SolarLoss: Turning Solar Output Anomalies into Reviewable Evidence
SolarLoss helps solar operations and maintenance teams investigate when a solar installation produces less power than expected. It compares measured inverter output with a weather-adjusted estimate of healthy output, then presents the difference alongside data coverage, uncertainty, and evidence limitations. This gives technicians and researchers a clearer starting point for deciding what to inspect next, instead of relying on a power graph alone.
When a solar PV system produces less power than expected, the first question is not always “Which component failed?” It is often “What does the available data actually tell us, and what should we check next?”
SolarLoss is a product built to help answer that question.
It compares measured output with a modelled expectation, shows the evidence and its gaps, and gives an operator a place to review a documented event.
The goal is to make an investigation easier to follow without presenting a statistical signal as a confirmed diagnosis.
The problem: an output gap needs context
An inverter’s lower-than-expected output can have several explanations. Shading is one possibility, but weather, curtailment, equipment state, sensor quality, or a change in the comparison stream can also affect the measurements.
A useful review therefore needs more than an alert: it needs the measured signals, the assumptions behind the expected value, data coverage, and a clear account of uncertainty.
SolarLoss explores this workflow using a published rooftop PV dataset from La Réunion. The study includes two inverter lines, irradiance and temperature measurements, and documented shading interventions.
One line is the target; the other is a comparison stream. Because the lines are not identical or independent, the comparison line is not ground truth for what the target would have produced without shading.
What the retrospective study found
The analysis trained a weather-assisted linear model on earlier healthy observations. It used comparison-line power, global and diffuse tilted irradiance, and ambient temperature to estimate expected target-line power. The intervention labels were held out from prediction and used to evaluate the retrospective results.
The selected detector flagged 23 of 23 documented shading episodes at this installation. That is a bounded result from one site and a retrospective experiment; it does not establish general accuracy or performance on future events. Simple baselines also flagged all 23 episodes.
The detector also produced 153 healthy alert runs over 1,794.5 supported healthy hours, or 0.0853 runs per hour. The study’s pre-set limit was 0.05 per hour, so the detector received a LIMIT research verdict. In practical terms, the event results support further investigation, while the healthy-period alerts are too frequent to call the detector ready for operational use.
For the example event
shading_01 , the signed expected-minus-measured generation deviation was about 0.397 kWh, with valid measurements covering 98.28% of the labelled interval. This is a comparison with a modelled expected stream. It is not verified causal shading loss, recoverable electricity, or a savings estimate. The voltage and current signals add operating context, but they do not identify a fault by themselves.From research output to an investigation view
SolarLoss has three separate parts. The research pipeline downloads and checksum-checks the pinned source data, aligns observations, fits and evaluates the model offline, and writes versioned results. A separate review service validates a review artifact and serves its supplied values. The React application presents those values, timelines, evidence, and report details to a reviewer.
This boundary matters: the browser and API do not rerun the experiment or create a scientific verdict. The research preview is a development-only path for committed study results. A local review integration can serve the specifically approved
shading_01 artifact, while AWS deployment remains postponed and no live deployment is claimed here.The architecture keeps the scientific workflow inspectable and lets the review application remain small. It also makes the handoff explicit: a change to scientific metric meanings needs a reviewed artifact contract before the serving layer can present it as an approved result.
What We learned
The most useful result was not a single detection percentage. It was seeing how event coverage, healthy false alarms, model error, and missing measurements change what can responsibly be said about a case. A detector can flag every labelled event in a retrospective dataset and still need more work before it belongs in an operations workflow.
The product follows the same principle. It leads with measurements and source context, keeps the expected series visibly modelled, exposes uncertainty and gaps, and presents technician checks as proposed investigation steps. It does not automatically diagnose a cause or prescribe a repair.
Current status and next work
SolarLoss is a research prototype with a local evidence-review workflow. The selected detector remains limited by the healthy false-alarm result. The study also does not establish transfer to other sites, prospective performance, a perfect reference stream, causal energy loss, repair recovery, financial savings, or long-term panel degradation.
The next research step is to evaluate the approach with additional sites and independently verified operating records, while preserving predeclared thresholds and clear separation between observations, model estimates, and field decisions. Any operational use would also need validation against real maintenance outcomes and a deployment review.
Data and further reading
The study uses version 2 of the La Réunion rooftop PV shading dataset , published by Graillet et al. in Data in Brief . The system and intervention context is described by Lebreton et al. in Entropy . The dataset is shared under CC BY 4.0 ; retain attribution when reusing it.
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article