TWI735385B - Method and sintering factory for predicting sulfur oxide - Google Patents

Abstract

A method for predicting sulfur oxide by a sintering factory includes: obtaining sintering data of a sintering machine and obtaining desulfurization data of a desulfurization tower; and training a machine learning model based on the sintering data and the desulfurization data to predict a sulfur oxide emission concentration of a chimney.

Description

Sulfur oxide prediction method and sintering plant

This disclosure is about the prediction method of sulfur oxides in sintering plants.

The sintering plant uses coke as fuel for the sintering process. The flue gas produced contains sulfur oxides and needs to be removed by a desulfurization tower before it can be discharged into the atmosphere. There are many determinants of sulfur oxide emissions, which affect each other and are not easy to predict and control. How to propose an accurate sulfur oxide prediction method is a topic of concern to those skilled in the art.

The embodiment of the present invention proposes a method for predicting sulfur oxides, which is suitable for a sintering plant. The sintering plant includes a sintering machine, a desulfurization tower, and a chimney. This prediction method includes: obtaining a plurality of sintering data of the sintering machine and obtaining a plurality of desulfurization data of the desulfurization tower; and training a machine learning model based on the sintering data and the desulfurization data to predict the monosulfur oxide emission from the chimney concentration.

In some embodiments, the sintering data includes sintering machine speed, process flue gas temperature, process reaction temperature, main fan current and fan suction pressure. Desulfurization data includes flue gas flow rate, circulating water volume, inlet sulfur oxide concentration, oxidation tank water intake, magnesium oxide consumption, circulating water pH, flue gas flow rate and booster fan current.

In some embodiments, the above prediction method further includes: obtaining sintering data and desulfurization data according to sampling frequency; training according to the sulfur oxide emission concentration at the first time point and the sintering data and desulfurization data at the second time point Random forest model, in which there is a time shift between the first time point and the second time point; and the time shift is gradually adjusted to retrain the random forest model to obtain the training result with the smallest error and the corresponding time shift, thereby minimizing the error The time shift of is set as a delay time.

In some embodiments, after training the machine learning model, the prediction method further includes: judging whether the error of the machine learning model is greater than a critical value; and if the error of the machine learning model is greater than the critical value, retraining the machine learning model based on new data.

In some embodiments, the aforementioned machine learning model is a regression forest, and the regression forest includes multiple regression trees. The sintering data and the desulfurization data form multiple feature vectors. The above prediction method further includes: solving the objective function represented by the following mathematical formula 1 in the training phase. [Math 1]

表示根據第i個特徵向量所預測的硫氧化物排放濃度

與真實數據

之間的誤差。

表示第k顆迴歸樹的複雜度，K為迴歸樹的個數。n為特徵向量的個數。 in

Indicates the predicted sulfur oxide emission concentration based on the i-th eigenvector

With real data

The error between.

Represents the complexity of the k-th regression tree, and K is the number of regression trees. n is the number of feature vectors.

。 [數學式2]

In some embodiments, the above-mentioned prediction method further includes calculating the weight of the j-th leaf node in the t-th regression tree according to the following mathematical formula 2.

. [Math 2]

表示前t-1棵迴歸樹所預測的數值，

是尋訪第t顆迴歸樹以後走到第j個葉結點的特徵向量所形成的集合，

為實數。 in

Represents the value predicted by the previous t-1 regression tree,

Is the set formed by the feature vector of the j-th leaf node after searching the t-th regression tree,

Is a real number.

From another perspective, the embodiment of the present invention provides a sintering plant, including a sintering machine, a desulfurization tower, a chimney, and a calculation module. The calculation module is used to obtain multiple sintering data of the sintering machine, multiple desulfurization data of the desulfurization tower, and train a machine learning model based on the sintering data and the desulfurization data to predict the sulfur oxide emission concentration of the chimney.

In the above prediction method and sintering plant, both sintering data and desulfurization data are used to predict the sulfur oxide emission concentration, which can achieve a higher accuracy rate.

Regarding the “first”, “second”, etc. used in this text, it does not particularly mean the order or sequence, but only to distinguish elements or operations described in the same technical terms.

The present disclosure proposes a prediction method of sulfur oxides, which is applicable to a sintering plant. FIG. 1 is a schematic diagram of the sintering plant according to an embodiment. The sintering plant includes storage tank 101, mixing barrel 102, raw material silo 103, litter silo 104, sintering machine 105, fan 106, desulfurization tower 107, circulating supplement 108, oxidation tank 109, magnesium hydroxide storage tank 110, desulfurization Nitrate tower 111, fan 112, chimney 113 and calculation module 130. The computing module 130 may be a personal computer, a notebook computer, a server, an industrial computer, or various electronic devices with computing capabilities, etc., which may include a central processing unit, a microprocessor, a microcontroller, a special application integrated circuit, and the like.

Here, the process of the sintering plant is simplified. First, the storage tank 101 stores various raw materials for sintering, such as fine iron ore, flux, fine coke, anthracite, burnt lime, etc. The present invention does not limit these raw materials. The above-mentioned raw meal will be sent into the mixing tank 102, and after being stirred, it will be sent into the raw meal silo 103. The raw meal and the litter in the litter bin 104 are sent to the sintering machine 105, and the sintered waste gas is sucked by the fan 106 and sent to the desulfurization tower 107. The desulfurization tower 107 is used together with the circulating supplement 108, the oxidation tank 109, and the magnesium hydroxide storage tank 110. The circulating supplement 108 is used to provide water, and the magnesium hydroxide storage tank 110 is used to provide magnesium hydroxide for chemical reaction. The flue gas generated by the desulfurization tower 107 will be sent to the denitration tower 111 and discharged from the chimney 113 through the operation of the fan 112. The general knowledge of the present invention should be able to manage the operation of the sintering plant, and will not be described in detail here. In addition, FIG. 1 is only an example of a sintering plant, and the prediction method proposed in this disclosure can also be applied to sintering plants with other settings.

In this embodiment, a plurality of sensors are provided in the sintering plant, and these sensors can be set at any position of the sintering plant to sense related values, and these values will be transmitted to the calculation module 130. The sensors 121 to 127 are shown in FIG. 1. The positions and numbers of the sensors 121 to 127 are only for illustration. Each sensor shown may include a plurality of different types of sensors. The location of the sensor is not limited to the location shown in FIG. 1, or the sensors 121 to 127 can also be built in related equipment. In this embodiment, the sensor 121 is provided in the sintering machine 105 to sense the process reaction speed and the process flue gas temperature of the sintering machine 105. In addition, the sintering machine 105 will also control its own speed (also called the sintering machine). Speed) is transmitted to the calculation module 130. The sensor 122 is used to sense the suction pressure (referred to as the fan suction pressure) and current (referred to as the main fan current) of the fan 106, and is also used to sense the temperature of the flue gas and the inlet sulfur oxide concentration. The sensor 123 is used to detect the flue gas flow rate and the pH value of the circulating water. The sensor 124 is used for sensing the consumption of magnesium oxide, and the sensor 125 is used for sensing the amount of circulating water and the amount of water entering the oxidation tank. The sensor 126 is used to sense the current of the fan 112 (referred to as a booster fan current). The sensor 127 is used to sense the flue gas flow rate and the sulfur oxide emission concentration. For the sake of simplicity, not all sensors are shown in FIG. 1.

The above-mentioned data can be divided into sintering data about the sintering machine 105 and desulfurization data about the desulfurization tower 107. Specifically, the sintering data includes sintering machine speed, process flue gas temperature, process reaction temperature, main fan current and fan suction pressure, etc. Desulfurization data includes flue gas flow rate, circulating water volume, inlet sulfur oxide concentration, oxidation tank water intake, magnesium oxide consumption, circulating water pH, flue gas flow rate, and booster fan current. However, the above data are only exemplary, and other sintering data and desulfurization data can also be added in other embodiments. The calculation module 130 can train a machine learning model based on the sintering data and the desulfurization data, so as to predict the concentration of sulfur oxide emissions in the chimney 113. The machine learning model used here can be a decision tree, a random forest, a multi-level neural network, a convolutional neural network, a support vector machine, etc. The present invention is not limited to this.

Fig. 2 is a flowchart of training machine learning according to an embodiment. Please refer to FIG. 2. First, the database 210 stores sintering data, desulfurization data, and sulfur oxide emission concentration. These data belong to time series data, such as one value per second. Here, a sampling frequency (for example, 1 hour) can be set first, and the average of each piece of data in every 1 hour can be calculated as the training sample, and these training samples form the data set 220. In other words, the data set 220 includes multiple training samples, and each training sample includes sintering data, desulfurization data, and sulfur oxide emission concentration at a certain point in time.

Next, in step 230, delete the abnormality or shutdown data of the sintering machine. Firstly, judge whether the sintering machine is abnormal or shut down. Here, tukey fences are used. After collecting all the speeds of the sintering machine, calculate the first quartile Q1 from the smallest row to the largest one. The three quartiles Q3 and the difference between the two Q3-Q1 are called interquartile range (IQR). If the sintering machine speed of a certain training sample is greater than Q3+1.5*IQR or less than Q1- 1.5*IQR, the training sample is judged to be abnormal, and the data for two hours before and after the abnormal time point will be discarded. In other embodiments, the average value and standard deviation can also be used to delete abnormal data, and the present invention is not limited thereto.

In step 240, the abnormal sulfur oxide emission data is deleted. Similarly, using the Dukai wall method, calculate the quartile Q1, the quartile Q3 and the interquartile range IQR of the sulfur oxide emission concentration. If the sulfur oxide emission concentration of a piece of training data is greater than Q3+1.5*IQR or less than or equal to 0, it is judged to be abnormal and the training sample is deleted.

In step 250, the delay time is searched. Due to the time difference between the sintering machine, the desulfurization tower and the flue gas from the chimney outlet, the sintering data, desulfurization data and sulfur oxide emission concentration of the same time cannot be used to train the machine learning model. FIG. 3 is a schematic diagram illustrating the search delay time according to an embodiment. Here, the sulfur oxide emission concentration at the first time point T _{1 is obtained, and then the sintering data and desulfurization data at the second time point T 2} are obtained, and a random forest model is trained based on these data, where the first time point T ₁ and There is a time shift 310 between the second time point T _2. In some embodiments, 80% of the training samples can be used for training, and the other 20% of training samples can be used for testing. The present invention is not limited to this. After the training is completed, a training result will be obtained, and the training result will contain an error, such as a root mean square error. After training, the time shift 310 can be adjusted gradually to retrain the machine learning model. Here, the sulfur oxide emission concentration is the target, so the second time point T ₂ can be gradually increased or decreased, and different time shifts 310 will correspond to different training results. After training multiple times, the training result with the smallest error and the corresponding time displacement 310 are obtained, and the time displacement 310 with the smallest error is set as the delay time. In some embodiments, the time shift 310 may be adjusted step by step in units of minutes. If the delay time is 60 minutes, it means that it takes about 60 minutes for the flue gas to flow from the sintering machine 108 and the desulfurization tower 107 to the chimney 113.

In the foregoing embodiment, the sintering data and the desulfurization data use the same second time point T ₂ , but in other embodiments, the two pieces of data may also use different time points. For example, referring to Figure 4, _{the sintering data at the third time point T 3} can be combined with the desulfurization data at the second time point T _{2 and the sulfur oxide emission concentration at the first time point T 1} to train the random forest model, where The third time point T _{3 is} before the second time point T ₂ , and the second time point T _{2 is} before the first time point T ₁ . There is a time shift 320 between the first time point T ₁ and the third time point T _3. In such an example, the time shifts 310 and 320 are both variables. The time shift 310 and 320 corresponding to the training result with the smallest error will be used as the aforementioned delay time.

Referring back to FIG. 2, in step 260, the data set is rearranged with the delay time calculated above, so that both the sintering data and the desulfurization data are matched to the corresponding sulfur oxide emission concentration. For example, if the delay time is 60 minutes, _{the sulfur oxide emission concentration at the first time point T 1} will be matched to the sintering data and desulfurization data at the time point (T ₁ -60), thereby forming a sorted A training sample, and these sorted training samples form a sorted data set 270.

The sorted data set 270 can be divided into three parts, which are a training set 281, a validation set 282, and a test set 283. The training set 281 and the validation set 282 are used to train the machine learning model 290. In step 291, the test set 283 is input to the trained machine learning model 290 for testing. In an experiment, the root mean square error of the test set 283 is 2.45, and the determination coefficient R square (R ² ) is 0.952.

Fig. 5 is a flowchart of a method in the inference stage according to an embodiment. 5, first obtain the sintering data and desulfurization data 520 from the sensors 510 of the sintering plant (sensors 121 to 127 in FIG. 2), and then input these sintering data and desulfurization data 520 into the above-mentioned training The machine learning model is used to make predictions (step 530). The predicted sulfur oxide emission concentration 540 is stored in the database 550. The predicted sulfur oxide emission concentration 540 can be used to control other processes or adjust the parameters of any device in the sintering plant. The present invention is not limited to this. . In addition, after the flue gas is discharged from the chimney 113, the ground truth of the sulfur oxide emission concentration can be obtained from the sensor 510, and these real data will also be stored in the database 550.

After a period of time, data of a specific length (for example, one month) can be selected from the database 550, including the above-mentioned real data and the predicted sulfur oxide emission concentration. Next, in step 560, the error of the machine learning model is calculated based on the real data and the predicted sulfur oxide emission concentration, for example, the root mean square error. In step 570, it is determined whether the error is greater than or equal to a critical value (for example, 5). If the result of step 570 is no, there is no need to retrain (step 580). If the result of step 570 is yes, then in step 590 the new data is added to the data set to retrain the machine learning model.

In some embodiments, the machine learning model used in the above embodiments is to add multiple regression trees (referred to as a regression forest). The difference is that a penalty term is added to the objective function to reduce the complexity. Specifically, assuming there are K trees, where K is a positive integer, these K trees can be summed up according to the following mathematical formula 1. [Math 1]

表示第i筆特徵向量，特徵向量可由上述的燒結數據與脫硫數據所組成。

是根據第i筆特徵向量所預測的硫氧化物排放濃度。

是第k顆迴歸樹。一顆迴歸樹具有一或多個中間節點，每個中間節點具有一個判斷式，用以判斷特徵向量中的某一個元素是否大於一特定數值，藉此繼續走向左子樹或是右子樹。迴歸樹的子節點則有一權重，即代表此迴歸樹的輸出。本領具有通常知識者當可理解迴歸樹，在此並不詳細贅述。 in

Represents the i-th feature vector, which can be composed of the above-mentioned sintering data and desulfurization data.

Is the predicted sulfur oxide emission concentration based on the i-th feature vector.

It is the kth regression tree. A regression tree has one or more intermediate nodes, and each intermediate node has a judgment formula for judging whether an element in the feature vector is greater than a specific value, thereby continuing to the left subtree or the right subtree. The child nodes of the regression tree have a weight, which represents the output of the regression tree. Those with ordinary knowledge should understand the regression tree, so I won't go into details here.

，在此實施例中目標函數表示為以下數學式2。 [數學式2]

During the training phase, the real data of sulfur oxide emission concentration is expressed as

In this embodiment, the objective function is expressed as the following mathematical formula 2. [Math 2]

表示預測的硫氧化物排放濃度

與真實數據

之間的誤差，可採用平方差、絕對值差或其他合適的誤差，本發明並不在此限。

表示第k顆迴歸樹的複雜度，例如為葉節點的數目，樹的深度等等。為了求解上述的目標函數，在此是一次求解一顆迴歸樹，進行多次循環以後得到迴歸森林，此循環可以表示為以下數學式3。 [數學式3]

...

Where n is the number of all feature vectors.

Indicates the predicted concentration of sulfur oxide emissions

With real data

The error between the squared difference, the absolute value difference, or other appropriate errors can be used, and the present invention is not limited thereto.

Represents the complexity of the k-th regression tree, such as the number of leaf nodes, the depth of the tree, and so on. In order to solve the above objective function, here is to solve one regression tree at a time, and the regression forest is obtained after multiple cycles. This cycle can be expressed as the following mathematical formula 3. [Math 3]

...

表示第t棵迴歸樹所預測的數值。換言之，t次循環所預測的數值

等於前t-1次循環所預測的數值

再加上第t次循環所得到的迴歸樹

所預測的數值。把數學式3代入至數學式2的目標函數，再透過泰勒展開式來近似函數

以後可以得到在求解第t顆迴歸樹時的目標函數，如以下數學式4所示。 [數學式4]

in

Represents the value predicted by the t-th regression tree. In other words, the predicted value of t cycles

Equal to the value predicted by the previous t-1 cycle

Plus the regression tree obtained in the t-th cycle

The predicted value. Substitute Mathematical Formula 3 into the objective function of Mathematical Formula 2, and then approximate the function through Taylor's expansion

Later, the objective function when solving the t-th regression tree can be obtained, as shown in the following mathematical formula 4. [Math 4]

The Taylor expansion in the above-mentioned Mathematical Equation 4 adopts the approximation of two series. When solving the t-th regression tree, the first t-1 regression trees have been determined. Therefore, the last two terms of Mathematical Formula 4 are constant terms, which do not affect the size of the objective function and can be ignored.

In some embodiments, the complexity of the regression tree can be calculated by the following mathematical formula 5. [Math 5]

為迴歸樹中葉結點的數目，

是第j個葉節點的權重。

與

為使用者可自行定義的實數。把數學式5代入數學式4，再刪除常數項以後可得到以下數學式6。 [數學式6]

in

Is the number of leaf nodes in the regression tree,

Is the weight of the j-th leaf node.

and

It is a real number that can be defined by the user. Substituting Mathematical Formula 5 into Mathematical Formula 4, and then deleting the constant term, the following Mathematical Formula 6 can be obtained. [Math 6]

表示在尋訪第t顆迴歸樹以後走到第j個葉結點的所有特徵向量所形成的集合。當求解第t顆迴歸樹時，權重

是所要求解的變數，數學式6可視為權重

的二次方程式，透過二次方程式的公式解可以求得最佳的權重

與目標函數的值如以下數學式7所示。 [數學式7]

in

It means the set formed by all the feature vectors of the j-th leaf node after searching for the t-th regression tree. When solving the tth regression tree, the weight

Is the variable to be solved, and Mathematical formula 6 can be regarded as the weight

The quadratic equation of, the best weight can be obtained through the formula solution of the quadratic equation

The value of the and objective function is shown in the following equation 7. [Math 7]

[數學式9]

Next, explain how to determine the structure of the regression tree. Starting from depth 0, a node must be determined each time, that is, an element is found from the feature vector for segmentation. After segmentation, a left leaf node and a right leaf node will be generated, the training target It is that the objective function can be minimized after segmentation. The values of the objective function before and after the division are as shown in the following equation 8. Here is to find the best division, so that the gain Gain in Mathematical formula 9 is the maximum. [Math 8]

[Math 9]

在分割後的目標函數的值，

為分割後所有葉節點的個數。

是分割前的目標函數的值，

是分割前所有葉節點的個數。搜尋最適當分割的演算法可參照以下表1所示的虛擬碼。 [表1] Gain

 0

,

for k=1 to m do    

,

    for j in sorted(I, by

) do        

,

       

,

       

)     end end Split with max score in

The value of the objective function after segmentation,

Is the number of all leaf nodes after splitting.

Is the value of the objective function before segmentation,

Is the number of all leaf nodes before splitting. The algorithm for searching for the most appropriate segmentation can refer to the virtual code shown in Table 1 below. [Table 1] Gain

0

,

for k=1 to m do

,

for j in sorted(I, by

) do

,

,

) end end Split with max score

The fifth row of Table 1 is to sort the k-th element in all the feature vectors first, and then take the j-th element after sorting to test the split result. After the loop of the 5th to 8th lines, the kth element in the feature vector is tested, and the best segmentation point for the kth element is found. After the 3rd to 10th loops, all the elements in the feature vector are tested. In the 11th row of Table 1, the regression tree is split by the split point corresponding to the maximum score.

In the above regression forest model, the accuracy of the inference can be increased by adding a penalty term (complexity of the regression tree).

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention. Anyone with ordinary knowledge in the relevant technical field can make some changes and modifications without departing from the spirit and scope of the present invention. The protection scope of the present invention shall be subject to those defined by the attached patent application scope.

101: storage tank 102: mixing barrel 103: raw meal silo 104: litter silo 105: sintering machine 106: fan 107: desulfurization tower 108: circulating supplement 109: oxidation tank 110: magnesium hydroxide storage tank 111: denitration tower 112: Fan 113: Chimney 121~127: Sensor 130: Computing Module 210: Database 220: Data Set 270: Organized Data Set 281: Training Set 282: Validation Set 283: Test Set 290: Machine Learning Model 230,240,250,260,291,530,560,570,580,590: step 310: time shift T _1: first time point T _2: second time point T _3: the third time point 510: sensor 520: data 540 desulfurization sintered data: sulfur oxides emission concentration prediction 550: database

In order to make the above-mentioned features and advantages of the present invention more comprehensible, the following specific embodiments are described in detail in conjunction with the accompanying drawings. [Fig. 1] is a schematic diagram showing a sintering plant according to an embodiment. [Fig. 2] is a flowchart of training machine learning according to an embodiment. [Fig. 3] is a schematic diagram showing the search delay time according to an embodiment. [Fig. 4] is a schematic diagram showing the search delay time according to an embodiment. [Fig. 5] is a flowchart of the method in the inference stage according to an embodiment.

210: database

220: Data Set

270: Organized data set

281: Training Set

282: Validation Set

283: Test Set

290: Machine Learning Model

230, 240, 250, 260, 291: steps

Claims (10)

A prediction method of sulfur oxides is suitable for a sintering plant, the sintering plant includes a sintering machine, a desulfurization tower and a chimney, and the prediction method includes: Obtain a plurality of sintering data of the sintering machine, and obtain a plurality of desulfurization data of the desulfurization tower; and According to the sintering data and the desulfurization data, a machine learning model is trained to predict the sulfur oxide emission concentration of the chimney. The prediction method according to claim 1, wherein the sintering data includes a sintering machine speed, process flue gas temperature, process reaction temperature, main fan current and fan suction pressure, The desulfurization data includes flue gas flow rate, circulating water volume, inlet sulfur oxide concentration, oxidation tank water intake, magnesium oxide consumption, circulating water pH value, flue gas flow rate and booster fan current. The forecasting method as described in claim 1, further including: Obtaining the sintering data and the desulfurization data according to a sampling frequency; Train a random forest model according to the sulfur oxide emission concentration at the first time point and the sintering data and the desulfurization data at the second time point, wherein the difference between the first time point and the second time point A time shift; and The time shift is gradually adjusted to retrain the random forest model, and the training result with the smallest error and the corresponding time shift are obtained, thereby setting the time shift with the smallest error as a delay time. The prediction method according to claim 1, wherein after training the machine learning model, the prediction method further includes: Determine whether the error of the machine learning model is greater than a critical value; and If the error of the machine learning model is greater than the critical value, the machine learning model is retrained according to the new data. The prediction method according to claim 1, wherein the machine learning model is a regression forest, the regression forest includes a plurality of regression trees, the sintering data and the desulfurization data form a plurality of feature vectors, and the prediction method further includes : Solve the objective function represented by the following mathematical formula 1 in the training phase, [Mathematical formula 1]

in

Indicates the predicted sulfur oxide emission concentration based on the i-th eigenvector

With real data

The error between

Indicates the complexity of the k-th regression tree, K is the number of these regression trees, and n is the number of the feature vectors. The prediction method described in claim 5 further includes: Solving the weight of the j-th leaf node in the t-th regression tree according to the following mathematical formula 2

, [Math 2]

in

Represents the value predicted by the previous t-1 regression tree,

Is the set formed by the feature vectors of the j-th leaf node after searching the t-th regression tree,

Is a real number. A kind of sintering plant, including: A sintering machine; A desulfurization tower; A chimney; and A calculation module for obtaining a plurality of sintering data of the sintering machine, obtaining a plurality of desulfurization data of the desulfurization tower, and training a machine learning model based on the sintering data and the desulfurization data to predict The concentration of sulfur oxide emissions from the chimney. The sintering plant according to claim 7, wherein the sintering data includes a sintering machine speed, process flue gas temperature, process reaction temperature, main fan current and fan suction pressure, The desulfurization data includes flue gas flow rate, circulating water volume, inlet sulfur oxide concentration, oxidation tank water intake, magnesium oxide consumption, circulating water pH value, flue gas flow rate and booster fan current. The sintering plant according to claim 7, wherein the calculation module is further used to execute: Obtaining the sintering data and the desulfurization data according to a sampling frequency; Train a random forest model according to the sulfur oxide emission concentration at the first time point and the sintering data and the desulfurization data at the second time point, wherein the difference between the first time point and the second time point A time shift; and The time shift is gradually adjusted to retrain the random forest model, and the training result with the smallest error and the corresponding time shift are obtained, thereby setting the time shift with the smallest error as a delay time. The sintering plant according to claim 7, wherein the calculation module is further used to execute: Determine whether the error of the machine learning model is greater than a critical value; and If the error of the machine learning model is greater than the critical value, the machine learning model is retrained according to the new data.

Images