Occupancy information, especially the number of occupants, can effectively guide the operation of building systems. Capturing real-time and accurate occupancy information is critical. Passive infrared (PIR) sensors and smart meters have been deployed extensively. To improve the occupancy detection accuracy, this study proposes an innovative occupant number detection approach involving a temporal sequential analysis of occupancy-related data using data fusion from PIR sensors and smart meters with a convolutional neural network (CNN) model. A university office was investigated as a case study. PIR sensors and smart meters were adopted to collect data. And CNN model was established with both various data types and historical data as multi-dimensional model inputs. The model outperformed than traditional ANN models with an improved accuracy by 26.8%. The proposed occupancy detection model was applied to optimize the control strategy of an outdoor air system, which helped to provide comfortable built environment and realize energy savings.