Multi-step fusion model for predicting indoor temperature in residential buildings based on attention mechanism and neural network

Abstract

Indoor temperature prediction is vital in HVAC system control, ensuring thermal comfort and energy efficiency. This study aims to propose a multi-step indoor temperature prediction model. Firstly, correlations between indoor temperature and environmental parameters are analyzed to select input parameters. Secondly, input parameters are processed by a convolutional neural network (CNN) and bi-directional long short-term memory network. Gramian angular fields are used to convert the indoor temperature data input to CNN into two-dimensional images, enabling model to extract spatial features and capture temporal dependencies. Dung beetle optimizer and attention mechanism enhance feature extraction, and point prediction results are obtained using a weight fusion method based on model error. Thirdly, the recursive multi-step prediction method is employed to extend the point prediction model into a multi-step prediction model. Finally, a case study on residential buildings in Hong Kong is conducted to demonstrate the model's applicability. The results show that the indoor temperature, outdoor temperature and indoor relative humidity strongly correlate with indoor temperature, and they are selected as input parameters. The MAE, MAPE, RMSE and R² of the fusion model achieve 0.0227, 0.015, 0.0497 and 0.9970. Compared with single models, they reduce by 83.2%, 68.7% and 69.7%. It accurately predicts indoor temperature for the next 4 h with MAE, MAPE, RMSE and R² of 0.0495, 0.003, 0.0627 and 0.9937. The proposed model significantly improves the accuracy of indoor temperature prediction, providing methodological support for intelligent building temperature control and safeguarding residents' thermal health.