Power Meter Data Collection: An RFP for Trusted 15-Minute Data

Power meter data collection is often specified as “read 30 Modbus meters every five minutes.” That sentence can deliver a connected system while still producing data that is unsafe for investment decisions. A CT or PT ratio may be applied twice, import and export signs may be reversed, meter clocks may drift, fifteen-minute boundaries may not match the tariff logic, and missing readings may be converted to zero. These defects are dangerous precisely because the dashboard keeps running and the charts still look plausible.

This guide shows how a factory in Thailand can procure power meter data collection as a defined, testable data product. It translates CT/PT ratio, sign convention, time, fifteen-minute demand and missing-data treatment into an RFP, FAT, SAT, a 90-day proof of concept and an acceptance evidence pack. The purpose is broader than lowering an electricity bill: it is to prevent bad data from ranking capacity, energy and production investments incorrectly.

Procure a data contract, not merely a meter connection

Meters, RS-485 cables, gateways, Ethernet switches, servers and dashboards are visible assets. Procurement documents therefore tend to concentrate on makes, quantities, protocols, retention and screens. Management decisions, however, depend on whether every number has an agreed meaning, time reference, transformation history and quality state.

A display of “1,250 kW” cannot be compared or approved until the buyer can answer:

• Has the meter already applied the CT/PT ratio, or will a gateway or server apply it?
• Is the value instantaneous power, a five-minute average, a fixed demand block ending at 00, 15, 30 or 45 minutes, or a moving window?
• Does a positive sign mean grid import, flow into a machine, or export from a generator?
• Is the authoritative timestamp created by the meter, gateway or server, and in which timezone?
• If one of three expected five-minute values is missing, is the fifteen-minute result missing, estimated or calculated from only two points?
• If a value is corrected later, can an auditor recover the original value, reason, mapping version and effective time?

The RFP should therefore require a version-controlled meter register, point map, data dictionary, transformation rules, quality codes, clock rules, interval rules and test evidence. Connectivity is the delivery mechanism. Decision-grade data is the deliverable.

Keep Thai tariff context separate from plant assumptions

Thailand’s Energy Regulatory Commission announced an Ft of 16.23 satang/kWh for September–December 2026 and stated an average electricity price of 3.95 THB/kWh. This is authoritative public context, but it is not a universal marginal electricity price for every factory. A plant’s invoice depends on its service area, tariff category, demand and time-of-use structure, Ft, taxes and contract conditions. An actual business case must use the plant’s current bills and applicable tariff.

For relevant large-user categories, MEA describes maximum demand in relation to the highest fifteen-minute integrated demand in the billing period. The phrase “fifteen minutes” is still incomplete for system design. Schneider Electric’s EM3570 documentation distinguishes Fixed block, Sliding block and Rolling block demand calculations. The RFP must identify which method, interval boundary and formula apply before anyone compares a factory dashboard with a utility bill.

Factories in PEA service areas should consult PEA’s official tariff material and the latest applicable publication. The May 2023 English schedule is useful for definitions and tariff structure, but a 2026 approval must not copy an old rate into a forecast. A factory submeter is also an operational instrument; it does not automatically acquire the metrological or contractual role of the utility billing meter.

Do not accept a Modbus power meter because “the register responded”

The Modbus Application Protocol Specification V1.1b3 defines 16-bit registers and protocol addresses relative to zero. Device manuals may use human-facing references such as 40001. If the difference between a one-based reference and a zero-based offset is not documented, a request can read an adjacent register and still receive a technically valid response.

Power, energy, current and power factor may also be stored as 32-bit integers or IEEE 754 floating-point values across two 16-bit registers. Word order, byte order, signedness, scale, decimal position, unit, invalid value and counter rollover must be frozen per device and firmware. “Modbus TCP supported” is not a data-quality specification.

The accepted point map should reference the exact serial number, firmware, test date, tester and evidence file. A generic manual attachment is not enough, because devices sold under the same family name can differ by option or firmware.

Manage CT ratio settings as history, not as one number

A ratio such as 400/5 appears simple, yet four states must be distinguished:

1. The primary and secondary ratings on the physical CT or PT.
2. The primary and secondary values configured in the meter.
3. Whether the communication register already returns a primary-referred value.
4. Any multiplier applied in the gateway, historian or BI layer.

The classic failure is double scaling: the meter has already converted to the primary value, and the server multiplies it by 80 again. The opposite failure produces one-eightieth of the real value. Both traces can still follow production activity, so a visual inspection may not detect the error.

CT direction and phase association also affect sign, power factor and phase power. A site with on-site generation must explicitly define import and export. Converting every negative value to an absolute value makes a chart easier to read but destroys the evidence needed to reconcile an energy balance.

Treat the CT/PT mapping as a versioned master with effective_from and effective_to. When a circuit, CT, meter or configuration changes, preserve the raw layer and recalculate a versioned normalized layer. Record which mapping version produced each management report. This makes a past investment decision reproducible instead of silently rewriting history.

Three checks belong in acceptance:

• Match the CT/PT nameplate, single-line or wiring drawing and meter setting.
• Compare simultaneous instantaneous readings with a suitable reference under an approved safe method; set project tolerance according to instrument class and operating condition.
• Compare the difference between start and end energy counters over a defined period, not only one instantaneous value.

Seven questions that make fifteen-minute demand unambiguous

“Store fifteen-minute demand” becomes implementable only when the project answers:

1. Are blocks fixed at 00, 15, 30 and 45 minutes?
2. Is the business timezone Asia/Bangkok, and is the database stored in UTC?
3. Does a 00:15 timestamp describe the interval that just ended or the interval that just began?
4. Is average kW derived from three five-minute averages or from the change in a kWh counter?
5. What happens when one five-minute slot is missing?
6. Does a late reading reopen a closed interval, and is the previous version retained?
7. Is the displayed series fixed block, rolling prediction, meter-internal demand or utility billing demand?

A clear fixed-block definition can derive average power from energy:

interval_energy_kWh = counter_end_kWh − counter_start_kWh

interval_average_kW = interval_energy_kWh × 60 ÷ 15 = interval_energy_kWh × 4

Counter rollover, reset, meter replacement, reverse flow and missing endpoints must be handled. If the project instead averages five-minute kW values, it should store the three expected source slots and their coverage. Two samples must not be published as an unqualified three-sample result.

Keep a fixed, settled fifteen-minute series separate from a rolling forecast used for control. A rolling value is useful for anticipating a peak, but it cannot replace the fixed interval needed for later reconciliation. See the related guide to demand control and peak power for the control layer; the control forecast, control command and settled interval should remain separate records.

Missing power data is neither zero nor “last value”

Zero filling makes a running machine appear stopped. Last-value carry-forward makes a failed connection appear healthy. Interpolation can help some analyses, but it becomes unauditable when it is indistinguishable from measurement. A robust design separates raw, normalized and derived layers.

At minimum, distinguish GOOD, MISSING, INVALID, SUBSTITUTED and ESTIMATED, or an agreed equivalent. An estimate should carry the method, input interval, execution time, actor or job and confidence. If measured data arrives later, retain the transition rather than erasing it.

An RFP should not stop at “less than 1% missing.” It must define the denominator and treatment of approved shutdowns, device outages, network outages, late values and duplicates. The following are example contractual thresholds, not laws or universal values:

• Completeness is the share of expected five-minute slots during in-scope operating time that contains an accepted measured state.
• Only approved planned outages are removed from the denominator.
• A missing slot is detected within fifteen minutes and routed to the responsible role within thirty minutes.
• Gateway NTP offset is within two seconds as an example; a source-meter clock difference above sixty seconds raises a warning.
• A settled fifteen-minute block normally requires all three five-minute slots; allowed estimation is listed separately.
• Any recalculation after monthly close creates a new version and preserves the old one.

Use IEC models to make a mapping—not to avoid one

IEC 62056-6-2:2023 defines DLMS/COSEM interface classes for representing meter information. IEC TR 61850-80-5:2026 provides guidance for mapping between Modbus and IEC 61850. These references are useful when the longer-term architecture connects energy metering, electrical equipment and higher-level management systems.

Nevertheless, “DLMS compatible” or “IEC 61850 integration available” does not guarantee the required object, unit, timestamp or quality. The RFP should name the required object or signal, access method, update cycle, clock source, quality, security, licence, gateway transformation and test. A standard can reduce ambiguity in a mapping table; it cannot replace the project mapping table.

For the physical selection around protocol, wiring distance, noise and maintainability, see factory IoT sensor selection. For deciding between local logging and a connected platform, see industrial data logger versus factory IoT. This article concentrates on accepting the meaning of the resulting data.

Twelve requirements for a power meter data collection RFP

1. Scope and decision purpose

List panels, circuits, equipment and decisions. “Energy saving” is too broad. Identify capacity ranking, peak monitoring, intensity, anomaly detection and energy balance, along with the grain and settlement time each use needs.

2. Meter register

Record make, model, serial, firmware, CT/PT, circuit, location, communication, supply and owner. Unknowns become dated survey deliverables, not an indefinite “to be confirmed after award.”

3. Point map

Freeze address, type, order, sign, scale, unit, update, invalid value, rollover and access. Select project points instead of merely attaching a manufacturer manual.

4. Time

Define UTC storage, Asia/Bangkok display, NTP source, source-versus-gateway priority, allowable offset, correction and late-arrival treatment. Thailand does not use daylight saving time, but servers must still be configured explicitly.

5. CT/PT and sign

Choose one authoritative scaling location. Verify configured values, physical nameplates, wiring and positive direction. Version every change with approver and effective time.

6. Fifteen-minute aggregation

Specify fixed or moving window, boundary, timestamp meaning, inputs, formula, missing condition, recalculation and version. Separate billing comparison, control and analytical series.

7. Quality and missing data

Define quality codes, detection, notification, retry, local buffer, replay, duplicate elimination, allowed estimation and monthly close. Include the completeness formula.

8. Cybersecurity

Define network zones, allowed flows, accounts, certificates, credential handover, remote support, logs, patch ownership, backup and recovery testing. An RS-485 gateway should not create uncontrolled Internet reachability.

9. Performance and retention

Test replay after an outage, concurrent backlog, retention, query response, export, expected growth and backup time—not only the normal 30-meter polling rate.

10. FAT and SAT

Include normal values, wrong address, word order, sign, scaling, disconnect, clock drift, restart, rollover, late data and duplication. Agree expected outputs and evidence before award.

11. Handover

Require as-built wiring/network drawings, meter register, point map, dictionary, configuration backup, source or export, test results, known issues, runbook and training record.

12. Change and acceptance

Define change request, impact review, approval and regression test. Acceptance is approval of the evidence pack and closure of critical defects—not simply seeing a chart.

FAT: test communication and meaning separately

FAT closes reproducible defects before field installation. Use three representative meters or models and a register-response replay when physical devices are unavailable. Keep protocol-success tests separate from semantic tests.

Screenshots alone are weak evidence. Preserve input conditions, configuration version, raw value, expected result, measured result, pass/fail, defect and retest as machine-readable CSV or JSON plus an approval-friendly report. If mapping changes later, the team should know exactly which tests to repeat.

SAT: close field-specific defects

SAT matches panel circuit labels to meter IDs and verifies CT direction, phase association, RS-485 polarity and termination, gateway ports, IP addresses and time sync. A point map that worked in the office can fail through duplicate Unit IDs, reversed wiring, stale panel labels or an unrecorded modification.

Where safe and operationally permitted, record a known equipment stop/start or load change and confirm that it reaches the intended meter at the correct time. Check energy balance across the incomer, transformer and major branches, while allowing for losses, unmetered circuits, interval differences and meter accuracy. A perfect 100% balance is not a sensible universal requirement; the project needs an explainable tolerance.

Comparison with a billing meter or invoice is useful but does not prove identity. Period, interval boundary, scale, demand method and meter role may differ. Align comparison conditions before deciding which instrument is wrong.

A seven-day continuous run is a practical example, adjusted to the factory’s operating pattern. Conditions not seen during that week—weekend shutdown, night shift, solar export or a planned high-load process—must be recorded as untested and carried into the 90-day PoC.

A 90-day PoC that moves from “visible” to “usable for decisions”

A PoC described as “show 30 meters on a dashboard within 90 days” tends to spend the final week completing connectivity. Divide the period into acceptance gates and require the team to use the data in a real decision pack.

During the shadow period, generate the same meeting material currently created from spreadsheets and invoices. When someone questions a peak, intensity or anomaly, the owner should be able to trace it back to raw input without calling the developer. Otherwise operational handover is incomplete.

The evidence pack should contain the final register and drawings; approved point map and rules; FAT/SAT inputs and outputs; 90-day completeness, late, duplicate, clock and estimation metrics; at least three independent raw-to-interval recalculations; backup and recovery proof; accounts and training; and a controlled list of remaining defects with owner and due date.

Recalculate the 30-meter, five-minute data model

The following is a TOMAS TECH illustrative planning case, not a customer result, market benchmark, quotation or guarantee. Assume 30 meters, a five-minute record and 250 operating days per year.

samples per meter-day = 24 × 60 ÷ 5 = 288

annual meter-samples = 30 × 288 × 250 = 2,160,000

There are 96 fixed fifteen-minute blocks each day:

annual meter-intervals = 30 × 96 × 250 = 720,000

This volume is modest for a modern database. The real design burden is retaining the expected source slots, input quality, aggregation formula, CT/PT version and interval boundary for each accepted result.

Sensitivity of a one-percent scaling error

Assume, solely for illustration, 12,000,000 kWh annual energy, a 1% scale error and the ERC public average of 3.95 THB/kWh:

apparent affected value = 12,000,000 × 1% × 3.95 = 474,000 THB/year

The public average is not the factory’s marginal price; replace it with an invoice-based value. Nor is THB 474,000 automatically recoverable cash. The calculation illustrates how a small semantic error can distort a baseline, line comparison or investment ranking while the utility invoice itself remains unchanged.

Illustrative cost, annual value and payback

Assume the 30 existing meters can be reused.

Assume annual operations of THB 180,000. Value is framed around decision quality and reconciliation rather than a promised tariff saving.

net annual value = 1,416,000 − 180,000 = 1,236,000 THB

simple payback = 1,500,000 ÷ 1,236,000 = 1.21 years

At only 50% realization:

net annual value = 1,416,000 × 50% − 180,000 = 528,000 THB

simple payback = 1,500,000 ÷ 528,000 = 2.84 years

Show both cases and replace every unit assumption with plant history. If the evidence is weak, make the purpose of the PoC to test these assumptions before a full rollout. The broader factory energy monitoring system guide explains how the accepted data can support improvement; this article defines the trustworthy input first.

Score vendor proposals on more than price

Do not compare RFP responses only by equipment price and number of dashboard screens. The following scoring model is an example.

Responses such as “covered by our standard,” “Modbus will make it work” or “AI fills missing data” need follow-up. Ask what the standard contains, which device and firmware were tested, how substituted values remain identifiable, how error is evaluated and who owns the final configuration.

Common failure patterns and how to prevent them

Accepting the system when a chart first appears

A visible chart proves communication and user-interface rendering, not semantic correctness. Recalculate known values, raw conversion, fifteen-minute aggregation and missing-data cases before acceptance.

Managing CT ratios separately in Excel, the meter and the server

No one knows which location is authoritative, and a change can be applied twice. Establish one mapping master with effective time and generate downstream configuration from it.

Leaving the timestamp label for fifteen-minute values undefined

If 00:15 may mean either interval start or interval end, equipment events can shift by fifteen minutes. Put mathematical interval notation and three worked example blocks in the specification.

Filling missing values with zero to close a monthly report

Consumption appears lower and an improvement may be falsely credited. Keep measurement and estimation separate and report missing data as a monthly quality KPI.

Treating the submeter as identical to the billing meter

Accuracy, interval, scaling, reading cycle and purpose can differ. Comparison is valuable, but tolerance and cause classification must be designed in advance.

Connecting all thirty meters before checking data quality

One configuration mistake is copied thirty times. Use staged gates: FAT on three representative meters, aggregation proof on ten and site acceptance on all thirty.

Frequently asked questions

Can all Modbus power meters be collected in the same way?

The transport is similar, but addresses, types, order, sign, scale, units, update and invalid states differ by device and firmware. Build a device-specific point map and prove it with known values. “Supports Modbus” is not an interoperability acceptance criterion.

Should the CT ratio be configured in the meter or the server?

Either architecture can work. The critical control is one authoritative scaling location. If the meter returns primary-referred values, do not multiply again. If it returns secondary values, apply a versioned server mapping. Reconcile nameplate, wiring, meter configuration, raw and normalized values.

Can three five-minute values simply be averaged into fifteen-minute demand?

That is valid only when all three are equivalent average-kW values in the correct fixed block. A counter-difference calculation, meter-internal demand and rolling demand may differ. Specify timestamp meaning, missing data, late arrivals and the utility comparison method.

How much missing power data may be estimated?

Conditional estimation can support operations, but it must never silently replace measurement. Define allowed duration and method and mark the result ESTIMATED or SUBSTITUTED. Billing comparison, warranty decisions and investment baselines may need to exclude estimates.

What is the most important acceptance requirement?

Require an independent reviewer to recalculate any selected fifteen-minute value back to raw input, including CT/PT, sign, time, missing-data treatment and calculation version. Availability and screen response do not prove semantic correctness.

Should all 30 meters be connected on day one of the PoC?

No. Freeze register and failure behavior on three representative meters, prove aggregation on ten, then expand to thirty. This prevents one mapping defect from being copied across the estate.

Can the 3.95 THB/kWh average be used directly in the investment case?

It is public context, not necessarily the plant’s applicable marginal value. Use recent invoices and the latest MEA or PEA tariff applicable to the customer. Treat the example above only as sensitivity analysis.

Conclusion: accept decision-grade data, not connected meters

Power meter data collection succeeds when CT/PT is applied once, import and export signs are consistent, fixed Bangkok-time intervals are reproducible, missing and estimated values remain distinguishable, and every management result can be traced through versioned transformations to raw input.

Putting the data dictionary, point map, clock and interval rules, quality codes, FAT/SAT, 90-day PoC and evidence pack into the RFP lets a buyer compare vendors on their ability to deliver correct data—not just hardware and screens. It also manages the risk of ranking factory investments with a convincing but incorrect chart.

TOMAS TECH can support factories in Thailand and ASEAN with installed-meter surveys, Modbus point maps, CT/PT, sign, time, interval and missing-data rules, and the RFP, FAT/SAT and 90-day acceptance plan. Discussion can begin before meter selection or when a team first questions whether existing energy data is reliable. Contact us.

References

• Thailand ERC: Ft and average electricity price for September–December 2026
• MEA: Electricity tariffs and maximum demand
• PEA: Electricity Tariffs May 2023
• Modbus Organization: MODBUS Application Protocol Specification V1.1b3
• Schneider Electric: Demand calculation methods, EM3570 series
• IEC 62056-6-2:2023: DLMS/COSEM interface classes
• IEC TR 61850-80-5:2026: Modbus / IEC 61850 mapping guidance
• Thailand ERC: Automatic Ft statistics