Traditional home energy management systems focus predominantly on managing energy within the household itself, often neglecting user comfort and the potential for bidirectional interaction with the distribution network. To further enhance the intelligence of home energy management, this paper proposes a two-layer hierarchical home energy management strategy. The upper layer considers system operational constraints and costs of the distribution network, whereas the lower layer aims to satisfy user electricity preferences and cost requirements as much as possible. Based on interaction with the distribution network, this strategy achieves optimal household energy management. To address the deficiencies and limitations of conventional distribution network power flow calculation and indoor thermal response algorithms in accurate modeling, two data-driven methods are introduced, namely the long short-term memory recurrent neural network (LSTM) and the online sequential extreme learning machine (OS-ELM). These methods aim to more accurately assess distribution network status and indoor thermal dynamics (LSTM for power flow estimation, OS-ELM for thermal response estimation). In addition, an efficient natural aggregation algorithm (NAA) is utilized to improve the solution speed and accuracy of the optimization process. To evaluate the effectiveness and scientific validity of the proposed algorithms, detailed comparative experiments are designed for the LSTM model in the upper layer, the dynamic thermal estimation algorithm proposed in the lower layer, and the overall two-layer energy management algorithm. Finally, experimental validation based on the IEEE 33-node system ultimately demonstrates that the proposed LSTM power flow estimation and OS-ELM-based thermal response estimation methods achieve significantly higher accuracy than conventional methods. The proposed strategy also outperforms traditional approaches remarkably in reducing users' electricity costs, ensuring distribution network security, and improving user comfort, thereby verifying its effectiveness and superiority. This method provides an effective technical support for addressing source-load coordinated optimization problems in communities with high penetration of distributed energy resources, and exhibits promising prospects for engineering applications.