S06 - Session O1 - Reliability of greenhouse climate dynamic models by uncertainty and sensitivity analyses, calibration and evaluation.

Authors: Irineo Lopez Cruz *, Efren Fitz-Rodríguez, Raquel Salazar-Moreno, Abraham Rojano-Aguilar

The improving of controlled environment horticulture systems such as greenhouses hinges on reliable dynamic mathematical models. To make models trustworthy is necessary not only the postulation of their structure but also performing uncertainty and sensitivity analyses, parameter estimation and model evaluation. In this work the benefits of such overall system's modeling procedure are demonstrated by performing a Monte Carlo and Generalized Likelihood Uncertainty Estimation (GLUE) uncertainty analyses on a two-state model of air temperature and humidity of a greenhouse. Furthermore, a standardized regression coefficients sensitivity analysis and the density-based PAWN sensitivity analysis, both being global sensitivity methods, were also applied to choose the most influential model parameters on the state-variables. Besides, the model was calibrated by the classical nonlinear least squares procedure which is a local search method and a Differential Evolution algorithm which is a kind of global optimization method. Finally, the model was evaluated by using an independent dataset and its performance was measured with several statistic measures such as MSE, RMSE, MAE and model efficiency. According to uncertainty analyses although both state variables are calculated accurately, distributions turned out to be highly skewed with a large kurtosis value, namely, quite different from a normal distribution. Because of the last result a density-based sensitivity method was needed. From both sensitivity analyses, the most influential model parameters turned out to be infiltration coefficient, the heat transfer coefficient of soil, the leaf boundary layer resistance, some physical constants such as soil density, and physical properties of the greenhouse (cover and ground area). Only the three most influential model parameters were estimated. For temperature RMSE values on calibration and evaluation were 0.96 and 1.26, respectively. Whereas for absolute humidity RMSE values were 2.03 and 2.12, respectively; and for relative humidity those values were 12.3 and 15.6, respectively.