Attention-LSTM architecture combined with Bayesian hyperparameter optimization for indoor temperature prediction

Accurate prediction of indoor temperature can provide more reference data for indoor thermal comfort assessment and the operational effectiveness of heating, ventilation and air conditioning equipment, making it possible to reduce unnecessary energy consumption while ensuring occupant comfort. This paper introduces a deep learning method to predict indoor air temperature. The aim is to explore the potential of a model combining LSTM with encoder-decoder and attention mechanisms in short-term forecasting and compare it with LSTM models and GRU models. The hyperparameters are optimized by TPE Bayesian optimization to facilitate the determination of various parameters in the deep model. The results show that compared with other commonly used time series prediction algorithms, the model has an advantage in the case of short-term time ahead prediction. The model can accurately predict the change trend of room temperature and maintain stability for a long time. The R-square of the prediction results is more than 0.9. This work has reference significance for the feasibility study of establishing an indoor temperature prediction model. • The Attention-LSTM architecture used to predict the room temperature. • Compare the used architecture with the LSTM architecture and GRU architecture. • TPE Bayesian hyperparametric optimization is used to determine the hyperparameters. • The architecture used is predicted to be more accurate and stable.