Photovoltaic generation differs from conventional dispatchable generation in two fundamental ways: its output depends directly on the weather, and a large share of installed capacity is distributed across many small and medium-sized systems. The result is a highly variable generation resource whose behaviour changes continuously across both time and space.
Limited observability
Many PV systems do not provide real-time production measurements to a common central platform. Available information may arrive with delays, be aggregated over large areas, or describe only a subset of installations. For forecasting purposes, this means that the real state of distributed generation is never observed with the same completeness as a small number of large conventional power plants.
SentinelPV addresses this limitation by asking whether a smaller network of high-quality observation points can provide enough information to infer the behaviour of the surrounding PV population. The project therefore treats monitoring strategy and forecasting methodology as parts of the same problem.
Spatial variability
Geographical distance alone is not sufficient to describe the similarity between PV systems. Two nearby locations can experience different cloud conditions, terrain-induced microclimates or local irradiance patterns, while sites that are farther apart may show similar production behaviour under recurring weather regimes. Topography, land cover, meteorology and historical production therefore become relevant variables for spatial modelling.
This is why SentinelPV combines geospatial analysis with production and weather data. The aim is to identify areas that are homogeneous enough to be represented efficiently, while preserving the local information that matters for forecasting.
Rapidly changing and atypical weather
Forecast errors tend to become more difficult during rapid cloud transitions, fog, unusual aerosol conditions and other atmospheric situations that are not fully represented by standard weather products. Saharan dust transport is an important example for solar applications because aerosols can alter the radiation reaching the ground and generate systematic deviations between conventional forecasts and actual PV production.
The project therefore includes specific atmospheric case studies in addition to average performance analysis. These cases are used to test whether hybrid physical and machine-learning methods remain robust when operating conditions differ from typical training periods.
From single sites to regions
A forecasting system designed for distributed PV must work at more than one spatial scale. SentinelPV studies forecasts for individual locations and their aggregation into zones or regions, analysing how uncertainty changes as the spatial scale becomes larger. This helps identify where detailed local modelling is necessary and where aggregation naturally smooths part of the variability.