New PhD student in SAPHiR project: Diego GRANTE - Impact of weather forecasts on the integration of photovoltaic and wind power into the electrical grid.

This thesis, carried out by Diego GRANTE and supervised by Christophe PAOLI and Rachel BAILE, is conducted within the computer system and ubiquitous system (SISU) team at the University of Corsica and is part of the SAPHiR project (ANR-21-CE04-0014-03). The objective of this work is to provide accurate forecasts of wind speed and solar radiation for time horizons useful to an electrical grid operator, such as EDF in France. 

After obtaining a Master's degree in Full-Stack Development (DFS) at the University of Corsica, Diego will leverage his development skills to strengthen his expertise in machine learning and more particularly in deep learning. 

 During the first months, he will be working on creating deep-learning models for predicting wind speed. These models will be trained using data from ERA5 (link), MeteoFrance (link), and SAPHiR (from a local sensors network). A middle-term objective will be to compare large models like GraphCast or Pangu-Weather with a custom local model. 

 Moreover, with the emergence of the explainability domain and the need for the decision makers to explain possible unexpected results, he will work on creating a custom explainable local model to assist a network manager, as EDF in France. For instance, Figure 1 illustrates the beginning of his work.  

Figure 1: Importance of Variables in the Temporal Fusion Transformer (TFT) Model during Encoding. 

Figure 1 shows preliminary results obtained using the Temporal Fusion Transformer (TFT) model with ERA5 data from 2018-2020 in the case of Corsica to predict (horizon h+1) the i10fg, u10, and v10 variables (see table below in figure 2 for the features and a simplified model description).

Figure 2: Features and model description 

With the help of explainability during the encoding phase, Diego will be able to work on feature optimization of the model. In this example, the Instantaneous 10m wind gust (i10fg) is the most important variable in terms of percentage. As the model is attempting to predict this feature, this observation seems logical. However, the 2m temperature (t2m) appears more important than one of the other variables we want to predict: the v-component of wind at 10m (v10). Even though it is well known that temperature is an important feature in wind forecasting, this result is surprising at first glance. Therefore, We can ask the variable selection impact on the architecture efficiency to predict v10.  

This issue will be explored in the coming weeks by Diego.