GridWatch — Industrial Energy Monitor
A custom three-phase metering node — our own PCB, our own firmware — streaming per-machine power data to a live dashboard.
±1%
Metering accuracy
64
Nodes on test bench
4-layer
Board
The build
Most factories get one electricity bill and no idea which machine caused it. GridWatch is our end-to-end answer: a four-layer board we designed around an ESP32 with current-transformer inputs, the metering firmware that runs on it, and the dashboard it reports to. It exists to show that we can take an industrial sensing problem from schematic to cloud without handing part of it to someone else.
This project was delivered under our Hardware & Embedded Projects service. From breadboard prototype to production-ready PCB.
- Scope
- Schematic, PCB, firmware, cloud dashboard
- Duration
- 9 weeks including fabrication lead time
- Stack
- ESP32 · KiCad · C / ESP-IDF · MQTT over TLS · Node.js · TimescaleDB
What made this hard
Sub-metering a plant sounds simple until you price it. Commercial three-phase meters run into the hundreds per point, which makes per-machine visibility uneconomic exactly where it would pay off most. The engineering problem is getting trustworthy measurements from cheap parts: current transformers are non-linear at the low end, mains-adjacent electronics need real isolation, and a reading nobody trusts is worse than no reading at all.
How we built it
- 1
Analogue front-end first
We designed the CT input stage around burden resistors sized for the expected current range, with anti-alias filtering ahead of the ADC. The isolation barrier between mains sensing and the logic side was drawn before anything else went on the board.
- 2
Four layers, not two
A dedicated ground plane and a separated analogue pour cost slightly more per board and removed most of the noise problems we would otherwise have chased in firmware. Cheaper to fix in copper than in code.
- 3
Sample properly, aggregate at the edge
Firmware samples all three phases continuously and computes true RMS, real power and power factor on-device. Only one-second aggregates leave the node. Streaming raw samples would have needed bandwidth nobody has on a factory floor.
- 4
Calibrate against a reference
Each board runs a two-point calibration against a reference meter, and the coefficients live in non-volatile storage. This is the step that turns a rough estimate into a number an engineer will act on.
How the pieces fit
CT clamps → isolated analogue front-end → ESP32 with continuous ADC sampling → per-second RMS/power/PF aggregates → MQTT over TLS → ingest service → TimescaleDB → React dashboard with per-machine breakdown and peak-demand view. Nodes buffer locally through a network outage and backfill on reconnect, so a dropped link costs no data.
Decisions we had to defend
Clamp-on CTs over inline shunts
Inline measurement is marginally more accurate and requires taking a machine offline to install. Clamp-on CTs fit around an existing conductor with no downtime, which is the difference between a retrofit that happens and one that gets postponed forever.
Wi-Fi first, LoRa as an option
Most plants already have coverage, and Wi-Fi keeps the bill of materials down. The firmware abstracts the transport so a LoRa module drops in for sites where it does not reach — a design allowance, not a rebuild.
Aggregate on-device
Sixty-four nodes streaming raw samples is a bandwidth problem and a storage problem. Computing on the node means one small message per second per device and a backhaul that scales linearly.
Where it ended up
The nodes hold within ±1% of a reference meter across the tested load range, and we have run 64 of them concurrently on the bench reporting once a second without loss. The full package — gerbers, firmware source, calibration procedure and dashboard — is what we hand over on a build like this, so the client can have more boards made without us.