Prediction of HVAC Operational Variables Using Recurrent Neural Networks for Advanced Control Strategies

Abstract

Heating, Ventilation, and Air Conditioning (HVAC) systems are major energy consumers in buildings, making accurate prediction of their operational variables essential for improving energy efficiency and occupant comfort. This study compares three recurrent neural network (RNN) architectures, namely standard RNN, Long Short-Term Memory (LSTM), and Gated Recurrent Unit (GRU) for multi-input, multi-output prediction of nine HVAC variables using high-resolution data from a real office building (2018-2020). A two-stage hyperparameter optimization varying network depth, width, learning rate, and batch size was applied. Results show that a shallow architecture with a single hidden layer of 16 nodes, learning rate of 0.001, and batch size of 64 offers the best balance between accuracy and generalization. While the standard RNN showed higher errors and lower stability, both LSTM and GRU performed well. The GRU achieved the lowest average rank (4.2) and mean validation loss (0.0049) across cross-validation folds, demonstrating superior consistency and reliability, and was selected for final evaluation. On the holdout test set, GRU consistently outperformed the other models, achieving R-2 > 0.81 for seven of nine variables, with a peak R-2 of 0.99 for mean indoor air temperature and an overall average R-2 of 0.80. These results highlight GRU's advantages in accuracy, stability, and efficiency, suggesting its suitability for intelligent building control. The study's scope is limited to three RNNbased architectures with a constrained hyperparameter search; future work should extend comparisons to broader model families and datasets.