改良型霍普菲爾類神經網路應用於中央空調系統冰水主機負載分配之最佳化

Optimal Chiller Loading of Central Air-conditioning System Using Improved Hopfield Neural Networks

陳武星(Wu-Hsing Chen)

指導教授 ： 張永宗

國立臺北科技大學/機電學院/機電科技研究所/博士(2011年)

摘要

中央空調系統無論在水側或空氣側探討其運轉最佳化方法有許多，由於霍普菲爾類神經網路能有效地求解組合式最佳化之問題，因此本論文提出霍普菲爾網路法來解水路系統多台冰水主機並聯運轉最佳化問題。在各種冰水主機負載分配最佳化之方法中，平均負載法最常應用但非最佳運轉方法。拉格蘭傑法對於冰水主機耗電與部分負載率組成曲線同時存在凹凸方向時，求解冰水主機負載分配最佳化會有發散問題。基因演算法求解過程應用複製、交配與突變等機制雖可克服拉格蘭傑法之缺點，但需編碼與解碼等運算，使程式之編寫相當麻煩。進化策略法簡化基因演算法之求解過程，採用突變與競爭技巧，省卻基因演算法的編碼、解碼與交配等操作，可避免誤差產生，使程式之編寫更容易，但由於其為隨機尋優的方法，每次執行的解結果將會有所不同，必須先取多次運算之平均值，避免不必要的盲目搜尋。連續型霍普菲爾演算法求解冰水主機運轉最佳化，經由改變主機負載率求解冰水主機負載分配最佳化。霍普菲爾法雖可以解決冰水主機耗電與負載率組成曲線同時存在凹凸方向之問題，但其類神經網路的S函數有曲線飽和與傾斜常數選擇不當之問題，本研究提出霍普菲爾線性模型來描述神經元之輸入-輸出關係，解決了上述S函數飽和之問題。本文所提出之方法分別應用於兩個案例，並分別與傳統之平均負載法、拉格蘭傑法與基因演算法比較，結果顯示本文之方法具有多項優越之性能。

關鍵字

冰水主機負載分配最佳化 ； 拉格蘭傑法 ； 霍普菲爾法 ； 線性模型

並列摘要

There have been many approaches to the investigation of the optimal operation of central air conditioning system based on either the water or air aspect of the system. Since the Hopfield Neural Network (HNN) can be effectively employed to solve combinatorial optimization problems, this dissertation proposes that the HNN be used to solve the water system multi-chiller parallel operation optimization issue. Despite not having the most optimized operation, Average Loading (AVL) is the most commonly employed approach among methods used to solve Optimal Chiller Loading (OCL) problems. When the Lagrangian Multiplier (LGM) method is applied to a curve composed of the power consumption and partial load ratio ( ) and exhibiting both concave and convex portions, divergence becomes an issue when attempting to solve the OCL problem. Although mechanisms such as reproduction, crossover, and mutation in the solution procedure of genetic algorithms (GAs) can overcome the shortcomings of the LGM method, the encoding and decoding operations complicate computer programming tremendously. Evolution strategy (ES) simplifies the solution process of GAs by adopting mutation and competitive techniques and dispensing with the encoding, decoding and crossover operations required for GAs, which helps both avoid errors and facilitate programming. However, since the technique is an optimization method based on randomness, each time the procedure is run, a different solution will result. It is therefore necessary to calculate the average of the solutions obtained from many different runs and to avoid unnecessarily aimless searches. The continuous HNN algorithm solves the optimal chiller operation problem by changing the chiller’s load ratio for its optimal solution. Although the HNN method can be used to solve the problem of a curve composed of the power consumption and load ratio exhibiting both concave and convex portions, the sigmoid function of its neural network suffers from the problems with curve saturation and inappropriate choice of incline constant. This dissertation proposes that the Improved Hopfield Neural Network（IHNN） be used to describe the input-output relations between neurons, thus solving the above saturation problem of the sigmoid curve. We provide two examples for which this method is applied, and comparisons are made with conventional AVL, LGM, and GA method. The results indicate that the approach adopted in this study possesses several excellent properties.

並列關鍵字

Hopfield ； Linear Model ； Optimal Chiller Loading ； Lagrangian Multiplier Method

參考文獻

[9] 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。

[16] 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，碩士論文，國立台北科技大學，台北，2004。

[17] 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，碩士論文，國立台北科技大學，台北，2004。

[1] 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. 95,1989, pp 663-675.

[3] Y.C. Chang,“A novel energy conservation method-optimal chiller loading,” Electric Power System Research, 2003, pp.221-226.