The thermal process of buildings is characterized by nonlinear and long time lag. Accurately predicting the mean room temperature of buildings is crucial for regulating heating systems. However, the current black box model determines inputs mainly through theoretical analysis without conducting correlation analysis. Moreover, the inputs do not sufficiently cover various factors influencing the mean room temperature, such as heating system operation parameters, outdoor meteorological conditions, time, and building physical parameters. To address these challenges, this study employs Spearman correlation index to select inputs for the mean room temperature long short-term memory (LSTM) prediction model. A composite indicator (the product of the building exterior surface area on different orientations with corresponding solar radiation on different orientations) is proposed and it is set as an input index. The results indicate that the proposed composite indicator improves prediction accuracy compared to using solar radiation alone, with errors decreasing 0.45%. The order of correlation strength between influencing factors and mean room temperature, from highest to lowest, is as follows: historical mean room temperature, water flow rate, supply water temperature, outdoor air temperature, solar radiation, return water temperature, wind velocity. These findings are essential for achieving predictive control and smart solutions in heating systems.