隨著能源轉型、再生能源併網比例提升以及工業用電成本持續攀升,如何透過智 慧化能源管理提升用電效率,已成為工業型微電網發展之重要課題。其中,儲能系統 可藉由跨時段能量搬移、削峰填谷與需量管理等方式,有效降低尖峰購電成本並提升 系統調度彈性。然而,儲能調度決策同時受到電池容量、充放電功率限制、荷電狀 態、時間電價以及契約容量等多重條件限制,使其成為一項具多時段耦合、非線性與 長期決策特性之複雜最佳化問題。 本研究以實際工業用戶年度用電資料為基礎,建立一套工業型微電網儲能控制模 擬平台,並導入磷酸鋰鐵電池之操作限制,建構符合工程實務之儲能系統模型。研究 首先設計兼顧時間電價與契約容量限制之 Rule-based 控制策略作為基準模型,其後 進一步導入深度強化學習方法—近端策略最佳化(Proximal Policy Optimization, PPO),建立智慧化儲能最佳化控制策略,透過與環境互動自主學習最佳充放電決策, 以最小化長期總電費成本。 在模型驗證方面,本研究採用 2016 年資料進行訓練,並固定模型參數,直接套 用於 2017 至 2019 年資料進行樣本外驗證,以檢驗模型泛化能力。研究結果顯示, 相較於傳統 Rule-based 控制策略固定約 0.56% 之節費率,所提出之 PPO 控制策略 於 2017、2018 與 2019 年分別達到 0.70%、0.95% 與 1.30% 之節費率,顯示其在不 同年度負載條件下仍具穩定之正向經濟效益。此外,契約容量敏感度分析結果指出, 適當提升契約容量可增加低價時段充電彈性,使儲能系統於高價時段釋放更多可用能 量,最高可達 1.512% 之節費率,展現人工智慧控制策略於跨時段能量配置上的優 勢。 綜合而言,本研究驗證以深度強化學習應用於工業型微電網儲能控制之可行性與 經濟效益,並證明智慧化控制策略相較傳統規則式方法,具備更佳之自適應能力、長 期最佳化潛力與實務應用價值,可作為未來整合再生能源、需求反應及智慧能源管理 系統之重要技術基礎。;Industrial microgrids increasingly rely on Energy Storage Systems (ESS) to improve energy efficiency and reduce electricity costs under Time-of-Use (TOU) pricing. However, ESS scheduling is constrained by battery capacity, charge/discharge limits, State of Charge (SOC), and contracted demand, making optimal control a complex sequential decision problem. This study proposes an artificial intelligence-assisted ESS control strategy based on Proximal Policy Optimization (PPO) for industrial microgrid energy management. A simulation environment was established using real industrial load data, incorporating practical operating constraints of Lithium Iron Phosphate (LFP) batteries, TOU pricing, and contracted capacity limits. A conventional Rule-based control strategy was used as the benchmark for comparison. The PPO model was trained using 2016 operational data and validated on out-of-sample datasets from 2017 to 2019 without retraining. Results show that the proposed method consistently outperformed the Rule-based strategy, achieving annual cost reductions of 0.70%, 0.95%, and 1.30%, respectively. Sensitivity analysis further indicates that proper contracted capacity settings can enhance charging flexibility and improve overall economic performance. The results demonstrate that deep reinforcement learning provides a feasible and effective approach for industrial ESS scheduling, offering better adaptability and higher long term optimization potential than conventional control methods.