在全球能源成本上升、淨零碳排與綠色供應鏈要求日益提高之背景下，高科技廠房之廠務系統節能已成為提升企業永續競爭力之重要課題。半導體與高科技製造廠房需長時間維持嚴格之溫度、濕度與潔淨度條件，使暖通空調系統（Heating， Ventilation， and Air Conditioning， HVAC）成為廠務端主要耗能來源之一。其中，中溫冰水系統（Medium Temperature Chilled Water， MTCW）涉及冰水主機、冰水泵、冷卻水泵與冷卻水塔等多項設備，其能耗受外氣條件、系統負載與控制參數共同影響，具有明顯之非線性耦合特性。MTCW實務上採用穩定、簡單與易維護之規則式控制，但多依賴固定設定值與工程經驗，較難即時掌握不同氣候與負載條件下之全域較佳操作點，不僅可能導致系統長期偏離節能運轉區間，更需要長時間在線調校才能完整啟動。 為解決新建或改造高科技廠房於初期缺乏完整歷史運轉資料之冷啟動問題，本研究建立一套基於數位雙生之人工智慧（Artificial Intelligence， AI）最佳化控制分析架構。研究首先透過氣象資料、負載生成假設與設備熱力學模型建立數位雙生物理能耗模型，生成涵蓋不同外氣條件、熱負載狀態與控制參數組合之合成訓練資料。接著分別建立隨機森林迴歸模型（Random Forest Regressor， RFR）與 Efficient Kolmogorov-Arnold Network（E-KAN）作為能耗推論模型，用以預測系統總耗電量，並作為後續控制參數搜尋與控制策略比較之基礎。 本研究設計四組控制策略進行比較，包括規則式控制作為基準組（Group R）、RFR 結合候選控制集合搜尋之機器學習策略（Group ML）、E-KAN 結合具上下界限制之有限記憶 BFGS 最佳化方法（limited-memory Broyden-Fletcher-Goldfarb-Shanno with bounds， L-BFGS-B）之深度學習策略（Group DL），以及 E-KAN 結合近端策略優化（Proximal Policy Optimization， PPO）之強化學習控制策略（Group RL）。為評估各策略於不同氣候條件下之節能效果與工程可行性，本研究建構真實夏季、真實冬季、合成夏季與合成冬季等四組 168 小時閉迴路測試情境，並以系統總耗電量、相對節能率、控制參數變化與邊界合規性作為主要評估指標。其中，邊界合規性以冰水回水溫度及冷卻水側溫度是否滿足操作上限作為判斷依據，以避免控制策略在追求節能時犧牲系統供冷能力與運轉可靠度。 研究結果顯示，RFR 與 E-KAN 皆具備作為中溫冰水系統 AI 能耗推論模型之能力。於控制策略比較方面，相較於 Group R，其他三種 AI 控制策略於四組測試情境中皆可降低系統總耗電量。整體而言，Group DL 於多數情境下呈現較佳節能效果，Group ML 具備穩定正節能與工程可實作性，Group RL 則展現強化學習控制策略之節能潛力，但其邊界合規表現仍受 reward 設計與安全約束處理影響。綜上所述，本研究透過一致之數位雙生測試環境，比較不同控制策略於節能效果、控制穩定性與邊界合規性上的差異，可作為高科技廠房中溫冰水系統導入 AI 最佳化控制之工程評估基礎。
關鍵字：數位雙生、中溫冰水系統、能耗推論模型、隨機森林迴歸、E-KAN、L-BFGS-B、近端策略優化、邊界合規性;In response to rising global energy costs， net-zero carbon targets， and increasing green supply chain requirements， energy saving in facility systems has become an important issue for improving the sustainability competitiveness of high-tech manufacturing plants. Semiconductor and high-tech manufacturing facilities must continuously maintain strict temperature， humidity， and cleanliness conditions， making heating， ventilation， and air conditioning (HVAC) systems one of the major energy consumers in facility operations. Among them， medium temperature chilled water (MTCW) systems involve multiple pieces of equipment， including chillers， chilled water pumps， cooling water pumps， and cooling towers. Their energy consumption is jointly affected by outdoor air conditions， system cooling load， and control parameters， resulting in significant nonlinear coupling characteristics. Although conventional rule-based control is stable， simple， and easy to maintain， it often relies on fixed setpoints and engineering experience， making it difficult to capture globally preferable operating points under varying seasons and load conditions. As a result， the system may operate away from the energy-efficient region for extended periods. To address the cold-start problem of insufficient historical operating data in newly constructed or retrofitted high-tech facilities， this study develops a digital twin-based AI optimization control framework for energy inference and control strategy comparison in MTCW systems. First， a digital twin physical energy model is established using weather data， load generation assumptions， and thermodynamic equipment models. The model is used to generate synthetic training data covering different outdoor air conditions， cooling load states， and control parameter combinations. Subsequently， a Random Forest Regressor (RFR) and an Efficient Kolmogorov-Arnold Network (E-KAN) are developed as energy inference models to predict total system power consumption， which serves as the basis for subsequent control parameter search and control strategy comparison. Four control strategies are designed and compared in this study: conventional rule-based control as the baseline group (Group R)， Random Forest Regressor (RFR) combined with candidate control set search as the machine learning strategy (Group ML)， Efficient Kolmogorov-Arnold Network (E-KAN) combined with the limited-memory Broyden–Fletcher–Goldfarb–Shanno with bounds (L-BFGS-B) continuous optimization method as the deep learning strategy (Group DL)， and E-KAN combined with Proximal Policy Optimization (PPO) as the reinforcement learning strategy (Group RL). To evaluate the energy-saving performance and engineering feasibility of each strategy under different climatic conditions， four 168-hour closed-loop testing scenarios are constructed， including real summer， real winter， simulated summer， and simulated winter conditions. The main evaluation indicators include total system energy consumption， relative energy-saving rate， control parameter variation， and boundary compliance. Boundary compliance is assessed based on whether the chilled water return temperature and cooling-water-side temperature satisfy their operating upper limits， thereby preventing control strategies from sacrificing cooling capacity and operational reliability while pursuing energy savings. The results show that both RFR and E-KAN are capable of serving as AI-based energy inference models for the MTCW system. In terms of control strategy comparison， the other AI-based control strategies reduce total system energy consumption in the four testing scenarios compared with Group R. Overall， Group DL shows better energy-saving performance in most scenarios， Group ML demonstrates stable positive energy savings and practical engineering implementability， while Group RL shows the potential of reinforcement learning-based control but remains affected by reward design and safety constraint handling in terms of boundary compliance. Overall， this study compares different control strategies under a consistent digital twin testing environment in terms of energy-saving performance， control stability， and boundary compliance. The results provide an engineering evaluation basis for applying AI optimization control to MTCW systems in high-tech manufacturing facilities. Keywords: Digital Twin， Medium Temperature Chilled Water System， Energy Inference Model， Random Forest Regressor， E-KAN， L-BFGS-B， Proximal Policy Optimization， Boundary Compliance.
