Greenhouse temperature modeling: a comparison between sigmoid neural networks and hybrid models

Greenhouse operation and inside climate strongly depend on the outside weather. This implies that at least a year of data collection is required to cover the whole operational domain. Greenhouse-climate models calibrated with data limited to only a small region of the operating domain (weather and control), may therefore, produce erroneous predictions when applied to unfamiliar conditions. A comparison is made between the performance of three types of models trained with several seasonal sub-sets of data: (1) black-box (BB) sigmoid neural network (NN) trained only with in situ data, (2) hybrid physical-RBF (radial basis function) model, and (3) sigmoid neural network trained with a combination of in situ data and synthetic data generated with a physical model (termed ‘prior-K sigmoid model’). The BB sigmoid model gives the best predictions within the training domain, but performs very badly outside it. On the other hand, the hybrid and prior-K sigmoid models produce useful predictions over the whole operating domain, although they are slightly less accurate within the training domain. © 2003 IMACS. Published by Elsevier B.V. All rights reserved.