Energy efficiency is becoming more important issue, especially in buildings sector where 15–20% of total energy in developed countries is used. One course toward higher efficiency is to improve HVAC (Heating, Ventilation and Air Conditioning) systems in building, especially by implementation of advanced control methods. Efficiency of these control methods (especially Model Predictive Control) significantly depends on quality of thermal models, where quality can be defined by accuracy and complexity. This paper investigates relationships between accuracy and complexity in thermal models produced with RC (Resistance-Capacitance) equivalent method, with parameters improved using optimization and estimation from data. Four experiments using simulation data regarding importance of selection of error function, selection of training data, selection of model complexity and reduction of complexity are conducted and explained.