Hybrid modeling approaches for accurate greenhouse climate prediction: Combining mechanistic models and LSTM neural networks

Abstract

An accurate greenhouse climate model is crucial for controller design, energy consumption, and crop yield prediction. However, for a given greenhouse, considerable cost and time are required to study the thermal and mass transfer processes needed to construct an accurate greenhouse climate mechanistic model. To explore highly efficient modeling methods for greenhouse climate, this study proposes two hybrid modeling methods that combine mechanistic modeling with neural networks. The first method establishes a residual dataset for the greenhouse environment using mechanistic models and trains this residual model with an LSTM neural network. The second method employs LSTM neural networks and mechanistic models to predict greenhouse climate, then weights and combines the predictions from both models to achieve more accurate forecasting of greenhouse climate. In these two hybrid models, the mechanistic models use optimization algorithms for parameter identification and are validated with data from four different periods. A comparison of the results from the mechanistic model and the LSTM greenhouse climate prediction model shows that the neural network residual correction model exhibits better prediction accuracy and generalization capability in handling uncertain climate environment data. In contrast, the weighted fusion model places higher demands on the base models and shows considerable uncertainty in adaptability to different environments. The developed models in this study not only improve the prediction accuracy of greenhouse climate but also enhance the capability to handle complex and changing climatic conditions, thereby providing reliable decision-making support for greenhouse management and agricultural production.