Description du sujet
Climate change, the intensification of droughts such as the one occurred in the summer of 2022, the reduction in low-water flows in rivers and significant changes in the recharge dynamics of underground aquifers require more informed and efficient management of water resources by users, including irrigated farmers. Such management must be based on forecasts of surface and groundwater resource trends on different time scales – seasonal, monthly, weekly – to adjust or prioritize needs and anticipate periods of tension and shortage. Recently, Deep Learning and artificial intelligence have become crucial for developing numerical simulation and forecasting models in hydrology. After years of disappointing attempts (Gaume and Gosset, 2003) the performance of Deep Learning models has improved considerably, outperforming now traditional numerical models. This milestone in hydrology has recently been reached with the introduction of Long Short-Term Memory (LSTM) recurrent neural network models for simulation (Kratzert et al., 2019, 2018) and short-term forecasting (Saint-Fleur et al., 2024) of river flows. Despite encouraging initial results, many scientific questions remain to develop robust numerical models for operational decision-making.
The A3P (Anticipation, planning and management of agricultural withdrawals) project aims to develop AI-based decision support tools. Selected in 2024 as part of the France 2030 call for projects “Innovating to ensure the agro-ecological and food transition”, it will run from 2024 to 2028. The project partners are Aquasys, a French SME, leader in digital water management technologies; INRAE’s EMMAH research unit, which is specialized in the estimation of water requirements of cultivated plants and water consumption for irrigation ; MEOSS, an expert in satellite image processing; and Gustave Eiffel University, which is responsible for developing water resource AI-based forecasting models.
This doctoral thesis will focus more specifically on modeling the rainfall-runoff relationship in a forecasting context over medium-term timeframes (from 1 to 6 months) using Deep Learning methods. Indeed, flow forecasting on timescales longer than a week is a topic that has not yet been widely covered in the scientific literature.
The main questions to be addressed in this thesis are:
- Recognize the key variables and network structures best suited to forecast the rainfall-runoff relationship in catchments, taking into account forecast horizons. Given the inertia of the rainfall-runoff relationship in watersheds, we may consider assessing the added value of composite variables (feature engineering), such as cumulative effective rainfall, basin characteristics, etc.
- Evaluate the robustness of learning strategies based on local (regional) or global (national) data for neural networks, and possibly fine tuning of global models in the context of forecasting (Kratzert et al., 2019) ;
- Evaluate the potential of physically informed neural network modeling techniques in forecasting (Borate et al., 2023; Saint-Fleur et al., 2024) ;
- The coupling with ensemble monthly of seasonal meteorological forecasts and evaluation of the performance of the complete ensemble forecasting chain,
- Improve the consistency of forecasts produced on a set of measuring stations located in the same geographical area (region). On this point, we need to draw on the extensive literature dedicated to the consistency of neural network forecasts in disciplines other than hydrology.
| Month No. | Steps |
| 0 – 6 | Bibliography (state of the art in the application of AI to hydrology and hydrogeology in particular), familiarization with the team’s modeling tools and methods and, with the support of the Aquasys team, creation of a national dataset: rainfall-runoff, meteorological and associated hydrological series and metadata from BRGM (groundwater) databases over the entire territory of France. |
| 6 – 12 | Data mining and experimental design for testing variable combinations and network structures, first exploratory tests. |
| 12 – 18 | Implementation of the experimental design, analysis of the results and writing of a first scientific article on the modeling and forecasting of medium terms hydrometeorological forecasting. |
| 18 – 24 | Implementation of physically-based or conceptual models on the considered set of rivers and water bodies, combined with neural networks previously developed (proposed physics based informed AI methods (Borate et al., 2023) or others). Depending on results, write a second article. |
| 24 – 30 | Bibliography on regional training and spatial consistency (geographic constraints). Implementation and testing of fine-tuning methods, including geographical consistency constraints for measurement stations belonging to the same zone/region. Analysis of results and preparation of an article. |
| 30 – 36 | Finalization of publications and formatting of dissertation. |
The thesis will be based on the use of generic libraries (Scikit-learn, Pytorch Numpy, Pandas, Scipy, …) or specific libraries (https://github.com/kratzert/lstm_for_pub) of the Python language. Part of the calculations will be performed on a supercomputer: the GLICID platform in the Pays de Loire region, or Jean Zay if required. Data will be extracted from national databases and prepared by the aQuasys team as part of the A3P project.
Prise de fonction :
02/12/2024
Nature du financement
Financement public/privé
Précisions sur le financement
BPI France
Présentation établissement et labo d’accueil
Université Gustave Eiffel – Laboratoire Eau et Environnement
The thesis will be carried out in GERS department of Gustave Eiffel University, located at Nantes, and supervised by Eric Gaume (CESAAR senior researcher directing research in hydrology) and Florian Surmont (PhD Engineering Science, data scientist). GERS is one of the leading European team in the fields of Geotechnics, Environment, Natural Hazards and Earth Sciences. More precisely, the Water and Environment team (GERS-EE), made up of 39 members, conducts research in hydrology, hydraulics and water resource management including flash floods, flood and drought risk assessment, hydrological forecasting modelisation… GERS-EE is involved in many national research projects, from fundamental research to high TRL projects.
The GERS department (175 people) of the Gustave Eiffel University is dedicated to research in the fields of geotechnics, the environment, natural hazards and earth sciences.
The EE team has 39 members, including 17 research managers, 10 technicians, 10 PhD students and 2 work-study students. In particular, it conducts research and development in hydrology, hydraulics and water resource management: exceptional floods, flood and drought risk assessment, development of digital flood forecasting models and natural risk management. It has coordinated and/or participated in national research projects aimed at improving rainfall-runoff forecasting models, notably the ANR PICS and MUFFINS projects.
Teams involved and key researchers :
– Éric Gaume: Director of the GERS department and senior researcher and expert in hydrology, working on the understanding and forecasting of flash floods, flooding and low-water conditions.
– Bob Saint-Fleur: PhD in hydrology, research engineer, working on hydrological modelling using AI.
– Florian Surmont: PhD in engineering sciences, data scientist, research engineer specialising in time series modelling using AI.
– Pierre NICOLLE: research engineer in hydrology and hydrogeology, modelling and forecasting, involved in the development of the Prémhyce national low-water forecasting platform.
Site web :
Intitulé du doctorat
PhD in Earth and Planetary Sciences
Pays d’obtention du doctorat
France
Etablissement délivrant le doctorat
UNIVERSITE DE NANTES
Ecole doctorale
Matière, molécules, matériaux – 3M
Profil du candidat
Requirements
Research Field – Hydrology and Artificial Intelligence
Education Level – Master’s Degree
Skills/Qualifications
– Solid experience with computer programming / scripting (Python)
– First experience in data anlysis and machine learning modelling is necessary.
– Sound and quantitative understanding of hydrology and/or hydrogeology
– Applicants must be proficient in both written and oral English
– Applicants must be able to work independently and in a structured manner and demonstrate good collaborative and communication skills.Date limite de candidature
31/10/2024
