Application of machine learning techniques to weather forecasting
MetadataShow full item record
Weather forecasting is, still today, a human based activity. Although computer simulations play a major role in modelling the state and evolution of the atmosphere, there is a lack of methodologies to automate the interpretation of the information generated by these models. This doctoral thesis explores the use of machine learning methodologies to solve specific problems in meteorology and particularly focuses on the exploration of methodologies to improve the accuracy of numerical weather prediction models using machine learning. The work presented in this manuscript contains two different approaches using machine learning. In the first part, classical methodologies, such as multivariate non-parametric regression and binary trees are explored to perform regression on meteorological data. In this first part, we particularly focus on forecasting wind, where the circular nature of this variable opens interesting challenges for classic machine learning algorithms and techniques. The second part of this thesis, explores the analysis of weather data as a generic structured prediction problem using deep neural networks. Neural networks, such as convolutional and recurrent networks provide a method for capturing the spatial and temporal structure inherent in weather prediction models. This part explores the potential of deep convolutional neural networks in solving difficult problems in meteorology, such as modelling precipitation from basic numerical model fields. The research performed during the completion of this thesis demonstrates that collaboration between the machine learning and meteorology research communities is mutually beneficial and leads to advances in both disciplines. Weather forecasting models and observational data represent unique examples of large (petabytes), structured and high-quality data sets, that the machine learning community demands for developing the next generation of scalable algorithms.