From Forecast to Monitoring: How Precise Nowcasting Data Reveals System Faults

A precise PV model based on the exact physics of solar cells is not only the foundation for a good forecast, but also the starting point for any serious monitoring. Many forecast providers now use AI-based predictions trained on system data. These, however, have the decisive disadvantage that different faults get learned in together — for example, shading cannot be cleanly separated from soiling or degradation.

Only when you know what a fault-free system should deliver under the given weather conditions can you judge whether the system is actually delivering its specified performance. With the launch of nowcasting, our /forecast endpoint has moved a decisive step closer to reality — and is therefore suitable for an initial target-actual comparison, without requiring an on-site sensor.

TL;DR

• Nowcasting makes the forecast monitoring-capable — recalculated every 10 minutes based on current satellite data.
• A real comparison shows a surprisingly good match across three cloudy days, including a brief deviation around midday that traces back to a curtailment of the system.
• Trends such as soiling or degradation can already be derived from this data basis — automatic detection as a feature is in the works.

Forecast Meets Reality: Three Days of a pvnode User

A pvnode user who has been using our nowcasting plan since launch provided us with the measured performance data of their system. We placed it, unchanged, against the output of our /forecast endpoint and compared the two over three days.

All three days were variably cloudy. This is exactly the scenario in which classic weather models with hourly or three-hourly updates traditionally struggle, because cloud movements are simply too fast for the update frequency. Our nowcasting update every 10 minutes shows its strengths here: the short-term drops, the recoveries during gaps in the clouds, and the general shape of the day are mapped surprisingly precisely. Looking at the chart longer, you see an almost congruent curve over large stretches of the day.

Around midday on the second and third day, a deviation stands out that traces back to a curtailment of the system. This detail only becomes visible at all because the rest of the curve fits so well — on an imprecise forecast, such a gap might have vanished in the noise.

How Precise Does Monitoring Actually Need to Be?

Classic monitoring with an on-site sensor is maintenance-intensive and expensive. In most cases, such a sensor also only provides the global irradiance at the location. Soiled or defective sensors render the entire monitoring useless, because the reference itself is wrong — and often without anyone noticing.

A far more pragmatic alternative is observed, high-resolution satellite data combined with a precise PV model, as found in our historical data via /recent. They draw on actual satellite observations, so they aren't a model of the future but a look back at what really happened in the sky. No sensor and no maintenance — and for most monitoring questions more than sufficient. Especially for sunny periods, this data is extremely useful for analyses of shading, soiling and degradation.

What nowcasting has changed: the forecast itself now also comes surprisingly close to reality for many of these questions. By definition, it's a prediction and therefore never quite as exact as the historical look back via /recent. Its decisive advantage is that it's available live. Anyone who wants to know in near real time whether a system is currently delivering what it could deliver now gets that answer from a single API request.

What the Data Can Show — and What It Can't

As soon as the forecasted value is close enough to reality, the difference to the actual generation itself becomes a signal. A system that systematically stays below its forecast over several weeks points to a change — typically soiling, degradation of individual strings, or other technical problems.

Shading can already be configured manually in the pvnode API through several options — from automatic horizon calculation to detailed string shading. Even more convenient, however, is when shading, degradation or other effects are detected automatically from the comparison with the measured generation. This allows underperformance to be uncovered and, where possible, specifically remedied. Alternatively, the result can be used to calibrate the forecast or the /recent data to the individual system — for example, in the case of irreversible degradation or permanent shading influences that systematically reduce performance. We're currently working on exactly such a feature. If you'd like to be among the first, feel free to write to us via the feedback button in the dashboard or by email at info@pvnode.com.

Nevertheless, a forecast remains a forecast, even with nowcasting. The cloud picture is often complex, and satellite data has a physically limited resolution of around 2 to 3 km². If a system sits right under the edge of a cloud for an extended period, nowcast and reality can temporarily diverge noticeably. Anyone working with the forecast difference should therefore think in somewhat longer time frames. To reliably detect faults or outages, clear, sunny days are particularly suitable. In addition, trends over weeks are more robust than individual data points. For the comparison shown, we therefore deliberately chose a cloudy period and not a perfect sunny day — almost every model works on those. It only becomes meaningful on difficult days. That helps to judge the true forecast quality and to set the right expectations for forecasting.

Get Started Yourself

If you're not using pvnode yet, all new accounts start on the Pro + Nowcasting plan with five locations for testing. This gives you immediate access to the same data quality shown in the comparison above. Integration into Home Assistant, Solectrus, EVCC or your own application is done in just a few minutes via our Forecast API.

If you have questions or feedback, feel free to use the feedback button in the dashboard. And if you've made interesting comparisons between forecast and reality yourself — we always love real data from the field.