. Monthly data may be suitable for a broad production model; shift data may better distinguish batch operations; interval power and flow are needed to understand short control transitions. More rows do not necessarily mean more independent information. Thousands of consecutive samples from one operating state do not represent a seasonal production cycle.

Keep the physical service comparable: accepted production output, required pressure at critical users, air-quality requirements, operating schedule and continuity. Lower energy obtained by starving a production line is a service failure, not an efficiency improvement. Conversely, a justified reduction in excess pressure can be the project itself and should not automatically be adjusted out. [U2]

4. Freeze the energy boundary before choosing an equation

A baseline must specify what the meter includes. Use the same boundary in both periods or reconcile a documented difference. Do not compare a motor-only baseline with a reporting total that includes dryers, cooling pumps and ventilation. Do not add a dryer submeter to an incoming station meter that already includes the dryer.

Generator-fed electrical kWh can be included in an electrical-use comparison when metered consistently, but identify the supply split for downstream carbon work. Direct diesel-compressor fuel remains a separate activity unless a defensible conversion and reporting boundary are explicitly established. Litres and kWh are not added together.

5. Build an aligned, reviewable input dataset

Each analysis row should represent the same interval across energy, production, hours and contextual variables. A calendar-month electricity reading cannot be paired with a production report covering different closing dates without reconciliation. Record whether the timestamp marks an interval’s beginning or end. Apply meter multipliers once and retain raw readings before conversion.

Prefer meter-register differences or integrated interval energy. A momentary loaded-power reading multiplied by an entire month’s hours can materially misrepresent unloading, standby and variable-speed operation. Keep operational zeroes distinct from communication failures. If a meter resets or is replaced, reconstruct the discontinuity from evidence rather than treating the negative increment as negative consumption. [U3, U4]

Classify records as observed, corrected with trace, estimated, excluded or unresolved. Store exclusion reason, original value, revised value and responsible reviewer. A high-energy point may be a genuine process event rather than bad data. Excluding it simply because it weakens the regression is not an acceptable cleaning rule.

6. Select explanatory variables for the decision being assessed

A useful variable explains legitimate differences in service demand or external conditions. Its role depends on the project. Production operating hours may be a suitable adjustment when shifts change independently; compressor running hours may fall because controls improved and therefore be part of the benefit. Review this distinction before fitting anything.

Do not use revenue as a convenient substitute for production where prices or product value change independently of air demand. Record the physical reason for every included driver. A statistical association alone does not establish that the variable is an appropriate adjustment.

7. Choose the simplest defensible adjustment method

Do not require regression where matched operating periods answer the question. Equally, do not force every plant into a kWh-per-unit ratio. Select the method from the operating evidence and preserve its limits. Single metrics, ratios and models answer different questions. [U2; R1]

Record rejected alternatives and the practical reason for the choice. A simpler transparent model that predicts representative periods adequately is preferable to an elaborate model that reproduces historical noise but fails when the plant changes.

8. Why fixed demand makes simple ratios misleading

Suppose energy over equal-length periods follows E = a + bQ, where Q is production, a is an estimated fixed-period component and b is incremental energy per production unit. Dividing by production gives E/Q = a/Q + b. Even without any physical improvement, the ratio falls as output increases because the fixed component is spread across more units.

In compressed-air systems, a fitted intercept can reflect several effects: persistent air use, treatment auxiliaries, standby power, schedule and unmodelled behaviour. It is not automatically a measured leakage load. Do not present the intercept as “leak energy” without independent evidence.

A fixed-period intercept also cannot be applied unchanged to arbitrary period lengths. A model developed for equal 30-day periods needs duration treatment before use with 7-day or 31-day totals. An alternative structure can include calendar hours and production hours explicitly, provided the data distinguish their effects.

9. Model equations, units and calculation direction

For a one-driver baseline, define predicted period energy as E_hat = a + bQ. Here E_hat and a are in kWh per stated period; Q is production in the stated unit per period; b is kWh per production unit. Record the fitted coefficients without premature rounding and the unit used when fitting.

A possible multivariable form is E_hat_i = beta0 + betaQ × Q_i + betaH × H_i + betaT × T_i + betaM × M_i. This is a candidate form, not a universal compressor model. H may represent required service hours, T a documented environmental measure and M a coded production mode. Each term needs engineering justification; correlated or redundant terms must not be retained merely because software accepts them. [R2]

For a supported operating-mode model, E_base_routine = sum(P_base,j × H_report,j). P_base,j is representative pre-project kW for mode j; H_report,j is the reporting period’s hours of equivalent required service in that mode. The product is kWh. Preserve the distinction between observed reporting energy and modelled pre-project energy; the second is not a meter reading.

10. Fit the model without using the result to choose the rules

For ordinary least squares with one driver, b = sum[(Q_i − Q_mean)(E_i − E_mean)] / sum[(Q_i − Q_mean)^2], and a = E_mean − b × Q_mean. The fit minimises the sum of squared residuals. A zero denominator means production did not vary, so the slope cannot be identified from that dataset. [R3]

Predefine the baseline window, inclusion rules, candidate drivers and evaluation method. Keep training observations separate from later validation observations where practical. Time-ordered checks can reveal drift that a random split hides. Document why observations are independent enough for the proposed uncertainty calculation; frequent sampling alone does not justify independence.

If multiple predictors are used, inspect rank, correlation and coefficient stability. A negative production coefficient is a reason to investigate the model, data or confounding variables—not permission to claim that making more product saves energy. An intercept can be negative in a local empirical fit, but predictions must remain physically credible over the approved range. Do not clamp negative outputs to zero and call the model validated.

11. Residual checks: test how the model fails

Define residual e_i = E_observed,i − E_hat,i. A positive residual means the observation used more energy than predicted; a negative residual means less. Inspect residuals against time, predicted energy, production and operating mode. Curvature can indicate a missing nonlinear relationship; a widening spread can indicate changing variance; a step can indicate a configuration or measurement change. Consecutive patterns can indicate dependence or omitted operating structure. [R4]

Let SSE = sum(e_i²), n be accepted observations and p the number of fitted coefficients, including the intercept. Residual standard error s = sqrt[SSE/(n − p)] requires n > p. Define any normalised error metric explicitly. In this worksheet, CV(RMSE)_df = 100 × s/E_mean; an RMSE calculated with divisor n is a different quantity. [R3]

R² = 1 − SSE/SST, with SST = sum[(E_i − E_mean)²], is undefined when SST is zero. A high R² does not prove causation, a suitable comparison boundary or predictive accuracy outside the observed range. Do not impose a generic pass threshold. Record the project’s accepted diagnostics, tolerance and uncertainty basis before using the model for a savings claim.

12. Record the valid range and the operating envelope

Store minimum and maximum observed values for each driver, but also review combinations. A reporting row can fall within each variable’s individual range while representing a combination never observed—for example, maximum production at minimum service hours. A rectangular range check is a useful warning, not complete evidence of interpolation.

New product grades, higher-pressure users, changed dryer duty or a different base/trim arrangement can break an old relationship. Identify those rows separately. Do not silently extrapolate, discard them or fit a replacement model to post-project data and describe it as the original baseline. Use a qualified engineering estimate, targeted comparison or new baseline with a documented bridge where the evidence warrants it. [U2; R2]

Present eligible-period results and excluded-period coverage separately. A result for the measured weekday campaign must not become an annual result for weekends, shutdowns and future expansion without a defensible operating-hours model.

13. Separate routine and non-routine adjustments

Routine adjustment applies the accepted baseline model to reporting-period external drivers. It does not rewrite historical meter data. Non-routine adjustment addresses a supported structural difference not already captured—for example, an independently metered auxiliary added inside the station boundary for an unrelated purpose.

Use this explicit sign convention: E_base_adjusted = E_base_routine + A_NR. A positive A_NR increases the energy the old system would have needed at the reporting boundary; a negative A_NR reduces it. Record the event date, affected equipment, calculation, supporting evidence, uncertainty and approval. State whether the adjustment applies to all reporting rows or only a defined interval.

Never apply the same change twice. If new production is already represented by the model’s production input, do not add an independent “expansion allowance” for the same load. A separately metered addition can sometimes be handled either by adding it to the comparison baseline or by removing it from reporting energy, but the chosen service boundary and percentage denominator must remain consistent.

14. Do not adjust away the improvement you are measuring

Choose variables according to the project’s causal boundary. If leak repairs reduce required compressor flow, evaluating the old generator only at the lower post-repair flow measures generation performance at that flow; it does not capture the complete benefit of eliminating unnecessary demand. If sequencing reduces unloaded hours, fixing the baseline to the shorter post-project compressor runtime can erase part of the intended saving.

The same issue arises with pressure. For a pressure-optimisation project, compare old-system operation and the accepted improved system while both meet the required user service. Automatically normalising both to the new lower header setpoint can remove the mechanism that produced the saving. For an unrelated production change that genuinely requires higher pressure, pressure-regime adjustment may instead be necessary.

This is a new implementation safeguard derived from the source distinction between production differences, control changes and interacting measures. It is not a claim that statistical software can establish causation without engineering evidence. [U2]

15. Time weighting, missing periods and aggregation

For valid interval-average power, E = sum(P_i × delta_t_i), with P in kW and delta_t in hours. Sum interval kWh directly when that is the measured quantity. Never sum kW readings as if they were energy, or multiply an accumulated kWh register by hours again.

Retain valid observed hours, explicitly estimated hours and unresolved hours separately. Do not fill a long logger gap with the preceding value or treat it as zero. A justified short-gap estimate remains an estimate in the export. Missing production can make an otherwise valid energy observation unusable for adjustment.

Match the aggregation level to the model. An hourly intercept applied once to a monthly total produces a different result from applying it to each hour. Apply nonlinear or segmented models at their approved interval before summing where aggregation changes the result. Retain chronological data even when a demand-duration summary is useful; sorted duration curves cannot demonstrate timing or residual autocorrelation.

16. Calculate savings and name the denominator

For the eligible reporting period, S = E_base_adjusted − E_report. Positive S is lower reporting consumption than the adjusted baseline; negative S is excess consumption relative to that baseline. Do not suppress negative values or relabel them as zero saving. When E_base_adjusted > 0, adjusted reduction (%) = 100 × S/E_base_adjusted.

Show three distinct quantities: raw baseline-period energy, adjusted baseline energy for the reporting service, and measured reporting-period energy. The first and third describe actual totals for their own periods. The second supplies the counterfactual comparison. They should not share a label such as “actual baseline” that hides the model.

An additional fixed-condition comparison can evaluate accepted before-and-after models at a common reference duty. Label it normalised performance, not actual reporting-period avoided energy. It answers a different question. Do not multiply the common-duty percentage by an unrelated annual bill. Annualisation, cost conversion and carbon conversion each need their own declared period and inputs. [U2]

17. Quantify uncertainty without manufacturing confidence

For S = B − R, a combined standard-uncertainty expression is u(S)² = u(B)² + u(R)² − 2 × Cov(B,R). Use covariance only on a defensible basis. Shared meters, calibration effects or fitted parameters can make independence inappropriate. Add separate adjustment uncertainty only when it is not already inside u(B).

A model’s mean-response uncertainty is not the same as prediction uncertainty for another individual observation. For ordinary independent-error regression, prediction uncertainty includes residual variation in addition to uncertainty in the estimated mean. Statistical intervals require assumptions, degrees of freedom and the selected coverage level. Do not attach “95% confidence” to an arbitrary plus/minus range. [R5]

For a sum of predicted periods, shared model coefficients create covariance among predictions. Do not automatically divide residual error by the square root of the number of logged rows. Report the interval-level or aggregate method actually used. Model uncertainty, meter uncertainty, imputation uncertainty and non-routine-adjustment uncertainty can overlap; document the budget rather than adding all labels indiscriminately.

18. Worked example A — lower consumption, worse adjusted performance

Illustrative assumptions: equal-length operating periods, unchanged accepted air service per production unit, and an already established baseline E_hat = 12,000 + 0.8Q. The coefficients are illustrative, not a universal plant rule. The baseline reference period produced 60,000 units and used 60,000 kWh. Reporting production fell to 40,000 units while measured consumption was 46,000 kWh.

The raw total fell substantially, but the reporting period used 2,000 kWh more than the model expected for its production. The worksheet should show an adverse adjusted difference, subject to model validity and uncertainty. It should not issue a 23.33% efficiency-improvement claim.

This example also shows why a lower bill is not enough to demonstrate project success. Production loss, schedule changes and tariffs must not be mistaken for compressor-system efficiency. No customer measurement or achieved outcome is implied.

19. Worked example B — fit a baseline and retain its residuals

The following six synthetic equal-duration baseline periods demonstrate the arithmetic. Their small number is for readability, not a recommendation for field-data sufficiency. All share the same illustrative service conditions. Ordinary least squares gives a = 12,000 kWh/period and b = 0.8 kWh/unit.

The mean observed energy is 64,000 kWh. SSE is 4,480,000 kWh²; p = 2 and n = 6. Residual standard error is 1,058.30 kWh; RMSE with divisor n is 864.10 kWh; CV(RMSE)_df is 1.65%; R² is approximately 0.996016. These numbers describe this constructed fit. They do not certify a production model, prove independent residuals or establish a universal acceptance criterion.

For a reporting period producing 75,000 units, the adjusted baseline is 72,000 kWh. With measured reporting consumption of 63,000 kWh, the candidate saving is 9,000 kWh, or 12.50%. Comparing instead with B3’s raw 60,400 kWh would show a 2,600 kWh increase and miss the change in required output.

Retain every residual, the full-precision coefficients and the 40,000–90,000-unit observed range. A reporting input of 110,000 units must trigger an out-of-range warning, even though the equation returns a number. Holdout performance, service checks and uncertainty remain separate requirements.

20. Worked example C — mode weighting with a boundary addition

Synthetic pre-project observations support three service modes. Reporting-period hours refer to required production/service states, not compressor runtime changed by the project. Assume the modes remain comparable and all 200 reporting hours are represented.

An unrelated new ventilation auxiliary was added inside the incoming meter boundary. Its independent submeter records 1,200 kWh during this period; assume it would consume the same energy with or without the compressed-air improvement and introduces no interaction. The agreed non-routine adjustment is therefore +1,200 kWh.

Adjusted baseline = 18,360 + 1,200 = 19,560 kWh. With reporting meter consumption of 17,200 kWh, candidate saving = 2,360 kWh, or 12.07%. Omitting the addition gives only 1,160 kWh. Adding it twice gives an unsupported benefit.

The adjustment is supported here by deliberately stated illustrative assumptions. In a real plant, a new pneumatic production load cannot automatically be treated as an independent electrical addition: its effect on compressor dispatch, pressure, treatment and fixed losses must be established. Record uncertainty and interaction limits with the adjustment.

21. Worked example D — coverage and uncertainty restrict the claim

A synthetic reporting window contains 240 scheduled hours, but only 216 hours have aligned valid energy and production observations. Raw coverage is 90%. For those covered hours only, adjusted baseline energy is 100,000 kWh and observed energy is 95,000 kWh. The numerical difference is 5,000 kWh, or 5% of the covered-period adjusted baseline.

Assume the aggregate baseline standard uncertainty is 4,000 kWh and reporting-energy standard uncertainty is 2,000 kWh, with independence justified for this illustration. Then u(S) = sqrt(4,000² + 2,000²) = 4,472.14 kWh. A declared coverage multiplier of 2 gives an expanded interval of 5,000 ± 8,944.27 kWh, extending approximately from −3,944 to +13,944 kWh. No particular confidence probability is asserted for that multiplier.

The interval crosses zero. Report an indicative difference whose positive direction is not resolved by this uncertainty budget, not a verified 5% saving. The missing 24 hours also prevent treating the covered-period result as the complete 240-hour result. Do not scale by 240/216 unless a separate representative-gap model is explicitly justified and reported as estimated.

22. Editable baseline definition and reporting worksheet

Complete the following records for each project. Duplicate the reporting row for additional periods. “Unknown” is a permitted data state; it is not a numerical zero. The tables are manually editable in Word and do not automatically recalculate.

Attach the aligned dataset, exclusion log, fitted/predicted values and residuals. Retain an unmodified source export so another reviewer can recreate the table. Where no adequate baseline exists, record the missing evidence and collect a supported comparison; do not populate the worksheet with a supplier’s generic savings percentage.

23. Interpret the result before sending it to a financial model

A positive point estimate can be useful even before independent verification, but its state must be explicit. Separate a preliminary estimate from an accepted project result. Explain what changed, how the counterfactual was built, which service checks passed and which uncertainties remain.

A persistent negative adjusted difference deserves investigation. Possible explanations include a real efficiency deterioration, altered product mix, higher required duty, bypassed treatment, changed meter scaling, control overrides or a baseline that no longer represents the plant. The worksheet should not diagnose a mechanical fault from the residual alone. Use the diagnostic and condition-monitoring pages for the appropriate follow-up.

Tariff escalation or reduced electricity prices cannot make physical kWh savings larger. Keep overlapping leak, pressure and sequencing projects reconciled at the station boundary rather than adding independent claims that contain the same reduction. [U2]

24. Preserve the original baseline and monitor persistence

Freeze the accepted project baseline before reporting savings. Keep the input snapshot, model equation, coefficient precision, variable roles, range, diagnostics and adjustment convention. A later correction should create a traceable revision with the original and revised results, not silently replace the history.

A post-project operating baseline can support ongoing monitoring, but retain its separate purpose. It should not replace the old counterfactual while claiming cumulative project benefits. Likewise, an equipment health baseline after overhaul belongs to condition monitoring; it does not automatically reset the project’s energy accounting. [U2; R6]

Review residual trends after stable operation and after meaningful changes. Record whether a change is a meter issue, unrelated expansion, altered required service, project deterioration or an unresolved event. A monitoring reset that removes every adverse residual can conceal lost savings. Preserve the cumulative series, explained discontinuities and any revised model’s effective date.

25. Practical questions and direct answers

Is lower electricity consumption proof of better efficiency?

No. Consumption can fall because production or required operating hours fell. Compare the reporting period with the energy the previous system would have needed for the same legitimate service. Retain the raw totals as well as the adjusted result.

Can I use kWh per tonne or per finished component?

Yes, when the production measure is meaningful and fixed demand or product mix does not distort the comparison. If fixed loads are material, a base-plus-production model or a mode-based comparison may be more informative. Do not assume that a falling ratio proves a physical improvement.

How much historical data is enough?

There is no universal row count. The dataset must represent relevant production, schedule, environmental and control states with enough independent information for the chosen model and checks.

Should I adjust for compressor running hours?

Only after checking the purpose. If improved controls reduce unnecessary runtime, that reduction is part of the project benefit. Adjusting the baseline to the shorter runtime can hide it. Required plant service hours are often a different and more suitable variable.

Should leakage-repair savings be normalised to post-repair airflow?

Not automatically. Post-repair airflow may be lower precisely because leakage was removed. Comparing generator efficiency at the new flow can be useful, but it does not by itself measure the full demand-reduction benefit. Define whether the assessment concerns generation efficiency or total station energy for unchanged useful service.

Does a high R² mean the baseline is accepted?

No. Review residual behaviour, meaningful coefficients, operating range, independent prediction and measurement boundaries. A fit can look excellent while adjusting away the intervention or extrapolating into an unsupported regime. R² is one descriptive statistic, not an approval decision.

What happens when a new production line is connected?

First determine whether the model already represents its demand and product mix. Otherwise document a non-routine adjustment, supported separate analysis or a revised baseline with an explicit bridge. Do not add both a production adjustment and an expansion allowance for the same service.

Can missing logger data be treated as zero?

No. Confirmed zero demand or stopped equipment can be valid; a missing signal is not. Keep observed, estimated and unresolved time separate. Report a covered-period result where appropriate instead of presenting an incomplete dataset as a full-period measurement. [U4]

What does a negative saving mean?

Reporting energy exceeded the adjusted baseline. Preserve the sign and investigate operating changes, measurement issues and performance. It may be an adverse result or an invalid comparison; the calculation alone does not determine the cause.

Can this worksheet automatically verify savings?

No. The written worksheet records and calculates an assessment using entered evidence. Actual verification also requires a suitable baseline, trustworthy measurements, comparable service, justified adjustments and review. The Word tables are editable records, not an installed or runtime-tested application.

IMPLEMENTATION APPENDIX — NOT PUBLIC PAGE COPY