The state of physical entity in the Internet of Things (IoT) has an obvious time-varying characteristic. Preliminarily selecting candidate entities by predicting their current state when searching match entities from massive ones can effectively reduce the communication overhead of IoT search system. The existing methods are all based on shallow learning theories whose performances are very limited. Thus, a high-accuracy entity state prediction method (HESPM) based on deep learning theory is proposed. The model of HESPM is built by utilizing the deep belief network. Then the contrastive divergence algorithm is adopted to train the model. Therefore, the dynamic evolution trend of entity state can be accurately perceived and the future entity state can be precisely predicted. Simulation results demonstrate the effectiveness of HESPM in enhancing the prediction accuracy and communication overhead performances.