Exploring Complexity in Meteorological Data

Enhancing Weather Forecasts Through Deep Learning-Based Post-Processing

Abstract

Weather forecasting plays an important role in all of our lives. It impacts everything from small decisions, like how to dress for the day, to great decisions, like which measures to take in preparation for an extreme weather event. And since they are the basis of so many decisions, it is crucial that the forecasts provided to the public are reliable and accurate. The task of modelling the weather is highly complex, as the models need to represent everything from local phenomena to global systems, over several time scales. And because these weather models are optimized for overall performance, the forecasts generally go through different types of post-processing – a sort of fine-tuning before reaching the public. Traditionally, this post-processing has consisted of modelling the statistical relationship between model output and observations, using traditional statistical methods. However, these relationships can be complex, and this, in combination with the availability of large data sets has made researchers turn their attention to more modern approaches, from the world of computer science. This thesis seeks to answer the question: How can we best exploit the potential that lies in the complex nature of meteorological data, using Deep Learning? The work explores different levels of complexity of deep neural networks, the benefits of combining heterogeneous input data in a multimodal neural network, and the value of providing ensembles of weather forecasts to a neural network, all with a view to identify the best ways of extracting the potential within the data. Two novel model architectures are presented in the thesis: the tower network and the ensemble conditional Generative Adversarial Network (cGAN). The tower network is used in the exploration of multimodality, as well as in an investigation into the importance of complexity in a neural network, in comparison with simpler and more complex networks. The ensemble cGAN is an extension of the cGAN, where ensembles of weather forecasts are provided as input. This architecture is compared to a traditional cGAN, and various configurations are tested and compared. The thesis finds that the combination of data sources in a multimodal network provides great value beyond the use of a single data source. However, increasing the complexity of the network does not necessarily translate into improved performance. The exploration of ensemble data used in the training and testing of neural networks reveals that there is value to be gained from the ensembles. The effect we observed was limited, but there are indications that more attention should be given to this research question in future work.