CN102176221A

CN102176221A - Coke furnace temperature predicting method based on dynamic working conditions in coke furnace heating and burning process

Info

Publication number: CN102176221A
Application number: CN2011100639488A
Authority: CN
Inventors: 吴敏; 雷琪; 曹卫华; 焦国华; 安剑奇; 刘建群
Original assignee: Central South University
Current assignee: Central South University
Priority date: 2011-03-16
Filing date: 2011-03-16
Publication date: 2011-09-07
Anticipated expiration: 2031-03-16
Also published as: CN102176221B

Abstract

The invention discloses a coke furnace temperature predicting method based on dynamic working conditions in a coke furnace heating and burning process. The method is characterized by comprising the following steps: 1, acquiring gas flow and coke furnace temperature data in various historical working conditions, and establishing a sample bank; 2, calculating the similarity of a current working point and a sample in the sample bank on the basis of an included angle of Euclidean distance and variation trend of the current working point and the sample in the sample bank; and 3, selecting a plurality of sample structural learning sets with maximum similarity; establishing a learning-set-based local linear model by adopting an iterative least square method; and calculating an output value corresponding to the current working point as a coke furnace temperature predicted value in the coke furnace heating and burning process. The coke furnace temperature predicting method based on dynamic working conditions in the coke furnace heating and burning process has high prediction accuracy, and has the online adaptive capability.

Description

Coke oven temperature prediction method in coke oven heating combustion process based on dynamic working conditions

Technical Field

The invention belongs to the field of coke oven heating combustion control, and relates to a coke oven temperature prediction method in a coke oven heating combustion process based on dynamic working conditions.

Background

The steel is used as an important basic raw material and strategic material for the development of national economy and national defense war industry in China, is widely applied to various industries such as machinery, electronics, building materials, traffic, aerospace, aviation, national defense war industry and the like, and has a very important position in the development of national economy. The coke oven is one of the extremely important industrial ovens in the coal chemical industry, is key equipment for producing coke, in the steel industry, the coke quality directly influences the production cost of enterprises, the product quality and the economic benefit of the enterprises, and the optimal control of the heating combustion process of the coke oven plays an important role in reducing the production cost of the coking enterprises and improving the economic benefit.

The coke oven temperature is an important parameter in the heating and burning process of the coke oven, if the flame path temperature is unstable and fluctuates greatly, the coke is heated unevenly, and the local coke formation can cause black smoke when the coke is discharged, thereby directly influencing the coke quality and the service life of the oven body. However, in actual production, due to the unique structure of the coke oven, the installation of the thermocouple is difficult, the temperature of the flame path is as high as about 1300 ℃, and the installed thermocouple is easy to damage due to high temperature, so that the method has very high cost and difficult maintenance, and is rarely adopted in practical application. Therefore, real-time detection of the coke oven temperature is very difficult.

The method is a dynamic prediction method based on a local model. When modeling complex industrial processes, it is always desirable to be able to cover all the operating condition data, so a global modeling method is generally adopted. However, the global modeling method generally has the problems of long learning time, poor generalization capability, difficulty in online updating of the model and the like. The heating and burning process of the coke oven is a complex heat transfer and chemical change process, the structure of the oven body of the coke oven is complex, the operating environment is severe, the detection means are few, the working conditions are changeable, and the process model is very difficult to establish by adopting a global modeling method. Aiming at the characteristics of the heating combustion process of the coke oven, a new method and a new technology are researched to establish a high-precision coke oven temperature dynamic model based on the thought of local modeling, and the method has extremely important significance for the optimal control of the heating combustion process of the coke oven.

Disclosure of Invention

The invention aims to solve the technical problem of providing a method for predicting the temperature of a coke oven in the heating and burning process of the coke oven based on dynamic working conditions, wherein the method for predicting the temperature of the coke oven in the heating and burning process of the coke oven based on the dynamic working conditions has high prediction precision and online self-adaptive capacity.

The technical solution of the invention is as follows:

a coke oven temperature prediction method in a coke oven heating combustion process based on dynamic working conditions comprises the following steps:

step 1: acquiring historical gas flow and coke oven temperature data under various working conditions, and establishing a sample library

u_i-2、u_i-3The gas flow rate of the i-2 th cycle and the gas flow rate of the i-3 th cycle, y_i-1、y_iThe coke oven temperature of the i-1 th cycle and the coke oven temperature of the i-th cycle, x_i＝{u_i-1，u_i-2，y_i-1The input variable of the ith period is input;

step 2: calculating the similarity between the current working point and the samples in the sample library based on the Euclidean distance between the current working point and the samples in the sample library and the included angle of the change trend;

and step 3: selecting a plurality of samples with the maximum similarity to construct a learning set; establishing a local linear model based on a learning set by adopting an iterative least square method; and calculating an output value corresponding to the current working point as a predicted value of the coke oven temperature in the heating combustion process of the coke oven.

In step 2, for the current working point x_q＝[u_q-1，u_q-2，y_q-1]The working point corresponding to a sample in the sample library, namely the input variable x corresponding to the sample_i＝{u_i-1，u_i-2，y_i-1}，

Computing

Wherein,

the value of alpha is 0.3-0.5, and the value of alpha is related to the heat value of the coal gas. The coal gas has high heat value and quick influence on the temperature, the influence of the coal gas in a relatively short period on the temperature is increased, and alpha can be larger at the moment. θ_iIs Δ x_qAnd Δ x_iThe included angle between them;

if cos (theta)_i) < 0, i.e. Δ x_qAnd Δ x_iIs greater than 90 degrees, indicating x_qAnd x_iDifference in variation tendency ofLarger, will x_iThe corresponding samples are discarded;

if cos (theta)_i) Not less than 0, the similarity of 2 working points is

Where γ is a weight, γ is determined experimentally by taking a value [ γ ] between 0 and 1, and γ is 0.7 experimentally for the problem of the present application. D_i＝||x_q-x_i||₂，i＝1，2，L N。

The step3 is:

step a: to s_iSorting from big to small, and sequentially selecting corresponding k maximum s from a sample library_iCorresponding k groups of data, (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3R denotes the real number field, l has a value range of [ k ]_min，k_max]Integer between, k_min＝3，k_max＝30，k＝28；X_lIs composed of a plurality of x_i＝{u_i-1，u_i-2，y_i-1Sample set of [ k ]_min3 is a requirement for a least squares solution, and the minimum data cannot be less than 3. k is a radical of_max30 is that the maximum number of regression samples is 30, and this 30 is a larger number of samples taken, and it can be generally determined by simulation that when the number of samples is greater than a certain number, increasing the number of samples has little influence on the accuracy of regression, and the 30 is taken in this application. k 28 is the value of l which is determined to be the best accuracy in practical terms according to the following step b, in combination with the following step, i_opt】。

Step b: let l equal 3, then have P₃＝W₃X₃，v₃＝W₃y₃，

According to

Calculating e₃(ii) a Wherein, W_iRepresents fromThe upper left corner of the middle part takes a matrix formed by i multiplied by i elements; (X)₃，y₃) Is (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3Middle S_i3 groups of data with the largest sequence from large to small; p_l＝W_lX_l，v_l＝W_ly_l，y_jIs y_lThe (j) th element of (a),

and

are respectively a matrix P_lAnd X_lRow j of (1);

step c: according to the sorted samples, adding one sample in turn, updating l to l +1, and calculating P_l＝W_lX_l，v_l＝W_ly_lThen, then

According to

Equation e_l；

Step d: if l is less than or equal to k_maxAnd c, repeating the step c; otherwise, all e's are compared_l，l＝k_min，...，k_maxFind out e_lThe value of l corresponding to the minimum value is l_optI.e. l_opt＝arg min(e_l) Obtaining the model output at the moment

x_q＝[u_q-2，u_q-3，y_q-1]，

Is a query point x_qAnd (4) outputting an estimated value of the corresponding model, namely a predicted value of the coke oven temperature in the heating combustion process of the coke oven.

Has the advantages that:

the method for predicting the coke oven temperature in the coke oven heating combustion process based on the dynamic working conditions, disclosed by the invention, is used for finding a field data set which is most similar to a historical data set in the historical data set by considering the factors of working condition change aiming at certain input data, and predicting the output of a system by utilizing a modeling method. The model is only generated at the current input sample point, with a higher accuracy. And the model can be provided with online self-adaptive capacity by adding new measurement data to the data sample set or deleting old measurement data from the data sample set online.

The coke oven temperature prediction method based on the coke oven heating combustion process of the dynamic working condition is a coke oven heating combustion process self-adaptive modeling method based on the dynamic working condition, and compared with a global modeling method, the method has the characteristics of high precision, wide coverage working condition range and self-adaptation; compared with the existing local learning modeling method, the method considers the trend of process change when selecting the modeling sample, and better meets the requirements of industrial process modeling.

Drawings

FIG. 1 is a schematic diagram of the similarity of k- Δ VNN vectors.

FIG. 2 is a schematic diagram of predicted results and actual measured results based on the k- Δ VNN method.

FIG. 3 is a graph of the error between predicted and measured results based on the k- Δ VNN method.

Detailed Description

The invention will be described in further detail below with reference to the following figures and specific examples:

example 1:

the main body of the coke oven is generally composed of 50-100 heating units, each heating unit comprises a carbonization chamber, a combustion chamber and a regenerative chamber, and the carbonization chamber and the combustion chamber are separated by only one wall. The carbonization chamber is a place where coal is isolated from air for dry distillation, and the combustion chamber is a place where coal gas is combusted. Each combustion chamber comprises a certain number of vertical flame paths, wherein every two vertical flame paths are used as a pair to form a gas passage, and two ends of the gas passage are respectively connected with the regenerative chambers. The heating coal gas and the air are mixed and combusted in a flame path of the combustion chamber to generate heat, and the hot waste gas mainly conducts heat by radiation at high temperature and transmits the heat to the coal material in the coal carbonization chamber by means of convection heat transfer.

The coke oven temperature is the average value of the temperature measuring flame path temperatures of all combustion chambers of the whole oven and is an important process parameter of the heating combustion process of the coke oven. The flow rate of the coal gas directly influences the temperature of the coke oven, but the same flow rate of the coal gas can also cause great difference of the temperature of the coke oven, which is mainly because the current working conditions of the heating and burning process of the coke oven are different, so the working conditions are also important factors influencing the temperature of the coke oven.

The working condition of the heating and burning process of the coke oven is closely related to the properties of the coal material in the current oven and the heat value of the current coal gas. In the heating process of the coke oven, firstly, a production plan is made according to production requirements and production conditions, and then, the operations of coal charging, coke pushing, coke blocking and coke quenching are carried out through production plan scheduling. During the heating process of the coke oven, most heat is taken away by coke cakes, and the coal in each coking stage in the coke oven can be kept approximately equivalent after the coke cakes are discharged normally according to a production plan.

However, the situation that coke is not pushed for a long time and the like can be caused due to the faults of mechanical equipment and electrical equipment, and the temperature of the coke oven is increased quickly due to the increase of coal at the end of coking; on the other hand, after the trouble is eliminated, the coking chambers that need to be discharged must be collectively treated in order to recover the normal production, which results in a large amount of coal in the initial stage of coking, and the overall temperature of the coke oven tends to decrease. In the heating combustion process of the coke oven, a gas heat value instrument is not installed generally due to the severe production environment, but the heat value of blast furnace gas for heating greatly fluctuates, so that the trend of the temperature change of the coke oven is different.

From the above analysis, it can be seen that the working conditions reflect the trend of process changes, and due to the complexity of the working conditions, the time lag and inertia characteristics of the heating combustion process of the coke oven become complex. Therefore, the gas flow rate including the previous cycle or several cycles, which affects the coke oven temperature, cannot be immediately reflected on the coke oven temperature at the present time. By analyzing the relationship between the on-site coke oven temperature and the historical data of the heating gas flow, the slow change of the coke oven temperature and the blast furnace gas flow u of the previous period and the previous period can be seen_i-1、u_i-2Has a large correlation with the previous time temperature y_i-1Correlation, therefore, N sets of historical process data are selected to build a sample library, X_N＝[x₁，y₁]，[x₂，y₂]，...，[x_N，y_N]Wherein

as input vector of the process model, from the gas flow u of the previous and second periods_i-1、u_i-2Temperature y at the previous moment_i-1Composition i.e. x_i＝[u_i-1，u_i-2，y_i-1，](ii) a Output variable

The current coke oven temperature.

The modeling method adopted by the patent does not train the sample in advance until the output corresponding to a certain input value needs to be estimated, and the input value is called as a query point. For query point x_q＝[u_q-1，u_q-2，y_q-1，]Finding out the samples from the existing N groups of sample setsForming a new sample set for modeling from similar data, establishing a mapping f, and obtaining the output of the corresponding query point

The specific process comprises two steps: (1) selecting similar data; (2) and establishing a local model.

(1) Selecting similar data

The basic principle of this patent modeling is that similar inputs produce similar outputs. How to select and x_q＝[u_q-1，u_q-2，y_q-1]The establishment of a learning set by samples with similarity is a main factor for determining the accuracy of the model.

The coke oven heating combustion process is a steady-state process which changes slowly, and the change trend of the process is very important to influence the temperature of the coke oven. The conventional method generally adopts x_qAnd x_iDistance of (2) reaction data, i.e. x_qThe closer the distance the greater the similarity compared to the N samples in the sample library.

In order to improve the accuracy of the model, the method takes the trend of the change of the heating combustion process of the coke oven into the criterion of selecting the samples, selects the samples in the sample set and combines the samples with x_qWhen the samples are similar, not only the distance between the two samples but also whether the trend of the change of the two samples is similar or not are considered. Input vector x in sample set_i＝[u_i-1，u_i-2，y_i-1，]The input variable corresponding to the previous cycle is x_i-1＝[u_i-2，u_i-3，y_i-2，]，

The change trend of the coke oven temperature in the sample relative to the coal gas flow can reflect the current working condition of the coke oven. By using

Represents a sample x_iAnd query point x_qA trend of change. The trend of the change is influenced by the gas flow u of the previous period_i-1Gas flow u in the first two periods_i-2Wherein α represents the gas flow u of the previous cycle_i-1The influence degree of the coke oven temperature is due to the large time lag characteristic of the heating combustion process of the coke oven_i-2Influence on the temperature of the coke oven is greater than u_i-1Therefore, α is 0.3. x is the number of_qAnd x_iThe cosine values of the included angles of the distance and the trend of the temperature change of the coke oven are expressed as

Wherein, theta_iIs Δ x_qAnd Δ x_iThe included angle therebetween.

If cos (theta)_i)＜0，Δx_qAnd Δ x_iIs greater than 90 degrees, x_qAnd x_iThe trend of change of (c) is relatively large, and the sample is discarded.

If cos (theta)_i) And the similarity is the weighted sum of the distance and the included angle. x is the number of_qAnd x_iSimilarity s_iIs defined as:

as can be seen from FIG. 1, if cos (θ)_i)＞0，Δx_qAnd Δ x_iIs less than 90 degrees, cos (theta)_i) The closer to 1, Δ x_qAnd Δ x_iThe smaller the included angle is, the greater the similarity of the change trends of the two is.

Similarity s_iAnd judging through the Euclidean distance between the input vector and the sample vector and the included angle information of the process change trend.Following the Euclidean distance d_iIs increased and decreased, cos (theta)_i) Angle θ with process trend_iIs increased and decreased. Thus, the Euclidean distance d between the input vector and the vector in the sample set_iThe smaller, cos (. theta.)_i) The greater the similarity s_iThe larger.

(2) Establishing a local model

Because the local model is required to be established for each query point, the calculation amount is large, and a model which is relatively simple is usually selected. Globally, the heating combustion process of the coke oven has complex nonlinearity, but the trend of the temperature change of the coke oven is similar near a certain working condition of the operation of an object, and the characteristic of the heating combustion process of the coke oven can be approximated by a linear model with enough accuracy. The local model is only used for describing the function mapping relation near the query point, and a simple model can also ensure certain precision. Therefore, an Auto Regressive evolution (ARX) model is used to build a local model of the coke oven heating combustion as follows:

y_k＝x^T _k-1B

wherein,

is the model output at the kth sampling instant, x_k-1Is a regression vector, B is a parameter of regressionAs follows:

x_{k - 1} = [y_{k - 1}, L, y_{k - n_{y}}, u_{k - n_{d} - 1}, L, u_{k - n_{d} - n_{u}}], - - - (5)

in the above formula, n_yAnd n_uIs an integer related to the order of the model, n_dIs the process skew. In this application, n_d＝0，x_k-1Gas flow u from the previous cycle or two cycles_i-1、u_i-2Temperature y at the previous moment_i-1Composition n_y＝1，n _u2 or x_i＝[u_i-1，u_i-2，y_i-1，]。

For the ARX model described above, the problem translates into a correlation data set (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3Establishing a mapping f and, from this mapping, for an input vector x_q＝[u_q-1，u_q-2，y_q-1，]To determine the corresponding predicted value of the coke oven temperature

。

To s_iSorting from big to small, selecting the largest group of data (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3，l＝[k_min，k_max]，k_min3. Let W_lIs a diagonal matrix in which the diagonal elements w_iThe magnitude of the contribution of each sample to the local modeling, since the data is sorted,in order to obtain a specific value of l and verify the quality of the model in time, the method adopts a cross-validation method: the 'de-error' of the model is calculated, namely one sample is removed from all modeling samples, the model is modeled by using the remaining samples, and the model at the moment is verified by using the removed sample.

De-one error is expressed as

Wherein

P_l＝W_lX_l，v_l＝W_ly_l，y_jIs y_lThe (j) th element of (a),and

is a matrix P_lAnd X_lRow j of (2). Based on the de-one error, the problem is translated into determining the optimal magnitude of l, l_optSo that the mean square error of a de-one error is minimized, i.e.

l_opt＝arg min(e_l)

The specific calculation process is as follows:

step 1: to s_iSorting the selected maximum group of data from large to small, (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3，l＝[k_min，k_max]，k_minDetermining k as 3_max＝30。

Step 2: let l equal 3, P₃＝W₃X₃，v₃＝W₃y₃Then, then

Calculating e according to equation (1)₃。

Wherein (X)₃，y₃) Is (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3Middle S_i3 groups of data with maximum order from large to small

Step 3: adding samples in sequence according to the sorted samples, l ═ l +1, and P_l＝W_lX_l，v_l＝W_ly_lThen, then

Calculating e according to equation (1)_l；

Step 4: if l is less than or equal to k_maxRepeating step 3; otherwise, all e's are compared_l，l＝k_min，...，k_maxFind out e_lThe value of l corresponding to the minimum value is l_optI.e. l_opt＝arg min(e_l) Obtaining the model output at the moment

x_q＝[u_q-2，u_q-3，y_q-1]，

In a conventional modeling method, after a model structure and parameters are determined, when the working condition of a system changes, modeling needs to be performed again. According to the coke oven heating combustion process adaptive modeling method based on the dynamic working conditions, a local estimation model is reestablished for each input data, so that when sample data changes, the local model can be changed along with the change of the sample data, and the coke oven heating combustion process adaptive modeling method is adaptive to the continuously changing working conditions.

Application example:

the invention is applied to an optimized control system of a coking plant of a certain steel plant, collects data of two months, and establishes coal gas flow u including the first 1 and 2 periods of input through filtering the coal gas flow_i-1、u_i-2And the coke oven temperature y of the preceding cycle_i-1The output is the coke oven temperature y_iThe sample library of (1). After two weeks of operation, fig. 2 is an operation effect graph, and fig. 3 is an error graph of the model. It can be seen that the model error of the invention is small, the maximum is 13.035 ℃, the error is 87% within +/-7 ℃, and conditions are provided for the optimization control.

Based on the same data, simulation is carried out by adopting a neural network global modeling method, as shown in Table 1, the maximum error is 142.2329 ℃, and RMSE (root mean square error) is more than 10 times of that of the invention, which is mainly because the working condition range of the heating combustion process of the coke oven is wide, and the neural network method cannot cover the working conditions, so that the model error is very large. Compared with the traditional local model, the RMSE and MAXE (maximum absolute error) of the invention are greatly improved.

TABLE 1

As for the heating combustion process of the coke oven, a neural network method is adopted to carry out modeling on the heating combustion process of the coke oven, and the modeling is actually a global modeling method, but because the working conditions of the heating combustion process of the coke oven are complicated and changeable, the neural network is very difficult to cover all the working condition data, so the generalization capability of the model is very poor. Meanwhile, when the neural network modeling method is used for updating the model on line, data needs to be collected and retrained, occupied system resources and time are long, and the problem that the modeling model is difficult to update exists. Therefore, when modeling is performed by using a neural network, the accuracy of the model is low.

The model established by the traditional local modeling method is mainly characterized in that similarity is judged according to the distance between a query point and data in a sample library, and similar data are selected to establish the local model. The coke oven heating combustion process is a continuously-changing slow process, and the changing trend of the coke oven heating combustion process is an important aspect of the process, so that the method considers the changing trend of the coke oven heating combustion process into the criterion of selecting samples, the similarity of the selected samples is larger, and the accuracy of the established model is greatly increased. Meanwhile, for each input data, a local estimation model is required to be established again, so that when sample data changes, the local model can be changed along with the sample data, and the method adapts to the continuously changing working condition.

Claims

1. A coke oven temperature prediction method in a coke oven heating combustion process based on dynamic working conditions is characterized by comprising the following steps:

u_i-2、u_i-3The gas flow rate of the i-2 th cycle and the gas flow rate of the i-3 th cycle, y_i-1、y_iRespectively in the i-1 th weekCoke oven temperature of phase and coke oven temperature of i cycle, x_i＝{u_i-1，u_i-2，y_i-1The input variable of the ith period is input;

2. The method for predicting the coke oven temperature in the coke oven heating combustion process based on the dynamic working conditions of claim 1, wherein in the step 2, for the current working point xq ═ [ uq-1, uq-2, yq-1 ═ uq-1]The working point corresponding to a sample in the sample library, namely the input variable x corresponding to the sample_i＝{u_i-1，u_i-2，y_i-1}，

Computing

Wherein,

alpha is 0.3-0.5 and theta_iIs Δ x_qAnd Δ x_iThe included angle between them;

if cos (theta)_i) < 0, i.e. Δ x_qAnd Δ x_iIs greater than 90 degrees, indicating x_qAnd x_iHas a large variation trend, x is set_iThe corresponding samples are discarded;

if cos (theta)_i) Not less than 0, the similarity of 2 working points is

Wherein gamma is weight, gamma takes a value between 0 and 1, d_i＝||x_q-x_i||₂，i＝1，2，L N。

3. The method for predicting the coke oven temperature in the coke oven heating combustion process based on the dynamic working conditions of the claim 1 or 2, wherein the step3 is as follows:

step a: to s_iSorting from big to small, and sequentially selecting corresponding k maximum s from a sample library_iCorresponding k groups of data, (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3R denotes the real number field, l has a value range of [ k ]_min，k_max]Integer between, k_min＝3，k_max＝30，k＝28；x_lIs composed of a plurality of x_i＝{u_i-1，u_i-2，y_i-1The sample set of { fraction };

step b: let l equal 3, then have P₃＝W₃X₃，v₃＝W₃y₃，

According to

Calculating e₃(ii) a Wherein, W_iRepresents from

The upper left corner of the middle part takes a matrix formed by i multiplied by i elements; (X)₃，y₃) Is (X)_l，y_l)，y_l∈R^l×1，X_l∈R^l×3Middle S_i3 groups of data with the largest sequence from large to small; p_l＝W_lX_l，v_l＝W_ly_l，y_jIs y_lThe (j) th element of (a),

and

are respectively a matrix P_lAnd X_lRow j of (1);

According to

Equation e_l；

x_q＝[u_q-2，u_q-3，y_q-1]，