Uncertainty quantification and sensitivity analysis for resilient infrastructure systems

Lead Research Organisation: University of Bristol
Department Name: Civil Engineering

Abstract

Computational modelling provides a vital tool to support infrastructure decisions, allowing to evaluate risks and benefits of different infrastructure options on a virtual system (or "digital twin") before committing to a particular design. Model outputs though are conditional on a range of uncertain assumptions and input data, due to our incomplete or imperfect knowledge of the drivers and the properties of the system being modelled. When models are used for long-term planning, the uncertainty about the current properties and drivers of the system is compounded with further uncertainty about how these will evolve in the future. For example, when planning water infrastructure for drought resilience, we need to make a set of uncertain assumptions about the way that future climate will affect water sources and how changes in the economy, society and lifestyle will affect future water demand. Another example is the energy sector, where the growing contribution of intermittent sources such as wind and solar introduces unprecedented levels of uncertainty to the quantification of energy production potential, both in the short and long term.

Overconfidence in model results and insufficient consideration of the breath of possible futures is a key obstacle to infrastructure resilience. If models are used to inform large investment decisions, they must be trustworthy and defensible. Decision-makers need to be made aware of the uncertainties affecting model output(s) and the critical assumptions that define the scope of validity of the model. Also, as uncertainty about the future is irreducible, modelling for resilience should aim at identifying designs that achieve an acceptable performance across a wide range of future scenarios, rather than designs that are optimal under any particular scenario.

Uncertainty Quantification and Sensitivity Analysis (UQ&SA) is a set of generic methods that can be used to analyse the propagation of uncertainties in model and thus improve the model's construction, validation, and use for decision-making under uncertainty. UQ&SA are "model-agnostic" methodologies, meaning that they are applicable to any mathematical model regardless of the specific application domain. The goal of this project is to set the foundations for integrating UQ&SA functionalities in the DAFNI platform. We believe this is very important for DAFNI to become a platform that not only enables users to share, combine and execute models, but also enables and promotes best practices for responsible modelling.

To achieve the project goal, we will develop DAFNI Workflows that use functionalities already existing in DAFNI for uncertainty propagation (the "Loop" functionality, which allows repeated executions of the same model against different input set-ups) and integrate them with existing open-source software packages for UQ&SA. In these Workflows, we will use two simple "proof-of-principle" models from the water and energy sector, so to ensure feasibility of the project within the limited timeframe, but also to produce training materials for current and future DAFNI users to learn "by example". Indeed throughout the project we will run a series of seminars and tutorial sessions on UQ&SA for early-career researchers, and develop recommendations for the DAFNI technical team on future developments needed in order to scale-up the applicability of UQ&SA to more complex models.

Publications

10 25 50