Two-stage MLP-lookup table model for predicting heat pump power in greenhouses

Energy costs account for a significant proportion of greenhouse operating expenses; thus, high-fidelity predictive tools are increasingly important for optimizing energy consumption. Although machine learning models demonstrate high accuracy within training ranges, their applicability to diverse operational conditions remains limited. This study developed and compared two prediction approaches: a two-stage multilayer perceptron-lookup table (MLP-LUT) model and a standalone multilayer perceptron (s-MLP) model for forecasting electrical heat pumps (EHP) energy consumption. The MLP-LUT model first predicts greenhouse temperature and humidity and then estimates power consumption through manufacturer performance mapping, whereas the s-MLP model directly predicts consumption. Bayesian optimization was used for hyperparameter tuning. The robust generalization performance of both models underwent evaluation across diverse operating conditions, including variations in the setpoint temperature, location, control strategies, and equipment model. At baseline, both models achieved comparable accuracy, with C V R M S E h r values of 4.05% and 4.04%, respectively, satisfying the American Society of Heating, Refrigerating, and Air-Conditioning Engineers (ASHRAE) criteria. For generalization testing across various setpoint temperatures, locations, and control strategies, both models retained stable prediction accuracy. However, when the EHP units are replaced, the s-MLP model exhibits severe degradation, whereas the MLP-LUT model maintains C V R M S E h r of less than 4.0%. The MLP-LUT framework offers resilience to hardware substitution by separating environmental predictions from equipment-specific performance mapping. In contrast, the s-MLP approach is constrained to static configurations. The present work establishes practical guidelines for the system selection and provides a foundation for the development of optimal greenhouse control strategies.