Climate and weather computer models attempt to predict the evolution of the atmosphere/ocean filtered to the Spatio-temporal resolution of the model. Such filtering is essential to make the predictions computationally affordable. Data Assimilation is the science of using models and observations of the atmosphere/land/ocean/ice/waves to estimate the actual evolution of the filtered atmospheric/oceanic state and the model error. Data assimilation provides the initial conditions for weather forecasts for a few hours to a few years ahead. In the last few decades, most of the biggest improvements in weather forecasting accuracy have been due to improvements in data assimilation that have reduced the error in the initial conditions used for making forecasts.
Because both models and observations are imperfect, there are fundamental limitations on our knowledge of the state of the Earth’s ocean and atmosphere in the past, present and future. In Ensemble Data Assimilation and Forecasting, instead of making a single state estimate, one makes a set or ensemble of state estimates. The aim is to create an ensemble that samples the probability distribution of possible past, current, and future states by rigorously accounting for the uncertainty in observations, models and the imperfections of the data assimilation scheme itself.
The data assimilation and ensemble forecasting systems currently used at the world’s major forecasting centres are far from optimal. They are highly sub-optimal in their treatment of trace-gas concentrations, aerosols, clouds, precipitation, and ice. Errors in model representations of these variables are known to be the major source of error in weather and climate forecasts; hence, the need for progress in these areas is urgent. Our group is striving to develop new, computationally affordable data assimilation and ensemble forecasting methods that will overcome these limitations. We are also developing new statistical, machine learning approaches that can remove predictable elements of the errors in ensemble forecasts.
Renewable energy from wind turbines and photovoltaic cells is highly weather dependent. To effectively manage electricity grids that use renewable energy sources, predictions of the wind and sunshine are required for a few minutes to a few years ahead. Ideally, these forecasts are in the form of probability distributions so that the electricity grid managers can better manage weather forecast uncertainty. As atmospheric scientists, we strive to reduce forecast uncertainty while accurately quantifying it. In one project, we are using numerical weather models and remote sensing tools to dynamically observe the wind field just before it reaches a wind farm, to give advance warning of changes of wind speed on timescales as short as 5 minutes and up to 24 hours. We are also studying unique night-time wind conditions and how they influence the hub-height wind speed and power output at a wind farm in Southeast Australia. In another project, we are developing advanced statistical techniques to improve probabilistic forecasts of conditions 3 weeks ahead.