【正文】 getable is (5)where is produced by frostbiting, while by deteriorating .When temperature increases , decreases and increases .Both of them are related to the storage time , thus (6)where decreases and increases respectively when the temperature increases , denotes the time of entering the storage, while denotes the storage time, then we have (7)For fruit or vegetable, its optimum storage temperature should satisfy the following equation (8)Let the gravity of fruit or vegetable be ,its storage loss , then the total storage loss in a unit time interval is (9)Let denote the optimum storage temperature in general .It should satisfy (10)that is, (11)The calculation of in above formulae with traditional methods is time consuming. Hence we use an RBF neural network to acplish the solution of . This RBF neural network is the first part of the twostage RBF neural network proposed in the paper .It has only one output , ,and inputs, that is , and ,.hidden units are used here .Equation(11) is used to produce enough training samples. THE ONLINE PREDICTION OF THE COLD STORAGE TEMPERATUREOne of the key problems of the optimum control over the storage temperature is how to predict the temperature accurately. Because of their robustness ,the prediction methods based on neural networks have attracted more and more attentions. BP neural network is a kind of earlier used neural network for this purpose .But its training time is usually too long, and it has many local minimum points. Thus the RBF neural network has attracted more and more attention thanks to its higher training speed. This paper employs a twostage RBF neural network to predict the storage temperature..In the prediction process, the coupling relation between the temperature and the humidity should be taken into account. The paper selects the output variables in a way that the set of the variables include the temperature variables and the humidity variables at the same time. The choosing of the input variables should be taken into account no matter whether the control is performed or not, with the following two different cases involved:Case l: Automatic control system is offSuppose that there are R operating variables of the cold storage and state variables .Consider a time window posed of time points, (12)Use and to denote the values of and at time point respectively . Let (13) (14)Where . The task of the prediction is to determine of (14)according to the vector of (13) .For the current time ,all of the measured results can be used to construct the inputs of the prediction network. Suppose that all of the operating variables and state variables can be got by measuring ,and their values in the future are unknown. To construct a prediction sample ,the related time should satisfy .Otherwise, unknown values would be contained in the sample which would be unreasonable.Suppose that enough samples have been got .First, calculate the parameters of the hidden units, then calculate the prediction value of the storage temperature.Case2 :Automatic control system is onAt this time, the set of the input variables only contains the environmental temperature, humidity and quantum of the stored fruits and vegetables ,etc. Any of the input variables doesn’t appear in the control algorithm ,while the prediction variables are the stable values of the state variables. The nonlinear mapping function of the RBF neural network is used to design the stable models. When the stable values of the state variables have been obtained, the control algorithm is used to calculate the temperature of the storehouse, thus the set of the predicted variables wouldn’t contain any variable to be controlled. That’s why the set of the predicted variables and the set of the controlled variables under Case2 are different from those under Case1. ONLINE OPTIMUM CONTROL OF THE COLD STORAGE TEMPERATUREThe mon PID control algorithm of a variable unit takes the following form （15）Where and are the initial value and the current value of the controlled variable respectively . is the difference between the assigned value and the real value of the control object, that is （16）where and are the real value at time point and the assigned value of the control object respectively. Write equation (15) in the incremental form ,then we have （17）Where is the integral coefficient, is the differential coefficient .Write the above equations in another form, then we have （18）Under the case of having got the predicted value of the controlled variable ,equations(17)and(18)should be changed .Let denote the current time ,and suppose that the predicted values at the instants and of variable with RBF neural network are and respectively ，Let （19）Combine the historic values with the predicted values of the variable to calculate the right side of equation(18).Let（20） （21） （22） In this way ,equation (18) is changed into the following form （23）The values of in above equations should satisfy ， ， （24）Hence there are only 6 independent coefficients to be determined. Choose them as the condition to determine them is that they should let the mathematical expectation of get its minimum, that is ,we have the following equation （25）with the following constraint conditionAll of the initial values of can be chosen as .6. APPLICATIONThe methods proposed in the paper have been used for the optimum control over the temperature of a cold storage for fruits and vegetables. Table 1 lists the daily storing losses of the fruits and vegetables before and after the methods