Coke oven gas reaction temperature prediction method based on bidirectional inference model

Abstract

The invention relates to a coke oven gas reaction temperature prediction method based on a bidirectional inference model, and belongs to the field of coke oven gas prediction. The method specifically comprises the steps of collecting factory coke oven gas data, preprocessing the data collected in the step S1, cleaning the data preprocessed in the step S2, storing the data, dividing the data into a training set and a test set, establishing a bidirectional reasoning modeling method, and reconstructing and evaluating a result. According to the method, the consistency of each feature is ensured by scaling the data, redundant features are eliminated, and the processed data are stored in a standardized format. The normal form not only can ensure that high-quality data is provided, but also has expandability, can meet the requirement of a machine learning model for a large amount of data, and not only can predict the concentration of H2 and CH4 in the coke oven gas, but also can deduce the optimal reaction temperature according to the required product concentration.

Description

Technical Field

[0001] The invention relates to the field of predicting coke oven gas, and in particular to a method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model. Background Art

[0002] Coal is an important contributor to global energy consumption and plays a vital role in promoting economic growth and ensuring the stability of the industrial market. Coal reserves account for the vast majority of the world's proven energy reserves and dominate the energy market. However, with the increasing awareness of green energy and environmental protection, coal has gradually lost its traditional dominant position due to its high energy consumption and environmental pollution problems. Fortunately, the development of coal coking technology has provided a feasible solution to alleviate the contradiction between coal utilization and environmental protection. Coal coking not only reduces the pollution emissions of traditional coal burning, but also provides the chemical industry with abundant raw material resources such as coke, coal gas and coal tar. Among these products, coke oven gas has received widespread attention as a by-product of coking. On the one hand, coke oven gas is a hydrogen-rich gas, so it is of great significance in the field of hydrogen production. It provides a reliable conversion guarantee for coal to clean energy. On the other hand, coke oven gas can be effectively converted into high value-added chemicals such as methanol and synthetic ammonia.

[0003] However, coal is a heterogeneous solid composed of various organic and inorganic substances. Under high-temperature carbonization conditions, various components in coal undergo different thermal decomposition reactions. These reactions interact and influence each other, making the coal coking process extremely complicated. Unfortunately, current theoretical chemical methods cannot fully simulate or derive the correlation between these reactions. The lack of an accurate description of the correlation between the reactions makes the exploration of optimal operating conditions, coal ratios, coal properties, and the prediction of product concentrations more complicated. Therefore, researchers have to use trial and error methods to systematically study the operating conditions one by one, consuming a lot of manpower and time costs in the exploration of coal coking. However, there are still two problems that hinder the application of machine learning modeling methods in the field of coal chemical industry: 1. The above methods are unidirectional in predicting product generation and cannot capture the correlation between reactants for bidirectional modeling. Specifically, these models can only predict product distribution and cannot fully simulate the reaction process to determine the optimal operating conditions, coal ratios, coal properties, etc.; 2. Previous methods directly use raw data, which contains a lot of noise, thereby interfering with the performance of the model. In addition, machine learning modeling methods usually require a lot of data. Therefore, to ensure the performance of machine learning modeling methods, a unified and easily scalable data processing paradigm is essential. Therefore, a modeling method for simulating the correlation of coal coking reactions is urgently needed to improve productivity and reduce research costs. Summary of the invention

[0004] In order to solve the above problems, the present invention proposes a coke oven gas reaction temperature prediction method based on a bidirectional reasoning model.

[0005] To implement the above method, the technical solution of the present invention is as follows:

[0006] A method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model, the specific steps comprising:

[0007] S1. Collect coke oven gas data;

[0008] S2, preprocessing the data collected by S1;

[0009] The maximum and minimum normalization method is used for normalization processing, so that the data value is in the interval [0,1]. The data is preprocessed by data normalization technology. Data normalization is a data preprocessing method widely used in machine learning. It helps to scale the data to ensure the consistency of each feature, thereby reducing the differences caused by different numerical ranges and units. Excessive data differences may cause the machine learning algorithm to mistakenly over-focus on features with larger values. The specific expression is as follows:

[0010]

[0011] In the formula: x represents a set of data that needs to be normalized, x i represents the i-th element of x, represents the normalized data;

[0012] S3, cleaning the data preprocessed by S2;

[0013] In order to improve the computational efficiency and the robustness of the model, the present invention eliminates the method of redundant features. In order to quantify the change of input features, the present invention calculates the rate of change of data. For N-dimensional input features {x1, x2, ...., x n}, where N = 1, 2, ... n, the specific expression for calculating the rate of change is as follows:

[0014]

[0015] Where: N represents the dimension of the input feature, x in represents the n-th dimension feature in the i-th element, x n+1 Represents the n-th dimension feature in the i+1-th element;

[0016] S4, storing the data cleaned in S3;

[0017] In order to facilitate the expansion of the data set, the processed data is stored in a standardized format. After the format is standardized, the data is stored in JavaScript Object Notation (JSON) format. JSON is a lightweight data exchange format that is easy to read and write, and is also easy for machines to parse and generate;

[0018] S5, dividing the data stored in S4 into a training set and a test set;

[0019] The division of the training set and the test set is achieved by dividing the data into five subsets of equal size, four of which are used for model training and the remaining subsets are used for model testing. This process is repeated five times, and a different subset is selected as the test set in each iteration, while the remaining subsets are used for training.

[0020] S6. Establish a bidirectional reasoning model;

[0021] The five-layer fully connected neural network model based on the autoencoder architecture is constructed, and the network model is mainly composed of an encoder and a decoder.

[0022] The specific steps to establish are as follows:

[0023] The five-layer fully connected neural network model based on the autoencoder architecture is constructed, and the network model is mainly composed of an encoder and a decoder.

[0024] The specific steps to establish are as follows:

[0025] S6.1. Determine the inputs of the neural network, including: operating conditions, coal ratio, coal properties, H2 and CH4 concentrations;

[0026] S6.2. Determine that the output of the neural network, including: operating conditions, coal ratio, coal properties, H2 and CH4 concentrations, are consistent with the input;

[0027] S6.3, make the five-layer fully connected neural network learn the connection between input values ​​and output values. The specific steps are as follows:

[0028] S6.3.1. Randomly set some of the data on operating conditions, coal ratio, coal properties, H2 and CH4 concentrations to zero, and then input them into the neural network, and reconstruct the original data, including the inputs that were randomly set to zero;

[0029] S6.3.2. Calculate the error between the reconstructed data and the original input, and update the model according to the back-propagation algorithm;

[0030] S6.3.3, repeat S6.4.1-S6.4.2 until the error between the reconstructed data and the original input does not fluctuate drastically, and the learning of the deep network model is completed;

[0031] S6.3.4. Input coal ratio, coal properties, operating conditions, coal ratio, H2 and CH4 concentrations in sequence, and set the value to be predicted to 0. The model will predict the data with a value of 0 to obtain the prediction result and complete the prediction of the deep network model.

[0032] S6.4. The five-layer fully connected neural network model after learning is trained through the training set, and the test set is used to test the trained network model.

[0033] S7, perform reconstruction and evaluate the results;

[0034] The reconstruction is performed by compressing the input with some values ​​randomly set to 0 into a latent space representation through the encoder, completing and reconstructing the input with some values ​​randomly set to 0 from the latent space representation through the decoder, and using MSE to calculate the similarity between the regenerated input and the original input without some values ​​randomly set to 0.

[0035] MSE is the average of the sum of squared errors. MSE is more sensitive to outliers and can better measure the generalization ability of the modeling method. Its expression is as follows:

[0036]

[0037] Where: y i represents the i-th original data, represents the model output result of the i-th data, and G is the number of this set of original data;

[0038] The original data refers to the input without some values ​​randomly set to 0, and the output result is the reconstructed data, that is, the complete original data after completion and reconstruction;

[0039] The evaluation results include four indicators, including mean absolute error (MAE), root mean square error (RMSE) and correlation coefficient (R 2 ), the specific expressions of the root mean square error and the correlation coefficient are as follows:

[0040] The expression of MAE is as follows:

[0041]

[0042] Where: y is the input without randomly setting some values ​​to 0, is the output of the model, G is the number of original data, and y i represents the i-th original data, Represents the model output result of the i-th data;

[0043] RMSE is the expected value of the squared error of the prediction relative to the label, calculated by the square of MSE. The purpose of introducing RMSE is to supplement MAE to observe the dispersion of model errors. The significant difference between RMSE and MAE indicates that there are great differences in errors in different situations. The expression of RMSE is as follows:

[0044]

[0045] Where: y i represents the i-th original data, Represents the model output result of the i-th data;

[0046] R 2 Often referred to as goodness of fit, R 2 The higher the value, the closer it is to 1, indicating a better fit of the model. 2 The expression is as follows:

[0047]

[0048] Where: TSS represents the total variance, RSS represents the residual variance, ESS represents the explained variance, and the specific expressions of TSS, RSS, and ESS are as follows:

[0049]

[0050]

[0051] Where: represents the mean of y.

[0052] Beneficial Effects of the Invention

[0053] 1. This paper proposes a coal chemical data processing paradigm that scales the data to ensure the consistency of each feature, removes redundant features, and stores the processed data in a standardized format. This paradigm not only ensures the provision of high-quality data, but also has scalability to meet the needs of machine learning models for large amounts of data.

[0054] 2. The present invention proposes a method for predicting the reaction temperature of coke oven gas based on a bidirectional inference model. Due to their inherent limitations, XGBoost, random forest and FNN algorithms cannot perform bidirectional inference analysis, while the bidirectional inference modeling method can not only predict the concentrations of H2 and CH4, but also infer the optimal reaction temperature based on the required concentrations of H2 and CH4.

[0055] 3. The present invention randomly sets the input value to 0, and then inputs the output value into the model to allow the model to reconstruct the original input. Because the input of the model is the missing condition, and the prediction result is the complete condition, through this asymmetric design, the model not only mines the correlation between the components, but also learns the ability to predict missing values.

[0056] 4. Finally, the five performance evaluation results are averaged to obtain the final evaluation result. Compared with the traditional method of randomly deleting part of the data in the validation set and test set, the five-fold cross-validation method can reduce the randomness of the model performance evaluation and provide a more reliable performance estimate. In addition, training and evaluating the model on multiple different training sets and test sets helps to better understand the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is the overall framework diagram of the present invention;

[0058] Figure 2 Schematic diagrams of four network schemes for machine learning of the present invention, wherein (a) is a fully connected neural network (FNN), (b) is an AutoEncoder method, (c) is a bidirectional reasoning model, and (d) is an asymmetric design network model;

[0059] Figure 3 The distribution of predicted values ​​and target values ​​of the four models;

[0060] Figure 4 The influence of different input features on the prediction of product concentration, among which (a) is the influence of XGBoost, (b) is the influence of random forest algorithm, (c) is the influence of FNN algorithm, and (d) is the influence of bidirectional inference model;

[0061] Figure 5 To predict the reaction temperature from the product concentration. DETAILED DESCRIPTION

[0062] The present invention is further described in detail below in conjunction with specific embodiments. Figure 1 As shown, a method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model specifically includes the following steps:

[0063] S1. Collect coke oven gas data;

[0064] S2, preprocessing the data collected by S1;

[0065] The maximum and minimum normalization method is used for normalization processing, so that the data value is in the interval [0,1]. The data is preprocessed by data normalization technology. Data normalization is a data preprocessing method widely used in machine learning. It helps to scale the data to ensure the consistency of each feature, thereby reducing the differences caused by different numerical ranges and units. Excessive data differences may cause the machine learning algorithm to mistakenly over-focus on features with larger values. The specific expression is as follows:

[0066]

[0067] In the formula: x represents a set of data that needs to be normalized, x i represents the i-th element of x, represents the normalized data;

[0068] S3, cleaning the data preprocessed by S2;

[0069] In order to improve the computational efficiency and the robustness of the model, the present invention eliminates the method of redundant features. In order to quantify the change of input features, the present invention calculates the rate of change of data. For N-dimensional input features {x1, x2, ...., x n}, where N = 1, 2, ... n, the specific expression for calculating the rate of change is as follows:

[0070]

[0071] Where: N represents the dimension of the input feature, x in represents the n-th dimension feature in the i-th element, x n+1 Represents the n-th dimension feature in the i+1-th element;

[0072] S4, storing the data cleaned in S3;

[0073] In order to facilitate the expansion of the data set, the processed data is stored in a standardized format. After the format is standardized, the data is stored in JavaScript Object Notation (JSON) format. JSON is a lightweight data exchange format that is easy to read and write, and is also easy for machines to parse and generate;

[0074] S5, dividing the data stored in S4 into a training set and a test set;

[0075] The division of the training set and the test set is achieved by dividing the data into five subsets of equal size, four of which are used for model training and the remaining subsets are used for model testing. This process is repeated five times, and a different subset is selected as the test set in each iteration, while the remaining subsets are used for training.

[0076] S6. Establish a bidirectional reasoning model;

[0077] The five-layer fully connected neural network model based on the autoencoder architecture is constructed, and the network model is mainly composed of an encoder and a decoder.

[0078] The specific steps to establish are as follows:

[0079] The five-layer fully connected neural network model based on the autoencoder architecture is constructed, and the network model is mainly composed of an encoder and a decoder.

[0080] The specific steps to establish are as follows:

[0081] S6.1. Determine the inputs of the neural network, including: operating conditions, coal ratio, coal properties, H2 and CH4 concentrations;

[0082] S6.2. Determine that the output of the neural network, including: operating conditions, coal ratio, coal properties, H2 and CH4 concentrations, are consistent with the input;

[0083] S6.3, make the five-layer fully connected neural network learn the connection between input values ​​and output values. The specific steps are as follows:

[0084] S6.3.1. Randomly set some of the data on operating conditions, coal ratio, coal properties, H2 and CH4 concentrations to zero, and then input them into the neural network, and reconstruct the original data, including the inputs that were randomly set to zero;

[0085] S6.3.2. Calculate the error between the reconstructed data and the original input, and update the model according to the back-propagation algorithm;

[0086] S6.3.3, repeat S6.4.1-S6.4.2 until the error between the reconstructed data and the original input does not fluctuate drastically, and the learning of the deep network model is completed;

[0087] S6.3.4. Input coal ratio, coal properties, operating conditions, coal ratio, H2 and CH4 concentrations in sequence, and set the value to be predicted to 0. The model will predict the data with a value of 0 to obtain the prediction result and complete the prediction of the deep network model.

[0088] S6.4. The five-layer fully connected neural network model after learning is trained through the training set, and the test set is used to test the trained network model.

[0089] S7, perform reconstruction and evaluate the results;

[0090] The reconstruction is to compress the input with some values ​​randomly set to 0 into a latent space representation by an encoder, complete and reconstruct the input with some values ​​randomly set to 0 from the latent space representation by a decoder, and use MSE to calculate the similarity between the regenerated input and the original input without some values ​​randomly set to 0;

[0091] MSE is the average of the sum of squared errors. MSE is more sensitive to outliers and can better measure the generalization ability of the modeling method. Its expression is as follows:

[0092]

[0093] Where: y i represents the i-th original data, represents the model output result of the i-th data, and G is the number of this set of original data;

[0094] The original data refers to the input without some values ​​randomly set to 0, and the output result is the reconstructed data, that is, the complete original data after completion and reconstruction;

[0095] The evaluation results include four indicators, including mean absolute error (MAE), root mean square error (RMSE) and correlation coefficient (R 2 ), the specific expressions of the root mean square error and the correlation coefficient are as follows:

[0096] The expression of MAE is as follows:

[0097]

[0098] Where: y is the input without randomly setting some values ​​to 0, is the output of the model, G is the number of original data, and y i represents the i-th original data, Represents the model output result of the i-th data;

[0099] RMSE is the expected value of the squared error of the prediction relative to the label, calculated by the square of MSE. The purpose of introducing RMSE is to supplement MAE to observe the dispersion of model errors. The significant difference between RMSE and MAE indicates that there are great differences in errors in different situations. The expression of RMSE is as follows:

[0100]

[0101] Where: y i represents the i-th original data, Represents the model output result of the i-th data;

[0102] R2 Often referred to as goodness of fit, R 2 The higher the value, the closer it is to 1, indicating a better fit of the model. 2 The expression is as follows:

[0103]

[0104] Where: TSS represents the total variance, RSS represents the residual variance, ESS represents the explained variance, and the specific expressions of TSS, RSS, and ESS are as follows:

[0105]

[0106] Where: represents the mean of y.

[0107] The model of the present invention is compared with the existing XGBoost (Extreme Gradient Boosting), Random Forest algorithm, and FNN algorithm by forward reasoning. The results are shown in Table 1:

[0108] Table 1: MAE, MSE, RMSE and R of CH4, H2 and CH4+H2 predicted by four models 2 contrast

[0109]

[0110] It can be seen that the results of the forward reasoning of the present invention and the existing model are consistent. Figure 3 , Figure 4 It can also be seen that the model of the present invention has the same effect as the traditional model in terms of forward reasoning;

[0111] like Figure 3 As shown, a scatter plot is shown with the true value of the test data on the x-axis and the predicted value on the y-axis. To help with visualization, a black line representing y=x (the best fit line) is added, indicating that the closer the points are to this line, the better the performance of the model. The same conclusion can be drawn from Table 1, all four modeling models show good performance.

[0112] Figure 4 The feature import parameters are visualized and analyzed to quantify the impact of each component on the prediction task. As can be seen from the figure, the white areas represent values ​​less than 0.001, indicating that their impact is small. Compared with other machine learning methods, the importance coefficients of each component in the bidirectional inference model show a more uniform distribution. XGBoost, random forest and FNN algorithms cannot perform bidirectional inference analysis due to their inherent limitations. Figure 4It can be concluded that C and fixed carbon are the main factors affecting CH4 concentration, while fixed carbon is the most important factor affecting H2 concentration. C plays an important role in both CH4 and H2 concentrations, and fixed carbon is the second most important. The main factors affecting the prediction task have been identified. By taking advantage of the unique advantages of the bidirectional inference model, the optimal reaction temperature can be inferred based on the required CH4 and H2 concentrations.

[0113] from Figure 4 It can be seen that the reaction temperature will affect the concentrations of CH4 and H2. Figure 5 The present invention predicts the reaction temperature diagram from the product concentration, Figure 5 It can be seen that the horizontal axis represents the H2 concentration, and the vertical axis represents the CH4 concentration. Blocks of different colors represent predicted temperatures; dark blocks represent higher predicted temperatures, and light blocks represent lower predicted temperatures. The higher the temperature, the higher the concentrations of CH4 and H2. When the temperature rises from 1005.5°C to 1007.5°C, the concentrations of CH4 and H2 increase from 0.24 to 0.32 and 0.56 to 0.64, respectively, indicating that temperature has a significant effect on the concentrations of CH4 and H2. Moreover, after one training, the present invention can complete multiple tasks, for example: simultaneously reverse prediction of the optimal temperature and forward prediction of CH4 and H2 concentrations, while other traditional models can only perform forward prediction of CH4 and H2 concentrations in one training.

[0114] In summary, the bidirectional reasoning ability of the bidirectional reasoning model is demonstrated, which goes beyond the one-way reasoning limitation of the traditional modeling method. The optimal operating conditions can be obtained according to the required CH4 and H2 concentrations, which can greatly reduce the time and labor costs of researchers.

Claims

1. A method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model, characterized in that: The following steps are involved: S1. Collect coke oven gas data; S2, preprocessing the data collected by S1; S3, cleaning the data preprocessed by S2; S4, storing the data cleaned in S3; S5, dividing the data stored in S4 into a training set and a test set; S6. Establish a bidirectional reasoning model; S7. Perform reconstruction and evaluate the results.

2. The method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model according to claim 1, characterized in that: The preprocessing in step S2 uses the maximum and minimum normalization method to perform normalization processing so that the data value is within the interval [0,1]. The specific expression is as follows: In the formula: x represents a set of data that needs to be normalized, x i represents the i-th element of x, Represents normalized data.

3. The method for predicting the reaction temperature of coke oven gas based on bidirectional reasoning modeling according to claim 1 is characterized in that: x i contains N-dimensional features, and the data cleaning described in step S3 is to calculate the change rate of the N-dimensional features. The specific expression is as follows: Where: N represents the dimension of the input feature, x in represents the n-th dimension feature in the i-th element, x n+1 Represents the n-th dimension feature in the i+1-th element.

4. The method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model according to claim 1, characterized in that: The division of training set and test set described in step S5 is achieved by dividing the stored data into five subsets of equal size, four of which are used for model training and the remaining subsets are used for model testing. This process is repeated five times, and a different subset is selected as the test set in each iteration, while the remaining subsets are used for training.

5. The method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model according to claim 1, characterized in that: The bidirectional reasoning model described in step S6 is constructed by a five-layer fully connected neural network model based on the autoencoder architecture. The neural network model is mainly composed of an encoder and a decoder. The specific establishment steps are as follows: S6.

1. Determine the inputs of the neural network, including: operating conditions, coal ratio, coal properties, H2 and CH4 concentrations; S6.

2. Determine that the output of the neural network, including: operating conditions, coal ratio, coal properties, H2 and CH4 concentrations, are consistent with the input; S6.3, make the five-layer fully connected neural network learn the connection between input values ​​and output values. The specific steps are as follows: S6.3.

1. Randomly set some of the data on operating conditions, coal ratio, coal properties, H2 and CH4 concentrations to zero, and then input them into the neural network, and reconstruct the original data, including the inputs that were randomly set to zero; S6.3.

2. Calculate the error between the reconstructed data and the original input, and update the model according to the back-propagation algorithm; S6.3.3, repeat S6.4.1-S6.4.2 until the error between the reconstructed data and the original input does not fluctuate drastically, and the learning of the deep network model is completed; S6.3.

4. Input coal ratio, coal properties, operating conditions, coal ratio, H2 and CH4 concentrations in sequence, and set the value to be predicted to 0. The model will predict the data with a value of 0 to obtain the prediction result and complete the prediction of the deep network model. S6.

4. The five-layer fully connected neural network model after learning is trained through the training set, and the test set is used to test the trained network model.

6. The method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model according to claim 1, characterized in that: The reconstruction described in step S7 is to compress the data input into a potential space representation by the encoder, reconstruct the data input from the potential space representation by the decoder, and use MSE to calculate the similarity between the reconstructed data and the original data.

7. The method for predicting the reaction temperature of coke oven gas based on a bidirectional reasoning model according to claim 1, characterized in that: The evaluation result described in step S7 includes four indicators: mean absolute error, mean square error, root mean square error and correlation coefficient.

Images