Accurate prediction of occupancy in buildings can significantly improve the performance of building energy simulation and further facilitate building design and system operation. Considering the temporal dependency of occupancy, Hidden Markov Model has been widely used to effectively predict occupancy behavior. However, the traditional Hidden Markov model that uses time-independent transition probability matrix is difficult to accurately describe the dynamic variation of occupancy as well as its correlation with environmental parameters. Such a model would greatly reduce occupancy prediction accuracy. To address this issue, an inhomogeneous Hidden Markov Model based on state transition was proposed. In this model, time-dependent transition probability matrices were calculated to capture the temporal dependency of occupancy at different time periods. Meanwhile, probability distribution of environmental parameters was calculated based on state transition instead of state only, aiming at rationally describing the correlation between occupancy and environmental parameters. The method was applied to predict the occupancy of a Belgian office. The results demonstrated that the effectiveness of the proposed approach and the prediction accuracy were improved significantly.