Product
Available todayEvery five-minute interval measured against a model of your own plant. The gap between what it should have made and what it made, separated into causes you can act on.
Conceptual illustrations — not product screenshots.
Who it's for
The same platform answers all three, because they are asking about the same gap from different sides.
Defend the return on a capital asset over a twenty-five year life.
Monthly reporting says output was below budget. It does not say how much of that was weather, how much was soiling, and how much was equipment you could have fixed.
Underperformance separated into causes, each quantified in energy — so the recoverable portion is distinguishable from the portion that was never yours to recover.
Keep a distributed fleet running with a finite crew and a finite number of site visits.
The work queue is assembled from inverter alarms, ticket age, and whoever called last. Gradual degradation never raises an alarm at all.
A queue ordered by modeled consequence, with the evidence attached, so a truck roll is justified before it is dispatched.
Explain performance to lenders, investors, and O&M contractors who will all dispute it.
Numbers are assembled by hand from vendor exports, which makes them slow to produce and difficult to defend line by line.
Every figure traceable from the headline back through the model to the raw interval data it came from.
Modules
They ship together because none of them works properly alone — attribution without a calibrated baseline is guesswork, and a baseline nobody acts on is a chart.
Expected generation computed for every interval from the plant's own configuration and the conditions measured at the site, then compared against what was actually produced.
The gap between expected and actual is decomposed into named causes rather than reported as a single shortfall.
Anomalous events are classified against patterns seen across the fleet and across history, rather than flagged as generic deviations.
Generation forecasts calibrated per site, with systematic bias corrected as the platform learns how each plant responds to conditions.
Heterogeneous inverter brands, module vintages, and site designs are mapped into one schema so plants can be compared on equal terms.
Operational, investor, and lender reporting generated from the same models that drive the alerts, so the two never disagree.
The modeling
A PAN file describes how a specific module behaves across irradiance and temperature. An OND file does the same for an inverter. Together with the string topology and the weather measured at your site, they produce an expected output for every component, every interval.
That model runs before any machine learning sees the data. The reason is not purity — it is that a model trained directly on your operating history learns your existing faults as normal behaviour, and then defends them.
What is left after the physics is accounted for is the residual. That is where learning is applied, and it is the only part of the signal where learning can tell you something you did not already know.
Specification and conditions produce an expectation. Expectation minus measurement is the residual. The residual is what gets attributed.
Loss attribution
Each is invisible at the plant meter and each is missed by onboard alarms for a different reason. They are detected separately and quantified separately.
Dust, pollen, salt, and industrial deposition accumulate on module glass and reduce transmitted irradiance.
It is gradual and fleet-wide, so it never trips a threshold. It is usually only recognised after a cleaning cycle shows how much was being lost.
Estimated continuously as the slow component of the residual between modeled and measured output, separated from fast effects and from weather.
One string in a combiner underperforms — shading, a failed connector, a damaged module, a degraded diode — and drags the group with it.
At the plant meter the loss is a rounding error. At the string it is total. Nothing between the two is looking.
String-level current and voltage compared against sibling strings on the same inverter under identical conditions, so the outlier is identified rather than averaged away.
MPPT tracking or conversion efficiency slips out of specification slowly, without ever entering a fault state.
Onboard alarms are tuned to catch failure, not decay. An inverter running two percent below spec reports itself as healthy.
Conversion efficiency modeled per unit and tracked over time, so a unit diverging from its own history and its siblings surfaces before it fails.
Breakers trip, combiners fail, trackers stall, connections degrade — sometimes with no fault code raised anywhere.
A silent failure produces zero output from a component nobody is watching, and the loss compounds until someone visits the site.
Expected output is computed for every component independently, so a component producing nothing is detected on the interval it stops, with or without a fault code.
inverters · one isolated
Conceptual illustration — not a product screenshot.
Getting started
No new instrumentation. We work with the telemetry and documentation you already hold.
We establish read access to your telemetry — vendor API, historian, SCADA export, or file drop. Whatever exists is what we work with; no new instrumentation is required.
From you: A technical contact and read credentials. NDA and data-handling terms are signed before this step.
Plant configuration is loaded: PAN and OND files, string maps, inverter and combiner topology, tracker geometry, and the weather-station layout.
From you: The as-built documentation you already hold. Gaps are normal and we work around them.
The physics model is built per plant and calibrated against your historical generation, so expected output reflects this asset rather than a generic one.
From you: Nothing. This is our work.
We run the platform against a historical period and compare its findings to your own records and your field team's knowledge. Anything it cannot justify gets fixed before go-live.
From you: A few hours from someone who knows the sites well enough to challenge the output. This step is the whole point.
Missing items are normal. Sparse or partial data extends validation rather than blocking it — we will tell you honestly what the platform can and cannot conclude from what you have.
Versus inverter monitoring
The difference is whether anything in the system knows what the plant was supposed to be doing.
| Dimension | Inverter monitoring | Peryx |
|---|---|---|
| What it tells you | What happened | Why it happened, and what it is costing |
| Baseline | Nameplate ratings or a rolling average of recent behaviour | A physics model of this specific asset under the conditions actually observed |
| Data resolution | Averaged on ingestion to keep storage and dashboards cheap | Retained at the frequency the asset produced it |
| Loss attribution | One aggregate shortfall figure | Separated into named causes and quantified independently |
| Detection | Threshold alarms, which gradual degradation stays underneath | Deviation from expectation, which catches decay while it is still slow |
| Mixed fleets | One portal per vendor, none of which agree on 'normal' | Brands, vintages, and sites normalised into one operating picture |
| Output | A dashboard someone has to remember to open | Ranked findings with evidence, delivered into your existing tools |
What it tells you
Inverter monitoring
What happened
Peryx
Why it happened, and what it is costing
Baseline
Inverter monitoring
Nameplate ratings or a rolling average of recent behaviour
Peryx
A physics model of this specific asset under the conditions actually observed
Data resolution
Inverter monitoring
Averaged on ingestion to keep storage and dashboards cheap
Peryx
Retained at the frequency the asset produced it
Loss attribution
Inverter monitoring
One aggregate shortfall figure
Peryx
Separated into named causes and quantified independently
Detection
Inverter monitoring
Threshold alarms, which gradual degradation stays underneath
Peryx
Deviation from expectation, which catches decay while it is still slow
Mixed fleets
Inverter monitoring
One portal per vendor, none of which agree on 'normal'
Peryx
Brands, vintages, and sites normalised into one operating picture
Output
Inverter monitoring
A dashboard someone has to remember to open
Peryx
Ranked findings with evidence, delivered into your existing tools
Mixed vendors, sparse instrumentation, an anomaly nobody can explain. The first call is with someone who can discuss the detail.