A method for modelling greenhouse temperature using gradient boost decision tree

Abstract

An accurate environment model is a fundamental issue in greenhouses control to improve the energy consumption efficiency and to increase the crop yield. With the increase of agricultural data generated by the Internet of Things (IoT), more feasible models are necessary to get full usage of such information. In this research, a Gradient Boost Decision Tree (GBDT) model based on the newly-developed Light Gradient Boosting Machine algorithm (LightGBM or LGBM) was proposed to model the internal temperature of a greenhouse. Features including climate variables, control variables and additional temporal information collected within five years were used to construct a suitable dataset to train and validate the LGBM model. An adaptive cross-validation method was developed as a novelty to improve the LGBM model performance and self-adaptive ability. For comparison of the predictive accuracy, a Back-Propagation (BP) Neural Network model and a Recurrent Neural Network (RNN) model were built under the same process. Another two GBDT algorithms, Extreme Gradient Boosting (Xgboost) and Stochastic Gradient Boosting (SGB), were also introduced to compare the predictive accuracy with LGBM model. Results suggest that the LGBM has best fitting ability for the temperature curves with RMSE value at 0.645℃, as well as the fastest training speed among all algorithms with 60 times faster than the other two neural network algorithms. The LGBM has strongly potential application prospect on both greenhouse environment prediction and real-time predictive control.