Hluboké učení pro predikci časových řad

Abstract

This bachelor’s thesis focuses on the application of deep learning methods for time series forecasting, specifically for modeling natural gas consumption. The aim of the thesis is to analyze different neural network architectures, implement selected models, and experimentally evaluate their predictive performance on real-world data. The theoretical part introduces the fundamentals of deep learning and time series analysis, with an emphasis on methods suitable for sequential data processing. The practical part covers data preprocessing, model design, and experimental evaluation. Selected architectures were implemented and compared, including Multilayer Perceptron (MLP), Convolutional Neural Networks (CNN), and Long Short-Term Memory (LSTM) networks. The experiments focus on analyzing the impact of input features, historical window size, and prediction horizon on model accuracy. The results show that LSTM models achieve the best performance due to their ability to capture long-term dependencies in the data. Furthermore, the study demonstrates that appropriate feature selection plays a crucial role in prediction accuracy, while redundant features may negatively affect model performance.