Bad Data,
Confident Answers

Energy analysis often begins too late. A spreadsheet arrives with interval values, a utility bill, equipment ratings, weather, a tariff, and vendor performance data. The analyst starts modeling. The more disciplined sequence begins one step earlier: identify the evidence boundary, audit the inputs, and write down what the data cannot support.

Section 01The model is not the evidence

A model is a transformation. It turns inputs and assumptions into outputs. It does not establish that the meter belongs to the site, that a timestamp represents local clock time, that a blank interval means zero load, or that a nameplate rating is an operating measurement. Those are evidence questions. If they are left unresolved, the model can become a machine for expressing uncertainty with impressive precision.

The U.S. Department of Energy’s current measurement and verification guidance treats data collection, validation, and proper application as separate obligations. For whole-facility work, it calls for metered energy, relevant independent variables such as weather, and information about unrelated site changes. It also says the start and end dates of those streams must align.¹ That is not paperwork around the analysis. It is the analysis’s foundation.

The audit should therefore precede technology selection. It asks six basic questions: whose data, for what physical boundary, measured in which unit, over what time basis, with what missing or estimated values, and reconciled to which independent record. Until those questions have answers, the correct output is not a preferred path. It is a data request.

Section 02Start with the meter map

The most consequential data error is often not a bad reading. It is the right reading from the wrong boundary. A campus may have a master meter and several submeters. A tenant may receive an allocation rather than a utility bill. A service agreement may cover two buildings while an operating plan changes only one. A gas account may include process fuel and space heat. A rooftop solar meter may sit behind the revenue meter while a spreadsheet treats it as if it were outside the boundary.

Build a meter map before building a load profile. For every series, record the service address, account and service-agreement identifier, meter identifier, commodity, physical coverage, ownership, unit, interval length, multiplier, source system, extraction date, and the person who can explain it. Then draw the relationship among revenue meters, submeters, generation meters, storage meters, process sensors, and control-system points.

This is where authorization matters as well as identity. PG&E’s Energy Data Hub distinguishes customer downloads from ongoing third-party access and lets customers choose the scope of data shared. Its current resources can provide electric interval data and, through Green Button, electric or gas interval and billing data for an authorized service agreement.² The existence of data does not erase the owner’s duty to control who receives it and for what purpose.

A meter map can reveal that the study boundary is incomplete. That is useful. It is better to state that one production line is unmetered than to force a sitewide total into a line-level conclusion. The missing boundary becomes an explicit measurement task, not a hidden assumption.

Section 03Time is an input, not a label

Interval files create the appearance of chronology. A timestamp column can still conceal several conventions: interval beginning versus interval ending, local time versus Coordinated Universal Time, standard time versus daylight time, and duplicated or missing clock hours at seasonal changes. A daily total may be a sum of energy intervals or an average of demand intervals. These are not interchangeable.

The Energy Information Administration defines advanced metering infrastructure as recording usage at least hourly and making data available at least daily.³ A particular commercial file may be finer than that, but “smart meter” alone does not establish its granularity. DOE’s utility-rate guidance notes that available interval data may show hourly or 15-minute consumption and monthly peaks.⁴ The file’s actual reading type must control.

Time conventions also change when data move between systems. PG&E’s current third-party API instructions specify historical request windows in Zulu time and return interval and billing data at the individual usage-point level.⁵ Weather files, tariff calendars, and operating logs may use different time conventions from the meter export. A load file in prevailing local time cannot be paired blindly with a standard-time series or a calendar whose interval orientation has not been stated.

A time audit documents timezone, daylight-time treatment, interval orientation, leap-day treatment, reading duration, and any clock correction. It then tests continuity. Every duplicate, gap, overlap, and irregular interval is counted before the file is resampled. Resampling first can hide the defect.

Section 04Units and aggregation can reverse the result

Energy and demand answer different questions. Kilowatt-hours measure energy over a period. Kilowatts describe a rate of use, commonly averaged over a defined interval for billing. Gas may arrive as therms, energy content, or volume under stated conditions. Fuel inventory may record deliveries rather than consumption. Thermal sensors may report temperature without flow, making useful heat unknowable.

The audit must preserve native units, multipliers, and sign conventions before conversion. A negative value may represent export, a correction, storage discharge, or bad data. A zero may mean no use, a disconnected channel, a communication failure, or a placeholder. A monthly peak cannot be reconstructed from monthly energy. A 15-minute peak can disappear when a series is averaged to an hour.

Reconciliation catches many of these errors. Sum interval energy over each billing period and compare it with billed consumption after accounting for exact start and end dates, meter multipliers, estimates, and adjustments. Compare calculated interval peaks with billed demand, recognizing that tariff ratchets or billing rules may create differences. If the two records do not reconcile, preserve the discrepancy and investigate it. Do not “scale the curve to the bill” without recording why.

Small unit errors can become material business errors. As an explicitly illustrative example, a one-cent-per-kilowatt-hour mistake applied to ten million kilowatt-hours becomes $100,000 per year. That arithmetic is exact; the example is not a claim about any site. It shows why a tariff unit, escalation convention, or energy total deserves the same scrutiny as a capital-cost line.

Section 05Missing, estimated, and abnormal are different

A data-quality report should distinguish at least four conditions: truly measured, utility-estimated, analyst-imputed, and unavailable. It should also preserve abnormal operations rather than deleting them automatically. A shutdown, heat wave, maintenance outage, temporary generator run, production trial, vacancy period, or equipment failure may be exactly the event that determines resilience or capacity need.

DOE recommends reviewing 12 to 36 months of utility bills to identify anomalies and seasonal variation when establishing an energy baseline, while emphasizing that the measurement period must cover the relevant operating conditions.¹ More history is not always better. A long record that crosses a major process change may describe several different facilities under one account number.

For every gap, record its start, duration, affected channel, likely cause, treatment, and decision impact. Short gaps may be imputed for an annual energy estimate if the method is disclosed. The same imputation may be unacceptable for peak sizing or outage analysis. The rule is decision-specific: a treatment allowed for one question does not become valid for every question.

Outliers require the same discipline. Test them against bills, operator logs, weather, alarms, production records, and adjacent meters. Delete only when evidence supports deletion. If the cause cannot be established, run the decision both with and without the point and show whether the conclusion changes.

Section 06Operating context completes the load

A meter records the result of operations, not their cause. Weather, occupancy, production, batch schedules, sanitation cycles, charging windows, irrigation, water flow, and equipment availability can all move energy use. DOE’s 50001 Ready guidance calls for relevant variables, static factors, operational criteria, past and present consumption, and estimates of future use. It also requires a defined data-collection plan covering location, owner, frequency, storage, and analysis method.⁶

The owner should build an event ledger beside the interval file. Each operational change receives a date range, description, affected equipment or area, evidence source, and likely direction of impact. The ledger prevents a model from treating a new production shift as weather sensitivity or a shutdown as permanent efficiency.

Future load deserves its own register. Separate committed projects, approved but unscheduled changes, credible scenarios, and ideas. State the evidence behind each: signed equipment purchase, board-approved plan, production forecast, engineering design, or interview. Do not blend future additions into the historical series and call the result measured.

Calibration also belongs here. DOE’s current 50001 Ready guidance identifies equipment, method, tolerance, frequency, responsibility, and records as core parts of a calibration program.⁶ NIST’s measurement guidance similarly treats uncertainty as a combination of repeatability, calibration information, and possible systematic effects.⁷ A sensor’s decimal places are not proof of its accuracy.

Section 07Grade evidence before using it

Every material input should carry an evidence grade. A simple four-level register works well:

• Measured: a traceable reading from an identified meter or sensor, with time basis, unit, multiplier, and known calibration state.
• Documented: a current bill, tariff, executed service record, equipment submittal, contract, or operator log that states the fact directly.
• Modeled: a value derived from measured or documented inputs using a disclosed method.
• Illustrative: an assumption used to test sensitivity, labeled so it cannot be mistaken for a site fact.

Nameplate data and vendor literature are documented evidence of a stated rating, not measured site performance. An operator interview is useful evidence of practice, not a timestamped record of every hour. A public weather station is measured evidence at that station, not necessarily at a roof, intake, or process location. The grade describes what the evidence can support, not whether the source is respectable.

Each modeled output should inherit the weakest material input that can change its conclusion. A recommendation driven by an illustrative gas price or an unverified utility date is not “bankable” because the rest of the workbook uses measured data. It is a conditional answer with a named verification task.

Section 08The technology-neutral data test

Different paths fail on different missing inputs. Auditing the evidence before choosing a platform keeps the comparison honest.

No row can be completed from annual bills alone. Conversely, no row requires perfect data before any decision can be made. The right standard is fitness for the question. Monthly bills may support an early annual-cost screen. Interval and operating data are needed for peak, dispatch, and coincidence. Site measurements and written provider terms are needed before equipment-specific claims become firm.

Section 09The pre-model release gate

A useful audit ends with a decision, not a cleanliness score. Release the model only when the evidence is sufficient for the claim it will make. The following gate is deliberately plain:

1. IdentityEvery data series is tied to an exact site, service agreement, meter or sensor, physical boundary, and responsible source.
2. TimeTimezone, interval orientation, daylight-time treatment, leap-day handling, and continuity are documented and tested.
3. UnitsNative units, multipliers, demand averaging periods, signs, and conversions are preserved and reviewed.
4. CompletenessMissing, estimated, imputed, duplicated, and abnormal values are counted, explained where possible, and assigned a treatment.
5. ReconciliationInterval totals and peaks are compared with bills, operating logs, production records, and adjacent meters where available.
6. ContextWeather, occupancy, production, outages, schedule changes, and future loads are aligned to the same period.
7. Evidence gradeMeasured, documented, modeled, and illustrative inputs are visibly distinct, with current source dates.
8. Decision sensitivityThe owner knows which unresolved input can change the recommendation and what evidence would close it.

A failed gate does not always stop the study. It changes the output. The paper may support a screening range, identify a metering campaign, or compare scenarios conditionally. What it cannot do is promote an estimate to a fact because a decision meeting is scheduled.

Section 10Confidence should follow the evidence

The last page of an energy analysis should show more than economics. It should show the evidence register, open data defects, sensitivity of the recommendation, and the exact facts that must become firm before capital moves. That makes the work easier to challenge and easier to update.

Good input discipline does not favor the utility, efficiency, storage, solar, generation, thermal integration, or waiting. It favors the owner. Each path gets to compete on the evidence it actually needs. Weak data can no longer create a quiet advantage for the technology whose assumptions are easiest to hide.

The standard is not certainty. Energy decisions occur under changing loads, tariffs, weather, equipment availability, and schedules. The standard is traceability: every important number has an identity, a time, a unit, a source, a treatment, and a stated limit. Once those are visible, the model becomes what it should have been all along: a tool for judgment, not a substitute for it.