Benchmarking different approaches to convert surface solar irradiance into PV power production : a case study with an operational forecast system for a roof-top PV farm

Photovoltaic production of a given solar power plant is mainly correlated with solar irradiance reaching the solar panel and to a lesser extent with air temperature and wind speed. Therefore, available energy in a building with roof-top photovoltaic (PV) panels in self-consumption can be highly variable due to cloud cover stochastic behaviour. Accurate irradiance forecast within the next hours are useful to help the energy management system of the building to cope with this variability and thus to maximize the consumption of the locally generated electricity at the expense of grid energy, then reducing financial and environmental costs of the overall building energy consumption.

Any solar energy forecast solution presents several sources of uncertainty at each main steps of the process: cloud forecast, radiative transfer of cloud and aerosols, irradiance conversion into power. Reducing the uncertainty of PV power modelled from irradiance forecast is a specific issue because it depends on certain conditions of the forecast application. PV performance models convert irradiance into PV power if PV cells characteristics are known and correctly specified by manufacturer, which is not always the case. Moreover, unexpected shadowing, panel surface soiling or aging of materials cannot be easily taken into account. If a consistent historical record of PV power data is available, a model output statistics (MOS) helps to decrease these induced uncertainties. If only real-time data are available, adaptive calibration using Kalman filter can improve the accuracy. If both historical and real-time data access are available, various machine learning approaches can set up more accurate MOS.

The start-up incubator of Institut Polytechnique de Paris is a building equipped with a roof-top PV farm with a total capacity of 17 kWp, made up with 53 using 8 different technologies with several tilted angle and some panels are equipped with reflectors. Minute PV power data have been continuously recorded from July 2020 onwards, collocated with the irradiance measurements of the Palaiseau BSRN station. This exceptional testbed enabled us to benchmark the several approaches for PV power modelling. We implement reference algorithms (linear regression, Kalman filter and autoregression) simulating respectively the availability data conditions (historical, real-time, both) and compared them with more complex machine-learning approaches. Results according to data availability, season, solar zenith angle, air temperature and irradiance variability are discussed. One of the outcome is that availability of real-time power measurements improves significantly forecasting results in most weather situations.