GAIA: tailor-made support for operators in greenhouse horticulture

For decades, greenhouse horticulture in the Netherlands is leading the world in terms of food and flower production in greenhouses. Every day the greenhouse operator makes many decisions to control the greenhouse in such a way that the crop can grow under the best conditions. These decisions are often based on personal experience. Increasingly, the operator uses decision-support systems to do this.

However, these systems assess greenhouse processes in a different way and thus lead to different recommendations for greenhouse control, depending on the intended objective. In addition, the implicit knowledge of the operator is usually not taken into account in such systems.

Deployment of Artificial Intelligence

It is a challenge to develop an Aritificial Intelligence (AI) system that assesses different processes for controlling greenhouses and includes the operator’s experience. The operator shares his decisions and reasoning with the AI system and also learns from the AI system. In time, the quality of the advice improves: the system becomes more autonomous and partly takes over the control of the greenhouse.

AI is applied to make the best choices for greenhouse control, such as indoor climate, energy and water consumption, CO2 production and crop growth. However, sometimes conflicting goals need to be reached. For example: temperature and relative humidity in greenhouses must be within a certain range, energy and net water consumption must be as low as possible, crop development must be as fast as possible whereas the quality needs to be acceptable.

Learning process

To this end, the Greenhouse AI Accumulator (GAIA) is being developed; it allows various underlying existing subcontrollers to work together to achieve a good balance through a collective control strategy. Therefore, the system must receive input from sensors in the greenhouse, including data on crop development.

The use of existing and new AI technologies, such as Model-based Predictive Control and/or Reinforcement Learning, are examined to determine whether they can provide the correct advice. In addition, advice will always be accompanied by a good explanation through explainable AI techniques. The operator can learn from the decisions of the system and will also be enabled to give feedback with substantiation about the correctness of an advice. The feedback will be used by the AI system to create improved future advice. In this way, the system learns from the operator, who makes decisions based on ‘green fingers’ founded on years of experience: ‘co-learning’.

What challenge does it solve?

The aim of the project is to support the greenhouse operator in making the correct decisions to achieve maximum yield and product quality in greenhouses at minimum cost and impact on the environment. In the longer term, this should lead to largely autonomously functioning greenhouses, making food production in greenhouses easier and more feasible in large parts of the world.

What does the use case deliver?

This AI development helps the Dutch greenhouse horticulture sector to stay ahead in the development of high-tech greenhouses. The application of AI systems is especially important for high-tech greenhouses at locations in the world with less agronomic expertise. In those places, in-depth expertise will more often be located at a distance (‘remote control’), thefore it is required that one can rely more on autonomous systems.

The intended results are:

  • A proof-of-concept of a first version of GAIA in which AI techniques are applied and fine-tuned for optimal greenhouse control;
  • A mechanism for explainability and a hybrid-AI approach for co-learning between operator/grower and GAIA;
  • Experimental results of the GAIA proof-of-concept for a specific greenhouse to test the usability for the operator/grower;

Results 2021

A prototype of GAIA has been realized, based on a hybrid framework which includes an AI Neural Network component to predict the greenhouse climate, a physical plant model to predict the plant state, and a knowledge graph to take in the feedback and expertise from the grower.

GAIA works with any given set of objectives. For concreteness, two main objectives were defined, namely maximizing the yield and minimizing the energy cost. An Model-based Predictive Control optimization algorithm is responsible for finding the optimal control setpoints which can involve multiple objectives (optimal crop growth and minimal energy cost). GAIA works with the climate and energy-management settings, and uses its built-in plant model to predict the crop-mass. These predictions can be compared with measurements of the crop mass on a regular basis.

The grower has the possibility to add his knowledge and expertise in the form of control rules. They express relationships between parameters in the greenhouse, regarding plant, climate, and outside weather. By adding rules to the system, the knowledge of the grower is explicitly added and can be used as guidance for the GAIA setpoint calculation. The rules from the grower are stored in a knowledge base in terms of the CGO (Common Greenhouse Ontology). This ontology is used because it formulates the weather, the climate in the greenhouse, and the control setpoints in a structured and common manner.

To train GAIA’s AI model, simulated timeseries generated by SIOM, a greenhouse simulation model, have been used. In addition, SIOM was also used to evaluate how well GAIA performs. The evaluation showed that the AI model is able to generate reliable predictions for the whole growing period in terms of optimal setpoints for day and night temperatures, maximum CO2 concentration, maximum relative humidity, and the control of the energy screen.

To show how the GAIA system works and to present the results, a mock-up dashboard was developed. Besides the generated setpoint predictions, also the effects of using these setpoints on the performance objectives is shown graphically. The grower has the option to give feedback to the advice given by the system and to add control rules to the knowledge base. After this, the GAIA system updates its calculations and presents the newly obtained results.

Next steps

The work to be carried out in 2022 has as main goal to make the current GAIA prototype practically applicable in existing greenhouses or small test compartments with actual growers or crop managers. To realize this goal, the following objectives are set for 2022. The explanation of the advice given by GAIA will be improved and extended. This will be done in such a way that the grower can understand what the reason was for the advice to apply changes in greenhouse control. Besides an advice on greenhouse control, this also could give the grower new insights, beyond his own experience. By feeding the reaction of the grower to GAIA in return, continuous two-way learning between the operator and the GAIA system will be realized and existing knowledge and expertise of the grower can be taken into account. The last objective is to make the step from training GAIA with simulated data to real greenhouse data. In addition, it is intended to test and evaluate GAIA in a setting with an actual greenhouse and a grower.

Parties involved

This project is part of the Appl.AI programme of TNO and is partly funded by the start impulse the NL AIC received from the Dutch government for research and development of AI applications. GAIA is developed by TNO in cooperation with Ridder and Hortivation (through the adjacent project AGROS, co-funded by TKI “Tuinbouw en Uitgangsmaterialen”).

Share via: