Title:Optimal Lighting Control in Greenhouses Using Bayesian Neural Networks for Sunlight Prediction

Abstract:Controlling the environmental parameters, including light in greenhouses, increases the crop yield; however, the electricity cost of supplemental lighting can be high. Therefore, the importance of applying cost-effective lighting methods arises. In this paper, an optimal supplemental lighting control approach is developed considering a variational inference Bayesian Neural Network (BNN) model for sunlight prediction. The predictive model is validated through testing the model on the historical solar data of a site located at North Carolina ($R^{2}$=0.9971, RMSE=1.8%). The proposed lighting approach is shown to minimize electricity cost by considering the BNN-based sunlight prediction, plant light needs, and variable electricity pricing when solving the underlying optimization problem. For evaluation, the new strategy is compared to: 1) a Markov-based prediction method, which solves the same optimization problem, assuming a Markov model for sunlight prediction; 2) a heuristic method which aims to supply a fixed amount of light. Simulation studies are conducted to examine the electricity cost improvements of the BNN-based approach. The results show that the BNN-based approach reduces cost by (on average) 2.27% and 43.91% compared to the Markov prediction-based method and the heuristic method, respectively, throughout a year.