AI for Power Quality

What AI Does in Power Quality Today

Most AI in power quality products does three things:

• Flags outliers. Marks unusual values in trended RMS and harmonic metrics.
• Infers causes. Proposes a likely physical cause from those summaries, for an engineer to investigate.
• Ranks captures. Scores triggered waveforms for severity, so the worst appear first.

Each saves an engineer time, and each sees only what the meter chose to keep. When the cycles that explain an event fall outside those captures, or sit in another meter’s record, no model can recover them.

Where AI on Summaries Runs Out

It Cannot See What Was Never Recorded

Traditional power quality data is summaries, typically the 3-second and 10-minute aggregates defined by IEC 61000-4-30, plus waveforms only around a trigger. A model working from that data inherits its gaps. A sag just above the threshold, the cycles that led up to an event, and a disturbance no trigger watches for leave nothing to analyze. The continuous point-on-wave primer covers why those gaps decide investigations.

The record also lacks a baseline. With only the violations kept, there is almost no normal waveform to compare an event against, and nothing of the weeks before it. Many failures show first as sub-threshold change: distortion creeping up as a capacitor ages, or a sag that deepens a little with each motor start. A trigger is binary, so a problem that grew for months enters the record on the day it crossed the threshold, looking sudden. The investigation then hunts for a sudden cause that never existed.

It Cannot Relate Events Across Devices

A root cause often lies somewhere other than where the symptom appeared: a motor drive on another feeder, a capacitor bank switching upstream, a generator responding to a load step. Relating them takes records from every point, on one timebase, covering the same cycles. Triggered windows on different meters rarely line up, so the evidence that links cause to effect is missing exactly where it is needed.

It Finds What It Was Taught

A model trained on past events recognizes what it has seen. A GPU cluster swinging its power in step with a training job is a new kind of load, and it produces signatures that were not in the training set. A model with no physics grounding either forces them into a familiar class or reports an anomaly it cannot explain.

It Runs in the Cloud

Most AI power quality analysis runs in a vendor’s cloud, which brings risks of its own. The facility’s operating data sits outside its trust boundary, where its protection depends on the vendor’s security rather than the facility’s own. The answers wait on an upload, so they arrive after the fact. When the power fails, the site’s routers and uplink can go down with everything else, and the analysis goes dark during the very event it exists to explain.

A cloud-only service also caps what the AI can read. A seven-channel sensor recording continuously at 32 ksps produces roughly 60 GB of raw samples a day. Where the site’s network and policy allow, the record can stream offsite, but for sensitive processes, defense sites and much critical infrastructure, it must remain on site. Analysis that needs the full record must run where the record is allowed to be.

What the Gaps Cost

While a cause stays unknown, the problem recurs, and each recurrence is lost production: a tripped line, a scrapped batch, a process restarted from the beginning. Equipment keeps absorbing unreported sub-threshold stress, and wears out early. When the cause finally must be found, the usual answer is a post-event investigation: a power quality specialist installs portable analyzers, the site waits weeks for the problem to come back, and incurs losses all the while. A continuous record turns that campaign into a query.

What AI Needs to Find a Root Cause

• A gap-free, time-synchronized waveform record from every monitored point, so that every cycle is available and events at different points can be put in order
• Physics as the authority: deterministic signal processing the AI reasons over rather than reinvents, and conclusions checked against how the power system can actually behave
• Cited evidence, so that every answer points to the cycles, channels and time range an engineer can verify
• Operational context: equipment state, topology and maintenance history, on the same timeline as the measurement
• Analysis where the data lives, so that the full record is in reach and the answer arrives while it can still be acted on

How Energy Quotient Applies AI

EQ Wave™ is a Waveform Measurement Unit (WMU) that records Continuous Point-on-Wave (CPOW) data at 32 ksps on every voltage and current channel, gap-free. EQ Coherence™ keeps that record and computes RMS, harmonics, sequence components and standards metrics on every cycle as it arrives. EQ Syntropy™ reasons across that output from every device and answers in plain language.

At 24-bit resolution and a fixed sample rate, one sensor’s record also carries the signatures of equipment beyond its point of connection: the interharmonics of a motor drive on another feeder, capacitor switching upstream, a resonance building in the network. EQ Syntropy analyzes those signatures at full resolution and relates them to events at other sensors and to operating data, so it can reason about equipment no sensor is attached to.

An engineer can ask why a drive tripped at 2:14 a.m., and get the answer in EQ Sight™ backed by the cycles that show it.