有效的冷鏈乳製品配送需在多項現實作業限制下，同步最佳化三維貨物裝載與多站式車輛路徑規劃。然而，現有多數整合方法多將裝載問題簡化為重量容量限制，或以序貫方式而非耦合方式處理裝載與路徑決策，使得溫度相容性、利潤導向訂單接受及碳排放成本等因素鮮少於單一整合框架中被共同納入考量。本研究提出一套三層式整合決策支援框架，將具現實作業限制之三維裝載模組與基於自適應大鄰域搜尋法（Adaptive Large Neighborhood Search， ALNS）之含時間窗容量限制車輛路徑問題（Capacitated Vehicle Routing Problem with Time Windows， CVRPTW）進行緊密耦合。裝載模組明確考量貨物穩定性、溫控區域相容性及依配送路徑卸貨順序之可及性，而路徑規劃則以利潤最大化為目標，同時納入訂單收益、配送成本、棄單懲罰與碳排放成本。此外，本研究導入近端策略最佳化（Proximal Policy Optimization， PPO）強化學習代理人，透過裝載與路徑層之雙向資訊交換，加速箱體放置決策並提升求解效率。 以臺灣北部一家中型乳製品配送公司為模擬情境，針對 10、100 與 1，000 筆訂單三種規模進行運算實驗。結果顯示，載重利用率皆高於 89%，卸貨可及性指標（UAS）高於 0.978，準時交貨率介於 91.7% 至 100% 之間；總配送利潤分別達 107，267、1，043，969 與 9，151，732 美元，計算時間介於 0.12 秒至約 31 分鐘，顯示本框架具備支援日常配送排程之實務可行性。敏感度分析進一步指出，需求水準為影響系統績效之主要因子，且呈現非單調之利潤變化特性，而時間窗寬度對配送成本具中度影響，車輛租賃成本對整體利潤影響相對有限。本研究之主要貢獻在於同步整合溫度相容性、異質三維裝載、配送時間窗、利潤驅動訂單接受及碳排放核算等五項關鍵限制維度，提升整合式冷鏈物流最佳化於實務應用之完整性與可行性，並超越多數現有研究之涵蓋範疇 關鍵詞： 冷鏈物流、三維裝載、車輛路徑、ALNS、PPO、碳排放;Efficient cold chain dairy distribution requires the simultaneous optimization of three-dimensional cargo loading and multi-stop vehicle routing under complex operational constraints. However， most existing integrated approaches either simplify loading as a weight capacity constraint or address loading and routing sequentially rather than jointly， leaving temperature compatibility， profit-oriented order acceptance and carbon emission accounting insufficiently considered within a unified framework. This study proposes a three-layer integrated decision support framework that couples a realistic 3D loading module with an Adaptive Large Neighborhood Search based capacitated vehicle routing problem with time windows. The loading module explicitly enforces cargo stability， temperature zone compatibility and route-aware unloading accessibility， while the routing layer maximizes distribution profit by jointly considering order revenue， delivery cost， dropped order penalties and carbon emission costs. A Proximal Policy Optimization reinforcement learning agent is further incorporated to accelerate box placement decisions through bidirectional information exchange between the loading and routing modules. Computational experiments based on a medium-sized dairy distributor in northern Taiwan were conducted across three order scales， including 10， 100 and 1，000 orders. Results show payload utilization consistently exceeding 89 percent， unloading accessibility scores above 0.978 and on-time delivery rates ranging from 91.7 percent to 100 percent. Total distribution profits reached USD 107，267， USD 1，043，969 and USD 9，151，732， respectively， with computation times ranging from 0.12 seconds to approximately 31 minutes， demonstrating practical feasibility for daily operational planning. Sensitivity analysis indicates that demand level is the dominant performance factor and exhibits nonmonotonic profit behavior， whereas time window width moderately affects distribution cost and vehicle rental cost has minimal influence on profitability. The main contribution of this study lies in the simultaneous integration of temperature compatibility， heterogeneous 3D loading， delivery time windows， profit-driven order acceptance and carbon emission accounting， thereby advancing the practical applicability of integrated cold chain logistics optimization beyond the scope of most existing studies. Keywords: Cold Chain Logistics， 3D Loading， Vehicle Routing， ALNS， PPO， Carbon Emissions
