在全球淨零碳排與能源轉型趨勢下，提升住宅能源使用效率已成為智慧家庭發展的重要議題。本研究針對家庭能源管理系統中住宅負載波動與資料稀缺性問題，提出 TCNplus 短期負載預測與儲能系統排程方法。TCNplus 以時間卷積網路（Temporal Convolutional Network， TCN）為基礎，結合多尺度擴張因果卷積、輸入特徵注意力、門控機制，以及多戶預訓練與目標住戶微調策略，以提升住宅負載預測能力。實驗結果顯示，TCNplus 於多項評估指標具有最佳表現，並可利用未來負載資訊改善儲能系統充放電排程，降低住宅用電成本。在系統實作方面，本研究將模型與排程流程部署於 NVIDIA Jetson Xavier NX，並透過 ECHONET Lite 完成儲能系統控制封包與設備回覆驗證，建立由負載預測、能源排程至實際控制之家庭能源管理系統應用架構。;As the global transition toward net-zero emissions accelerates， improving residential energy efficiency has become increasingly important in smart home development. To address the challenges of highly variable residential loads and data scarcity in Home Energy Management Systems (HEMS)， this study proposes TCNplus for short-term load forecasting (STLF) and battery energy storage system (BESS) scheduling. TCNplus is built on the Temporal Convolutional Network (TCN) and incorporates multi-scale dilated causal convolutions， input feature attention， gating mechanisms， multi-household pretraining， and target-household fine-tuning to improve residential load forecasting performance. Experimental results show that TCNplus achieves the best performance across multiple evaluation metrics. The resulting forecasts provide future load information to improve BESS charging and discharging schedules， thereby reducing residential electricity costs. For system implementation， the forecasting and scheduling workflow is deployed on an NVIDIA Jetson Xavier NX， and ECHONET Lite is used to validate BESS control packets and device responses. The proposed framework integrates load forecasting， energy scheduling， and device control into a HEMS.
