Deep Neural Network based Optimal Control of Greenhouses

TL;DR

This work tackles energy-efficient greenhouse climate control under variable weather by proposing a two-level control architecture that couples a day-ahead forecast-based optimizer with real-time NMPC. A physically grounded Venlo greenhouse model links climate states $T$, $C$, and $H_a$ to tomato crop growth via the CG4 model, with four actuators regulating heating, cooling, ventilation, and CO$_2$ supply. To enable real-time deployment on low-cost hardware, NMPC data are used to train a Deep Neural Network that approximates the NMPC tracking policy, reducing online computation while maintaining performance under disturbances. The approach is validated via simulations demonstrating robust tracking of references and potential applicability in resource-constrained or remote settings where energy efficiency is critical.

Abstract

Automatic control of greenhouse crop production is of great interest owing to the increasing energy and labor costs. In this work, we use two-level control, where the upper level generates suitable reference trajectories for states and control inputs based on day-ahead predictions. These references are tracked in the lower level using Nonlinear Model Predictive Control (NMPC). In order to apply NMPC, a model of the greenhouse dynamics is essential. However, the complex nature of the underlying model including discontinuities and nonlinearities results in intractable computational complexity and long sampling times. As a remedy, we employ NMPC as a data generator to learn the tracking control policy using deep neural networks. Then, the references are tracked using the trained Deep Neural Network (DNN) to reduce the computational burden. The efficiency of our approach under real-time disturbances is demonstrated by means of a simulation study.