FUNCTIONAL MACHINE LEARNING WITH APPLICATION TO DAY-AHEAD FORECASTING OF NATURAL GAS FLOW DATA

Abstract

In this thesis, we propose Long Short-Term Memory (LSTM) model for forecasting functional time series for which the underlying data are defined as functions in continuous space. The functional data can be decomposed based on either fixed parametric basis functions through the Fourier expansion or data-based basis functions through functional principal component analysis (FPCA). In our proposed model, the LSTM is implemented based on the Fourier expansion coefficients or principal components scores of the functional data. The functional machine learning techniques are applied to real-world data, the hourly natural gas flow in Germany and compared with two alternative models – Persistence model and the LSTM directly applied to the original discrete data. We found that the Fourier based LSTM has slightly better prediction accuracy compared with the Persistence model and the original data based LSTM model. However, the FPCA based LSTM does not perform well in our case.