A greenhouse control with feed-forward and recurrent neural networks

https://doi.org/10.1016/j.simpat.2007.06.001Get rights and content

Abstract

Greenhouses are classified as complex systems, so it is difficult to implement classical control methods for this kind of process. In our case we have chosen neural network techniques to drive the internal climate of a greenhouse. An Elman neural network has been used to emulate the direct dynamics of the greenhouse. Based on this model, a multilayer feed-forward neural network has been trained to learn the inverse dynamics of the process to be controlled. The inverse neural network has been placed in cascade with the neural model in order to drive the system outputs to desired values. Simulation results will be given to prove the performance of neural networks in control of the greenhouse.

Introduction

Greenhouses are considered as complex processes. In fact, they are non linear, multi-input multi-output (MIMO) systems, they present time-varying behaviors and they are subject to pertinent disturbances depending generally on meteorological conditions [1], [2], [3], [4], [5]. All these make difficult to describe a greenhouse with analytic models and to control them with classical controllers. In [2] the author has presented a model of a greenhouse using the energy balance. The proposed model is then used to try simulation on the greenhouse climate (temperature and hygrometry) with optimal control in a part of day. In [3] the author has proposed a greenhouse model including the crop transpiration. Then he showed a comparison between optimal and predictive control on the considered greenhouse in a part of day. In [5] the authors have described the application of model predictive control (MPC) for temperature regulation in agricultural processes (a greenhouse). The MPC algorithm used here takes in account the constraints in both manipulated and controlled variables using an on-line linearisation with a very low computational burden. This MPC scheme is compared with an adaptive PID controller. In [17] the authors have proposed an application of fuzzy logic to identify and control of multi-dimensional systems. They describe a method to reduce the complexity of a fuzzy controller and they show an application on a real system (a greenhouse). In our case we opt to the use of neural networks to model and to control a greenhouse. A recurrent neural network based on an Elman structure [6], [7] is trained to emulate the direct dynamics [7], [8], [9] of the greenhouse and used as a greenhouse model and a multilayer feed-forward neural network [10], [11] trained to emulate the inverse dynamics of the considered greenhouse is applied as a controller [12], [13], [14], [15] to provide the control inputs to the greenhouse.

This paper is organized as follows: in Section 2, we describe the considered greenhouse. In Section 3 we present the architecture of the used Elman neural network to emulate the direct dynamics of the greenhouse and the results of the modeling step. In Section 4, we show, the training structure of feed-forward neural network to emulate the inverse dynamics of the greenhouse then the recall structure used for control following with simulation results and comments. Finally, a conclusion and prospects are given in Section 5.

Section snippets

Greenhouse description

The considered greenhouse is a classical one with glasses armatures and defined by, a surface with 40 m2 and a volume with 120 m3. It is equipped with sensors allowing measurements of the internal and external climates. The internal climate defined by the internal temperature and the internal hygrometry constitute the greenhouse outputs. In order to control the internal climate the greenhouse is equipped with a set of actuators composed with a heater functioning in the on/off mode with 5 kw power,

Greenhouse neural modeling

Greenhouses are classified as complex processes. So it is very difficult to obtain kinetic models that represent the whole dynamics of the system. For this we have resort to advanced techniques to resolve such problem. In our case we have chosen a resolution with neural network and precisely an Elman structure [6], [7].

Greenhouse neural control

Now the real greenhouse is replaced by the described Elman neural network model above. To control the greenhouse we need a controller able to take with the complexity of the system. The multilayer feed-forward neural network with an input layer, an output layer and one hidden layer can be used as solution to control such process [7], [12], [13].

Conclusion and prospects

In this paper we have used an Elman neural network to emulate the direct dynamics of a greenhouse. The obtained model has been used next in closed loop control using a multilayer feed-forward neural network. This last is trained to emulate the inverse dynamics of the greenhouse and then used as a nonlinear controller with feedback state to provide the control actions for the process. The simulation results show that neural networks strategies give good performances when controlling complex

References (17)

  • M.Y. El Ghoumari et al.

    Non-linear constrained MPC: real-time implementation of greenhouse air temperature control

    Computers and Electronics in Agriculture

    (2005)
  • J.L. Elman

    Finding Structure in Time Cognitive Science

    (1990)
  • K.J. Hunt et al.

    Neural networks for control systems – a survey

    Automatica

    (1992)
  • V.C. Gaudin, Simulation et commande auto-adaptative d’une serre agricole. Ph.D. Thesis, University of Nantes,...
  • L. Oueslati, Commande multivariable d’une serre agricole par minimisation d’un critère quadratique. Ph.D. Thesis,...
  • M. Souissi, Modélisation et commande du climat d’une serre agricole. Ph.D. Thesis, University of Tunis, Tunis,...
  • F. Fourati, Contribution à la commande neuronale de systèmes dynamiques complexes: Application à une serre agricole....
  • D.T. Pham et al.

    Neural Networks for Identification, Prediction and Control

    (1995)
There are more references available in the full text version of this article.

Cited by (0)

View full text