Non-Invasive Monitoring for Grid Assets: From Sensor Data to Operational Decisions
Quick Overview
Problem: Condition data from non-invasive sensors arrives reliably and changes no operational decision.
Common failure points: No named decision owner for the data, clamp-on measurements treated as metered-grade, accuracy loss at low current and temperature extremes, multi-supplier schema fragmentation, OMS/DMS integration and OT security classification scoped after procurement.
Where it appears: Distribution feeder monitoring programs, dynamic line rating deployments driven by FERC Order 881 compliance and Order 1920 planning obligations, substation thermal and partial discharge monitoring, underground vault joint monitoring.
Engineering focus: Measurement-chain validation across the real operating envelope, edge processing and data normalization, the OT integration path, security classification, acceptance testing against the decision rather than against sensor uptime.
Production Failure Scenario
The deployment went in cleanly. Clamp-on sensors across a set of distribution feeders, no circuit interruption, installation finished ahead of schedule.
Within months every unit was reporting current, voltage, conductor temperature, and a fault-current snapshot on every protective operation. Uptime was above target. The dashboard was current.
And the average time to locate a fault had not moved.
The sensors were not the problem. The data landed in a supplier portal that distribution operators did not have open during an outage, in a schema the outage management system could not consume, and nobody had written down which decision the numbers were supposed to change.
The programme had bought telemetry. It had not yet engineered the path from a measurement on the asset to a decision in the control room.
Wrong Assumption
The assumption is reasonable on its face: if sensors install without interrupting the circuit and the data arrives reliably, asset visibility improves. Grid deployments break it in two places. First, non-invasive does not automatically mean outage-free or metered-grade — access requirements depend on the device, the mounting point and the voltage class, and a field-installed clamp-on sensor carries an error budget that a calibrated instrument transformer does not. Second, a measurement only becomes visibility when it reaches the system where the decision is made, in a form that system can act on. Until the measurement chain is validated across the real operating envelope and the data path into OMS, DMS or work management is designed rather than assumed, the result is a reliable stream of numbers that no operational process consumes.
Why It Fails
"Non-invasive" names a constraint, not a technology. Four measurement classes satisfy it and they behave differently. Contact but non-interrupting devices — split-core CTs, HFCTs on earth straps and cable screen bonds, surface RTDs and thermal patches — touch the asset without breaking the circuit. Proximity and field-based devices — Rogowski coils, tower-mounted LiDAR, thermal cameras at a standoff — avoid contact entirely. Remote and airborne methods cover corridors periodically rather than continuously. Distributed fiber optic sensing turns installed cable into the sensor. Writing "non-invasive" into a specification without naming the class is where the first mismatch happens.
The distribution feeder gap is a placement problem, not a coverage problem. The feeder head is measured at the substation bus and the customer end by AMI. Where few measurement points exist between them, fault location becomes a patrol problem, load distribution along the feeder stays invisible, and power quality events go unrecorded. Adding points at selected sectionalizing locations compresses the fault search space from the whole circuit to a segment — but only if the deployed device has the measurement channels, sample rate and event-capture behaviour for the phenomenon in question. A load-profile sensor is not a power quality instrument, and substituting one for the other is a common and expensive error.
The accuracy figure applies to the chain, not the sensor. A split-core CT with imperfect closure, on the wrong phase, or with inconsistent contact produces a plausible number with an uncharacterised error. A sensor that harvests energy from line current has a minimum current below which it does not energize or does not measure accurately — which is exactly the condition on a lightly loaded feeder section at night. Temperature cycling across a year drives calibration drift. The chain is most trustworthy at the operating point already understood and least trustworthy at the extremes that justified the programme.
The data path into operations is scoped last. Backhaul cost, storage, normalization across supplier formats, OMS and DMS integration, and the OT security classification are design-phase decisions that are routinely treated as integration-phase details. Promwad’s work on power metering and power quality systems and on IEC 61850 and smart grid protocol integration sits in exactly this layer, between the measurement and the system that has to consume it.
In deployment these rarely arrive one at a time. An accuracy question at light load undermines confidence in the fault-location use case; inconsistent schemas across two suppliers mean no one trusts the combined view; and an unresolved security classification stops the integration that would have made the data operational. Each one alone is manageable. Together they produce a programme that reports sensor uptime because it has no other metric to report.
Hidden System Complexity
asset condition → sensor physics and mounting → measurement chain and calibration → edge processing → backhaul → normalization and schema → platform and alerting → OT integration (OMS / DMS / work management) → operator decision
A symptom seen at the operator level usually originates several layers up. "The sensor data disagrees with the meter" is read as a sensor defect, when the cause is a commissioning step that never validated placement and closure. "The alerts are noise" is read as a threshold problem, when the cause is that two suppliers define an event differently and the platform treats both as the same field.
Distributed sensing adds a layer that is easy to skip. Fiber in OPGW or ADSS measures conditions along its own route; translating that into power conductor temperature depends on the cable configuration and a thermal model. Tower-mounted non-contact systems track conductor position with LiDAR, and in the AES deployment the system derives average conductor temperature from measured sag. In both cases a modelled quantity enters the chain, and the model is part of the measurement whether or not the specification acknowledges it. What the platform can show is bounded by what the firmware at the grid edge measures and timestamps, which is the argument made in detail in embedded control and telemetry at the grid edge.
Failure Patterns
Scenario 1. Feeder sensors report reliably for two quarters with uptime above target. Average time to locate a fault is unchanged, because fault-current snapshots sit in a supplier portal with no mapping into the OMS fault model and no role accountable for acting on them during an event. The programme metric defaults to sensor uptime.
Scenario 2. Sensor readings and metered readings disagree on a lightly loaded feeder section. The sensor is replaced; the disagreement persists. The cause is the accuracy curve against load on a self-powered device, which was never requested at specification and never validated at commissioning — so neither reading can be ruled out.
Scenario 3. Three suppliers across three procurement cycles feed one operational view. Timestamp conventions, units, quality flags and event definitions differ. Operators stop trusting the combined dashboard and revert to the systems they already used, and a planned integration into the DMS stalls when the OT security classification turns out never to have been resolved.
Grid Monitoring Engineering
Non-invasive monitoring failures — uncharacterised accuracy at the operating points that matter, schema fragmentation across suppliers, data that never reaches the system where decisions are made — are structural, not sensor-selection problems. Closing them takes measurement-chain specification, edge and platform engineering, and an integration path designed alongside the hardware. Promwad develops metering and power-quality electronics, grid-edge firmware, substation communication hardware and the platform layer around them, including custom sensor development, industrial network design, and edge AI where anomaly detection has to run close to the asset.
Engineering Experience Across Metering, Grid-Edge and Substation Platforms
Where a Measurement Accuracy Claim Actually Comes From
A European client needed a power quality analyser specified to IEC 61000-4-30 Class A — ±0.1 percent voltage measurement accuracy, ±1 percent current, ±10 mHz frequency, RMS calculation in half-cycle steps, and compliance analysis against IEC 61000-4-7 and IEC 61000-4-15.
Those figures are not a property of the analogue front end. They are a property of the whole chain, and the chain has to be partitioned so that the measurement loop is never competing with device management, logging and connectivity for the same execution budget.
Promwad built the device on an asymmetric multiprocessing split using OpenAMP. The RTOS side owns power quality measurement with deterministic timing, with the mathematics on CMSIS-DSP against the IEC series. The Linux side owns device management, logging, connectivity and the Qt user interface. Two hardware versions were delivered, with the ADC variant running the full measurement chain through a DSP coprocessor. The build system is Yocto.
The reason this matters for grid monitoring is that it is the same question a utility should ask of a field sensor, in a different package: does the stated accuracy hold at the operating points that justified the deployment, or only at rated conditions on a bench? A supplier who can answer that has characterised the chain. A supplier who quotes a single figure has characterised the front end.
Solution Approach
Step 1: Name the decision and the system of record before the sensor. "Improve feeder visibility" is not a decision. "Reduce the patrol segment after a lockout from the full feeder to one identified section, acted on by the distribution operator in the OMS" is. Write down the decision, the role that owns it, and the system in which it is made. Sensor selection follows from that sentence, and a programme that cannot produce it is not ready to procure.
Step 2: Specify the measurement chain across the real operating envelope. Required accuracy at minimum load, not at rated current. Installation validation procedure and periodic accuracy review interval. Behaviour at temperature extremes. What the device does when backhaul is lost and for how long it buffers. Where a modelled quantity enters the chain — conductor temperature from sag, conductor conditions from fiber — state the model and its validity range as part of the specification.
Step 3: Design the data plane and the OT security classification together. Cellular backhaul carries a recurring per-device cost that compounds across a fleet, alongside storage, normalization across supplier formats, and integration with OMS, DMS and work management. Size that explicitly rather than as an adjunct to the hardware line. Resolve the OT data path, network segmentation and the BES Cyber System classification question in the same exercise, because they constrain each other — and because secure boot, signed firmware update and key storage are silicon-selection decisions, which is why embedded security engineering belongs in the architecture phase.
The acceptance test is where a programme discovers whether it bought visibility or telemetry. For a fault location case it is concrete: replay a recorded event, or run a coordinated secondary injection test on the protection path, and confirm that the operator sees the candidate section, the event time and a data quality indication inside the OMS — then measure time to localization against the pre-deployment baseline. A test that only confirms the sensor is reporting confirms the half of the system that was never in doubt.
Real Trade-Offs
Conductor-mounted DLR sensing measures temperature, sag, tension and current directly and draws power from the line, which is easier to defend to a rating methodology review. It also requires access to the energized conductor and gives a point measurement, so siting against the thermally limiting span matters. Tower-mounted non-contact sensing removes the access requirement and changes deployment economics, at the cost of a conductor temperature derived from geometry rather than measured.
Distributed fiber gives continuous coverage along its route and removes the sensor-siting question entirely, but only where suitable fiber already exists on the corridor, and conductor conditions still arrive through a model. Point sensors go anywhere and leave the spans between them uninstrumented.
Denser deployment improves fault localization, and every additional point adds recurring backhaul cost, storage, normalization effort and attack surface. The density that makes the fault-location case work is not automatically the density the data plane budget supports.
Continuous thermal monitoring catches gradual degradation that an annual infrared survey samples once a year and usually misses, at the cost of fixed installation, ongoing alert triage, and a false-positive rate that has to be managed. Related to that, automated thresholds scale and generate alarm fatigue on noisy signals — partial discharge above all — while expert interpretation is accurate and does not scale. Most working programmes use thresholds for triage and route PD-band findings to a specialist.
Processing at the sensor cuts bandwidth cost and latency but fixes the analytics in firmware, which is the trade examined in embedded AI in power-grid monitoring. Raw backhaul keeps analytics flexible and grows the data plane bill every month the fleet stays deployed.
Typical Grid Monitoring Engineering Tasks
Grid-Edge Firmware and Sensor Electronics
Metering front ends, power quality measurement, energy-harvesting and low-power nodes, deterministic measurement partitioning on AMP and RTOS architectures.
Protocol and Substation Communication Integration
IEC 61850, DLMS/COSEM, PLC and cellular telemetry paths, and teleprotection interfaces where timing and framing errors are not tolerable.
Measurement Chain Specification and Validation
Defining required accuracy across the real operating envelope including minimum load, commissioning validation procedure, drift review intervals, and the validity range of any modelled quantity in the chain.
Data Plane and OT Integration Engineering
Gateway, schema normalization across suppliers, alerting rules, dashboards built around operator workflow, and the integration path into OMS, DMS and work management with its security classification resolved.
Qualifying Symptoms
- Sensor uptime is reported as the programme metric, because no operational metric has moved since deployment.
- Sensor readings and metered readings disagree and no one can establish which is right.
- Measurement quality is good at normal load and visibly degrades at light load or temperature extremes.
- Data from two suppliers cannot be shown on the same chart without manual reconciliation.
- Monitoring data lives in a portal that operators do not open during an event.
- A DLR business case is built on annual average uplift rather than uplift during the hours when the constraint binds.
- An integration into DMS or OMS has stalled on a security classification question that was never raised at architecture.
From Measurement to Operational Integration
At this point the work is measurement-chain and integration analysis, not another sensor evaluation. In practice: accuracy specified across the full operating envelope, a data plane sized before procurement, an OT path and security classification resolved at design time, and an acceptance test defined against the operational decision.
For transformers specifically the parameter set is larger and the consequence of failure is higher, and the diagnostic value of each parameter and how multiple dimensions are synthesised into a health index is covered separately in advanced transformer health monitoring beyond periodic inspection. Where the constraint is new wiring and planned outages rather than analytics, the sensing classes that avoid both are the subject of this article; where it is the firmware layer underneath, custom sensor development and industrial network design are usually where the programme actually unblocks.
Promwad’s Power Monitoring and Grid Engineering Experience
This class of problem shows up most in distribution feeder monitoring programmes, DLR deployments moving from pilot to fleet scale, and substation condition monitoring projects where the sensing was procured before the data path and the operational process were designed. Promwad’s work in this domain spans metering front ends, power quality measurement, substation teleprotection interfaces, and the gateway and platform layer that has to carry the result into operations.
FAQ
What does non-invasive monitoring mean for grid assets, and how is it different from traditional instrumentation?
Does FERC Order 1920 require utilities to deploy dynamic line rating?
How much additional capacity does dynamic line rating actually unlock?
Can thermal monitoring replace partial discharge monitoring in a substation?
What usually causes a non-invasive monitoring programme to stall after deployment?
Related Engineering Cases
- Power Quality Analyser Design: IEC 61000-4-30 Class A measurement on an OpenAMP Linux + RTOS split, with the measurement chain on the deterministic side.
- Three-Phase Power Meter Sensor Development: STM32L4 with STPM33/STPM34 metering front end, ST7580 power line communication, LoRa collection and NB-IoT transfer.
- IEEE C37.94 Multiplexer for IEC 61850 Teleprotection: FPGA, PCB and mechanical design for a substation teleprotection interface converter on Xilinx Spartan.
- ThingsBoard Server-Room Power Supply Monitoring: MQTT sensor gateway, threshold-based alerting and a real-time dashboard for server-room power supply parameters.