Předvídání v časových řadách

Abstract

This thesis addresses short-term forecasting of hourly natural gas consumption in the Czech Republic under non-stationary time-series conditions with structural changes. The theoretical part summarizes the main concepts of time series, deep learning methods, and foundation models for time series. The practical part focuses on the construction and preprocessing of a custom multivariate dataset and on the experimental comparison of selected forecasting approaches. Special attention is also paid to the impact of changing data conditions and structural changes after 2022 on the forecasting task.