|M.Sc Student||Lousky Marc|
|Subject||Integration of a Plant Model in a Greenhouse Decision|
|Department||Department of Civil and Environmental Engineering||Supervisor||Professor Raphael Linker|
|Full Thesis text|
This thesis is about the integration of a tomato plant model in a decision support system (DSS) developed by the Ministry of Agriculture. This DSS could allow a grower to compare his actual crop data with predictions based on the model. The crop growth model TOMGRO was used for this project. It describes the phenological development and increase in dry weight of various organs from planting till maturity. The original version of the model was implemented in the FORTRAN language, which is not suitable for internet technologies. We re-implemented the model in C#, a more recent, object oriented language, which offers many advantages for developing modular crop models. In order to calibrate the model, data was collected in a greenhouse at the Bsor Research Station during one season. The available experimental data did not allow calibration of all the parameters, and, using a sensitivity analysis, we determined that four parameters should be fitted. These four parameters were fitted using genetic algorithm and a good agreement was obtained between the model and measurements. Since pruning was performed in the greenhouse, we tried to improve the agreement by modifying the model to include the effect of pruning. Overall this change had very little influence on the predicted yield.
In the second stage of the study, we combined the plant model with a greenhouse climate model to perform an economic optimization of climate conditions.
The large number of state variables in TOMGRO making optimization virtually impossible, we chose to search for a sub-optimal solution based on a simplified model. A compact, single-state-variable, two-stage crop model was used and its parameters estimated by comparing its predictions to TOMGRO. This simple model was used to determine sub-optimal control strategy using Pontryiagin’s maximum principle. This method produces the so-called costate variable, the marginal value of the state variable of the simplified dynamic system. We normalized this costate with respect to the plant leaf area index, resulting in a rather constant costate. Since we have only one costate, varying rather slowly, and activating several control fluxes, it plays the role of a strategic set-point. We showed that while the crop model may be inaccurate or the weather may not be as expected, a corrective course can be estimated based on the available information by generating pairs of equivalent trajectories. It is further shown that this approach keeps the costs increases reasonable.