Applied machine learning in greenhouse simulation; new application and analysis

Abstract

Prediction the inside environment variables in greenhouses is very important because they play a vital role in greenhouse cultivation and energy lost especially in cold and hot regions. The greenhouse environment is an uncertain nonlinear system which classical modeling methods have some problems to solve it. So the main goal of this study is to select the best method between Artificial Neural Network (ANN) and Support Vector Machine (SVM) to estimate three different variables include inside air, soil and plant temperatures (Ta, Ts, Tp) and also energy exchange in a polyethylene greenhouse in Shahreza city, Isfahan province, Iran. The environmental factors which influencing all the inside temperatures such as outside air temperature, wind speed and outside solar radiation were collected as data samples. In this research, 13 different training algorithms were used for ANN models (MLP-RBF). Based on K-fold cross validation and Randomized Complete Block (RCB) methodology, the best model was selected. The results showed that the type of training algorithm and kernel function are very important factors in ANN (RBF and MLP) and SVM models performance, respectively. Comparing RBF, MLP and SVM models showed that the performance of RBF to predict Ta, Tp and Ts variables is better according to small values of RMSE and MAPE and large value of R2 indices. The range of RMSE and MAPE factors for RBF model to predict Ta, Tp and Ts were between 0.07 and 0.12 °C and 0.28–0.50%, respectively. Generalizability and stability of the RBF model with 5-fold cross validation analysis showed that this method can use with small size of data groups. The performance of best model (RBF) to estimate the energy lost and exchange in the greenhouse with heat transfer models showed that this method can estimate the real data in greenhouse and then predict the energy lost and exchange with high accuracy.