Analytics validation

Evidence should be published with its sample, version, and limitations.

ArraySignal tests deterministic expected-production, fault, data-quality, and loss-analysis behavior. This page reports only evidence supported by public repository artifacts; unavailable field metrics stay explicitly unavailable.

Methodology documented

Synthetic fault validation

Sample count: 10 deterministic simulated sites

The public reference portfolio exercises healthy operation, outages, persistent underperformance, possible soiling, communications gaps, and normal clipping through deterministic fixtures.

• Synthetic cases verify rule behavior and UI contracts; they do not establish field accuracy.
• The cases are simulated and are not customer sites.

Validation in progress

Real-weather and public-data validation

Sample count: 3 public dataset adapters

La Réunion, TECNALIA, and NREL PVDAQ adapters exercise real timestamps, telemetry quality, weather/model inputs, and analysis limitations without exposing customer data.

• The public adapters do not provide a complete field-confirmed answer key for every ArraySignal diagnosis.
• Cases with missing model inputs are withheld from unsupported expected-production or fault claims.

Not yet validated

Real-world pilot validation

Sample count: 0 published labeled pilot cases

No customer-identifiable pilot validation results or independently reviewed aggregate accuracy artifact is currently approved for public publication.

• Precision, recall, false-alert rate, and diagnosis calibration remain unpublished until sufficient independently labeled outcomes exist.

Public metric register

Unsupported accuracy numbers are not filled with estimates.

A metric is published only with an independently supportable value, sample count, and context. The current public release does not yet meet that bar for the metrics below.

Metric	Published result	Sample count	Required context
Precision	Not yet validated	Not published	Requires independently labeled incident outcomes.
Recall	Not yet validated	Not published	Requires independently labeled incident outcomes, including missed conditions.
F1 score	Not yet validated	Not published	Requires published precision and recall on the same labeled sample.
False alerts	Not yet validated	Not published	Requires a defined field observation period and confirmed outcomes.
Detection latency	Not yet validated	Not published	Requires externally established event start times.
Expected-model error	Not yet validated	Not published	Requires a published healthy validation window and benchmark protocol.
Lost-energy error	Not yet validated	Not published	Requires an independent lost-energy benchmark.

Versioned analysis

The rule and model version travels with the result.

Version identifiers below reflect the current production analysis contracts. A validation result from one version is not silently represented as evidence for another.

Site analysis: arraysignal-site-analysis-v1
Expected production: pvwatch-pvwatts-v1
Fault detection: pvwatch-detectors-v1
Loss estimation: pvwatch-loss-engine-v1

Explore your portfolio

Bring a real operating question to the conversation.

We’ll discuss telemetry coverage, portfolio priorities, and a practical evaluation scope—without presenting sample estimates as guaranteed outcomes.

Book a portfolio review