Building performance simulation often fails to predict accurately the real energy performance, mostly due to great uncertainties in the input data. Errors in computed performance are particularly significant in the case of existing buildings, for which the amount of information about intrinsic characteristics is low. However, efficient energy retrofit operations make necessary an accurate understanding of the initial state of a building using a calibrated prediction model. Several works have investigated the use of identification techniques for model calibration. The present paper investigates the use of such techniques to derive an energy audit procedure suitable to be an efficient aid for retrofit. In particular, we study here the possibility to determine the unknown thermal conductivity of the envelope and the internal gains based on temperature measurements only. We show how the adjoint method can be used to solve efficiently the inverse problem, while providing a fast method for computing model's sensitivities.