On-line optimal control of greenhouse crop cultivation.

Thus far, optimal control has primarily been investigated for seasonal crop growth optimization. On-line aspects have received much less attention. The decomposition between long term strategies and on-line control, however, is not trivial. Appreciable losses occur when set-points generated by seasonal optimization using nominal weather and the static physics assumption are put to the fully dynamic system. It is argued that the main product of seasonal optimization are not the state trajectories of the fast variables, but rather the state trajectories, and the co-state trajectories of the slow variables. The co-states represent marginal values of the slow variables. The actual on-line control requires the solution of a new dynamic optimization problem over a far shorter horizon, with a modified goal function incorporating the marginal value of all slow crop variables (not just the harvested parts). Feed-back and feed-forward to account for modelling errors and weather deviations can be accounted for by receding horizon optimal control, using the actual measured weather and fast states. An feasible economically optimal greenhouse crop control and operation system thus may consist of three major elements: (i) specification of constraints that pertain to unmodelled management aspects, possibly supplied by expert systems or other AI decision support, (ii) a solution to the seasonal slow crop response optimization using known long term nominal weather, and (iii) a receding horizon optimal control to generate short term controls responding to the actual weather and state, using crop state and co-state information from the slow optimization.