Weather or Hardware? How to Diagnose Solar Underperformance with Physics, Not Guesswork
The most common question in solar operations — and the discipline for answering it with physics instead of opinion, every interval, automatically. To separate weather-driven underperformance from hardware problems, compare actual generation against expected generation computed from measured irradiance and temperature at the same interval. If the plant meets its weather-adjusted expectation, the shortfall is environmental; if a gap remains, it is physical — and its pattern across time, inverters, and strings identifies the cause.
Monday's generation is fifteen percent below Friday's, and the morning call wants to know why. Someone says it was cloudier. Someone else remembers the derate alarms. The asset manager looks at the sky, effectively. This ritual plays out across the industry daily, and it is remarkable how often the answer is settled by seniority rather than by physics. The physics is not hard. It is just rarely operationalized — which is a gap worth closing, because every wrong answer to this question costs money in one of two directions: a truck rolled at a cloud, or a fault left running behind one.
What is the expected-generation method?
For every interval, compute what the plant should have produced given the weather it actually received: measured plane-of-array irradiance, module temperature, and the plant's electrical model. That expectation is the exoneration line. Actual generation at or near it means the weather explains the day — case closed, no truck roll, no meeting. Actual generation below it means something physical is subtracting energy, and the size of the gap is the size of the problem, in kilowatt-hours, convertible to dollars. The critical property is that this works on cloudy days. Naive comparisons — against yesterday, against the monthly budget, against a clear-sky model — all break the moment conditions vary, which is to say, they break on exactly the days the question gets asked. Expected-versus-actual from measured weather never blames a cloud for a combiner fuse, and never lets a fuse hide behind a cloud. The exoneration half of that sentence is underrated: a model that can definitively say 'the weather explains it' is where genuine false-positive suppression comes from — not from muting alarms, but from understanding the data well enough to close cases.
How do you localize the gap once it exists?
The pattern is the diagnosis. A gap uniform across every inverter points to something plant-wide or upstream — a meteorological sensor problem, grid curtailment, transformer losses. A gap concentrated in specific inverters against their fleet peers points at those units. Within a unit, string-current comparison finishes the job: strings wired in similar configurations should track together, and the deltas tell precise stories. Two series strings reading 8.1 and 6.9 amps — elevated deviation, worth a warning and a watch. Two series strings reading 6.0 and 0.1 amps — a delta that should never exist on that wiring, diagnosable from a desk as an open circuit or a blown combiner fuse, before anyone drives anywhere. Time signature adds the final layer. A gap that appears only on hot afternoons is thermal derating. Only at low sun angles: shading or soiling geometry. Stepwise and persistent: a discrete failure with a start date — go find the event. Growing slowly across weeks: soiling accumulation or developing degradation. Each signature routes to a different response, which is the entire point: the question was never 'is something wrong' but 'what, where, and is it worth a truck today.'
| Gap pattern | Probable cause family |
|---|---|
| Uniform across all inverters | Sensor error, curtailment, plant-level electrical |
| Specific inverters vs peers | Inverter fault, derating, DC field of that unit |
| Specific strings within a unit | Open circuit, fuse, connector, module group |
| Hot afternoons only | Thermal derating, cooling degradation |
| Low sun angles only | Shading, soiling edge effects |
| Step change with a date | Discrete failure — find the event |
How Ellume Vector answers the question every fifteen minutes
This method is not a monthly study in our platform; it is the resting state of the interface. Vector's physics engine computes the weather-adjusted expectation continuously and surfaces the comparison at every level of the plant.
- •The O&M dashboard's yield card shows today's production against a target calculated from current weather — on a representative fleet plant, 14,215 kWh against a 15,686 kWh weather-derived target, 90.6% attainment — so the expected-versus-actual gap is the first number of the morning, not the subject of the meeting.
- •The weather correlation page carries the actual-versus-predicted overlay — the model's output for current irradiance laid over the measured output — plus GHI-versus-power, temperature-versus-PR, and cloud-versus-generation correlations, and a seasonal hour-by-month heatmap where recurring patterns (the site that always sags on humid summer afternoons) become visible as shape.
- •The daily power curve is drawn against its expected envelope, and the generation heatmap — hours on one axis, days on the other — makes shading geometry, recurring faults, and grid events jump out of the grid.
- •When a gap is physical, the string health heatmap and per-unit peer comparisons localize it: the 6.0 A / 0.1 A series-string delta above is a real fleet example, and the platform had already written the diagnosis — open circuit or combiner fuse failure — into the event record before any human read it.
- •Every attribution links to its evidence: the physics rule, the raw rows, the highlighted deviation, and the full physics documentation, open to challenge. Weather exonerations are as auditable as fault detections.
From the Ellume fleet: in one month on a single plant, the weather-adjusted expectation model suppressed eleven raw alarms as environmentally explained — zero false dispatches — while simultaneously flagging three genuine string outages that mild, pleasant weather was helping to hide. Both halves matter: the same physics that stopped five wasted truck rolls ($3,500) also caught faults that a comparison-against-yesterday would have exonerated along with the clouds.
What does this change operationally?
Everything about dispatch quality. Weather-exonerated intervals generate no tickets, so the alarm stream regains its credibility. Physically-confirmed gaps arrive pre-localized — a string, not a site — with impact and fix-time attached, so the technician's day is planned by return on wrench time. And management reporting stops relitigating the weather: the monthly narrative becomes a loss ledger — this much environmental and unavoidable, this much asset fault, this much recoverable — which is the only version of the story an owner can act on, and the only version an O&M contract can be fairly settled against.
Frequently Asked Questions
- Does this require on-site weather instrumentation?
- Measured plane-of-array irradiance from a maintained sensor is the gold standard; well-validated satellite irradiance is a workable fallback with a known 3–5% uncertainty. What is not workable is judging plant performance with no irradiance reference at all.
- How large a gap is worth investigating?
- Persistence beats size. A 2% gap that survives a week of varied weather is a fault; a 6% single-interval gap is often transient. Set triggers on sustained deviation, and let the localization pattern set the priority.
- Can the expected-generation model itself be wrong?
- Yes — a good one is validated against clear-sky, fault-free periods and re-baselined for degradation annually. A model nobody validates drifts into either false alarms or false comfort; the second is worse, because it is silent. Vector's model validation and physics documentation are exposed for exactly this reason.
- How does this interact with curtailment?
- Grid curtailment appears as a uniform, plant-wide gap with a grid-event time signature — distinguishable from hardware by pattern, and essential to tag separately so it can be reconciled against the compensation terms rather than blamed on the plant.
- Is this the same as a digital twin?
- It is the useful core of one: a physics model of expected output under measured conditions, continuously compared to reality. The label matters less than the properties — measured-weather inputs, interval resolution, auditable rules, and attribution of every deviation.
