應用霍普菲爾類神經網路於冰水主機 負載分配之最佳化

Optimal Chiller Loading Using Hopfield Neural Network

李俊宏(Chun-Hung Li)

指導教授 ： 張永宗

國立臺北科技大學/機電學院/冷凍空調工程系所/碩士(2006年)

摘要

目前冰水主機的負載分配法有平均負載法、拉格蘭傑法、基因演算法，這些方法皆有其缺點。例如平均負載法並不是最佳運轉點、拉格蘭傑乘數法對於kW－PLR性能曲線為凸函數時，可求得最佳主機負載分配，但當kW－PLR性能曲線為凹凸函數共同存在時，拉格蘭傑乘數法卻無法求出最佳解。雖然基因演算法（Genetic Algorithm , GA）可用來克服此缺點，但其求解過程需應用到複製、交配、突變等機制及編碼、解碼等運算，使程式的編寫相當麻煩。
故本文提出霍普菲爾(Hopfield)法來求解以克服上述缺點。由於霍普菲爾採用陡降技巧，省去了GA 方法的編碼、解碼及交配等操作，不僅可避免誤差產生，且程式編寫更加容易。再分別與拉格蘭傑法、基因演算法以及平均負載法進行比較。
冰水主機負載分配最佳化之目的在於滿足空調系統負載需求條件下，決定每台冰水主機運轉之最佳負載率，而使系統總耗電量達到最小，而本文之研究方法為霍普菲爾，能比基因演算法有更高準確性的求出滿足負載需求下最佳的負載分配。

關鍵字

霍普菲爾 ； 類神經網路 ； 主機最佳負載分配 ； 拉格蘭傑乘數法

並列摘要

The chiller loading distribution methods include Average Loading (AVL) method, Largrangian Multiplier (LGM) method and Genetic Algorithm (GA) at present. Each of these methods has its own shortcomings. For example, AVL method may be the most popular used method but not the optimal one. Also, even the Lagrangian Multiplier Method（LGM） is able to get the optimal chiller loading distribution (OCLD) when the kW-PLR performance curve is convex function, it can not achieve the optimal solution when convex function and non-convex function exist together in the kW-PLR performance curve. Although the Genetic Algorithm (GA) method could overcome these shortcomings, its programming design is very complicated. It processes include reproduction, crossover, mutation, encoding and decoding etc.
On the other hand, the Hopfield method leaves out all of these processes in the GA approach and is able to adjust itself to figure out the optimal value. By using the Hopfield method, it can decrease the tolerance and make the programming much easier. As a result, the Hopfield is used as the main research method to compare with the methods of Average Loading (AVL), Largrangian Multiplier (LGM), and Genetic Algorithm (GA).
The main purpose of getting the OCLD is to not only minimize the system power consumption but also meet the HVAC system load and eventually achieve the optimal part load ratio (PLR) of each chiller. The results of the research paper indicates that the Hopfield method has better performance on finding out the most efficient distribution loading than the GA does.

並列關鍵字

Hopfield ； Artificial Neural Networks ； Optimal Chiller Loading ； Lagrangian Multiplier Method

參考文獻

[7] 黃傳興，「空調主機節約能源方法」，中國冷凍空調雜誌，第59-66頁，1992年6月。

[17] 江建樺，「應用模擬退火演算法於冰水主機負載分配之最佳化」，台北科技大學碩士論文，2004。

[18] 莊虔毅，「以基因演算法求解含凹凸特性曲線之冰水主機最佳負載分配」，台北科技大學碩士論文，2004。

[28] 張智星著，「MATLAB程式設計與應用」，清蔚科技股份有限公司，2002。

[3] J.E. Braun, S.A. klein, J.W. Mitchell and W.A. Beckman, “Applications of optimal control to chilled water systems without storage”, ASHRAE Transactions. Vol. 95. , pp 663-675, 1989.