Estimation of greenhouse CO2 concentration via an artificial neural network that uses environmental factors

Abstract

In order to improve photosynthesis efficiency and crop growth, it is important to predict CO2 concentration as well as CO2 consumption in greenhouses. The objective of this study was to predict greenhouse CO2 concentration via an artificial neural network (ANN) that incorporated environmental factors. Temperature, relative humidity, atmospheric pressure, solar radiation, and CO2 concentration were measured every 10 min over a 6-month period in a greenhouse located in Boryeong, Korea. Measured environmental data were used to train the ANN. Among the 14,866 data points used in the experiment, 10,000 and 4866 data points were used for training and testing, respectively. An ANN with an input layer with input neurons, two hidden layers with 32–2048 neurons, and an output later with one neuron was selected. A rectified linear unit was used as the activation function in each node of the ANN. An ANN structure that included 256 neurons in the hidden layers showed the highest test accuracy (R2 = 0.97) was selected from all the structures, while multivariate linear regression showed lower test accuracy than the ANN (R2 = 0.78). The ANN accurately estimated CO2 concentration in the greenhouse using big data for changing patterns of the inside environmental factors without vent position data. Furthermore, it is possible to estimate crop CO2 consumption in greenhouses with this ANN using the change in greenhouse CO2 concentration.