Long-range temperature forecasting correction techniques Using machine learning

Abstract

We present solutions to four problems emerging in data-driven long-range weather prediction that were explored as part of an M.Eng Thesis. These problems are related to long-range prediction using a network of observing stations and climate indicators. The first problem relates to the correction of phase error in long-term temperature forecasts. The second problem involves the task of using correlated observed and proxy signals to update each other to improve forecasting accuracy. The third problem relates to the use of deep learning in the problem of predicting the future value of near oscillators. The fourth problem relates to the discovery of new, finer scale oscillation signals using Representation Learning based Dimensionality Reduction techniques. Together, our proposed solutions enable the use of inference and learning for data-driven long-range weather forecasting using context from the global climate system.