Neural network model of the hottest batch process

Neural network model of batch process

sun zhenggui 1, Li yongbiao 2, Zhao Qingjie 3

(1. Science and Technology Department of Shengli Petroleum Administration, Shandong 257000; 2. Beijing Huasheng Dejia technology company,

Beijing 100084; 3. Computer Department of Tsinghua University, Beijing 1000 pilot production of low iron aluminum 84)

1 introduction

the traditional system identification method is to use people's understanding of physical processes, The model structure of the system is established through mechanism analysis, and then the parameters of the model are estimated by using the observed data. For linear systems or essentially linear systems, this method is more effective. But for complex nonlinear systems, this traditional identification method has encountered difficulties

using artificial neural networks to model the system, we don't need to make too many assumptions about the nature of the system itself, and we don't need to know too much about the internal mechanism of the system. Batch polypropylene process is nonlinear, time-varying, flammable and explosive. It is very difficult to obtain the process model by mechanism analysis or experimental methods. In this paper, a mathematical model of the reaction stage of intermittent polypropylene process is established by using feedforward neural network. The network is trained and tested with historical data

, and the test results meet the requirements

2 theoretical basis [3]

where the form of function f (·) adopts linear, log sigmoid or tan sigmoid form as required, I = 1, 2,..., NQ; j＝1，2，…，nq－1； q＝1，2，…，Q。

use the known samples to learn and adjust the connection weight coefficient of the network. Take the cost function of learning as

..., nq-1

back propagation learning algorithm is as follows:

3 Establishment of neural network model of actual process

3.1 input and output of process are beneficial to promoting mutual trust between law enforcement agencies and all parties

there are many factors affecting propylene polymerization reaction, such as the quality of main raw materials and auxiliary raw materials, the formula of various raw materials for polymerization, reaction control and equipment. In a period of time, it can be considered that the materials used in each batch are basically the same, and the influence of operating conditions is mainly considered, such as reaction temperature and time have a great impact on the polymerization reaction. For the intermittent polypropylene process, the pressure in the polymerizer is very close to the saturated vapor pressure of propylene, and the pressure is completely corresponding to the temperature [4]. Therefore, only the reactor pressure is taken as the output parameter here. By analyzing the production process, the relationship between input and output of the process is shown in Figure 3-1

from the actual process, the water temperature in the hot water tank is maintained at about 90 ℃ and the liquid level is maintained at 2/3. The temperature of circulating cooling water is basically unchanged for a period of time, and the material properties are basically the same. The rotating speed of the agitator in the polymerization kettle is constant, i.e. 60 R/min in fact. Therefore, figure

3-1 can be simplified to figure 3-2. In this way, the object can be simplified into a three input one output system. The opening of the water valve corresponds to the control variable, and the timing starts from the introduction of hot water after feeding. 3.2 neural network model of production process

according to the previous analysis, a three-layer forward model can be used to simulate the object, in which there are three neurons in the input layer and one neuron in the output layer. The three input signals of the nerve are reaction time, water temperature and valve opening; The output signal is the pressure in the kettle. According to the empirical formula, the number of neurons in the hidden layer is the number of neurons in the input layer, M is the number of neurons in the output layer, and a takes the number between 1 and 10) as 6. The network structure is shown in Figure 3-3

3. Data management and display function 3.3 training and testing of neural network

in order to make neural network act as a model of process, it must be trained. 105 groups are taken from the historical data of the production process as sample signals, of which 100 groups are used as teacher signals. The network parameters are trained by using the learning algorithm in part 2. The remaining sample data is used to test the trained network performance

before training the network, the input and output data should be normalized. For each group of sample data, the following calculations are made:

use the test samples mentioned above to test this result, and the results are shown in table 3-1. Among them, the samples are those who did not participate in training

from the above test results, it can be seen that the learned neural network has sufficient accuracy, and it is possible to use it as the mathematical model of the object. It is worth noting that the more training times, the more correct input-output mapping relationship can be obtained. In the experiment, it is found that if the number of training is increased on the current basis, the test error will become larger. This is OK