I have up to two funded studentships available for 2024/25:
1. Forecast Uncertainty on Days to Decades: A Case Study with the National Grid
The National Grid ESO uses weather and climate data on timescales of days to decades to understand risks in their network. The key risks are days when there is a mismatch between supply and demand, with consequences in the case of either too much or too little supply: the grid must be balanced. Both supply and demand are a function of weather conditions. With increasing reliance on renewable energy, supply is a function of wind speed and solar production, while demand is a function of temperature.
The ultimate goal of this project is to develop a theoretical framework whereby we can make best use of all available forecast data from days to many decades. Key to the problem is quantifying uncertainties, and understanding where they come from.
For example, when it comes to short-range forecasts, we have data at the km scale, providing us with the detail needed to predict supply and demand. But on seasonal to climate timescales models have relatively low resolution. This impacts how accurately we know wind speed and solar information - yet we need this on the scale of the wind and solar farms. How can we use detail from weather forecasts to give us information needed on seasonal and climate timescales? One possibility is to use novel ideas from machine learning (ML) to 'fill in the gaps'. Can we train a ML emulator to (probabilistically) increase the resolution of climate model projections towards that of weather forecasts? Further work could consider the relative role of climate model biases and representativity in decision making, for example making use of initialised seasonal-to-decadal forecasts which show increasing bias with lead time. Only with an improved understanding of forecast uncertainties and their sources can we hope to reduce these uncertainties, directly impacting users of weather and climate data.
This project will be in collaboration with the National Grid ESO. For more information on the National Grid ESO, including an example of their use of forecast data, please see:
National Grid ESO Control Centre
National Grid ESO Winter Outlook
This DPhil project is funded as part of the Seamless Uncertainty Quantification for Earth System prediction (SUQCES) project, funded by the Leverhulme Trust
2. Improving probabilistic weather and climate forecasts using observational constraints
One of the foundational principles of the scientific method is the need to assess and quantify uncertainty. Weather and climate prediction are not exempt from this requirement, and forecasts should include an estimate of the uncertainty in the forecast.
Making a weather or climate forecast involves combining information about the current state of the Earth-system with a computer model which represents the equations of motion describing the system. The computer model propagates the state of the Earth-system forward into the future, while making any necessary assumptions about future forcing (e.g. greenhouse gas concentrations). Potential errors in the starting conditions, unknowable future forcing, and approximations made when building the computer model of the Earth-system can introduce uncertainty into the prediction. These uncertainties must be accounted for to produce the reliable probabilistic forecasts essential for policy makers, industry and the humanitarian sector.
This project focuses on the uncertainty introduced into the forecast due to limitations in the forecast model. This model uncertainty is a substantial source of uncertainty on both weather and climate timescales, so it is imperative that it is correctly accounted for in forecasts. However, despite its importance, there is no consensus on how to best account for model uncertainty. At best, model uncertainty is represented in a heuristic or ad hoc manner, while at worst it is ignored completely. There is also a disconnect between the weather and climate communities, with different approaches used for the same model on different timescales.
This project seeks to provide a new understanding of the nature of model error and how to best represent this as a source of uncertainty in weather and climate models. The student will do this by making use of the large database of forecasts archived at ECMWF. The database includes operational forecasts made at a range of resolutions (i.e. using different pixel sizes in the atmosphere and ocean). Through comparing the forecasts with observations at short lead times (a day or less) we can measure the error in the forecast, and seek to understand the statistics of this error. This project would pioneer combining model forecasts with observations in this way, allowing us to characterise the uncertainty in our forecast model.
Exploration and assessment of the database will allow us to characterise the nature of model error. Are current techniques for representing model uncertainty consistent with these characteristics? Can we rule out some techniques as not fit for purpose? Can we identify new approaches which better represent the statistics of model error? How does changing the degree of pixilation in our forecast model affect error growth in the model? Understanding the answer to these questions will enable us to improve the fidelity of our weather and climate models
This DPhil project is funded as part of the Seamless Uncertainty Quantification for Earth System prediction (SUQCES) project, funded by the Leverhulme Trust