End-to-end Machine Learning quantification of hydrological uncertainties: from climate to flood risk management

New advances in Machine Learning (ML) using deep learning strategies are enabling the creation of so-called emulators to efficiently propagate uncertainties across chains of models. In addition, recent steps incorporating ML in climate and hydrology modelling allow the representation of key processes (e.g. precipitation, runoff) with high fidelity at a low computational cost.

The unique modelling of the entire hydrological processes over Japan at our partner RIKEN-CCS in Kobe provides an opportunity to combine, utilise, and enhance the ML research at UCL to this context. Uncertainties are crucial to model, and stem from various sources: climate scenarios, precipitation patterns and extremes, runoff, and the risk appetite and constraints of decision-makers. Novel investigations can thus be undertaken to quantify uncertainty and shape decisions about the dimensioning against flood disasters, the operation of sustainable water supplies, as well as dam electricity production.

The project will seek to make progress in creating more efficient ML techniques to tailor the climate predictions and the uncertainty representation to a specific area of risk management. It will also allow the creation of decision tools for disaster mitigation, and infrastructure optimisation.

Our partner RIKEN-CCS in Kobe, Japan has created a granular model of hydrology over the entire country of Japan accounting for snow and extreme events (typhoons) in addition to regular precipitation. The model incorporates runoff and river flows and can function in real time. The collaboration consists of including a layer of climate modelling at the top, with ML helping with the computational costs, and making the entire chain of modelling able to propagate uncertainties to the decision-making step. Travel to RIKEN-CCS in Japan will help the student in these tasks.

In recent years, ML models have been shown to provide accurate, conceptually simple and computationally efficient modelling of hydrological processes. However, the translation of these improvements into actionable insights for water management operation planning remains largely unchartered territory. The inclusion of uncertainties in these ML models requires additional investigations, e.g. the use and extension of Deep Gaussian Processes. The student will seek to make progress in creating more efficient ML techniques to tailor the climate predictions and the uncertainty representation to this area of risk management.

The main research questions will relate to optimising certain objectives under uncertainties through the hydrological chain. In practical terms, dimensioning of water management infrastructure includes (non-exhaustively), elevating dam walls to increase storage capacity, levee infrastructure management, river bed engineering while infrastructure operation refers to dam operation and operation of early flood warning systems. The focus is still open to the interests and skills of the student.