Univariate machine learning models applied in photovoltaic power prediction using Python

ABSTRACT

Photovoltaic field has attracted the attention of researchers around the world because of the intermittence of this energy since its integration into the electricity grid poses problems in maintaining the balance between consumption and production. In fact, forecasting PV (photovoltaic) energy sources would make it possible to anticipate the availability of production sources and thus facilitate network management. Indeed, the PV production depends on many parameters such as the meteorological parameters, irradiance, and installation conditions of PV modules. However, this study concerns the prediction of photovoltaic production using machine learning techniques applied for the first time on the Beni Mellal’s photovoltaic station. These methods make it possible to predict the production of photovoltaic electricity by programming a computer to learn from previous data of the parameters on which the photovoltaic production depends and then to predict future ones, that is to say direct forecasting. First, a three-year meteorological parameter (including just ambient temperature and irradiance) and panel parameters (panel temperature) database is used to train the algorithm and a three different months database to test it; all, using supervised learning models based on three approaches – programmed by Python: Support Vector Regression, Decision Tree Regression, and Random Forest. These approaches aim in first time to analyze the correlation between meteorological parameters and photovoltaic production and then, in other time, to predict PV production directly. The results have shown that the global inclined irradiance and the panel temperature are the data most correlated with the photovoltaic production with a correlation coefficient of 0.88 and 0.61, respectively, so that led us to rely, on the first level, on global inclined irradiance and panel temperature, and then on other meteorological data, on the second level to predict PV output. It is also concluded in this study that Random Forest was the best used approach with a mean square error of 0.18.

KEYWORDS:

• Machine learning
• meteorological conditions
• PV power prediction
• PV production
• Python

Acknowledgements

This work is part of the ‘Propre.Ma’ project, which has been sponsored by IRESEN (Morocco) since February 2014. In the name of all the rest of the consortium members, the authors would like to thank IRESEN Board and Scientific Committee for financing this through the INNO’PV Program as well as its General Manager, Badr IKKEN, and the rest of the staff for their daily efforts. Authors are grateful to the members of the ‘Propre.Ma’ consortium who are essential for territorial extension of the project.