Deep Learning for clustering of weather and atmospheric dispersion features

Summary

Dimensionality reduction alleviates computational burdens faced in various tasks from design optimization to predictive modelling. Hence, popular dimensionality reduction approaches aim at projecting the data onto a subspace spanned by linear and non-linear functions obtained from the compression of a dataset of solution snapshots. This thesis approaches the problem of obtaining meaningful low-dimensionality representations of high-dimensional data, referred to as dimensionality reduction, by the design and application of suitable deep neural network configurations. Moreover, in the applications targeted by this dissertation it is crucial for the algorithms to learn the underlying data manifold in order to facilitate further processing. The novelty of the proposed technique is to design auto encoders that make use of weather variables and temporality, in order to learn enhanced data representations for weather data clustering. We examine a plethora of end-to-end trainable models improving clustering procedure on multiple years of weather data in the European continent. A number of real-world applications could benefit from such models, for instance anomaly detection in renewable energy sources utilizing weather features from wind- flow and solar energy in order to develop monitoring systems. Furthermore, scenarios including environment and climate change could be developed for atmospheric inverse analysis systems to estimate emissions or gas flux sources. Our knowledge in meteorology for agricultural purposes could be enhanced from such frameworks by collecting dominant meteorological patterns. Among the various ways of learning representations, this thesis focuses on deep learning methods: those that are formed by the composition of multiple non-linear transformations, with the goal of yielding more abstract – and ultimately more useful –representations. We present alternative non-linear representation learning methods, such as Convolutional and LSTM (Long Short Term Memory) networks and discuss advantages and disadvantages, focusing on extracting reusable and robust deep features. These methods are known to be able to disentangle the data manifolds, i.e., to uncover the latent dimension(s) along which the temporal and/or spatiotemporal aspect of the data. However, their application on the clustering of weather data has not been sufficiently studied before. The models are evaluated and compared to alternative methods demonstrating competitive results. They are shown to map input data into a new space where the data are linear separable. Hence, the extracted representations can be applicable in a plethora of domains, such as in discovering robust weather patterns, developing emergency response applications in atmospheric physics, radiology, environmental research, healthcare and in other application areas.