CN1635050A - Coke oven coking production automatic heating method - Google Patents

Abstract

The invention relates to the coking production of coke ovens in the technical field of industrial heating, and belongs to the automatic heating technology of coke ovens. The technology can realize the optimal setting and adjustment of gas flow and flue suction, and stabilize the straight temperature of the coke oven flue. The invention is a coke oven automatic heating technology based on a feedback structure, including: (1) based on a self-correcting flue temperature model, the accurate estimation of the coke oven flue temperature is realized; (2) the intelligent fault-tolerant adjustment of the gas flow rate realizes the (3) The flue suction automatic adjustment method based on the precise mathematical model of the flue, while changing the gas flow, the gas flow data and the expected flue gas oxygen content The volume data is input into the flue model, and the output of the flue model corresponding to this input is the new flue suction setting value. The invention can improve the reliability and adaptability of coke oven production, effectively reduce energy consumption and environmental pollution, improve coke quality and prolong the service life of the furnace body.

Description

Coke oven coking production automatic heating method

technical field

The invention relates to the technical field of industrial heating and belongs to coking oven coking production technology.

Background technique

The coke oven is an important production equipment in the metallurgical and energy industries. It produces coke, gas and chemical products through high-temperature dry distillation of coal. As a typical industrial heating process, the coke oven is a complex system with large time delay, large inertia, strong nonlinearity, variable parameters and multiple disturbances. Among them, the straight-line temperature of the coke oven is one of the very important technical indicators in the coking production process. , directly related to coke quality and furnace life. Coke oven heating is to adjust the straight-line temperature by adjusting the gas flow and flue suction. The goal is to ensure the optimization of combustion efficiency through suction adjustment while stabilizing the furnace temperature.

Traditional artificial heating systems rely on experience and manually measured flue temperature to regulate gas flow and flue suction. Due to the long interval of manual temperature measurement (generally 4 hours) and the limitations of the operator's experience, the temperature adjustment effect cannot be guaranteed, the furnace temperature fluctuates greatly, and the combustion efficiency is low. The automation of coke oven heating (automatic heating) helps to avoid the uncertainty of manual heating and improve the operation and operation level of coke ovens.

Coke oven automatic heating has three structures, namely feedforward structure, feedback structure and feedback + feedforward structure. The feedforward structure directly controls the total heat supply according to the heating model. The disadvantages are large investment, poor adaptability, difficult maintenance, and inability to overcome various random disturbances. The feedback structure aims to stabilize the straight temperature of the flue, and adjusts the gas flow and flue suction by comparing the flue temperature with the target temperature to achieve heating control. Compared with the feedforward structure, the feedback structure can overcome various random disturbances, has less investment, low maintenance cost, and good adaptability. However, it is difficult to detect the temperature of the flue on-line, so it is technically difficult to implement automatic heating control based on the feedback structure. The essence of the feedback + feedforward structure is to add feedforward compensation to the measurable disturbance on the basis of the feedback structure.

Although the automatic heating based on the feedback structure has many advantages, due to the particularity of the coke oven production process, it is difficult to detect the temperature of the coke oven flue on-line, which has become the biggest technical obstacle to the implementation of the feedback automatic heating. In order to solve this problem, the soft sensing method is mostly used at present, that is, the temperature of the flue is estimated through the functional relationship between the temperature of the regenerator and the temperature of the flue, and automatic heating is realized accordingly. The fire path model is key to this approach. In order to establish a model, the traditional method requires collecting a large amount of data of the regenerator temperature and the corresponding manual temperature measurement of the flue, and then obtains the model parameters through certain mathematical means. However, in the actual production process, changes in the calorific value of gas, coal moisture, coke oven conditions and other factors will lead to changes in the actual relationship between the regenerator temperature and the flue temperature, and the time and magnitude of the changes are unpredictable. The traditional method ignores the change of the flue path model, resulting in large errors in the estimated flue temperature. In addition to causing deviations in the observed furnace temperature, the estimated results of the flue temperature with errors lead to problems in gas flow regulation.

Gas flow adjustment generally adopts a proportional adjustment algorithm, namely:

Q＝k _Q (T _sp -T _h )+Q ₀

Among them, Q and Q ₀ represent the gas flow rate and its steady-state value, k _Q is the proportional gain, T _sp is the set value of the flue temperature, and T _h is the estimated value of the flue temperature. Thus, the gas flow regulation is based on the deviation between the estimated flue temperature and the set target temperature. However, there are inevitably estimation errors in the flue temperature estimated by the flue model, which may lead to two abnormal situations: 1) The actual furnace temperature is lower than the target temperature, but the estimated flue temperature is higher, so the decrease 2) The actual furnace temperature is higher than the target temperature, but the estimated fire path temperature is lower, so the gas flow rate is increased to make the actual temperature higher. This essentially turns the furnace temperature regulation into a disturbance to the furnace temperature, which causes large fluctuations in the furnace temperature and further deteriorates the furnace temperature regulation. This is a very serious problem with traditional feedback automatic heating methods. The reason for this problem is the low accuracy of flue temperature estimation and poor fault tolerance of the regulation algorithm.

In addition, while changing the gas flow rate, the air volume entering the coke oven must be changed to ensure the best air-fuel ratio, reduce energy consumption, and avoid environmental pollution. The air volume is adjusted by the suction of the flue, and the goal of the adjustment is to control the oxygen content of the flue gas. Oxygen content can be measured online by zirconia, but zirconia has short service life, high failure rate, and poor reliability. Therefore, it is generally not used for closed-loop control, but only for detection and monitoring. At present, the industrial heating process including coke oven heating is generally adjusted manually or by a rough model established based on experience, and the combustion effect cannot be guaranteed. Because automatic heating needs to change the gas flow frequently, if the suction cannot be adjusted effectively, it may cause excess air to take away heat or insufficient air to cause incomplete combustion of gas to pollute the environment and waste energy, thereby greatly reducing the combustion efficiency of gas and possibly Cause gas flow regulation to lose its effect.

In conclusion, in order to fundamentally improve the automatic heating of coke oven based on the feedback structure, the following three key issues must be solved:

(1) Improve the accuracy of flue temperature estimation;

(2) Enhance the fault tolerance of gas flow adjustment;

(3) The flue suction is automatically adjusted.

The traditional automatic heating technology has not effectively solved these three problems or has major limitations, which makes the reliability and adaptability of the coke oven heating feedback control poor, and the effect has a large gap with the requirements of industrial production.

Contents of the invention

The purpose of the present invention is to overcome the defects of the prior art, aiming at the three key problems in the coke oven heating feedback control, respectively provide new solutions, and realize a new type of automatic heating technology based on the feedback structure on this basis .

In order to achieve the above object, the technical solution of the present invention is to provide a coking oven coking production automatic heating method, which includes:

(1) Flue temperature estimation method based on self-correcting flue model;

(2) Intelligent fault-tolerant gas flow adjustment method;

(3) An automatic adjustment method for flue suction based on the precise mathematical model of the flue;

Among them, the self-correcting flue path model is to continuously collect the temperature data of the regenerator and the corresponding flue temperature data obtained by manual measurement, construct a "sliding data window" data set and modify the flue path model parameters to track changes due to system environmental factors The resulting change in the functional relationship between the regenerator temperature and the flue temperature improves the accuracy of the estimation of the flue temperature and enhances the reliability of automatic heating.

In the method, the functional relationship between the temperature of the regenerator and the temperature of the flue path is the flue path model.

In the described method, the sliding data window is a first-in-first-out data queue, and its window width is the length of the queue; the window width determines the size of the data set used to correct the fire path model, and the specific focus Furnace correlation, which characterizes the change law of coke oven characteristics and environmental factors, is the basis for the self-calibration of the flue path model.

Described method, its described window width, the determination method of its optimal sliding data window width is as follows:

Let the historical data set arranged in chronological order be Z={z ₁ , z ₂ , K, z _l }, where l is the data length, and the subscript corresponds to the chronological order, $z_i=(T_x^i,T_h^i)$ is the i-th data pair, T _x is the temperature data of the regenerator, T _h is the infrared temperature measurement data of the flue path, and the flue path model takes the first-order polynomial form T _h = a ₁ T _x + a ₀ . Assume that the width of the sliding data window is a positive integer m, and then set Θ=(a ₁ , a ₀ ) ^T , A ^T =(T _x , 1), then the fire path model can be written in vector form T _h =A ^T Θ;

According to the m data pairs in the sliding data window, the relationship is as follows:

$\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>}\end{array}\right]<span\; class="notranslate">\; =</span>\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\end{array}\right]<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; \&Theta;</span>}$

make $h=\left[\begin{array}{c}{T}_h^{\left(1\right)}\\ m\\ T_h^{\left(m\right)}\end{array}\right],{\mathrm{\&Phi;}}^T=\left[\begin{array}{c}{A}_1^T\\ m\\ A_m^T\end{array}\right],$ Then there is H= ^ΦT Θ, and the model parameters are calculated according to the least squares algorithm: Θ=( ^ΦΦT )ΦH; the data window slides in the data set Z, and the corresponding parameters can be established as the fire path model of Θ according to the aforementioned method; Each sliding can establish a flue path model, through which the flue path temperature at the next moment outside the sliding window can be predicted ${\hat{T}}_h^{m+1}=A_{m+1}^T\mathrm{\&Theta;},$ estimate error $e_{m+1}=T_h^{m+1}-{\hat{T}}_h^{m+1};$ Define the cost function: $J\left(m\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{e}_i,$ Then the m that makes J the smallest is the optimal sliding data window width; since the cost function curve is descending-minimum-rising-stable with a single minimum value, the minimization process can use the enumeration method, starting from 1 , until a minimum value is obtained, m corresponding to the minimum value is the optimal sliding data window width;

After the m value is determined, whenever new data is collected in actual use, the data queue in the sliding window is automatically updated according to the first-in-first-out method, and the parameters of the fire path model are recalculated and updated according to the data set in the window. Realize automatic correction of flue temperature model.

In the method, the automatic correction of the temperature model of the fire path is realized, and the process is: delete the earliest data in the window, and shift forward in sequence, and store new data at the end of the queue.

Described method, its described (2) intelligent fault-tolerant gas flow regulation method, is to regulate the gas flow based on the variation of flue temperature, i.e. trend; for:

u(k)=u(k-1)+Δu(k)

$\mathrm{<span\; class="notranslate">\; \&Delta;u</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; sp</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; sp</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span>$

Among them, u and Δu represent the gas flow and its increment, k and k-1 represent the current moment and the previous moment, and the superscript represents the parameter corresponding to the time, is the flue temperature estimated according to the above steps, K _p represents the proportional gain, T _sp represents the set target flue temperature, u ₀ represents the offset of the gas flow; u ₀ is manually set or heated before automatic heating Manual modification during the automatic control process.

The described method, its (3) flue suction automatic adjustment method based on the precise mathematical model of the flue, is to establish an accurate mathematical model of the flue, and realize the flue flue with the goal of controlling the oxygen content of the flue exhaust gas through the model. Suction automatic adjustment method: While changing the gas flow, input the gas flow data and the expected flue gas oxygen content data into the flue model, and the output of the flue model corresponding to this input is the new flue suction setting. Fixed value; the significance of the flue suction is the flue suction required to make the oxygen content of the flue exhaust gas reach the desired value under the current gas flow rate.

Described method, its described establishment flue precise mathematical model, is to adopt the neural network method of artificial intelligence field, this method collects historical data, comprises coal gas flow rate, flue exhaust gas oxygen content and flue suction, and trains neural network , to establish a nonlinear mathematical function relationship expressed in the form of a neural network; the input of the flue model is the gas flow and the oxygen content of the flue gas, and the output of the flue model is the suction of the flue.

In the method, the precise mathematical model of the flue is set up, and its specific steps include: (1) data preprocessing; (2) constructing a training sample set and a test sample set; (3) using the training sample set to train a neural network; (4) Test the neural network using the test sample set.

In the method, the system environment factors include the condition of the coke oven, the calorific value of the gas, the moisture content of the coal entering the furnace, and the suction.

The method, the intelligent fault-tolerant gas flow adjustment method, can improve the correctness of gas flow adjustment in the case of errors in flue temperature estimation, avoid furnace temperature fluctuations caused by incorrect adjustments, and improve heating control effects.

In the method, the oxygen content of flue gas required for establishing an accurate mathematical model of the flue can be obtained from zirconia detection or laboratory analysis results.

Description of drawings

Fig. 1 Schematic diagram of gas flow control structure;

Fig. 2 Gas flow regulation process curve diagram.

Detailed ways

The present invention includes three basic contents: (1) self-correcting flue temperature model for on-line detection of flue temperature; (2) intelligent fault-tolerant gas flow adjustment method; (3) automatic adjustment of flue suction based on accurate flue model method.

1 Flame temperature detection

Install two thermocouples on the mechanized coke side of the two regenerators corresponding to the lower part of each vertical flue of the coke oven. If the temperature uniformity of the whole coke oven is good, only a few regenerators in the middle can be installed (generally should be more than 8). Ten minutes after the reversing of the coke oven, after the temperature of the descending regenerator tends to be stable, the temperature data of each descending regenerator on the coke side of the machine is collected and geometrically averaged, and the average value is used as the machine-side regenerator of the entire coke oven. temperature and the coke-side regenerator temperature.

The temperature data of the regenerator on the coke side and the temperature data of the fire channel on the coke side obtained through infrared temperature measurement corresponding to the time (that is, the straight-line temperature) constitute a data pair, and several data pairs form a data set. Through this data set, the mathematical relationship between the regenerator temperature and the flue temperature can be established, that is, the flue path relationship model. The present invention adopts a first-in-first-out queue to store the data sets required for building a model. The queue is called "sliding data window", and the length of the queue is called "width of sliding data window". Whenever new data arrives, the earliest data in the window is deleted and shifted forward in sequence, and the new data is placed at the end of the queue. This data update process is called the sliding of the data window. The significance of adopting the sliding data window is to enable the data set of the model to fully reflect the current working conditions of the coke oven in time, improve the accuracy of the model thus established, and avoid model errors caused by invalid data.

The method of building the fire path model is as follows:

Let the width of the sliding data window be m, and take the first-order polynomial form T _h =a ₁ T _x +a ₀ as the model of the fire path relationship, and then set Θ=(a ₁ , a ₀ ) ^T , A ^T =(T _x , 1), then the fire path model can be written in vector form T _h = A ^T Θ. According to the m data pairs in the sliding data window, the relationship is as follows:

$\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>}\end{array}\right]<span\; class="notranslate">\; =</span>\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\end{array}\right]<span\; class="notranslate">\; \&Center\; Dot;</span>\mathrm{<span\; class="notranslate">\; \&Theta;</span>}$

make $h=\left[\begin{array}{c}{T}_h^{\left(1\right)}\\ m\\ T_h^{\left(m\right)}\end{array}\right],{\mathrm{\&Phi;}}^T=\left[\begin{array}{c}{A}_1^T\\ m\\ A_m^T\end{array}\right],$ Then there is H=Φ ^T Θ, and the model parameters are calculated according to the least square algorithm:

Θ＝(ΦΦ ^T )ΦH

The sliding of the data window and the estimation of model parameters constitute the process of continuous correction of the fire path model, and the fire path model established by this is called the self-correcting fire path model.

After the flue model is established, the flue temperature can be estimated whenever there is new regenerator temperature data, ${\hat{T}}_h=a_1{T}_x+a_0,$ In this way, the online detection of the temperature of the flue can be realized. The flue temperature detection cycle based on the model can generally be shortened to half an hour, while the artificial infrared measurement method is generally four hours. Faster detection of flue temperature means that it can reflect the working status of the coke oven in a more timely manner, and provide reference information for heating control, so as to quickly suppress temperature fluctuations and overcome the influence of various interference factors.

The width of the sliding data window is the basis of the self-calibration of the fire track model, which has an important impact on the reliability and accuracy of the model. This data is related to a specific coke oven, and represents the comprehensive change law of the coke oven itself and various related disturbance factors. The determination method is as follows:

Let the historical data set be Z={z ₁ , z ₂ , K, z _l }, l is the data length, and the subscript corresponds to the time sequence, $z_i=(T_x^i,T_h^i)$ is the i-th data pair, and the fire path model adopts the first-order polynomial form T _h =a ₁ T _x +a ₀ . Assuming that the width of the sliding data window is a positive integer m, then the data window slides in the data set Z to establish a fire path model (parameter is Θ) according to the aforementioned method. Each sliding can establish a flue path model, through which the flue path temperature at the next moment outside the sliding window can be predicted ${\hat{T}}_h^{m+1}=A_{m+1}^T\mathrm{\&Theta;},$ estimate error $e_{m+1}=T_h^{m+1}-{\hat{T}}_h^{m+1}.$ Define the cost function: $J\left(m\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{e}_i,$ Then the m that makes J the smallest is the optimal sliding data window width. Since the form of the cost function curve is descending-minimum-rising-stable and has a single minimum value, the minimization process can adopt the enumeration method, starting from 1 until a minimum value is obtained, and the m corresponding to the minimum value is Optimal sliding data window width.

The self-calibration model is realized by programming in the controller, and its input is the average data of the regenerator temperature on the coke side and the infrared temperature measurement data of the coke side flue temperature, and the output is the estimated value of the coke side flue temperature. The sliding data window width is determined in the preparatory stage before automatic heating is implemented. Data can also be re-collected, calculated and modified during automatic heating if required.

2 gas flow adjustment

The method described in 1 can quickly and accurately detect the coke oven flue temperature, which reflects the current actual temperature of the coke oven. If the temperature is different from the set target temperature, you need to increase or decrease the gas flow, that is, increase or decrease the heat supply, so that the temperature of the fire path rises or falls. This process is gas flow adjustment.

The gas flow regulating method of the present invention is a quantitative calculation formula, which is programmed in the controller, whose input is the coke side fire path temperature obtained from the fire path model, and the output is the optimized setting value of the coke side gas flow. The set value is sent to the gas valve control circuit (included in the basic level control system of the coke oven), so that it can change the opening degree of the regulating mechanism installed on the gas pipeline, so that the actual gas flow supplied follows the optimal setting value.

As shown in Figure 1, it is a schematic diagram of the gas flow control structure. The gas flow adjustment algorithm in digital mode is as follows:

u(k)=u(k-1)+Δu(k)

$\mathrm{<span\; class="notranslate">\; \&Delta;u</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; sp</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; sp</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{\mathrm{<span\; class="notranslate">\; k</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span>$

Among them, u and Δu represent the gas flow and its increment, k and k-1 represent the current moment and the previous moment, and the superscript represents the parameter corresponding to the time, ${\hat{T}}_h^{\left(k\right)}=a_1^{\left(k\right)}{T}_x^{\left(k\right)}+a_0^{\left(k\right)},$ ${\hat{T}}_h^{(k-1)}=a_1^{(k-1)}{T}_x^{(k-1)}+a_0^{(k-1)}$ is the flue temperature estimated by the flue temperature model, K _p represents the proportional gain, T _sp represents the set target flue temperature, u ₀ represents the offset of the gas flow. u ₀ Manual setting before automatic heating or manual modification during heating automatic control. During the period when the flue temperature model is not updated, the same model parameters make the error in the same direction as the flue temperature, so the influence of the model error can be reduced, and the change of the flue temperature can be used to calculate the control amount, and the accuracy of flow regulation can be improved. sex. If the flue temperature model is updated at the k-th time, the new model can realize a more accurate estimation of the flue temperature, and generate a corrected control variable through the difference with the estimated result at time k-1, compensating the previous generation due to the estimation error The cumulative deviation of the control effect.

The adjustment process shown in Fig. 2 is as follows: T _h and T _h ′ represent the actual flue temperature and simulated flue temperature respectively, T _sp represents the set target flue temperature, and the flue temperature model is updated at time k. According to the adjustment process of the trend (see the solid line in Figure 2): at time k-2, according to the downward trend of the flue temperature, the gas flow increment q(k-2)>0, the total gas flow increases, which makes k The rise of the estimated flue temperature and the actual flue temperature at time -1; the rising trend of flue temperature at time k-1 makes the gas flow increment q(k-1)<0, the total gas flow decreases, and the actual flue temperature tends to The flue model is updated at time k, and the simulated flue temperature is closer to the actual flue temperature and shows a downward trend. According to this trend, the gas increment q(k)>0 at time k, and the total gas flow rate is higher than that of the previous one. The time increases, which will compensate for the gas flow calculation deviation caused by the high temperature of the simulated flue, so that the actual flue temperature will rise and get closer to the set target temperature.

The gas flow regulating method of the present invention has the characteristic of intelligent fault tolerance, and the so-called fault tolerance means that the correctness of the flow regulating direction is always maintained when there is an error in the temperature of the flue path. Its significance is to enhance the reliability of adjustment and avoid the deterioration of coke oven temperature caused by wrong adjustment. The reason why it has this characteristic is that this regulation method calculates the gas flow rate essentially according to the variation of the flue temperature, that is, the trend of the flue temperature. Compared with the traditional adjustment method based on the difference between the flue temperature and the target temperature, the trend more accurately reflects the change of the coke oven temperature, so the adjustment method of the gas flow thus obtained has better fault tolerance. In this adjustment method, the estimation error of the flue temperature only affects the adjustment range, and the correctness of the adjustment direction is greatly enhanced.

3 flue suction adjustment

While increasing or decreasing the gas flow according to the method described in 2, the air supply must be adjusted so that the gas entering the coke oven is fully combusted and the combustion efficiency is the highest. The adjustment of the air volume is realized by adjusting the suction of the flue, that is, changing the opening of the regulating flap in the coke side branch flue. Its function is to change the air volume entering the coke oven and assisting the combustion of gas. Flue suction adjustment must avoid two situations, namely (1) Insufficient air volume causes excess gas and is discharged into the atmosphere; (2) Excessive air causes exhaust gas to take away a lot of heat, these two situations directly lead to energy waste and Great reduction in combustion efficiency.

Whether the air volume is appropriate or not should be measured by analyzing the oxygen content in the coke-side exhaust gas. Insufficient or excess air volume will lead to too low or too high oxygen content. Oxygen content data can be obtained in two ways: one is exhaust gas sampling and laboratory analysis; the other is to install zirconia in the coke side flue for online detection.

The invention is a flue suction adjustment method based on an accurate mathematical model, which includes two parts: establishing an accurate flue suction model and adjusting the flue suction based on the model.

The flue suction model in the present invention refers to the precise mathematical relationship among gas flow, flue suction and oxygen content of exhaust gas. The model adopts the feedforward neural network structure in the field of artificial intelligence. The input of the model is the gas flow rate and the oxygen content of the exhaust gas, and the output of the model is the flue suction. Collect gas flow, exhaust gas oxygen content and flue suction three kinds of historical data to construct a data set, use the data set to train the neural network to obtain the nonlinear functional relationship between the three quantities hidden in these data, that is, the model. If only the main pipe gas flow is controlled, then a flue suction model can be established; if the gas flow at the coke side is controlled separately, the flue suction models on both sides of the coke need to be established separately.

Steps to build a model (see relevant literature on neural networks for details):

(1) Data preprocessing: remove abnormal data and normalize data;

(2) Construct a training sample set and a test sample set: the two are used to train the neural network model and test the neural network model respectively;

(3) Neural network training: programming on the computer, the result is a neural network model with a certain structure and parameters;

(4) Neural network test: programming on the computer is used to verify the accuracy and reliability of the neural network model obtained in step (3);

After the above steps, a nonlinear mathematical function relationship expressed in the form of neural network can be established, that is, the flue suction model. The model can be re-collected, established and updated before the automatic heating is implemented, or according to the needs during the automatic heating process.

In this automatic heating technology, whenever the gas flow rate is changed as described in 2, the flue suction is adjusted based on the model, the process is as follows:

(1) Normalize the new gas flow setting value and the expected exhaust gas oxygen content data;

(2) input the data after normalization into the flue suction model;

(3) The output of the flue suction model is denormalized, and the obtained data is the new set value of the flue suction, which is sent to the flue flap control loop (included in the basic stage of the coke oven control system);

(4) The flue flap control circuit adjusts the flap opening according to the new set value, and then adjusts the air volume entering the coke oven.

The entire adjustment process is realized by programming in the controller. After calculating the optimal set value of the new gas volume and suction, it should be sent to the gas valve control circuit and the flue flap control circuit according to the requirements of "double cross" operation.

The model-based flue suction adjustment method essentially also provides a flue suction adjustment structure that does not require zirconia. The oxygen content data needed to build the model can come from the laboratory, and the flue suction adjustment only needs the gas flow data and the expected flue gas oxygen content data (set manually), so that automatic heating gets rid of the dependence on zirconia . The flue suction adjustment without zirconia simplifies the structure of the suction adjustment system, reduces investment and maintenance costs, and improves the reliability of suction adjustment.

Implementation conditions: (1) The coke oven production process is normal;

(2) Already have a basic level control system;

(3) The performance of the detection instrument and the execution structure is good.

definition:

Flue path model: a functional relationship between the temperature of the regenerator and the temperature of the flue path;

Sliding data window: a first-in-first-out queue used to store the data sets needed to build the fire path model;

The width of the sliding data window: the length of the queue;

Sliding of the data window: It is the process of updating the data. Whenever new data arrives, the earliest data in the window is deleted and shifted forward in sequence. The new data is placed at the end of the queue;

Flue suction model: precise mathematical function relationship between gas flow, flue gas oxygen content and flue suction;

a preparatory stage

1 Install the thermocouple. Install two thermocouples on the mechanized coke side of the two regenerators corresponding to the lower part of each vertical flue of the coke oven. If the temperature uniformity of the whole coke oven is good, only a few regenerators in the middle can be installed (generally should be more than 8).

2 Collect the temperature data of the regenerator on the coke side of the machine and the infrared temperature measurement data. Ten minutes after the reversing of the coke oven, after the temperature of the descending regenerator tends to be stable, the temperature data of each descending regenerator on the coke side of the machine is collected and geometrically averaged, and the average value is used as the machine-side regenerator of the entire coke oven. temperature and the coke-side regenerator temperature. The infrared temperature measurement data of the flue temperature is obtained by manual measurement. The temperature data of the regenerator on the coke side and the temperature data of the fire channel on the coke side obtained by infrared temperature measurement corresponding to the time (that is, the straight-line temperature) constitute a data pair, and several data pairs constitute a temperature data set Z={z ₁ , z ₂ , K, z _l }, l is the data length, the subscript corresponds to the time sequence, $z_i=(T_x^i,T_h^i)$ is the i-th data pair, T _x represents the temperature of the regenerator, and _Th is the temperature of the fire path.

3 Collect gas flow Q, flue suction P and flue exhaust gas oxygen content ρ ₀ three kinds of historical data to form a flue suction data set. The three data must correspond in time. If the gas flow at the coke side is controlled separately, the gas flow at the coke side is collected separately; otherwise, only the gas flow at the main pipe is collected. If mixed gas is used for heating, it is necessary to collect the flow of various gases before mixing. 4 Data analysis and processing, including determining the width of the sliding data window and establishing the flue suction model.

1) Determine the sliding data window width

Let the temperature data set established in step 2 be Z={z ₁ , z ₂ , K, z _l }, where l is the data length, and the subscript corresponds to the time sequence, $z_i=(T_x^i,T_h^i)$ is the i-th data pair, T _x is the temperature data of the regenerator, T _h is the infrared temperature measurement data of the flue path, and the flue path model takes the first-order polynomial form T _h = a ₁ T _x + a ₀ . Assume that the width of the sliding data window is a positive integer m, and then set Θ=(a ₁ , a ₀ ) ^T , A ^T =(T _x , 1), then the fire path model can be written in the vector form T _h =A ^T Θ. According to the m data pairs in the sliding data window, the relationship is as follows:

$\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>}\end{array}\right]<span\; class="notranslate">\; =</span>\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\end{array}\right]<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; \&Theta;</span>}$

make $h=\left[\begin{array}{c}{T}_h^{\left(1\right)}\\ m\\ T_h^{\left(m\right)}\end{array}\right],{\mathrm{\&Phi;}}^T=\left[\begin{array}{c}{A}_1^T\\ m\\ A_m^T\end{array}\right],$ Then there is H= ^ΦT Θ, and the model parameters are calculated according to the least square algorithm: Θ=(ΦΦ ^T )ΦH. Sliding the data window in the data set Z can establish the corresponding fire path model (parameter is Θ) according to the aforementioned method. Each sliding can establish a flue path model, through which the flue path temperature at the next moment outside the sliding window can be predicted ${\hat{T}}_h^{m+1}=A_{m+1}^T\mathrm{\&Theta;},$ estimate error $e_{m+1}=T_h^{m+1}-{\hat{T}}_h^{m+1}.$ Define the cost function: $J\left(m\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{e}_i,$ Then the m that makes J the smallest is the optimal sliding data window width. Since the form of the cost function curve is descending-minimum-rising-stable and has a single minimum value, the minimization process can adopt the enumeration method, starting from 1 until a minimum value is obtained, and the m corresponding to the minimum value is Optimal sliding data window width.

If the width of the sliding data window needs to be re-determined for some reason during the automatic heating process, the data can also be re-collected, calculated and modified according to the aforementioned operations.

The determined m value is sent to the correction part of the fire path model (see 3 for details) to adjust the queue length and model parameter calculation.

2) Establish the flue suction model

The model adopts the feedforward neural network structure in the field of artificial intelligence. The input of the model is the gas flow rate and the oxygen content of the exhaust gas, and the output of the model is the flue suction. The process of building the model is the process of training the neural network, and the training data comes from the flue suction data set in step 3. The method of building the model is as follows:

(1) Data normalization

Transform the data in the flue suction data set into the same interval. Let x be the original data, y be the normalized data, and the normalized interval is [a _min , a _max ], then the calculation formula is as follows:

$\mathrm{<span\; class="notranslate">\; the\; y</span>}<span\; class="notranslate">\; =</span>\frac{\mathrm{<span\; class="notranslate">\; x</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}{{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; max</span>}}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; min</span>}}}$

(2) Construct sample data set

Construct the input data set and output data set respectively according to the normalized data $I=\{x_i=[\hat{Q}\left(i\right),{\widehat{\mathrm{\&rho;}}}_0\left(i\right)],i=\mathrm{1,2},\mathrm{\&Lambda;},\mathrm{no}\},o=\{{\hat{P}}_i,i=\mathrm{1,2},\mathrm{\&Lambda;},\mathrm{no}\}.$ Then {I, O} constitutes a sample data set. Divide the sample data set into two parts {I ₁ , O ₁ } and {I ₂ , O ₂ } to be used as training sample set and test sample set respectively, where I ₁ ={X _j }, O ₁ ={P _j } , j{1, 2, Λ, n} and I ₂ ={X _k }, O ₂ ={P _k }, k{1, 2, Λ, n}, j≠k.

(3) Determine the type of neural network, network structure and train the neural network

Use the training sample set to train the neural network and build the flue suction model. See related literature for details.

(4) Verify the neural network model

Use the test sample set to verify the neural network model established in step (4), and test the accuracy and reliability of the model. See related literature for details.

The structure and parameters of the verified neural network model can be determined, and the model is used for the model-based flue suction adjustment process in the automatic heating process (see step 3 in II for details).

Two heating automatic control

The heating automatic control is implemented in a program through the controller programming, including three main operations, namely, estimating the temperature of the flue, adjusting the gas flow and adjusting the suction of the flue. The input of the whole automatic heating procedure is: (1) temperature data of each descending regenerator measured by thermocouples installed in the regenerator; (2) expected target flue temperature; (3) oxygen content of flue gas; (4) Initial test offset of gas flow. The output of the automatic heating program is: (1) the gas flow setting value of the gas valve control loop; (2) the flue suction setting value of the flue flap control loop.

The specific execution process of the three automatic heating operations is described in detail as follows.

1 Estimated flue temperature

Ten minutes after the reversing of the coke oven, after the temperature of the descending regenerator tends to be stable, the temperature data of each descending regenerator on the coke side of the machine is collected, and the controller performs a geometric mean of these data, and the average value is used as the machine temperature of the entire coke oven. side regenerator temperature and coke side regenerator temperature.

The parameter of the flue path model is Θ=(a ₁ , a ₀ ) ^T , then the formula for calculating the flue path temperature from the regenerator temperature and the flue path model is ${\hat{T}}_h=A^T\mathrm{\&Theta;},$ Where A ^T = (T _x , 1), T _x regenerator temperature. According to the flue model on the coke side and the regenerator temperature, the corresponding flue temperature on the coke side can be estimated respectively according to this calculation method.

2Adjust the gas flow

The gas flow adjustment algorithm in digital mode is as follows:

u(k)=u(k-1)+Δu(k)

$\mathrm{<span\; class="notranslate">\; \&Delta;\; u</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\stackrel{<span\; class="notranslate">\; ^</span>}{\mathrm{<span\; class="notranslate">\; T</span>}}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; K</span>}}_{\mathrm{<span\; class="notranslate">\; p</span>}}<span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; sp</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; sp</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span><span\; class="notranslate">\; +</span><span\; class="notranslate">\; [</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; -</span>{\mathrm{<span\; class="notranslate">\; u</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; k</span>}<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}<span\; class="notranslate">\; ]</span>$

是按照前述步骤估计得到的火道温度，K_p表示比例增益，T_sp表示设定的目标火道温度，u₀表示煤气流量的偏置量。u₀在投入自动加热前人工设定或加热自动控制过程中人工修改。Among them, u and Δu represent the gas flow and its increment, k and k-1 represent the current moment and the previous moment, and the superscript represents the parameter corresponding to the time,

is the flue temperature estimated according to the above steps, K _p represents the proportional gain, T _sp represents the set target flue temperature, and u ₀ represents the offset of the gas flow. u ₀ Manual setting before automatic heating or manual modification during heating automatic control.

The obtained gas flow u(k) is the new optimized set value of the gas flow, which is sent to the gas valve control loop (included in the basic level control system of the coke oven), and the valve control loop adjusts the valve opening. to make the gas flow follow the set value to change.

3 Adjust the flue suction

The flue suction is adjusted every time the gas flow setting value is adjusted, and the adjustment process is carried out according to the flue suction neural network model established in the preparation stage. The neural network model is actually a complex compound function, which is realized by programming. The calculation process of the required flue suction is essentially the processing of the input and output data of the function, as follows:

(1) Normalize the gas flow setting value and the expected exhaust gas oxygen content data,

Calculation formula: $\mathrm{the\; y}=\frac{x-a_{\min }}{a_{\max }-a_{\min }},$ The meaning and value of each parameter

The normalization formula is the same as the normalization formula when building the neural network model in the preparation stage.

(2) The flue suction model takes the normalized gas flow and exhaust gas oxygen content as input, and calculates the model output;

(3) The output of the flue suction model is denormalized, and the calculation formula is:

x=a _min +y·(a _max -a _min ), where the meaning and value of each parameter are the same as the normalization formula when establishing the neural network model in the preparation stage. The data obtained after denormalization is the new flue suction optimization setting value and sent to the flue flap control loop (included in the basic level control system of the coke oven), which is adjusted by the flue flap control loop The opening of the flap makes the flue suction change with the set value.

4 "Double crossover" operation of gas flow and flue suction

After the gas flow and flue suction are obtained in the aforementioned steps 2 and 3, the "double cross" operation is adopted in the process of sending the set value to the gas valve control loop and the flue suction control loop: (1) If the gas flow is to be increased , first increase the flue suction, delay 20 seconds before increasing the gas flow; (2) If you want to reduce the gas flow, first reduce the gas flow, delay 20 seconds and then reduce the flue suction.

Three fire channel model correction

The calibration process is programmed in the controller. The so-called calibration refers to the process of recalculating and updating the parameters of the flue path model after collecting the infrared measurement data of the new flue temperature (usually manually measured and manually input into the computer). This operation is triggered by an event, that is, new fire path infrared temperature measurement data is collected. The specific operation is as follows:

1 data window sliding

Whenever new data arrives, the earliest data in the window is deleted and shifted forward in sequence, and the new data is placed at the end of the queue. The significance of adopting the sliding data window is to enable the data set of the model to fully reflect the current working conditions of the coke oven in time, improve the accuracy of the model thus established, and avoid model errors caused by invalid data.

2. Calculation of fire path model parameters

The fire path model takes the first-order polynomial form T _h =a ₁ T _x +a ₀ , the width of the sliding data window is m, let Θ=(a ₁ ,a ₀ ) ^T , A ^T =(T _x ,1), then The fire path model can be written in vector form T _h ＝ A ^T Θ. According to the m data pairs in the sliding data window, the relationship is as follows:

$\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; )</span>}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; T</span>}}_{\mathrm{<span\; class="notranslate">\; h</span>}}^{<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; m</span>}<span\; class="notranslate">\; )</span>}\end{array}\right]<span\; class="notranslate">\; =</span>\left[\begin{array}{c}{\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\\ \mathrm{<span\; class="notranslate">\; m</span>}\\ {\mathrm{<span\; class="notranslate">\; A</span>}}_{\mathrm{<span\; class="notranslate">\; m</span>}}^{\mathrm{<span\; class="notranslate">\; T</span>}}\end{array}\right]<span\; class="notranslate">\; \&CenterDot;</span>\mathrm{<span\; class="notranslate">\; \&Theta;</span>}$

make $h=\left[\begin{array}{c}{T}_h^{\left(1\right)}\\ m\\ T_h^{\left(m\right)}\end{array}\right],{\mathrm{\&Phi;}}^T=\left[\begin{array}{c}{A}_1^T\\ m\\ A_m^T\end{array}\right],$ Then there is H=Φ ^T Θ, and the model parameters are calculated according to the least square algorithm:

Θ=(ΦΦ ^T )ΦH.

After calculating the new parameters of the flue path model, send the parameters to the automatic heating control program for estimating the flue path temperature (see step 1 in II for details).

Example:

1 preparation stage

Record, collect and organize relevant data, including regenerator temperature and corresponding flue temperature; gas flow, flue suction and corresponding flue gas oxygen content.

2 data processing stage

Process the data collected in the previous stage to establish the relevant parameters required by the automatic heating system, including the sliding data window width m and the initial fire path model; the neural network model of the flue suction.

3 automatic heating

The automatic heating process includes three basic operations:

(1) Measure the temperature at the top of the descending regenerator and estimate the current flue temperature according to the flue model;

(2) Calculate gas flow according to formula (3) and formula (4);

(3) Calculate the flue suction according to the flue suction model.

4 Modification of flue temperature model

If there is a new flue temperature (infrared temperature measurement or optical pyrometer temperature measurement) data input, find the corresponding temperature data at the top of the descending regenerator, update the data set in the sliding data window and correct the flue path model.

5 Modification of flue suction model

Once new data (gas flow, exhaust gas oxygen content and corresponding flue suction) are input, the neural network suction model can be corrected.

Claims (12)

1. a coking by coke oven is produced self-heating method, it is characterized in that, comprising:

(1) based on the fire path temperature method of estimation of self-checkign n. flue model;

(2) intelligent fault-tolerance gas flow control method;

(3) based on the flue suction force Automatic adjustment method of flue mathematical models;

Wherein, the self-checkign n. flue model is exactly the fire path temperature data that continuous acquisition regenerator temperature data and corresponding manual measurement obtain, structure " slip data window " data set merges revises the flue model parameter, to follow the tracks of because the variation of funtcional relationship between regenerator temperature that the change of environment of system factor causes and fire path temperature, improve the accuracy that fire path temperature is estimated, strengthen the reliability of heating automatically.

2. the method for claim 1 is characterized in that, funtcional relationship, i.e. flue model between described regenerator temperature and fire path temperature.

3. the method for claim 1 is characterized in that, described slip data window is the data queue of a first outer, and its window width is the length of formation; This window width has determined to be used to proofread and correct the size of the data acquisition of flue model, and is relevant with specific coke oven, characterized the Changing Pattern of coke oven characteristic and environmental factors, is the self-tuning basis of flue model.

4. method as claimed in claim 3 is characterized in that, described window width, and definite method of its best slip data window width is as follows:

The history data set that order is arranged according to time sequence is combined into Z={z ₁, z ₂, K, z _l, wherein l is a data length, subscript is corresponding with time sequence, $z_i=(T_x^i,T_h^i)$ Be that i data are right, T _xBe the regenerator temperature data, T _hBe quirk infrared measurement of temperature data, flue model is got single order polynomial form T _h=a ₁T _x+ a ₀If the width of slip data window is a positive integer m, make Θ=(a again ₁, a ₀) ^T, A ^T=(T _x, 1), then flue model can be write as vector form T _h=A ^TΘ;

Right according to the data of the m in the slip data window, following relation is arranged:

$\left[\begin{array}{c}{T}_h^{\left(1\right)}\\ M\\ T_h^{\left(m\right)}\end{array}\right]=\left[\begin{array}{c}{A}_1^T\\ M\\ A_m^T\end{array}\right]\&CenterDot;\mathrm{\&Theta;}$

Order $H=\left[\begin{array}{c}{T}_h^{\left(1\right)}\\ M\\ T_h^{\left(m\right)}\end{array}\right],$ ${\mathrm{\&Phi;}}^T=\left[\begin{array}{c}{A}_1^T\\ M\\ A_m^T\end{array}\right],$ H=Φ is then arranged ^TΘ is according to least-squares algorithm computation model parameter: Θ=(Φ Φ ^T) Φ H; Data window slides in data acquisition Z, just can set up the flue model that corresponding parameters is Θ according to preceding method; Slide each time and can set up a flue model, can estimate the outer next fire path temperature constantly of moving window by this model ${\hat{T}}_h^{m+1}=A_{m+1}^T\mathrm{\&Theta;},$ If predictor error $e_{m+1}=T_h^{m+1}-{\hat{T}}_h^{m+1};$ The definition cost function: $J\left(m\right)=\underset{i=1}{\overset{l}{\mathrm{\&Sigma;}}}{e}_i,$ Then make the m of J minimum be best slip data window width; Because this cost function curve form is declines-minimum-rising-stablize, has single minimum value, so its minimization process can adopt enumerative technique, since 1, up to obtaining a minimum value, the m corresponding with this minimum value is exactly best slip data window width;

Determined after the m value, in actual the use whenever collecting new data, just upgrade data queue in the moving window automatically according to the mode of first outer, and according to the data acquisition in the window by recomputating and upgrading the flue model parameter, realize the fire path temperature model from normal moveout correction.

5. method as claimed in claim 4 is characterized in that, described realization fire path temperature model from normal moveout correction, its process is: with the earliest data deletion in the window, and shift forward in turn, deposit new data at the least significant end of formation.

6. the method for claim 1 is characterized in that, described (2) intelligent fault-tolerance gas flow control method is based on the change amount of fire path temperature, and promptly trend is regulated gas flow; Intelligent fault-tolerance gas flow setter k under digital form manipulated variable constantly is:

u(k)＝u(k-1)+Δu(k)

$\mathrm{\&Delta;u}\left(k\right)=K_p[{\hat{T}}_h^{\left(k\right)}-{\hat{T}}_h^{(k-1)}]+K_p[T_{\mathrm{sp}}^{(k-1)}-T_{\mathrm{sp}}^{\left(k\right)}]+[u_0^{\left(k\right)}-u_0^{(k-1)}]$

Wherein, u and Δ u represent gas flow and increment thereof, and k and k-1 represent current time and previous moment, subscript represent be and the time corresponding parameters,

Be the fire path temperature that obtains according to the abovementioned steps estimation, K _pThe expression proportional gain, T _SpThe target fire path temperature that expression is set, u ₀The amount of bias of expression gas flow; u ₀Artificially modifying in artificial setting or the heating automatic control process before dropping into heating automatically.

7. the method for claim 1, it is characterized in that, described (3) are based on the flue suction force Automatic adjustment method of flue mathematical models, be to set up the flue mathematical models, and realize that by model with control stack gases oxygen level be the flue suction force Automatic adjustment method of target: when changing gas flow, with the stack gases oxygen level data input flue model of gas flow data and expectation, the output of the flue model of input correspondence therewith is the set(ting)value of new flue suction force; The meaning of this flue suction force is under current gas flow, makes the stack gases oxygen level reach the needed flue suction force of numerical value of expectation.

8. method as claimed in claim 7, it is characterized in that, the described flue mathematical models of setting up, it is the neural net method that adopts artificial intelligence field, this method is gathered historical data, comprises gas flow, stack gases oxygen level and flue suction force, and neural network training, set up nonlinear mathematics funtcional relationship with the neural network formal representation; The input of flue model is gas flow and stack gases oxygen level, and the output of flue model is flue suction force.

9. as claim 7 or 8 described methods, it is characterized in that, the described flue mathematical models of setting up, its concrete steps comprise: (1) data pre-treatment; (2) structure training sample set and test sample book collection; (3) use the training sample set neural network training; (4) use test sample set test neural network.

10. the method for claim 1 is characterized in that, described environment of system factor is the coke oven working of a furnace, and caloric power of gas is gone into stove moisture content of coal and suction.

11., it is characterized in that described intelligent fault-tolerance gas flow control method as claim 1 or 6 described methods, can estimate to exist under the situation of error at fire path temperature, improve the exactness that gas flow is regulated, avoid the wrong furnace temperature fluctuation that causes of regulating, improvement adds the thermal control effect.

12. as claim 7 or 8 described methods, it is characterized in that, the described stack gases oxygen level of setting up flue mathematical models needs, its data can detect from zirconium white, perhaps from the lab analysis result.

Images