粒子群演算法結合機器學習參數優化於考量輔具能源消耗之流線型生產排程

Abstract

隨著氣候變遷影響日益嚴重，如何在生產過程中有效降低能源消耗已成為一項重要課題。同時，加工過程中的能源系統穩定性亦不可忽視，特別是對能源消耗峰值的控制更是關鍵。此外，輔具資源限制在實際生產中相當常見，若在排程規劃時忽略此類限制，可能導致排程結果與實際運作出現顯著落差。因此，若能在排程過程中同時納入能源消耗與輔具資源的考量，將能使規劃結果更貼近現實，進而有效提升生產效率並降低能源浪費。 為因應上述挑戰，本研究開發一套粒子群演算法，並針對問題特性設計兩階段編碼機制：第一階段編碼代表工件的生產優先順序，第二階段則對應每個工件所搭配的輔具配置。此外，本研究亦提出結合多層感知器（Multilayer Perceptron, MLP）以最佳化粒子群演算法參數的方法，有效提升演算法的收斂表現。該方法亦具有良好的泛用性，可應用於其他需要參數調整的啟發式演算法中。

As the effects of climate change intensify, reducing energy consumption during production has become a critical challenge. Equally important is maintaining energy system stability throughout machining operations particularly by controlling peak demand. Moreover, auxiliary resource constraints are commonplace in real-world manufacturing; overlooking them during scheduling can create a wide gap between planned and actual shop-floor performance. By simultaneously accounting for energy use and auxiliary resource availability, scheduling decisions become more realistic, thereby boosting efficiency and curbing energy waste. To tackle these issues, this study develops a particle swarm optimization (PSO) algorithm with a problem-specific two-stage encoding scheme: Stage 1 encodes the production priority of each job, while Stage 2 assigns the corresponding auxiliary resource to each job. We additionally integrate a multilayer perceptron (MLP) to automatically tune PSO parameters, greatly improving convergence. This tuning framework is generic and can be transferred to other heuristic algorithms that require parameter calibration.