The aim of this study is to predict the thermal behavior of a low-energy building presented in the average indoor temperature using the classical regression methods such as the Recursive Least Square (RLS), the ordinary least square (OLS) and a machine learning model which is the Gated Recurrent Unit (GRU) using as input the outdoor temperature and the energy consumption of electric heater. Based on the evaluation metrics, all the methods have demonstrated a high precision which is closed to be 100%. On the other hand, based on the distribution of errors GRU presents a narrower distribution and a higher concentration of errors around zero, with errors mainly between −0.5 and 0.5, while the residuals obtained using the OLS method vary from −0.3 to 0.3, but with longer tails and increased variability which demonstrates the ability of GRU to study the variability existing in energy consumption time series.