Reduced Dimension Models for Weather Forecasting

Abstract

Weather forecasting is one of the most computationally intensive activities that is routinely undertaken. This project studies the possibility of reducing the dimensions of the models and data sets considered, while maintaining reasonably good predictions. Techniques dealing with the two problems separately, dimension reduction and forecasting, are applied on real data provided by the European Center for Medium-Range Weather Forecasting, and on an artificial data set generated using the Lorenz equations. A new algorithm is presented as an extension of the principal interaction patterns framework to neural networks, allowing a simultaneous optimization of the subspace basis for the data projection and the model considered to make predictions. Advantages and drawbacks of those methods are discussed, and conclusions are drawn from this study regarding the feasibility of reducing the dimensions in the forecasting problem.