Evaluation of CFD and machine learning methods on predicting greenhouse microclimate parameters with the assessment of seasonality impact on machine learning performance

Abstract

For cleaner and sustainable greenhouse crops production, it is essential to successfully manage the needs and resources. Thus the prediction of the greenhouse microclimate, especially the temperature and relative humidity is of great interest. The research done in this area is, however, still limited, and a number of machine learning techniques have not yet been sufficiently exploited. The objective of this paper is to evaluate two greenhouse modeling techniques (machine learning (Artificial Neural Networks (ANN), Support Vector Machine (SVM), Bagging trees (BG) and Boosting trees (BT)) and Computational Fluid Dynamics (CFD) methods and assess the impact of the seasonal changes on machine learning performances. The study was carried out in a commercial greenhouse located in Agadir, Morocco, and the experimental data were collected during October and March. Results show that all predictive models are capable of predicting the inside air temperature (Tin) and relative humidity (Rhin) of the greenhouse with a quite good precision (R>0.98, nRMSE<7%). However, the time required by machine learning models was much more less than the one required by CFD model. For this reason, machine learning models were selected for further analysis and assessment of seasonality impact on their performances. The analysis and assessment of seasonality impact on Machine learning models prove their efficiency in predicting Tin and Rhin with a good agreement. A “combined data” model, built from experimental data of the two months, is tested and proved its efficiency in predicting Tin and Rhin of March and October separately and at the same time (R>0.98, nRMSE <9%).