Towards self-tuned physics-informed dam monitoring system: the case of hydraulic phenomena

CONTEXT: Motivated by obvious safety reasons, monitoring hydraulic dams is a major concern for all hydroelectricity providers worldwide, and for this reason a lot of efforts have been dedicated to the development of algorithmic tools to that end since several years now. Among them, so-called HST and HTT statistical models are currently the most used ones, basically relating hydrostatic effect, ageing, and possibly temperature, to measured displacements [1]. The parameters of those models are generally determined on the basis of previous measurements (model calibration period), with no automatic updating. Any new measurement is then compared with the model forecast, and if a discrepancy is detected, a specific analysis has to be carried out to understand its origin. With the increased expansion of machine learning, many studies more and more focus on the use of such a tool, pretty efficient in matching HST/HTT models with data [2], but at the expense of physical insights. Another limitation in the usual practice is that most of the time, methods are proposed phenomenon by phenomenon, without accounting for possible couplings, nor multi-sensing facilities (which may yet be available).

In the specific case of EDF group, the Hydroelectric branch is, in a similar fashion, in charge of continuously monitoring specific measurements on dams, which are analysed in order to detect any behavioural anomaly by comparison with previous recordings. The approaches are thus facing same limitations as highlighted before. For this reason, and within the context of the long-term collaborative project which has recently started with Gipsa-lab about digital twins in hydropower production, EDF is a stakeholder in the present PHD topic.

PHD OBJECTIVES AND WORK: The main goal of the PHD is to address some of the limitations listed before in dam monitoring, and enhance monitoring systems accordingly, taking advantage of a system viewpoint [3]. More particularly, it is proposed to focus on the problem of analysing hydraulic phenomena in the body of dams or in their foundations. This issue is a key one in the analysis of monitoring data, and is common to all the types of dams which can be found (concrete gravity and arch dams, embankment dams). Many of such structures are equipped with piezometers or pressure cells distributed linearly from upstream to downstream, allowing to measure under pressure profiles and detect any change in hydraulic conditions.

The primary goal of this PHD is then to develop a 1D numerical model of the flow in the most critical zone of the structure at the interface with the foundation, usable for dynamical monitoring purposes.

This 1D model would be made up of sections of variable permeability, with values to be adjusted using available pointwise measurements. Boundary conditions would be given by upstream and downstream elevations, as well as drainage effects where there exist. The development of the model could also be connected with some more complex finite element modelling, and potentially take into account the delay effect observed in embankment dams.

The automatic adjustment of model parameters based on regular data acquisition is then to be constructed, so that analysis of possible changes can thus provide a direct physical explanation. This raises scientific issues related to parameter/state estimation in distributed parameter systems [4] (here mainly of a convection-diffusion nature), which could benefit from results for hyperbolic case [5], or when coupling Partial Differential Equations with Ordinary ones [6] for instance.

The study is intended to take advantage of EDF databases, for calibration of the model with real data, and subsequent investigations towards monitoring application.

Intermediate results should give rise to communication and publication proposals, and the overall work will ultimately be reported in a final thesis.

References

[1] B. Li, J. Yang, D. Hu, Dam monitoring data analysis methods: A literature review, Struct. Control Health Monitoring, 2020.

[2] J. Mwanza, P. Mashumba, A. Telukdarie, A Framework for Monitoring Stability of Tailings Dams in Realtime Using Digital Twin Simulation and Machine Learning, Procedia Computer Science 232, 2024.

[3] M. Blanke , M. Kinnaert , J. Lunze , M. Staroswiecki, Diagnosis and Fault-Tolerant Control, Springer 3^rd Ed., 2016.

[4] G. Evensen, Femke C. Vossepoel, P. J. van Leeuwen, Data Assimilation Fundamentals - A Unified Formulation of the State and Parameter Estimation Problem, Springer 2022.

[5] V.T. Nguyen, D. Georges, G. Besancon, State and parameter estimation in 1-D hyperbolic PDEs based on an adjoint method, Automatica, vol.67, pp.185-191, 2016.

[6] M. Mishra, G. Besançon, G. Chambon and L. Baillet, Observer design for state and parameter estimation in a landslide model, 21rst IFAC World Congress, Berlin, Germany, July 2020.

The PHD will take place at Gipsa-lab, as the University research center host (within Univ Grenoble Alpes, in Grenoble), in collaboration with EDF group.

Gipsa-lab is one of the major academic research groups in Systems and Signals in France, with about 400 people, including around 40% of PHD students (https://www.gipsa-lab.grenoble-inp.fr/en).

EDF is one of the leading groups worldwide in electricty production, transmission and distribution, supplying energy and related services to nearly 40 million client sites around the world (https://www.edf.fr/en/the-edf-group).

With a Master degree, or equivalent, candidates should have a strong background in systems and signals, with skills in physical modelling, data and learning, as well as simulation. Knowledge and/or experience in mechanical / hydraulic / civil engineering would be more than welcome. Skills in Matlab are also requested, as well as good communication abilities.