Three, summary of the invention
Intelligence control method of the present invention is: the intelligence control method of ethylene rectification tower in the ethylene unit, select to influence the process operation parameter of ethylene rectification tower overhead ethane concentration and tower still ethylene concentration, and with its normalization method, utilize fuzzy GMDH neural network model separately to calculate then, the model output valve is through after the anti-normalization method, online " rollings " correction is carried out in the model output of the manual analysis value of utilizing ethylene rectification tower overhead ethane concentration and tower still ethylene concentration after to anti-normalization method, thereby obtains the soft observed value of overhead ethane concentration and tower still ethylene concentration; At last ethylene rectification tower is inferred the control (see figure 1) in real time according to the soft observed value of overhead ethane concentration and tower still ethylene concentration.Promptly, realize the deduction control of overhead ethane concentration by the produced quantity of ethane concentration adjustment ethylene product in the extraction of cat head ethene; And promptly improve or reduce the ethylene concentration of tower still temperature in can the extraction of control tower still ethane by ethylene rectification tower in the ethylene unit being increased or reduces heating propylene flow.
Technical solution of the present invention:
The foundation of overhead ethane concentration neural network soft sensor model:
Group of methods (the GMDH of data processing, Group Method of Data Handling) be to propose by the Self-Organization Principle that USSR (Union of Soviet Socialist Republics) mathematician Ivakhnenko uses in the biological cybernetics, it is based on the polynomial expression theory, limit input variable self and mutual various combinations, therefore be considered to the description fully of nonlinear model, can reach the best of breed of describing practical problems.The local describing function of fuzzy GMDH network is selected non-single, selects to be equivalent to the partial descriptions of the RBF network of fuzzy model as the network the first layer, utilizes good part of RBF network and overall match characteristic, approaches desired output to the full extent.And polynomial function is adopted in other each layer partial descriptions of network, and the polynomial expression partial descriptions can be carried out the network structure identification faster, makes the output combination of adopting the RBF partial descriptions approach network output.So this network training process is very fast, has good match and generalization ability.For this reason, adopt fuzzy GMDH neural network to set up the neural network soft sensor model of overhead ethane concentration and tower still ethylene concentration here.
In DCS, obtain the processing parameter that influences ethylene rectification tower overhead ethane concentration, comprise tower top temperature T1, tower top pressure P1, tower still temperature T 3 and reflux ratio R (being the ratio of trim the top of column amount and cat head produced quantity); The soft-sensing model has here been considered the different time of above-mentioned 4 factors to the overhead ethane concentration affects, according to the concrete structure and the sampling location of ethylene rectification tower, choose the input variable of T1 current time value T1 (k), P1 current time value P1 (k), T3 current time value T3 (k), R current time value R (k), a R current time unit time value R (k-1), another unit time value R (k-2) (referring generally to preceding two unit time) before the R current time before here as neural network model; Choose the output variable of the manual analysis value of ethylene rectification tower overhead ethane concentration current time as fuzzy GMDH neural network model.
The overhead ethane concentration manual analysis Value Data of T1 (k), P1 (k), T3 (k), R (k), R (k-1), R (k-2) and current time is carried out normalized, the normalization method scope can be chosen for [0,1], [0.5,0.5], [1,1] etc., here its normalizing is arrived between [0.2,0.8].Method for normalizing is as follows:
Wherein, X is the actual measured value of each input variable, and x is the numerical value of each input variable after the normalization method, and [a, b] is the range ability of X.
In neural network model, the node number of input layer is i (i=6), and the hidden layer number of plies in middle layer is L (L=1~100), and each the number of hidden nodes is j (j=2~100), and the output layer node is k (k=1~100).In an embodiment: the hidden layer number of plies in middle layer is L (L=2), and each the number of hidden nodes is j (j=60), and the output layer node is k (k=1).
The some groups of real time datas that the industrial production collection in worksite is arrived are as the learning sample of ethylene rectification tower overhead ethane concentration neural network soft sensor model.Select fuzzy GMDH neural network that overhead ethane is predicted, its input node respectively corresponding ethylene rectification tower T1 (k), P1 (k), T3 (k), R (k), R (k-1), R (k-2) after the normalization method, and the output node correspondence the ethane concentration manual analysis value (see figure 2) after the normalization method.
In above-mentioned some groups of learning sample, select wherein part to organize data, and predict the neural network generalization ability with other sample as the sampled data neural network training, trained and predicated error is less one group of weights and threshold value.
After neural network weight and threshold value are determined, just can be by the continuous collection of on-the-spot real time data (each input variable data in the finger print type here), bringing neural network model after normalization method into calculates, and neural network model calculated output valve through after the anti-normalized, just obtain the neural network prediction value of ethylene rectification tower overhead ethane concentration, unit is ppm.
The foundation of ethylene distillation Tata still ethylene concentration neural network soft sensor model:
Here, the fuzzy GMDH neural network of same employing is set up tower still ethylene concentration soft-sensing model.
In this model, selecting ethylene distillation Tata still pressure P 2 (k), tower still temperature T 3 (k), the sensitive plate temperature T 2 (k) of current time is the main operational variable that influences ethylene rectification tower still ethylene concentration performance characteristic; Promptly choose P2 (k), T3 (k) after the normalization method, T2 (k) input variable as tower still ethylene concentration soft-sensing model.
P2 (k), T3 (k), T2 (k) and current k tower still ethylene concentration manual analysis Value Data is constantly carried out normalized, and normalized method is the same.
Carry out " big zone " orthogonal simulation test on every side at full scale plant " operating point ", obtained some groups of operating restraint variations ethylene distillation Tata still ethylene concentration operation sample greatly, with these sampled datas and industrial production collection in worksite to real time data combine, constitute the learning sample of ethylene distillation Tata still ethylene concentration neural network soft sensor model.Select fuzzy GMDH neural network that tower still ethylene concentration is predicted, its input node respectively corresponding ethylene distillation Tata still pressure P 2 (k), tower still temperature T 3 (k) and tower sensitive plate temperature T 2 (k), and the output node correspondence ethylene concentration manual analysis value (see figure 3).Select a certain amount of sampled data neural network training, predict the neural network generalization ability, trained and predicated error is less one group of weights and threshold value with other a certain amount of sample.
On the basis of existing instrument of ethylene rectification tower and DCS computer control system, by real-time, the continuous acquisition of DCS system to process data (each input variable data in the finger print type here), bringing neural network model after normalization method into calculates, after the anti-normalized of model output valve process, just obtain the neural network prediction value of ethylene distillation Tata still ethylene concentration, unit is ppm.
Application module or advanced process administration module (Application Module or AdvancedProcess Manager at DCS, AM or APM) goes up the establishment that realizes control language by program circuit shown in Figure 6, real-time, continuous acquisition by data, just can obtain the real-time neural network prediction value of overhead ethane concentration and tower still ethylene concentration, can certainly carry out the calculating of real-time neural network prediction value with other on-line computer.
The correction of model:
Because have multiple interfering factors in the actual industrial production process, the neural network prediction value of the said products concentration and the manual analysis value of full scale plant unavoidably can produce certain deviation.For this reason, must be at set intervals, with manual analysis value (the several hrs analysis once usually) the Neural Network model predictive value is carried out online " rolling " and optimize correction, so that this neural network model adapts to the variation of commercial run performance characteristic and the migration of production status, finally obtain the final soft observed value of overhead ethane concentration, tower still ethylene concentration.
The foundation of the online deduction Controlling System of overhead ethane concentration:
The soft observed value of overhead ethane concentration after the compensation of lead-lag link, is inferred control to reflux ratio in view of the above in real time.The quantity of reflux of return tank of top of the tower liquid level control simultaneously with the influence of real-time adjusting feed loading to overhead ethane concentration, and by reflux ratio controller control produced quantity, realizes the deduction control of overhead ethane concentration, as shown in Figure 4.
The foundation of the online deduction Controlling System of tower still ethylene concentration:
The soft observed value of tower still ethylene concentration after the compensation of lead-lag link, is inferred control to ethylene rectification tower in view of the above in real time, comes control tower still ethylene concentration by regulating tower still reboiler heating propylene flow, as shown in Figure 5.
The described method of the application of the invention is implemented intelligent control to ethylene rectification tower in the ethylene unit, can reduce ethylene product loss in the discharging of ethylene distillation Tata still, reduce the energy consumption that heats propylene flow, reduction reflux ratio, the cat head ethene quality of stablizing, increase ethylene yield, reduces tower, make ethylene rectification tower be in the optimum operation operating mode.
Five, embodiment
The present invention is further illustrated below in conjunction with embodiment and accompanying drawing:
In DCS, obtain the main technique operating parameters that influences ethylene rectification tower overhead ethane fluctuation of concentration: tower top temperature, tower top pressure, tower still temperature, reflux ratio etc., it comprises before current time tower top temperature T1 (k), current time tower top pressure P1 (k), current time tower still temperature T 3 (k), current time reflux ratio R (k), the current time before 1 hour R (k-1), the current time 2 hours R (k-2) and current time overhead ethane concentration manual analysis value, and with these data normalizations to [0.2,0.8] scope, normalized method is as follows:
x=(X-a)/((b-a)×0.6+0.2
Wherein, X is an input variable, and [a, b] is the range ability of X, and x is the input after the normalization method.The tower top temperature variation range is taken as [28 ,-35], and unit is ℃; The variation range of tower top pressure is taken as [1.75,1.95], and unit is MPa (G); Tower still variation of temperature scope is taken as [5 ,-12], and unit is ℃; The variation range of trim the top of column ratio is taken as [3,10], no unit.
In 400 groups of real time datas of industrial production collection in worksite, utilize overhead ethane concentration manual analysis value to carry out neural network training as target value.Wherein preceding 300 groups of data are as learning sample, and the 100 groups of data in back are as forecast sample.By neural network model is carried out off-line training, obtain structure, weights and the threshold value of GMDH neural network.The action function that the calculating of GMDH neural network is adopted is a quadratic polynomial.
Application module or advanced process administration module (AM or APM at DCS, Application Moduleor Advanced Process Manager) goes up the establishment that realizes control language according to the program circuit of Fig. 6, real-time, continuous acquisition by data, the weights that train and threshold band are gone into neural network to be calculated, the ethane concentration value that obtain this moment is between [0.2,0.8]; For this reason, anti-normalization method is carried out in this neural network that calculates output, the soft observed value of the real-time neural network of ethane concentration that obtains cat head is in [0,1000] scope, and unit is ppm; At last, the neural network prediction value after utilizing recently the manual analysis value of overhead ethane concentration constantly to anti-normalization method is carried out online " rollings " and is optimized correction, obtains the final soft observed value of ethane concentration.
On the AM of DCS system or APM, realize the establishment of control language, control block diagram by Fig. 4 is built control module, and the full scale plant that carries out the characteristic test of industry spot Object Operations, Controlling System is actual puts into operation and the parameter testing of link such as dynamic process compensation, cat head ethylene concentration 〉=99.95% is guaranteed in the final deduction control that realizes overhead ethane concentration.
In like manner, according to the implementation method of the soft observed value of above-mentioned ethane concentration, can obtain the soft observed value of tower still ethylene concentration.And on the AM/APM of DCS system, realize the establishment of control language, structure by Fig. 5 is built control module, and the full scale plant that carries out the characteristic test of industry spot Object Operations, Controlling System is actual puts into operation and the parameter testing of each link, realize the deduction control of tower still ethylene concentration, reach control tower still ethylene concentration in the scope of 0.7-1.1%.Because the charging ethane of tower and the mixture of ethene evenly enter, generally do not control.And mainly control by heating of tower still and reflux ratio.
Optimized the operation of ethylene rectification tower by tower still and the advanced control techniques of cat head, improved the operation " elasticity " of ethylene distillation system, reduced tower still reboiler and overhead condenser load, reduce the load of propylene compressor, improved the throughput and the operational stability of ethylene distillation system.Control by tower still ethylene concentration, guarantee at the bottom of the tower that ethylene content is below 1.1% in the extraction ethane, and be controlled at set(ting)value ± 0.2% scope in, to be set in ethylene concentration be 0.3 in soft measurement when coming into operation as the scene, the soft measurement concentration range of the actual ethylene concentration that comes into operation illustrates that at (0.18-0.45) effect that comes into operation of soft measurement is more significant.Ethene Tata still reboiler obviously reduces, and before and after putting into operation according to advanced control techniques ethylene distillation Tata still reboiler is added the statistics of heat flux, and it on average adds heat flux by original 32000Nm
3/ h drops to 26000Nm
3/ h on average per hour reduces 6000Nm
3The average operation reflux ratio of tower is also reduced to present about 4.15 by original 4.4, saved energy expenditure greatly behind the advanced control system that therefore comes into operation.
When adopting DCS control, the control of overhead ethane concentration is by following control mode: the trim the top of column amount is controlled according to the return tank liquid level, determine reflux ratio according to the soft measurement ethane of cat head concentration, can calculate the produced quantity of ethene like this, set(ting)value as ethene extraction controller, by the extraction of ethene extraction flow director control ethene, finally reach the purpose that reduces overhead ethane again, when control, control by soft measurement output valve and reflux ratio controller tandem.The control of tower still is to come tandem tower still heating propylene flow according to the soft observed value of ethylene concentration, and then reaches the purpose of control tower still ethylene concentration.According to the concrete condition of production, the control corresponding parameter need be adjusted, and as when certain concrete field control, the soft measurement of cat head control ethane concentration infers that control adopts PID control, and its controlled variable is: k=0.5, T1=15, T2=7.5; Cat head ethene extraction controller adopts PI control, and its controlled variable is: k=1.1, T1=1.5; The tower still is inferred control employing PID control, and its controlled variable is: k=0.15, T1=20, T2=8; Tower still heating flow of propylene amount controller adopts PI control, and its controlled variable is k=1.35, T1=1.65.