Attention‐Based Spatial–Temporal Model for the Entire Indoor Space Temperature Estimation

This study proposes a novel attention‐based spatial–temporal model for indoor temperature estimation, which integrates positional encoding and an attention mechanism with a fully connected neural network. By jointly capturing temporal dynamics and spatial correlations, the model enables high‐resolution estimation of temperature fields across the entire indoor space. Using temperature data collected from 136 distributed sensors in an indoor space, along with time and coordinate information, the model effectively captures both short‐term and long‐term dependencies in temperature variations influenced by external environmental conditions. Ablation studies demonstrated that incorporating an encoder for the temperature series significantly improves prediction accuracy compared to models using encoders solely for time or coordinate data. The model was evaluated using multiple metrics, and results showed that it achieved a mean absolute error of 0.15 °C. It also exhibited robust performance across various sequence lengths and sensor configurations.