This paper studies the intelligent building thermal energy management system based on multi-source data fusion, explores the multi-source fusion method suitable for building thermal energy data, constructs an improved LSTM-CNN fusion model based on attention mechanism, formulates a thermal energy scheduling strategy based on heat flow prediction, and verifies the effectiveness of the system through experiments. The experiment uses the annual monitoring data of a smart office building and the relevant benchmark data set in the USA. The results show that the RMSE of the improved model for winter heat flow prediction is as low as 0.72 ℃, which is 21.3% lower than the traditional LSTM. The scheduling strategy achieves a 14.3% reduction in daily average energy consumption, the temperature compliance rate reaches 96.8%, and the system average response time is 0.8 second, which meets the real-time requirements and provides strong support for the efficient management of thermal energy in intelligent buildings.