隨著智慧家電、再生能源與儲能系統逐漸普及，家庭能源管理系統需在動態環境下兼顧用電成本、居住舒適度與設備安全。針對家庭設備同時包含離散排程與連續功率控制，以及傳統深度強化學習較難掌握極端安全風險的問題，本研究提出一套結合混合動作空間與條件風險值（Conditional Value-at-Risk， CVaR）的 SAC PLUS 控制方法，使代理人能共同學習可排程負載的啟動決策，以及空調與儲能系統的功率控制，並於策略更新時抑制尾端安全風險。本研究建構基於真實資料之家庭能源管理模擬環境，並透過一般測試情境與非連續高溫擾動情境，比較 SAC PLUS、SAC Hybrid 與 DDPG Baseline 的控制表現。實驗結果顯示，SAC PLUS 在兩種測試情境下皆能有效降低整體安全成本與尾端安全風險，並在電網成本略高於 SAC Hybrid 的情況下，仍維持優於 DDPG Baseline 的能源調度表現，顯示其能以可接受的電費成本代價換取較佳安全穩定性。此外，本研究將訓練完成之模型部署於 NVIDIA Jetson Xavier NX 進行邊緣推論，驗證其於低功耗裝置上具備即時運算能力；並進一步將 HVAC 控制結果結合 ECHONET Lite 應用情境，初步驗證邊緣端控制策略與智慧家庭設備通訊協定整合之可行性。整體而言，本研究所提出之 SAC PLUS 方法能兼顧能源調度、安全風險抑制與邊緣部署需求，可作為智慧家庭能源管理系統之可行控制方案。;With the growing adoption of smart appliances， renewable energy， and energy storage systems， home energy management systems must simultaneously consider electricity costs， occupant comfort， and equipment safety in dynamic environments. To address the coexistence of discrete scheduling and continuous power control in household devices， as well as the limited ability of conventional deep reinforcement learning to manage extreme safety risks， this study proposes SAC PLUS， a control method integrating a hybrid action space with Conditional Value-at-Risk (CVaR). This method enables the agent to jointly learn the activation decisions of schedulable loads and the power control of heating， ventilation， and air conditioning and energy storage systems while mitigating tail-end safety risks during policy updates. A home energy management simulation environment based on real-world data is constructed to compare SAC PLUS， SAC Hybrid， and the DDPG Baseline under a normal test scenario and an intermittent high-temperature disturbance scenario. Experimental results demonstrate that SAC PLUS effectively reduces overall safety costs and tail-end safety risks in both test scenarios. Although its grid cost is slightly higher than that of SAC Hybrid， SAC PLUS still maintains better energy scheduling performance than the DDPG Baseline， indicating that it can achieve improved safety stability with an acceptable electricity cost trade-off. Furthermore， the trained model is deployed on an NVIDIA Jetson Xavier NX for edge inference， confirming its real-time computational capability on a low-power device. In addition， the HVAC control results are further integrated into an ECHONET Lite application scenario， preliminarily validating the feasibility of integrating the edge control strategy with smart home device communication protocols. Overall， the proposed SAC PLUS method can simultaneously address energy scheduling， safety risk mitigation， and edge deployment requirements， providing a feasible control solution for smart home energy management systems.
