Cómo la inteligencia artificial resuelve la pérdida de datos en plantas solares
One of the most common, and most expensive, problems in the operation of photovoltaic plants is the loss of production data. Sensors that fail, communications that cut out, inverters stopped for maintenance. The result is the same: incomplete data series that complicate both plant maintenance and the forecasting of how much energy will be generated.
At Smarkia we have been researching this problem for some time, and the result of that work, a model based on Deep Learning, was published in Expert Systems With Applications, a scientific journal of international reference. We explain to you, in simple terms, what it consists of.
What is Deep Learning, in two sentences
Deep Learning is a branch of machine learning that uses multi-layered neural networks to identify complex patterns in data. It is the technology behind applications as diverse as diagnostic imaging or traffic forecasting, and it can also be applied to the reconstruction of incomplete energy data series.
Why solar production data is lost
The causes are varied: communication failures, inverter shutdowns, noise in measurements, atmospheric phenomena, or sensor problems. According to various studies in the sector, series with more than 5% of missing data already require some type of treatment before they can be analyzed with guarantees, and in practice, it is not uncommon to find losses of up to 40% in real solar production series.
This lack of data is not a minor issue: it affects both the operation and maintenance of the plant, hindering predictive maintenance, as well as the forecasts necessary to properly integrate solar energy into the electricity system.
What the Smarkia model brings to the table
The difference compared to previous solutions is that the model can be trained with a single production data series, even an incomplete one, without the need for a previous history or correlated external signals, such as solar radiation or temperature. This makes it especially useful in installations where long historical series or proprietary meteorological sensors are not available.
In the tests of the study, the model maintains good performance even with series that have up to 70% of missing data, with the best results around 50% loss, overcoming the results of other previous approaches.
Why this matters to those who operate solar plants
Beyond the academic component, this ability to reconstruct data in the event of sensor failures is today part of how Smarkia manages solar assets in production, something that Verdantix has explicitly recognized in its report Smart Innovators: Energy Management Software (2025), by pointing out that providers like Smarkia apply deep learning at the edge of the network to reconstruct missing solar generation data and maintain visibility of the installation even when sensors fail.
In practice, this translates into fewer gaps in the data, better generation forecasts, and less time wasted interpreting incomplete series, whether you manage a solar plant in Spain or in the United States.
If you want to consult the full technical detail, the paper is publicly available in Expert Systems With Applications.