Integration of Physical Knowledge in Statistical Photovoltaic Production Forecasting Models - Archive ouverte HAL Accéder directement au contenu
Pré-Publication, Document De Travail (Preprint/Prepublication) Année : 2022

Integration of Physical Knowledge in Statistical Photovoltaic Production Forecasting Models

Résumé

Photovoltaic (PV) production is characterised by a high variability originating from complex processes involving the combination of astronomical, meteorological, environmental, and technological factors. Irradiance reaching the PV panels is modulated through the Sun’s position in the sky dome, and depends on the composition of the atmosphere, the presence of surrounding obstacles, and also on the inclination angle and optical properties of the module’s cover. Then, during the conversion process into electric power, ambient conditions such as temperature impact negatively cell’s efficiency. This variability raises concerns in a context of high penetration of renewables, and requires efficient forecasting tools to guarantee a safe and profitable operation of the grid. In the literature, we observe a dichotomy between the nature of forecasting tools traditionally used: one can explicitly model the conversion laws at stake, or consider data-driven models that implicitly derive these relations. In an attempt to merge these two fields, we propose a simple physics-based model that converts irradiance into power-like feature as a pre-processing step before the integration into a Machine Learning (ML) model. This contributes to reduce the inferring effort by explicitly integrating plants technical properties within regression models. In contrast to ML models, it is common practice in the time-series forecasting domain to resort to clear-sky normalisation methods to provide stationary inputs. Typically, this process only clears the PV production signal from its dependency on the Sun’s path. We propose to go further by applying the conversion model in a clear-sky normalisation framework to remove dependencies on plant geometry and ambient environmental conditions. This approach turns out to slightly improve local stationary properties of production time series, and to impact positively the forecasting performances of the studied ML model in comparison with the consideration of raw features. In addition, we highlight that the ML model is able to derive, to some extent, the conversion laws when fed with a set of relevant features intervening in the conversion chain.
Fichier non déposé

Dates et versions

hal-03839668 , version 1 (04-11-2022)

Identifiants

  • HAL Id : hal-03839668 , version 1

Citer

Kevin Bellinguer, Robin Girard, Guillaume Bontron, Georges Kariniotakis. Integration of Physical Knowledge in Statistical Photovoltaic Production Forecasting Models. 2022. ⟨hal-03839668⟩
92 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More