The energy bill fell: real savings or simply less production?

Bills, consumption and efficiency measure different things

A lower bill may reflect changed tariffs, output, hours or billing periods. First define the claim. Reduced cost is a financial outcome, lower energy use is a physical outcome and doing equivalent work with less energy requires a performance comparison. None automatically proves the others. A savings assessment needs an explicit counterfactual: how much energy would have been used under comparable conditions without the intervention?

Publishing a large percentage before establishing that comparison is weak evidence. Work begins with boundaries, intervention dates, data ownership and operating changes rather than model selection. If several interventions occur together, their individual effects may not be identifiable. The assessment must state which combined effect it measures and avoid attributing the entire observed difference to one application, algorithm or piece of equipment without an appropriate design.

Align meter boundaries and time intervals

A main meter and a production-line submeter cover different loads. Auxiliaries, compressed air, lighting and idle consumption may sit inside or outside the selected boundary. Meter identity, multiplier, flow direction and intervals belong in the data contract. Shift-based production reports and hourly energy records need a defined alignment rule. Dividing mismatched boundaries or durations can produce a plausible-looking but physically misleading intensity metric.

Consumption derived from cumulative meters needs checks for reset, rollover and late records. Missing intervals must not become zero consumption; any estimation or exclusion is disclosed. Timezone conversion and day boundaries can materially affect daily models. Raw observations remain available and cleaning rules are versioned so the assessment can be reproduced using the same data revision instead of relying on a number displayed on a chart.

A baseline must represent operating conditions

Choose a pre-intervention baseline that represents relevant operating conditions. Production, product mix, operating hours and temperature need physical justification. Not every correlated variable belongs in the model, particularly if it is unavailable at the required time or affected by the intervention itself. A simple linear model may be sufficient. Lower training error from a more complex model does not necessarily produce more credible savings evidence.

Test the model on withheld historical periods and examine residuals by production level and season. Predicting at output levels never observed in the baseline introduces extrapolation risk. That region may require more data or a different method. The assessment states valid operating conditions and triggers for reevaluation. A visually close fitted curve cannot establish that all operating regimes are represented or that the model remains valid after structural changes.

Section references: U.S. DOE · Measurement and Verification Options ↗

Worked example: a 40 MWh raw difference becomes 10 MWh

Assume a monthly baseline E = 40 + 0.8 × Q, where E is MWh and Q is tonnes. At Q = 100, actual consumption is 110 MWh while the adjusted baseline predicts 120 MWh. The difference is 10 MWh, approximately 8.33% of the baseline prediction. Comparing with a previous month's 150 MWh gives a raw reduction of 40 MWh, but counting all of it as savings would be incorrect in this example.

The coefficients are illustrative, not fitted to observations. The code preserves negative differences instead of clipping deterioration to zero and summing only favorable periods. Aggregation across periods requires consistent boundaries and methods. An unweighted mean of percentages may differ from the ratio calculated using total energy. Reports therefore include numerator, denominator, period and baseline version so the result remains interpretable rather than being reduced to an isolated percentage.

const assert = require('node:assert/strict');
// Coefficients are illustrative inputs, not fitted from real observations.
function adjustedDifference({ tonnes, actualMWh, intercept, slope }) {
  for (const value of [tonnes, actualMWh, intercept, slope])
    assert(Number.isFinite(value) && value >= 0);
  const baselineMWh = intercept + slope * tonnes;
  assert(baselineMWh > 0);
  return {
    baselineMWh,
    avoidedMWh: baselineMWh - actualMWh,
    relativeDifference: (baselineMWh - actualMWh) / baselineMWh,
  };
}
const result = adjustedDifference({
  tonnes: 100, actualMWh: 110, intercept: 40, slope: 0.8,
});
assert.equal(result.baselineMWh, 120);
assert.equal(result.avoidedMWh, 10);
assert(Math.abs(result.relativeDifference - 1 / 12) < 1e-12);
assert.equal(150 - 110, 40); // Unadjusted difference: not savings evidence.
// Simplified independent standard-uncertainty components, not a CI.
const combinedStandardMWh = Math.hypot(4, 2);
assert(Math.abs(combinedStandardMWh - Math.sqrt(20)) < 1e-12);
assert.throws(() => adjustedDifference({
  tonnes: -1, actualMWh: 110, intercept: 40, slope: 0.8,
}));
const increase = adjustedDifference({
  tonnes: 100, actualMWh: 130, intercept: 40, slope: 0.8,
});
assert.equal(increase.avoidedMWh, -10);
console.log(result, { combinedStandardMWh });

How uncertain is the 10 MWh difference?

Meter and model uncertainty both affect the result. In a simplified example with independent standard-uncertainty components of 4 and 2 MWh, root-sum-of-squares gives about 4.47 MWh. That is not automatically a 95% confidence interval. Parameter estimation, autocorrelation, period aggregation and dependence require further analysis. If independence does not hold, the same formula cannot simply be applied. The uncertainty calculation must expose its assumptions and scope.

A small observed difference may be indistinguishable from measurement and modeling uncertainty. More data, better submetering or a narrower boundary may then be needed. Low average model error can coexist with large bias in the interval of interest. Reporting uncertainty clarifies which decisions the evidence can support. Investment-payback calculations should expose assumptions and sensitivity rather than turning one estimated difference into a definite payback date.

Section references: NIST · Measurement uncertainty ↗

Exceptional changes can invalidate a baseline

A new production line, product family, building extension or permanent shift change can alter the meaning of the baseline. Such changes should not disappear into residual error. A change log links events to assessment periods. Adjustments need a method, rationale and source data. Selecting a favorable baseline or retrospectively changing coefficients to improve the result undermines verification; alternatives and their effects should remain visible.

Cost optimization and efficiency can occur together. Moving load to cheaper hours can reduce payment without reducing total energy. Peak demand, energy and tariff effects therefore remain separate. Reductions achieved by lowering output quality or service level also differ from doing equivalent work more efficiently. Analytics does not replace operating knowledge; technical and operational teams need to review the physical meaning of changes together.

A traceable path from SCADA to analysis

The pipeline connects meter observations, quality flags, cleaned series, operating context and model version. A report item should be traceable to source records. Late data can trigger recomputation without erasing the version used in an earlier report. Anomaly outputs require separate interpretation: detecting a consumption deviation does not establish its cause or prove a savings opportunity. Central review needs enough context to investigate the result.

Live dashboards and period-end verification reports serve different purposes. Operations may need the latest observation, while reporting needs a closed period and defined data rules. Data ownership, model ownership and operating approval remain distinct. An unauthorized coefficient change must not silently rewrite historical results. This structure helps explain how a number was produced, both internally and to stakeholders, without implying certification or an independently audited savings claim.

Acceptance evidence for an energy-analytics pilot

A pilot needs more than attractive charts. Review boundaries, missing intervals, baseline period, excluded observations and error distributions. Trace at least one known operating change from raw data to the report. The same data and model version must reproduce the result. Document conditions where the model should not be used, distinguishing decisions supported by current evidence from those requiring additional field measurement.

The following checks are a starting point and should reflect site-specific risk and decisions. Numerical targets follow data assessment. When expected savings are small, resolution and uncertainty deserve attention before modeling begins. The objective is a comparison clear enough to support an investment or operating decision. Discovering that a model is unsuitable or that data are insufficient is also a valid research outcome and belongs in the report.

• Do meter and production boundaries and intervals align?
• Was the baseline tested on withheld pre-intervention periods?
• Are uncertainty and exceptional changes visible in the report?
• Can the result be reproduced from raw data through the reporting chain?

Gokbilge Tech: from energy data to decision evidence

Gokbilge Tech connects energy analytics, big data and forecasting research with software engineering and field-data knowledge. Enlogy and SCADA work provide product and system context. A project can define data-quality assessment, model comparison, versioned calculations and central review. The example is not a measured Enlogy saving at a particular site and does not claim that the product already includes this exact model.

A shareable meter sample, production schedule, intervention date and intended decision can support an initial discussion. If evidence is missing, measurement scope comes first. Savings percentages and payback periods are not marketing commitments made before assessment. A sound scope explains which figures an organization can rely on, why that reliance is justified and what further work would reduce the remaining uncertainty.