Předvídání v časových řadách

Abstract

This master thesis addresses the problem of forecasting energy consumption time series using deep learning architectures. The primary objective is to compare the performance of LSTM, GRU, and one-dimensional convolution architectures in predicting the consumption of two distinct commodities, namely natural gas and electricity. The theoretical part defines the specifics of time series and the operational principles of the selected neural networks. The practical part implements an extensive experiment in a long-term recursive forecasting mode, testing various model configurations, including the impact of Batch Normalization and Dropout. The thesis analyzes how forecasting accuracy depends on the nature of the predicted variable and the influence of external meteorological data, taking into account their spatial correlation with the consumption area. The results demonstrate that there is no universal configuration suitable for both types of energy series and identify key factors for maintaining model stability in long-term recursion.