【正文】 the controlled variables under Case2 are different from those under Case1. ONLINE OPTIMUM CONTROL OF THE COLD STORAGE TEMPERATURE The mon PID control algorithm of a variable unit takes the following form 0( ) { ( ) ( ) [ ( ) ( 1 ) ] }ksdcsiisTTu k K e k e i e k e k u== + + +229。 = + 182。 Xue Guoxin Jiangsu Institution of Petrochemical Technology, Changzhou 213016, (Received November 26, 1999) Abstract :In recent years ,advanced control technologies have been for the optimum control of a cold storage. But there are still a lot of shortings. One of the main problems is that the traditional methods can’t realize the online predictive optimum control of a refrigerating system with simple and valid algorithms. An RBF neural work has a strong ability in nonlinear mapping, a good interpolating value performance, and a higher training speed. Thus a twostage RBF neural work is proposed in this paper .Combining the measured values with the predicted values , the twostage RBF neural work is used for the online predictive optimum control of the cold storage temperature. The application results of the new methods show a great success. Keywords: RBF neural work, Cold storage, Online prediction, optimum control. 1. INTRODUCTION The predictive optimum control of cold storage temperature has found a wide application in a agricultural engineering, especially for keeping fruits and vegetables fresh by cold storage. All of the currentlyused temperature control units face the problems on how to choose the optimum temperature as the controlled object, how to predict the temperature variation of the refrigerating storehouse and how to realize the optimum control. A lot of study efforts have been made. The earlier methods were based on the Taylor’s series theory and the PID control algorithm[1,5].Later, Furrier’s transformation method, Chebyshev’s theory and knowledge based system were used and better results were got [2,3].In recent years ,BP neural works have been used for the optimum control of the cold storage temperature .A BP neural work has a good performance of nonlinear mapping, but it has too many local minimum points, and usually its training speed is too slow[2,5]. Hence it couldn’t be used for online control calculation conveniently .This paper proposes a two stage RBF neural work to realize the online optimum control of the cold storage temperature. The first stage is used to determine the current optimum refrigerating temperature of the system, and the second is used to predict the temperature values in the ing time points .Furthermore, an optimum problem is solved, whose solution is used to direct the action of the refrigerating system.. 2. A TWO STAGE RBF NEURAL NETWORK A twostage RBF neural work is adopted. The first stage is used to determine the optimum value of the cold storage temperature, and the second is used to predict the temperature. Generally, suppose that there are n input variables 1x ,… , nx and m output variables 1y ,… , my . Let 1( ,..., )Tnx x x= (1) 1( ,..., )Tmy y y= (2) Using RBF Neural Network for Optimum control of a Cold Storage where x denotes a point in the n dimensional input space nR ,while y denotes a point in the m dimensional output space mR ,Suppose that the number of the hidden units is H .Every hidden unit uses two parameters, one is scalar quantity ( 0)hs ,the other is vector ()hx .Suppose that the set of the training samples is ( ) ( ){ ( , )}kkx y k K＃ .Generally, n H K＃ should be satisfied. RBF neural works are based on the interpolating value performance of radiusbased functions. To improve this performance ,the following equation is used to calculate the j the output of an RBF neural work. ^ 221^ 221()e xp( ),1()e xp( )Hjhh hjH hh hxxy j mxxv ss=== ＃229。以這種方式實(shí)現(xiàn)溫度的在線優(yōu)化控制 ,并得到了滿意的結(jié)果。 每天總損失率可以根據(jù)以下公式計(jì)算1N iiil wl==229。表1列出了水 果和蔬菜的日常存儲(chǔ)損失之前和之后使用本文提出的方法。1 1 。D = + + + + + + + + （ 22） 用這個(gè)方法，方程（ 18）可以變成一下格式 2( ) ( ) ( ) ( )c i du k K e k K e k K e k儋D = D + + D （ 23） 上式中 1 1 1 2 2 2 3 3 3, , , , , , , ,a b g a b g a b g的值應(yīng)該滿足 { 1 1 11 1 1 10 , , 1a b ga b g+ + =＃ ， { 2 2 22 2 2 10 , , 1a b ga b g+ + =＃ ， { 3 3 33 3 3 10 , , 1a b ga b g+ + =＃ （ 24） 因此 系統(tǒng)中 只有 6獨(dú)立系數(shù)待定。 kt 表示當(dāng)前時(shí)間 ，并且設(shè)在 1kt+ 和 2kt+ 時(shí)刻變量 v 的 預(yù)測(cè)的值 分別是 ( 1)vk+ 和 ( 2)vk+ ，令 ( 1) ( 1)e k v kt+ = + ( 2) ( 2)e k v kt+ = + （ 19） 結(jié)合歷史值和變量的預(yù)測(cè)值計(jì)算方程 (18) 的右邊。 在線最優(yōu)控制的冷藏溫度 普通 PID 控制算法的一個(gè)變量單位需要以下 公式 0( ) { ( ) ( ) [ ( ) ( 1 ) ] }ksdcsiisTTu k K e k e i e k e k u== + + +229。任何輸入變量不出現(xiàn)在控制算法 ,而預(yù)測(cè)變量是穩(wěn)定狀態(tài)變量的值。為了 構(gòu)建一個(gè)預(yù)測(cè)樣本 ， 相關(guān)的時(shí)間 t 應(yīng)該滿足公式 0t t Q t?D。本文選擇輸出變量 , 在同一時(shí)間 內(nèi) 設(shè)置包括溫度變量和濕度變量 。但它的訓(xùn)練時(shí)間通常是太長(zhǎng) ,和它有很多局部最小值點(diǎn)。 Hn= 在這里作為 隱藏的單位使用 ，方程（ 11） 用于產(chǎn)生足夠的訓(xùn)練樣本。 =抖 (10) 就是 ( 1 ) * 0 ( 2 ) * 01[ , ] [ , ]{ } 0n i i i i i iiif T t t f T t tg TT=?? +=抖229。 = + 182。這兩個(gè)數(shù)據(jù)都和存儲(chǔ)時(shí)間 t 相關(guān)。 為一種特殊的水果或蔬菜就進(jìn)入冷藏 ， 它的最佳儲(chǔ)存溫度可以得到正交實(shí)驗(yàn)方法。 在線計(jì)算的冷藏溫度 選擇的目標(biāo)價(jià)值冷藏溫度 ,需要綜合考慮所有的因素。 ^ 221^ 221()e xp( ),1()e xp( )Hjhh hjH hh hxxy j mxxv ss=== ＃229。 假設(shè)的訓(xùn)練樣本集 是 ( ) ( ){ ( , )}kkx y k K＃ 。 采用 RBF 神經(jīng)網(wǎng)絡(luò)分為兩個(gè)階段。因此它不能方便 地 用于在線控 制計(jì)算。大量的工作研究了前面的方法是基于泰勒級(jí)數(shù)理論和 PID 控制算法 [1,5]。新方法的應(yīng)用效果顯示一個(gè)巨大的成功。的一個(gè)主要問(wèn)題是 ,傳統(tǒng)方法不能實(shí)現(xiàn)在線預(yù)測(cè)最優(yōu)控制制冷系統(tǒng)的簡(jiǎn)單而有效的算法。但仍有許多缺點(diǎn)。將測(cè)量值與預(yù)測(cè)值 ,兩級(jí) RBF 神經(jīng)網(wǎng)絡(luò)用于在線預(yù)測(cè)最優(yōu)控制的冷藏溫度。所有的 currentlyused 溫度控制單元面臨如何選擇最適溫度為控制對(duì)象的問(wèn)題 ， 如何進(jìn)行冷藏庫(kù)溫度的變化 ，和如何實(shí)現(xiàn)最優(yōu)控制。 BP 神經(jīng)網(wǎng)絡(luò)具有良好的非線性映射的性能 ,但它有太多的地方 并不是那么理想 ,通常 是 其訓(xùn)練速度太慢 了 ( 5)。此外 , 他的解決方案是用于制冷系統(tǒng)的直接行動(dòng) ， 一個(gè)最 難的 問(wèn)題是解決了 。每個(gè)隱單元使用了兩個(gè)參數(shù) ,一個(gè)是標(biāo)量 ,另一個(gè)是矢量 。為了改善性能 ,使用下列方程計(jì)算 出 RBF 神經(jīng)網(wǎng)絡(luò)的輸出 j 。通過(guò)這種方式 ,改進(jìn)的 RBF 神經(jīng)網(wǎng)絡(luò)具有更好的性能。然而 ,蒸發(fā)溫度顯然是在冷藏條件下的溫度對(duì)象的限制。當(dāng)環(huán)境溫度升高了， (1)iL 降低但是 (2)iL 會(huì)升高。 (7) 對(duì)于水果或者蔬