Background

Bad weather can be far more than just an inconvenience. In NZ, where primary production contributes $22.5B to our economy, bad weather can incur significant economic, environmental and social costs. Mitigating the impacts of severe weather largely depends on the quality and reliability of weather forecasts.

The most highly damaging extreme weather events (e.g. hail or intense rainfall) often occur over small areas, driving a need for higher-spatial-resolution forecasts. By solving the mathematical equations describing atmospheric processes, numerical weather prediction (NWP) models predict how the weather will change - this computationally demanding task requires supercomputers. Increasing the model resolution, so that they resolve local weather events, adds large financial costs.

We will apply artificial intelligence methods to develop a new way of generating weather forecasts, producing high-resolution forecasts at a fraction of current costs.

Research Aims

A neural network (NN) will be trained to learn how to generate weather at hyperlocal scales (several 100m) given data from a lower resolution NWP model. While the initial training may be computationally expensive, once trained, the NN can be applied to any NWP forecast to fill in the missing detail inside each grid-cell, at negligible cost. This cost reduction means that we can generate higher resolution forecasts than are currently available, and process many more forecasts to produce probabilistic risk assessments of rare but highly damaging events.

If successful, our fused-NN-NWP model will be incorporated into MetService’s NWP chain, delivering new hyperlocal weather forecasts, enhancing the ability of emergency managers to save lives and protect property, and industries to manage risks and minimise losses. The need for such forecasts will only increase as the frequency and severity of extreme weather events increase under climate change.