This paper shows how Nonlinear Finite Impulse Response (NFIR) models realized by artificial neural networks can be used for developing simulation models of the inside temperature of greenhouses. The proposed NFIR models use integrated variables to reduce the number of past values needed as inputs. Several NFIR models have been developed using past data following a systems identification methodology. All data have been obtained from a real greenhouse in Southern Spain dedicated to tomato crop. The NFIR models are later compared with a model based on first principles. The results obtained in the a posteriori application of the models to new real data show that the performance of the NFIR model with integrated variables compares well with that of a first principles model, although the generalization capabilities of the latter are superior.