Independent loss analysis for PV farms

Forum Energii calculated that in June 2026 PSE, the grid operator, curtailed 14.2% of the potential output of large PV farms (246.9 GWh). From January to September 2026 it curtailed 1,736.9 GWh of renewable energy in total, more than in all of 2025 (1,376 GWh). Sources: Forum Energii, Monthly Magazine, September 2026 (opens in a new tab), gramwzielone.pl, 9 July 2026 (in Polish) (opens in a new tab), gramwzielone.pl, 29 July 2026 (in Polish) (opens in a new tab).

The act of 18 September 2026 amends art. 9j: a generator with a type B, C or D module sends the operator structural and real-time data, and sends and carries out schedule data (para. 1a). The change applies from 31 October 2026 (art. 16). Art. 9je is a grid measure, not a fine: for a gross breach of the requirements on schedule data, real-time data or the operator's orders, the operator may, on the terms set in the network operation manual, restrict or halt the module's operation for no longer than 24 hours, with no compensation. The money fine is in art. 56 para. 1 point 19a: it covers failing the duties of art. 9j para. 1a, structural data included, runs from 1,000 PLN to 500,000 PLN and is one fine for all breaches in a calendar quarter (art. 56 para. 3 point 3 and para. 3d). Source: Journal of Laws (Dz.U.) 2026, item 1348 (in Polish) (opens in a new tab).

The average power loss in 2025 across PV sites in the US (a sample of 373 GW) was 5.08%. Inverters account for less than a quarter of equipment losses; strings, combiner boxes and trackers cause most of the rest. This is US data. Source: Raptor Maps (a company that sells inspections), 2026 Global Solar Report, via pv magazine USA, 25 February 2026 (opens in a new tab).

Helio is our own project, not a client deployment. The demo runs on two public datasets, and the simulation is always marked. 5.05 MWh of real losses in 2015, at least 2,272 PLN at an assumed price of 450 PLN/MWh. Source: one installation from the public NREL PVDAQ dataset (system 1204, a 78.6 kWp rooftop, 8 inverters with data, readings every 5 minutes). It is a lower bound: the data of inverter inv4 ends on 1 September 2015, while the outage was still on. 98.8% of that loss is one outage: inverter inv4 was down for 132 days, from 22 April 2015 to the end of its data. 9 of 9 detector windows overlap a labelled fault run (95% Wilson confidence interval: 70% to 100%). Test set: the FHC twin plants, 4 strings, readings every 30 seconds over 25 days within 17 June to 15 July 2025. The detector found 6 of 31 labelled fault runs, including 6 of 8 where the sun was up for at least an hour. Most of the rest are sensor errors and disturbances shorter than an hour: the detector looks for lost energy, so by design it does not catch them. The FHC set is single 400 W modules, so the result does not carry over to another farm. The price of 450 PLN/MWh is an assumption, not a market quote. The samples are small (9 windows and 31 runs). Both datasets are CC BY 4.0. Data: Deline et al. (2021), PVDAQ, NREL/OEDI, doi:10.25984/1846021 (opens in a new tab), Bradl et al. (2026), Zenodo, doi:10.5281/zenodo.18979877 (opens in a new tab).

An export of your inverter data: a CSV from your SCADA or the inverter portal, for a year or a quarter. We need the AC power of each inverter with a timestamp every 5 minutes or less, and at least four inverters. Instead of a file you can give a read-only token for your Zabbix. By default Zabbix keeps 31 days of history, so a year or a quarter needs the export or a longer history setting. We compare each inverter with the others on the same farm and check the result on your data. We measured detection accuracy on one lab installation, and the result does not carry over to another farm. The detector looks for lost energy, so by design it does not catch short disturbances and sensor errors. An inverter that delivers less than 15% below the others counts as healthy.

From an export of your inverter data we count the energy lost, in MWh and PLN, by cause, and list the problems with dates. Next to every number in the report stands the query that counted it, so you can check it. In the AI analysis you ask about your data in your own words, and the AI reaches your database through a server that can only read. Every query is logged, and limits on rows and time protect the database. The AI computes nothing: it describes the results of queries. We switch this on only with your written consent for a named AI provider. The dashboard and the analysis run on your infrastructure or in a private instance set up for you, under a data processing agreement that names the hosting provider. Data refreshes on a schedule, and you get a login, backups, documentation and training. In support we watch that the data keeps refreshing, install security fixes and develop the dashboard to fit what you need. Integrations: Zabbix, SCADA exports, inverter portals; we join what you already use and fit the tools to your system.

An export of your inverter data needs no access to your systems from us: you send a file. Remove the names and contact details of people from the file before you send it. If you prefer, you give a read-only Zabbix token for the hosts you choose, instead of a file. Before you send the file, we sign a non-disclosure agreement, yours or ours. It covers one thing: one analysis of your inverter data. We delete your data 30 days after we deliver the final report. You get a deletion note with the date and the scope. Your data reaches an AI provider only with your written consent. The consent names the provider, the product, the files and the provider's retention terms, and you can withdraw it. Your data and your configuration (names of sites, thresholds, energy price) stay yours.