【正文】 ence on ：22－24[18] Sharad Bhartiya. Benefits of factorized RBFbased NMPC. [J].Computers and Chemical ：90－93[19] Bengt Fornberg. Stabilization of RBFgenerated finite difference methods for convective PDEs. [J].Journal of Computational ：34－37 附錄A 1 neural network neural network principleThe human brain saves the information is distributional saves between the brain cells in the connection, but is not the preservation in brain cell in. The brain cell the function relations (for example drive and suppression) save through them between. The manpower simulates this kind of mapping relations the system to be called (artificially) neural network (ANN). The neural network is one has the highly nonlinear ultra largescale runon time dynamic system, is network which (neuron) the widespread interconnection forms by the massive processing unit. It is proposes in the modern neuroscience research results foundation, has reflected the brain function basic characteristic. But it is not the human brain real description, but only is its some kind abstract, the simplification and the simulation.The network information processing does mutually by the neuron between uses for to realize, the knowledge and the information memory performance for the circuit ponent interconnection the distributional physical relation, the network study and the putation decides in various neurons connection power department39。s dynamic evolutionary process. Among them, the neuron constituted the network fundamental operation unit, each neuron has own cuts off from the value, after if each neuron input signal is all the connected neuron output signal and the weighting sum. But the output signal is its net input signal nonlinear function. If the input signal weighting set is higher than its idle value, this neuron is then activated outputs the corresponding value. Is between the unit saves which in the artificial neural networks connects the weighted value array.The neural network work process mainly is posed by two stages:A stage is the work time. This time each connection weight is fixed, the puting element change of state, achieves the steady state in order to. Another stage is studies the time (autoadapted time, or design time). This time each puting element condition is invariable, each connection weight may revise (through study sample or alternative means). The preceding stage is quick, various units condition also calls the shortterm memory (S TM). Latter stage slow many, the power and the connection way also called records (L TM) for a long time.At present the neural network structure has near hundred kind of many, the algorithm is unable to register. According to the network characteristic, the neural network may divide into the static state and the dynamic two kinds approximately. Static network current output merely reflection current datain processing result. The dynamic network has the memory ability network, memory ability may be because the neuron transfer function is the differential or the difference equation causes。Also may be because the network output or the network internal condition feed back to the network input end produce. Below regarding some mon makes the brief introduction in the control system network architecture and the algorithm. Neural network structure type1. Neural network basic structureThe neural network is connects the constitution mutually by the massive simple neurons the plex network. Figure 21 is a typical monolayer neural network model, it has the R Uygur to input, S neuron. p is the RXI Uygur39。s input vector, the network level by weight matrix W(SxR), shuts value vector b(Sxl), the summation unit. With the transfer function arithmetical unit f position, S neuron output has posed Sxl Uygur39。s neural network output vector a. Among them, input level network weight matrix W and the valve value vector b concrete form is as follows:In the monolayer neural network foundation may the structure multilayer neural network. A typical three neural network model2. Neural network classificationNeural network type many and varied, they are from the different angle to the biology nervous system different level abstract and the simulation [18]. Generally speaking, after neuron model determination, a neural network characteristic and the function mainly are decided by network topology and the study method. Mainly may divide into the forward feed according to the network topology and the feedback two kinds “19,13”.(1) Feeds the network. Various neurons accept the preceding input, and outputs for next, does not have the feedback, the point to divide into two kinds, namely the input unit and the puting element, each puting element may have many inputs, but only then an output (it may take its input willfully pale pinkish purple to many other points). The usual forward feed network may divide into the different level, the ith input is only connected with a ith 1 output, the input and the output point and the outside are connected, but other intermediate levels are called the implicit strata. (2) Feeds the network. All points all are the puting elements, simultaneously also may accept the input, and to the outside output, may draw bees one not to have to the chart, like Figure 24(a), in which each connection arc all is bidirectional, also may draw bees like Figure 24(b) form. If the cargo certificate number is n, then each point has a n1 input and an output.The neural network work process mainly divides into two stages:The first stage studies the time, this time each puting element condition is invariable, connects respectively the weight may revise through the study。The second stage is the work time, this time each connection power is fixed, the puting element change of state, achieves some kind of stable shapeLooking from the functio