Ongoing Research Project

Navigating Towards Smart Sustainable Logistics: Integrating Artificial Intelligence and Real-World Data in Green Vehicle Routing

邁向智慧永續物流：結合人工智慧與真實數據於綠色車輛路徑規劃

Duration
2025–2028
Funder
National Science and Technology Council (NSTC)
Grant no.
114-2628-E-002-009-MY3

A three-year project building data-driven, emission-aware Green Vehicle Routing (GVRP) for last-mile logistics — fusing real traffic and road-gradient data with vehicle-propulsion, refrigeration, and refrigerant-leakage emissions, plus AI optimization — to cut delivery carbon while keeping routes operationally practical. Validated on a real northern-Taiwan cold-chain network.

Preliminary results

  1. Emission-aware routing cut cold-chain last-mile carbon by up to ~35% versus drivers' existing routes and consistently beat distance-minimizing VRP — confirming that the shortest route is rarely the lowest-emission one, since the greenest path avoids steep grades and congestion even when slightly longer.

    以排放為導向的路徑規劃，使冷鏈最後一哩碳排放較駕駛員既有路線最高降低約 35%，並穩定優於以距離最小化為目標的 VRP——證實最短路線通常並非碳排最低，因為最環保的路線會避開陡坡與壅塞，即使距離略長。

  2. Traffic and road gradient must be modeled together: the full model combining real Google-API traffic speeds with continuous road gradient gave the largest and most stable reductions across problem sizes, whereas gradient-only or traffic-only variants sometimes did worse than plain VRP.

    交通與道路坡度須一併建模：同時納入 Google API 真實交通車速與連續道路坡度的完整模型，在各種問題規模下皆取得最大且最穩定的減量；僅納入坡度或僅納入交通的版本，有時甚至不及一般 VRP。

  3. The framework integrates vehicle-propulsion (CMEM), refrigeration (ASHRAE), and refrigerant-leakage emissions with traffic and tunnel/bridge-smoothed elevation data, solved at scale — 223 convenience stores from a single Ruifang hub — by an Adaptive Large Neighborhood Search, and stays within ~3% of its optimum under ±5–15% emission-estimate uncertainty.

    框架整合車輛推進（CMEM）、冷凍（ASHRAE）與冷媒洩漏排放，並結合交通資料與經隧道／橋樑平滑處理的高程資料，以自適應大鄰域搜尋（ALNS）求解大規模問題（單一瑞芳場站、223 家便利商店）；於 ±5–15% 排放估算不確定性下，結果仍維持在最佳值約 3% 以內。

About this project

Reaching net zero in freight hinges on the last mile — the high-frequency, stop-heavy final leg from depot to customer that is among the most polluting and least efficient parts of any supply chain. Cold-chain last-mile delivery is harder still: refrigerated trucks burn fuel both to move and to stay cold, yet most Green Vehicle Routing Problem (GVRP) studies simplify away the very things that drive emissions — real traffic speeds, road gradients, refrigeration load, and refrigerant leakage.

This three-year project builds a data-driven, emission-aware GVRP that closes that gap, in partnership with a Taiwanese third-party logistics (3PL) operator.

Path-level emission modeling. Rather than approximating emissions from straight-line distance, the framework estimates them along the actual driven path. Vehicle-propulsion emissions use the Comprehensive Modal Emissions Model (CMEM) — sensitive to speed, load, and gradient — while refrigeration load (ASHRAE-based) and R404A refrigerant leakage are modeled explicitly. Real traffic-speed profiles and continuous road elevation come from the Google Directions and Elevation APIs, with cubic-spline smoothing to strip tunnel and bridge artifacts.

Solved at real scale. The problem is cast as an energy-minimizing GVRP and solved with a custom Adaptive Large Neighborhood Search (ALNS) metaheuristic, validated on a real northern-Taiwan cold-chain network — 223 convenience stores served from a single Ruifang distribution centre, where a ~100 m elevation drop into the Taipei basin makes terrain genuinely matter.

A three-year arc. Year one (above) establishes the last-mile, single-objective foundation. Year two extends to a bi-objective model — carbon and operating cost, with a marginal-abatement-cost lens — and to first-mile pickup-and- scheduling. Year three integrates both into a two-echelon (2E-GVRP) network and brings in AI-assisted optimization — reinforcement learning and learning-based methods — for large-scale, dynamic routing.

貨運要邁向淨零，關鍵在於最後一哩——由配送中心送達終端客戶的最後一段， 具高頻次、頻繁停靠的特性，是供應鏈中污染最高、效率最低的環節之一。而冷鏈 最後一哩更為棘手：冷藏車不僅要行駛、還須持續製冷，但多數綠色車輛路徑問題 （GVRP）研究往往簡化了真正驅動排放的因素——真實交通車速、道路坡度、冷凍 負載與冷媒洩漏。

本三年期計畫與臺灣一家第三方物流（3PL）業者合作，建構一套資料驅動、 以排放為核心的 GVRP 框架，以填補上述落差。

路徑級排放建模。 框架不以直線距離近似排放，而是沿實際行駛路徑估算。 車輛推進排放採用全面性車輛排放模型（CMEM）——對車速、載重與坡度敏感 ——並明確納入冷凍負載（依 ASHRAE 準則）與 R404A 冷媒洩漏。真實的 交通車速剖面與連續道路高程取自 Google Directions 與 Elevation API， 並以三次樣條平滑移除隧道與橋樑造成的雜訊。

以真實規模求解。 問題建構為能源最小化的 GVRP，並以客製化的自適應大鄰域 搜尋（ALNS）演算法求解，於臺灣北部實際冷鏈配送網路驗證——單一瑞芳場站 服務223 家便利商店；瑞芳與臺北盆地間約 100 公尺的高程差，使地形成為不可 忽視的因素。

三年研究軸線。 第一年（如上）建立最後一哩、單目標的基礎。第二年延伸為 雙目標模型——同時考量碳排放與營運成本，並引入邊際減量成本觀點——並擴展至 第一哩集貨與排程。第三年將兩者整合為兩階段（2E-GVRP）網路，並導入 人工智慧輔助最佳化——強化學習與學習式方法——以因應大規模、動態的路徑規劃。

Top