ENAC Seminar Series by Dr J. P. Matos
Event details
Date | 18.07.2019 |
Hour | 13:00 › 14:00 |
Speaker | Dr José Pedro Matos |
Location | |
Category | Conferences - Seminars |
13:00 – 14:00 – Dr José Pedro Matos
Consultant for the hydropower sector, Stucky SA, Switzerland
Using machine learning and efficient computing to model uncertainty in hydraulic infrastructures
Uncertainty should be at the core of decision-making, particularly so when addressing critical assets such as hydraulic infrastructures. Notwithstanding, uncertainty is often hard to quantify, which has led generations of engineers to shy away from explicitly evaluating it. Recent developments in machine learning techniques and platforms for efficient computing constitute extremely valuable tools in the quantification of uncertainty. Exploiting these developments, the presentation will highlight two applications where uncertainty plays a central role.
The first application lies at an intersection between machine learning and multi-objective optimization. It explores a breakthrough algorithm capable of making reliable probabilistic predictions based on observed data. Notions of probabilistic forecasting will be introduced and several examples, illustrating the broad applicability of the model, will be discussed. They include the operational flood forecasting model employed at the Rogun dam (to be the highest in the world), an inflow forecasting system for the Kariba reservoir (the largest in the world by volume), the estimation of suspended sediment concentrations on the Upper Yangtze River, and the prediction of the euro-dollar exchange rate (with bittersweet results).
The second application addresses the risk associated with large dams. It provides a framework capable of estimating risk while accounting for the most important sources of uncertainty, be it aleatoric or epistemic. It can reproduce the complex chains of events that may lead to failure and evaluate losses when a failure does occur. Through efficient computing, the most likely paths to failure are found in a dynamic simulation including interactions between hazards, dam components and the reservoir. For each simulated failure, multiple dam-break waves are generated and propagated downstream, acting on built infrastructure in non-deterministic ways and culminating on individualized loss of life estimates.
Consultant for the hydropower sector, Stucky SA, Switzerland
Using machine learning and efficient computing to model uncertainty in hydraulic infrastructures
Uncertainty should be at the core of decision-making, particularly so when addressing critical assets such as hydraulic infrastructures. Notwithstanding, uncertainty is often hard to quantify, which has led generations of engineers to shy away from explicitly evaluating it. Recent developments in machine learning techniques and platforms for efficient computing constitute extremely valuable tools in the quantification of uncertainty. Exploiting these developments, the presentation will highlight two applications where uncertainty plays a central role.
The first application lies at an intersection between machine learning and multi-objective optimization. It explores a breakthrough algorithm capable of making reliable probabilistic predictions based on observed data. Notions of probabilistic forecasting will be introduced and several examples, illustrating the broad applicability of the model, will be discussed. They include the operational flood forecasting model employed at the Rogun dam (to be the highest in the world), an inflow forecasting system for the Kariba reservoir (the largest in the world by volume), the estimation of suspended sediment concentrations on the Upper Yangtze River, and the prediction of the euro-dollar exchange rate (with bittersweet results).
The second application addresses the risk associated with large dams. It provides a framework capable of estimating risk while accounting for the most important sources of uncertainty, be it aleatoric or epistemic. It can reproduce the complex chains of events that may lead to failure and evaluate losses when a failure does occur. Through efficient computing, the most likely paths to failure are found in a dynamic simulation including interactions between hazards, dam components and the reservoir. For each simulated failure, multiple dam-break waves are generated and propagated downstream, acting on built infrastructure in non-deterministic ways and culminating on individualized loss of life estimates.
Practical information
- General public
- Free
Organizer
- ENAC
Contact
- Cristina Perez