Demand-controlled ventilation has been proposed to improve indoor air quality and to save energy for ventilation. It is important to estimate occupancy in a building precisely in order to determine adequate ventilation airflow rates, especially when people are the major source of indoor contaminants such as in office buildings. In this paper, we investigate occupancy estimation methods using a dynamic neural network model based on carbon dioxide concentration in a space. We conducted an experiment in a single room to measure carbon dioxide concentration and actual occupancy continuously in the room. We trained and tested the dynamic neural network model TDNN (time-delayed neural network) by varying the number of tapped delay lines and the number of neurons. Networks were trained using the first-day data and results were obtained for the rest of the days. The estimated results were compared with the actual number of occupants measured by a number counter installed at the entrance door. The root mean square (RMS) errors were obtained depending on system parameters. The dynamic model with tapped delay showed smaller errors in general than conventional static neural network models. The RMS errors were reduced, as the tapped delay line increased up to 15 minutes for the present experiment. The time delay has been found to be related to the dispersion time of contaminants in the space, which is again related to the dimensions of the space and the source locations relative to the sensor locations. Further research is needed to include the effect of the concentration in the adjacent rooms and the effect of other contaminants such as humidity and particle concentrations.