Learning hidden relationship between environment and control variables for direct control of automated greenhouse using Transformer-based model

Abstract

Climate change poses a significant threat to agricultural sustainability and food security. Automated greenhouse systems, which provide stable and controlled environments for crop cultivation, have emerged as a promising solution. However, traditional rule-based greenhouse control algorithms struggle to determine optimal control variables due to the complex relationships between environmental variables. In response, we propose a Transformer-based model, Trans-Farmer, which predicts the control variables by considering the complex interactions among environmental variables. Trans-Farmer leverages the attention mechanism to learn the intricate relationships among the environmental variables. The encoder-decoder structure enables the translation of the environmental variables into the corresponding control variables, analogous to language translation. Experimental results demonstrate that Trans-Farmer outperforms baseline models across all the evaluation metrics, achieving superior accuracy and predictive performance. The attention maps of the encoder visualize how Trans-Farmer comprehends the complex interactions among the environmental variables. Additionally, the compact size of Trans-Farmer is suitable for application in general greenhouses with constrained microcontroller units. This approach contributes to the development of automated greenhouse management systems and emphasizes the potential of artificial intelligence applications in agriculture.