The SAPHiR project hosts a scientific discovery internship
The intern
Nicolas Ottavi-Nedelec is an undergraduate student at the University of Corsica where he is majoring in both mathematics and computer science as part of the “Licence sciences pour l’ingénieur”.
Having always had a passion for all physical sciences he had a late start, going back to college at the age of 30 after unsuccessful attempts in the past.
After a successful first couple of years he took an interest in the research that was going on at the university and for his final year internship asked to join, and was welcomed at, the Saphir project.
There he would work for four weeks in the month of April 2022 under the supervision of Professor Christophe Paoli on machine learning for time series forecasting with applications to solar energy production.
The work
The aim of project Saphir is to investigate weather phenomena using a combination of traditional numerical forecasting methods and machine learning tools.
It was Professor Paoli that first suggested looking into a recent explainable deep learning architecture from Google, a relatively complex model called Temporal Fusion Transformer or TFT. So Nicolas delved into the paper and tried to make sense of it.
Nicolas: “We wanted to accomplish two things during the internship, the first was to figure out how the model worked in some details and then actually use it to make predictions and see if that gave us any benefits.”
The first part proved the most challenging but also stimulating.
Nicolas: “I had never seen anything like it, the model seemed like it had a million parts and all of them complicated. We had barely got a look at neural networks in class and suddenly I have to figure this thing out. I loved it though, trying to see how each individual part worked and then how it fit into the whole. There was also a mathematical aspect that as a math major, I particularly enjoyed.”
Then, having gained a basic understanding of how the model worked, it was finally used on a concrete example. Data from a previous study using a more conventional LSTM model and conducted by one of Professor Paoli’s former student would be fed into the model and the results compared.
The model was trained on meteorological features captured by a Météo France weather station and the aim was to predict the amount of sunlight reaching the ground accounting for things like nebulosity, humidity, wind direction and speed, etc
Nicolas: “The accuracy was pretty much the same but the really interesting things came from the explainability features of the model. The model tells us what it pays attention to. It decides what importance to give to the variables and also, because it is a time series problem, to different points in time.”
The previous study had determined a list of the most important features in the data by trial and error and the TFT essentially agreed.
Nicolas: “It looks like it is doing some feature engineering of its own. It is quite remarkable.”
Conclusion
Nicolas: “Overall the internship was a success. There were some things we wanted to do that we just didn’t have time for, like using data from multiple weather stations or using the computing cluster at the university but we did what we set out to do.”
For Nicolas it was a positive first experience in the world of scientific research and it would lead to his employment as a research engineer at the university, while he continues his studies pursuing a master’s degree in computer science.
Nicolas: “I was too old for a regular co-op program but they wanted to work with me again, and so did I. There is so much more to do.”
While for project Saphir and professor Paoli it suggests potential new forays into explainable models.
