Hluboké učení pro predikci časových řad v energetice

Abstract

This master's thesis focuses on short-term hourly electricity consumption forecasting using deep learning methods. The work summarises the theory of time series, classical statistical models and modern neural network architectures ranging from the multilayer perceptron and one-dimensional convolutional network through the recurrent LSTM network to the Transformer model. On hourly consumption of 27 European countries over the period 2015-2024, linear regression as a baseline is experimentally compared with four deep learning models within a unified processing pipeline. Through a hyperparameter grid search (scaler type, look-back window length and architectural parameters), more than 8 000 configurations were trained on two NVIDIA RTX 4090 GPUs. For the 24-hour horizon the LSTM achieves the lowest MAPE of 6.53 %, followed by the Transformer (6.83 %), the convolutional network (10.33 %), the multilayer perceptron (10.36 %) and linear regression (12.66 %). The thesis further quantifies the impact of scaler choice, look-back window length and regularisation, identifies Montenegro as the most difficult country and shows that, thanks to a country embedding, the Transformer maintains the most balanced performance across countries. The contribution of the thesis is a unified experimental comparison of five architectures on a large multi-country dataset together with an analysis of model robustness against varying consumption volatility.