Description
Date depot: 23 mai 2023
Titre: Machine learning research for renewable energy forecasting and planning
Directrice de thèse:
Claire MONTELEONI (Inria-Paris (ED-130))
Domaine scientifique: Sciences et technologies de l'information et de la communication
Thématique CNRS : Intelligence artificielle
Resumé: The transition to renewable energy sources is crucial to mitigating or reducing the worst effects of climate change. Machine learning can help accelerate this transition in a variety of ways. This doctoral project will focus on improving the forecasts of solar and wind energy, towards accelerating the transition to renewable energy.
Overview
The goal of this doctoral thesis is to design, develop, and compare machine learning algorithms to model and forecast solar and wind energy potential, with a particular focus on West Africa. The large-scale meteorological fields available from physics-based climate models are at lower resolution than needed for this task. In addition, these fields naturally evolve slowly. As such, they are more predictable by physics-driven models than potential outputs of these renewable energy sources.
Approach
Given reanalysis data for the relevant meteorological fields, which is computed from past observations, subject to physical laws, we can train machine learning models to downscale the course resolution climate model outputs, using a supervised learning framework. In particular, the data from the physics-driven climate models, which is at course resolution in both time and space, will be used as the training data, which will be “labeled” by the fine resolution (in space and time) reanalysis data.
Since the data are in the form of spatial fields, we will compare the performance of convolutional neural networks (CNNs) for the task. Recent work has also shown the benefit of Conv-LSTM (which uses convolutions in a form of recurrent neural network) for downscaling one data field using other data fields for training (Harilal et al., 2021). Meanwhile, since there are non-linear relationships in these fields over both space time, as well as across different fields, we will also use self-supervised learning methods involving masking parts of the data, and learning attention-based models (e.g. Transformer; Vaswani et al., 2017) to estimate it from the remaining data. We will also compare to physics-constrained learning methods for the task, while being mindful that both the reanalysis data and the climate model data were already generated with some physical assumptions.
Dr. Monteleoni's team has demonstrated the promise of unsupervised deep learning methods for domain alignment for the general task of downscaling geospatial data (Groenke et al., 2020). We are currently developing an ML-based method for temporal downscaling. The proposed doctoral thesis project will apply and extend these methods to address the challenges below, in order to provide critical information for solar and wind-power planning.
Résumé dans une autre langue: Anglais
Doctorant.e: Boubakar Zourkalaini