A Predictive Greenhouse Digital Twin for Controlled Environment Agriculture

Controlled environment agriculture offers significant advantages for the efficient use of resources in food production, especially in hot desert climate regions due to the scarcity of arable land and water. However, farming practices such as hydroponics and aquaponics have high energy requirements for temperature control and present higher operational complexity when compared to traditional forms of farming. This study describes a Predictive Greenhouse Digital Twin (PGDT) that addresses these challenges through a dynamic crop yield assessment. The PGDT uses greenhouse measurements gathered through an IoT sensor network and a regression approach to multivariate time series forecasting to develop a model capable of predicting final crop yield as a function of the gathered measurements at any point in the crop cycle. The performance of the PGDT is evaluated with reference to forecasting algorithms based on deep and ensemble learning methods. Overall, deep learning methods show superior performance, with Long short-term memory (LSTM) providing a marginal advantage compared to Deep Neural networks (DNN). Furthermore, the models were deployed on an edge device (a Raspberry Pi-based gateway), where DNN demonstrated faster inference while delivering performance better than LSTM.