APPLICATION OF DEEP ESN NETWORKS AND WAVELET DECOMPOSITION FOR ENERGY CONSUMPTION FORECASTING IN BUILDINGS

Abstract

The aim of this study is to investigate the application of deep Echo State Networks (DeepESN) combined with wavelet decomposition for short-term forecasting of daily natural gas consumption in a building. The proposed approach includes extracting the temperature-related component using linear regression, wavelet decomposition of the residual signal, and independent forecasting of the resulting components using NAR, ESN, and DeepESN models. The experimental study was conducted using real data on natural gas consumption and outdoor air temperature from eleven heating seasons. The last season was used exclusively as an independent test set. Forecasting accuracy was evaluated using RMSE and MAE. The results showed that, among the investigated models, DeepESN provided the best forecasting performance for the complete time series, achieving an RMSE of 92.17 m³/day and an MAE of 55.52 m³/day. The results demonstrate the applicability of combining wavelet decomposition with DeepESN for short-term natural gas consumption forecasting.