To increase the share of renewable energy sources in the energy mix, buildings must provide energy flexibility to the grid. Model Predictive Control (MPC) has proven to be instrumental in activating the energy flexibility of buildings. MPC requires a predictive model to capture the thermal behavior of a building. These models require operational data from the building. The quality of these datasets affects the predictive model’s performance and, in turn, the energy flexibility. A dataset could be obtained by exciting the building using conventional controllers or randomly generated signals. Using a random excitation causes thermal discomfort for the occupants while using data from a traditional controller might not generate a dataset suitable to train predictive models. This study aims to assess the effect of the excitation signal of a dataset used for training an MPC's predictive model to harness a dwelling's energy flexibility. The under-study dwelling is equipped with underfloor heating coupled with an air-to-water heat pump. First, building is excited with three excitation signals: a Pseudo-Random Binary Sequence (PRBS), Rule-Based Controller (RBC) and Random Set-point Tracking (RST). Second, predictive models trained by each excitation signal are employed in an MPC framework to maximize the energy flexibility of a dwelling. The findings reveal that it is feasible to excite a building in a way that the thermal comfort of the occupants is guaranteed while the identification dataset is rich enough to generate a robust predictive model. Such predictive model would lead to high energy flexibility activation of the dwelling.