Gated recurrent unit models outperform other Machine learning models in prediction of minimum temperature in greenhouse Based on local weather data

Abstract

Obtaining the minimum temperature of greenhouse (Timin) in advance can determine the lower limit of crop growth. Some research thought that was feasible to install sensors in the greenhouse so that can obtain the input variables for predicting Timin. However, lots of expenditures would be spent for intensive greenhouse parks. Local weather data can be easily and economically obtained so that considered it as an input variable for predicting Timin. First, the Pearson correlation coefficients were used to select the relevant input meteorological variables and eight different input combinations were consequently constructed. Then, three generalized machine learning models (i.e. Random forest, RF; Support vector machine, SVM; and Multiple linear regression, MLR) and two deep learning models (i.e. Long-short term memory, LSTM; Gated recurrent unit, GRU) were used to predict Timin based on the eight different input combinations. The results showed that the RF and GRU model had the best prediction performance among the generalized machine learning and deep learning models, respectively. Deep learning models were not sensitive to the number of input variables. In the absence of sufficient meteorological factors as input variables, the GRU model generally had better prediction performance than the other models. The prediction ability of deep learning models was obviously superior to the generalized machine learning models when Timin > 28.5 °C or Timin < 13.9 °C, particularly for the GRU model. Most of the differences between predicted and observed value (Di) of deep learning models distributed between −1 °C and 1 °C. And most of the predicted Timin by RF, SVM and MLR models were lower than the actual Timin, while the deep learning models were relatively stable. Finally, consider the prediction accuracy in terms of lacking input variables and the stability of the model, we recommended deep learning models to predict Timin, especially the GRU model.