Within the realms of managing smart environments, especially within multi-user laboratory settings, predicting accurate occupancy faces a range of challenges. This study aims to address this challenge by utilising data gleaned from an IoT sensor network, and a digital twin paradigm. To enhance the accuracy of occupancy prediction Generative Adversarial Networks (GANs) has been implemented to generate synthetic data that can overcome the limitation of scarcity of the historical data that is gathered from the IoT sensor network. Long Short-Term Memory (LSTM) networks and Random Forest algorithms were used for the predictive modeling. In order to forecast occupancy rates, the models were trained on both synthetic datasets created using GANs and real datasets. The performances of the models found that the F1 score for the LSTM model was 0.801 without the synthetic data in the training and 0.802 with GAN for both the models. This strategy promotes more effective building management techniques in addition to optimizing the utilization of physical spaces.