Skip to Main content Skip to Navigation

Qualité et valeur des prévisions solaires probabilistes

Abstract : Solar forecasting is a major asset in many solutions for large-scale integration of solar power into electricity mixes. This work aims to prove that probabilistic forecasting, yet not frequently used at an industrial scale, is nonetheless for many applications. Since the scientific literature related to value of forecasts is rare, contrary to literature about quality, this work is devoted to establishing links between quality and value. If they can be generalized, they would demonstrate and help the understanding of added-value of probabilistic forecasting. In a first step, a framework for the assessment of quality of probabilistic forecasts is presented. It is constituted of metrics, graphic tools and reference benchmark models. After a bibliographiv review focused on probabilistic metrics, a new tool called “outcome matrix” designed for the comparison of probabilistic metrics is presented. It is firstly used to select the most appropriate scores for the research question, and then to quantify the potential benefits in terms of quality that the usage of probabilistic forecasting may bring. In order to complete the framework of quality assessment of probabilistic forecasts, we propose a new reference model to benchmark solar probabilistic forecasts called “CSDCLIM”. The main idea is to use at best all knowledge contained in the historical data of a site to generate a reliable forecast. The quality of this model slightly outperforms those of commonly used reference models. Finally, the benefits really obtained from the use of probabilistic forecasts are calculated for several probabilistic forecasting models from the literature and on six sites facing very different climatologies. This calculation is made with a new methodology called “MC-CRPS”, for day-ahead forecasts built by post-processing of ensemble forecasts provided by ECMWF. A second step is dedicated to the study of the value of forecasts from a user perspective. The value of a forecast depends on its usage, and it is then necessary to select a practical case study to quantify it. For industrials, it measures the gains brought by this forecast in a decision-making process. A theoretical research is contucted on the way to model the cost associated to decisions based on the forecast. A methodology for the evaluation of the value of a probabilistic forecast of continuous variables is proposed. It consists in the usage of a plot that combines two independant informations, namely the cost structure and the skills of the forecast for all the different probability levels. This methodology allows to consider separately the factors explicable by the quality of the forecast and by the market cost structure. Finally, this methodology is used for the analysis of the results obtained by different forecasting models for the economic optimization of the production of a solar powerplant on three different electricity markets. The results tend to prove that the cost structure has indeed a crucial role on the effective economic value of forecasts. Again, the gains of the probabilistic approach compared to the deterministic approach are evaluated and compared with quality gains. For an industrial usage, it must be noted that it is necessary to have also a forecasting model dedicated to market prices. A simple method of such a forecast has been implemented. On the considered cases, the values of the probabilistic forecasts outpace those of deterministic forecasts. However, the differences between these two approaches decrease as soon as the cost structure is itself uncertain.
Complete list of metadata
Contributor : Abes Star :  Contact Connect in order to contact the contributor
Submitted on : Monday, October 11, 2021 - 2:15:29 PM
Last modification on : Wednesday, October 13, 2021 - 3:34:25 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03373288, version 1


Josselin Le Gal La Salle. Qualité et valeur des prévisions solaires probabilistes. Physique Numérique [physics.comp-ph]. Université de la Réunion, 2021. Français. ⟨NNT : 2021LARE0016⟩. ⟨tel-03373288⟩



Record views


Files downloads