TL;DR
This paper presents a novel approach to optimizing greenhouse operations by integrating a machine learning model (LSTM) for predicting climate dynamics into a constraint programming (CP) framework. This allows for long-term economic optimization by considering the delayed effects of short-term climate control decisions, outperforming traditional grower strategies in a case study.
What Problem Does This Paper Address?
The agricultural sector faces increasing pressure to enhance productivity and sustainability, especially in controlled environments like greenhouses. While data-driven greenhouses offer potential for higher yields and prolonged cultivation, practitioners often rely on intuition and experience rather than sophisticated decision-making tools. Existing decision support systems (DSS) typically focus on short-term climate adjustments, neglecting their long-term economic consequences. Furthermore, complex mathematical models for greenhouse climate and crop growth are computationally expensive and difficult to integrate into real-time systems. This paper addresses the challenge of creating a DSS that can optimize for long-term economic profit by predicting the impact of immediate actions on greenhouse climate and crop growth, bridging the gap between data-driven insights and practical grower decision-making.
Methods
The core innovation lies in the integration of a Long Short-Term Memory (LSTM) neural network, trained to predict greenhouse climate, within a Constraint Programming (CP) framework.
- Machine Learning for Climate Prediction: An LSTM model was trained offline on historical greenhouse data to predict key internal climate variables (PAR-light, temperature, humidity deficit, CO2) based on outside weather and past control actions. This replaces complex, computationally intensive physical models.
- Constraint Programming for Optimization: The CP framework is used to model the entire optimization problem. This includes:
- Embedding the LSTM: The trained LSTM model is translated into CP constraints, allowing the CP solver to query its predictions. This is a key technical challenge, as NNs are procedural while CP is declarative.
- Crop Growth Model: A simplified, existing crop growth model (Lintul-3 adaptation) is also embedded as a CP function.
- Economic Model: A cost-benefit analysis is incorporated, linking crop growth to revenue and actions to costs.
- Decision Variables: The CP model determines sequences of control actions (e.g., heating, lighting, ventilation) to maximize predicted profit.
This "decision-focussed learning" approach leverages the predictive power of ML with the declarative modeling and search capabilities of CP.
Datasets
- Training Data: A dataset from a 150m² Venlo-style greenhouse in the Netherlands, covering the period 25/11/2020 – 20/9/2021. This data includes sensor readings for outside weather (radiation, temperature, absolute humidity, wind speed), greenhouse actions (heating, lighting, ventilation, shading), and inside climate (PAR-light, temperature, humidity deficit, CO2).
- Prediction Data: A subset of the above dataset from 5/8/2021 – 7/8/2021 was used for prediction and validation.
No new benchmarks were introduced; the focus was on applying the integrated system to a real-world use case.
Key Insights and Ideas
- Bridging the Prediction-Action Gap: The most significant insight is the successful integration of a learned predictive model (LSTM) into a decision-making framework (CP) to optimize for long-term economic outcomes. This allows the system to "look ahead" and understand the delayed consequences of current actions.
- CP as a Powerful Integration Tool: Constraint Programming proves to be a flexible and powerful tool for embedding complex ML models and other functional components (like crop models) into a unified optimization problem. Its declarative nature simplifies the modeling process.
- Data-Driven Optimization Outperforms Experience: The case study demonstrates that a well-designed DSS, even with a simplified crop model, can outperform an experienced grower in terms of profit, highlighting the potential of AI in agriculture.
- Challenges in ML-CP Integration: The paper emphasizes the conceptual differences between NNs and CP, and the practical challenges in translating NN operations into CP constraints efficiently, including handling scaling, capping, and sequential dependencies.
Key Results
- LSTM Prediction Accuracy: The embedded LSTM model achieved reasonable predictions for temperature and humidity deficit, with some challenges in predicting PAR light and CO2 due to sensor limitations and model complexity. The RMSE for climate prediction within the CP framework was 0.0665.
- Economic Optimization: In a one-day case study, the DSS achieved a higher profit (€0.40/m²) compared to the experienced grower (€0.29/m²). While the DSS incurred higher costs (€0.05/m² vs €0.01/m²), the increased revenue from better crop growth justified these costs.
- Runtime Limitations: A significant limitation identified is the runtime of the CP solver. The DSS required an 11-hour timeout to find a better solution than the grower, indicating that current performance is too slow for real-time practical deployment.
Limitations and Future Work
- Runtime: The primary limitation is the computational time required by the CP solver to find optimal solutions. This currently prevents real-time deployment.
- Crop Model Simplicity: The Lintul-3 adaptation used is relatively basic. More sophisticated crop models could improve prediction accuracy.
- LSTM Prediction Accuracy: While reasonable, the LSTM's prediction accuracy for PAR light and CO2 could be improved.
Future work includes:
- Improving Solver Performance: Techniques like initializing the CP solver with grower decisions or incorporating domain knowledge to reduce the search space are being explored.
- Enhancing Crop Model: Integrating a more advanced crop model, potentially also learned via ML and embedded in CP, is planned.
- Full Deployment: Connecting the DSS to greenhouse actuators for fully autonomous cultivation is the ultimate goal.
- Increasing Prediction Accuracy: Further improving the accuracy of the LSTM climate predictions.
Summary updated