Occupant energy-related behaviors such as activities, interactions, and occupancy are widely acknowledged as key drivers of building performance. These behaviors are highly individualized and influenced by factors such as building characteristics, lifestyle preferences, technological adoption, and demographic differences. The associated profiles, consisting of multiple behaviors performed continuously over time, are commonly used as inputs for building performance simulations (BPS) to model occupancy dynamics, predict energy demand, and optimize performance. Due to limited profiles’ acquisition and synthetization, traditional BPSs often rely on static or average assumptions that fail to account for individual contexts, causing bias and unreliability. To address the limitation, this study demonstrates a framework that incorporates time-use surveys to observe occupant energy-related behavior profiles and employs a Graph Attention and Long Short-Term Memory (GAT-LSTM) networks for profile prediction. It was validated through a case study using Swedish time-use survey data from six occupants. The GAT-LSTM model achieved high validation accuracy (96–98%) and demonstrated strong agreement between predicted and next-day behavior profiles, with Cohen’s Kappa values ranging from 0.479 to 0.638.