Predicting indoor temperature distribution with low data dependency using recurrent neural networks

Accurately predicting indoor temperature distribution can provide valuable reference data, helping residents independently adjust HVAC equipment around them to ensure comfort while reducing unnecessary energy consumption. This study proposes a prediction framework composed of two neural networks, enabling accurate indoor temperature distribution prediction with minimal training data in both temporal and spatial dimensions. The Dual-Stage Attention-Based Recurrent Neural Network calculates the importance ranking of feature values to enhance individual feature information and reduce training data volume, while Long Short-Term Memory is used to predict time-series features. From February to September 2022, 22 temperature sensors were installed in a target office to collect minute-by-minute indoor temperature data, which served as training and testing datasets. The results showed that, for short-term prediction, using data from five out of the 22 sensors collected over two weeks in winter (heating season) and summer (cooling season), the framework accurately predicted temperatures at the remaining 17 sensor locations, with a root mean squared error between 0.3 and 0.7. This study is significant for continuous indoor temperature prediction under low data volume conditions.