Trabajamos con simulación avanzada para afianzar la seguridad nuclear.

La Cátedra UNESCO en Métodos Numéricos lidera la innovación de vanguardia en el Sur global.

PhD Thesis Defense – «Evaluation of potential hazard due to off-stream reservoir failure using Machine Learning techniques» by Nathalia Silva Cancino

20240404
}
00:00
CIMNE

📆 Friday, October 11, 2024

🕐 11:30 am CET

📍 ETSECCPB, C/Jordi Girona 1-3, building C1, room 002, Campus North, Barcelona

ABSTRACT

Hazard classification for dams and off-stream reservoirs, which entails identifying potential damages in the event of structural failure, is a crucial tool for implementing local-level risk reduction plans. Consequently, national administrations have developed guidelines including suggested methodologies and tools.The modification of the Spanish Regulation of the Public Hydraulic Domain (Royal Decree 849/1986, of April 11), carried out through Royal Decree 9/2008, of January 11, obliges owners of off-stream reservoirs with a height of 5 meters or a capacity higher than 100,000 m3, whether public or private, to develop a classification study based on the potential risk of their failure (Articles 356 and 367). This represented a significant paradigm shift, under the former regulation, such a study was only required for dams exceeding 15 m in height or those with a height between 10 and 15 m and a capacity exceeding 100,000 m3.The procedure for this classification is time and resource-consuming, and in the specific case of owners of off-stream reservoirs, they may not have these assets. Therefore, this research proposes a simplified methodology to classify off-stream reservoirs, utilizing a surrogate Machine Learning (ML) model that is simpler and has a lower computational cost than conventional approaches. Additionally, the influence of two sources of uncertainties on hazard classification is analysed. This research is based on the generation of synthetic data. A specialized tool in Iber was developed to generate massive 2D hydraulic models of synthetic off-stream reservoir failures, which make all the processes of construction, calculation and extraction of results automatic.The first analysis was focused on the effect of selecting a breach parametric model on the hydraulic variables, the potential damages, and the hazard classification of the structures. Three common parametric models were compared, using a set of synthetic cases and a real off-stream reservoir. The results highlighted that the choice of the model has significant effects. Notably, the erodibility of the material exerts a high influence, surpassing that of the failure mode. The use of an inappropriate model or a lack of information regarding dike material can lead to overly conservative or underestimated results, consequently affecting hazard classification.The ML model constructed for the simplified methodology was a Random Forest classifier capable of identifying potential damages at any point in the vicinity of an off-stream reservoir. This ML model was trained using synthetic data, offering an estimation of potential damages by considering the physical characteristics of the structure, the surrounding terrain, and vulnerable areas. During a real case application, the simplified methodology achieved an accuracy rate of 91%. The simplified methodology allows owners and administration to obtain a pre-classification without the need to make a 2D hydraulic model, which saves time and money.Furthermore, an interface called ACROPOLIS was developed, integrating the ML model. Users can apply the simplified methodology through ACROPOLIS, which guides them step by step, providing the overall classification of the off-stream reservoir based on Spanish regulations.Finally, the analysis of uncertainties related to breach formation and the location of the breaking points in reservoirs was integrated. This involved comparing the current deterministic approach for hazard classification and a newly proposed fourth-step probabilistic approach that accounts for uncertainties in constructing hazard maps. The study revealed variations in classifications between scenarios, as different breaking points and breach formations generate diverse classifications that can affect emergency plans. Additionally, the proposed visualization can be used for various purposes, including tracking the evolution of categorization over time due to land use changes.

Committee

PHD CANDIDATE

Mr. Sergio JiménezMs. Nathalia Silva Cancino is a civil Engineer with a strong background in water resources, hydrology, and hydraulics. She has a wide experience in water management sector, working on research of the behaviour of the most important rivers in Colombia and focusing her research experience on the analysis of hydraulic structures under climate change impacts. PhD student in Civil Engineering (UPC), MsC in Water Science and Engineering (IHE Delft Institute for Water Education), BsC in Pontificia Universidad Javeriana (Colombia). She is part of the Machine Learning in Civil Engineering RTD group at CIMNE, part of the Machine Learning and models in Hydro-Environmental Engineering research cluster.

Añadir al calendario:

Compartir: