Solar energy production forecast using deep-learning techniques applied on meteorological satellite images
30 Août 2022
Catégorie : Doctorant
Solar electricity production forecast enables a better integration of solar energy into the grid and then increases its share in the energy mix. Geostationary meteorological satellites, such as Meteosat, observe the cloud layer in real time by producing an image representing the same part of the Earth surface every 15 min. Proven forecast methods using image processing techniques (such as, block matching and optical flow) can anticipate cloud motions and then deduce photovoltaic (PV) production forecast on a given location for the next hours. Despite of their limit to detect sudden cloud appearance/disappearance, these methods are generally more reliable than classical weather forecast models.
Recent deep-learning models, in particular convolutional neural networks (CNN) models, encouraged researchers to develop precipitation forecast methods from radar images. Rainfall forecasts are known to be very complex and demanding in accuracy in order to prevent weather hazard consequences (floods, storm…). Despite of heavy computational costs, these techniques are advantageous compared to classical weather forecast models. This research demonstrates that solar energy forecast using satellite images have several reasons to use the CNN models. Indeed, more than 20 years of homogenous satellite data at high frequency (15 min.) are available for training models. Cloudiness and irradiance are bounded physical quantities that avoid inconsistent training. Finally, cloud evolution at fine spatiotemporal scale is a consequence of stochastic phenomena that cannot be represented in the current physical models.
The objective of the work consists in the design of a new deep-learning model dedicated to PV production intraday forecast using visible channel of geostationary meteorological satellite. The main steps are:
· classification of various cloud spatiotemporal patterns observed in satellite data linked with identifiable weather situations (e.g. advection, convection, cold or hot front, depression etc.).
· Implementation of a CNN model already proven for precipitation forecast (e.g. U-net, convLSTM, …) on each weather situation classes
· For each of these situations, studying the added-values of external related data at synoptic (continental) scales (e.g. surface pressure, temperature field from weather model or reanalysis …)
· Quantify the added value of ground observations (typically the surface solar irradiance and the PV power production).
· Identifying the main features of a pertinent DL forecasting method dedicated to cloud cover behaviour at relevant time and space scale for intraday solar energy forecast.
· Education: MSc in data analysis, data science, image processing or equivalent.
· Computer skills: Python, proven experience in deep learning model programming using Keras library with training on CPU and GPU.
· Strong interest in Earth observation, solar energy, meteorology
· Excellent written and communication skills in English
· French speaking and writing skills are strongly encouraged
This work will take place at Laboratoire de Météorologie Dynamique (LMD), Ecole Polytechnique, Palaiseau, 25km from Paris, France, as part of the Energy4Climate interdisciplinary Center research actions (https://www.e4c.ip-paris.fr/#/en/ )
This PhD position is funded in collaboration with french industrial energy group.
The position is expected to start from November 2022.
To apply, please send CV and motivation letter to: Sylvain Cros - firstname.lastname@example.org Jordi Badosa – email@example.com
Cros, S., Badosa, J., Szantaï, A., & Haeffelin, M. (2020). Reliability predictors for solar irradiance satellite-based forecast. Energies, 13(21), 5566.
Berthomier, L., Pradel, B., & Perez, L. (2020). Cloud cover nowcasting with deep learning. In 2020 Tenth International Conference on Image Processing Theory, Tools and Applications (IPTA) (pp. 1-6). IEEE.
Reichstein, M., Camps-Valls, G., Stevens, B., Jung, M., Denzler, J., & Carvalhais, N. (2019). Deep learning and process understanding for data-driven Earth system science. Nature, 566(7743), 195-204
Su, X., Li, T., An, C., & Wang, G. (2020). Prediction of short-time cloud motion using a deep-learning model. Atmosphere, 11(11), 1151.