Coking coal blending method based on convolutional neural network

Abstract

The invention discloses a coking coal blending method based on a convolutional neural network, and the method comprises the following steps: S1, collecting original data which is historical coal blending data of a coking plant for many years, and using actual production data as a data source; s2, data cleaning and data set production: cleaning original data, removing irrelevant data and abnormal values, and generating a data set; s3, feature extraction and evaluation: extracting features from the original data; s4, algorithm construction: based on a deep neural network model, performing feature training by taking coal blending data as input and optimizing a proportioning model, the proportioning model comprising a plurality of convolutional layers and LSTM units which are combined together through a full connection layer in a final stage to generate output, and performing analysis and calculation on input available coal types by using the coal blending model to obtain a matching model; according to the coal blending model, the working experience and professional knowledge of a coal blending engineer are integrated, and the accuracy and scientificity of the coal blending work are further enhanced.

Description

Technical Field

[0001] The invention belongs to the technical field of coal coking, and in particular relates to a coking coal blending method based on a convolutional neural network. Background Art

[0002] With the rapid development of the chemical industry, improving the utilization rate of coal energy has become the main method for coal chemical enterprises to reduce production costs. In coking plants that carry out coal coking, some coking plants have begun to use various coal blending models to optimize coal blending in recent years. Compared with the previous use of pure manual experience to calculate the coal feeding scheme, the use of coal blending models can provide scientific suggestions in coal selection, coal ratio, inventory and usage analysis, coal procurement, etc., and has become an important way for the coking industry to increase costs and efficiency. The quality and cost of coke produced by coking plants mostly depend on the comprehensive proportion of mixed coal before entering the furnace. The production of coke is a complex chemical reaction, involving the principle of coal colloid layer superposition, interchangeability principle, co-carbon principle, coal-rock blending principle, etc., and various indicators of coal such as volatility, sulfur, ash, moisture, lithofacies, etc. may affect the quality of coke production. It is difficult to reduce the cost of coal feeding while ensuring the quality of coke by relying solely on manual experience. Therefore, how to balance the various indicators of coal entering the furnace and maximize production benefits (production benefits = coke benefits + chemical product benefits - coal cost) on the basis of producing qualified quality coke has become a crucial and urgent problem to be solved in the coking industry.

[0003] Chinese invention patent CN102585868A discloses a method for coking coal blending, which is to blend coal by a fixed proportion of coal entering the furnace. Although this method draws on past experience, it has the limitation of fixed thinking and cannot be further diverged and promoted. It is completely driven by historical data and does not absorb the professional experience of practitioners. The final results may have some differences, and the data accuracy deviation is large when different types of coal are involved in production.

[0004] Another Chinese invention patent CN118333229A discloses a precise coal blending algorithm and system, which uses a reinforcement learning algorithm with the following core constraints:

[0005] r0=-|T C -T0|

[0006]

[0007] Where T C is the combustion temperature and thermal efficiency in the actual combustion process, T0 is the combustion temperature and thermal efficiency required for coal blending, ARG c The highest calorific value in production, ARG safeis the highest data allowed by the safety parameters, and reward is the reward value. Through this constraint, the combustion efficiency required in actual combustion can be calculated and the thermal efficiency can be improved.

[0008] Another Chinese invention patent CN110295049A discloses a method for efficiently utilizing complex mixed coal for coking. This scheme establishes a mechanism model in the coking process and uses the reflectance index of the vitrinite group to guide coal blending. Since the coking process is a complex chemical reaction, it is very difficult to fully reflect the chemical process in this scheme. There are hidden dangers of incomplete consideration of some factors, which poses certain risks to coal blending.

[0009] Therefore, the present invention provides a coking coal blending method based on convolutional neural network to solve the problems raised by the above-mentioned background technology. Summary of the invention

[0010] In view of the problems raised by the above background technology, the purpose of the present invention is to provide a coking coal blending method based on convolutional neural network, and to build a coking coal blending model based on convolutional neural network. The coal blending model is used to analyze and calculate the available coal types input, and the coal type ratio scheme with the highest comprehensive benefit is obtained to guide business personnel in coal blending. This coal blending model incorporates the work experience and professional knowledge of coal blending engineers, and further enhances the accuracy and scientificity of coal blending work.

[0011] In order to achieve the above technical objectives, the technical solution adopted by the present invention is as follows:

[0012] A coking coal blending method based on convolutional neural network comprises the following steps:

[0013] S1: Collecting original data, which is the historical coal blending data of the coking plant for many years, and using actual production data as the data source;

[0014] S2: Data cleaning and data set production, cleaning the raw data, removing irrelevant data and outliers, and generating a data set;

[0015] S3: feature extraction and evaluation, extracting features from the raw data;

[0016] S4: Algorithm construction, based on a deep neural network model, uses coal blending data as input for feature training and optimizes the blending model, which contains multiple convolutional layers and LSTM units, and is combined with a fully connected layer to generate output.

[0017] It is further defined that the data cleaning and data set preparation in S2 specifically include selecting a method of deleting, interpolating, filling the mean or median with respect to missing values ​​in the data, and the data set includes a training set and a validation set.

[0018] It is further defined that the objective function of the optimization ratio model in S4 is as follows:

[0019] max P=P C +P P -W T XR T +B

[0020] Where P is the total profit per ton of coke, C For coke yield, the P P is the product income, X is the mass of a single type of coal, and W T is the unit price of a single type of coal, the R T is the coal-to-coke ratio of the coal type, and B is the adjustment parameter of the proportioning model.

[0021] Further defined, the chemical product income P P Including tar, crude benzene, sulfur, coal gas and ammonium sulfate.

[0022] It is further defined that the adjustment parameter B of the proportioning model includes the business experience of coal blending engineers, historical coal blending data learning experience and model self-learning experience.

[0023] After further qualification and determination of the core objective function, CNN is selected as the first layer of the hybrid network for data operations. The goal is to extract the short-term patterns and dependencies between different input variables of coal information. The convolution layer consists of multiple convolution filters with a width of m and a height of n, denoted as t. The tth filter scans the input matrix X to produce:

[0024] g t =tanh(a t *Y+b t )

[0025] In the formula, g t represents the output function, the a t is the input vector, Y represents the weight, and b t Indicates deviation.

[0026] Further defined, also includes an output matrix, the output matrix is ​​represented by q c , the output is input as an input variable into the LSTM unit for calculation. The calculation process includes the following steps:

[0027] f t =σ(x t Y f +h t-1 W hf +b r )

[0028] Wherein, the LSTM unit includes a forget gate ft , input gate i t and output gate o t , the σ is the activation function, the f t The range is between (0,1), the w hf is the weight of the forget gate, the x t is the input of the current layer at time t, the h t-1 is the output of the previous layer, and combining it as input yields:

[0029] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0030]

[0031] Among them, i t The range is between (0,1), the W i is the weight of the input gate, the b i is the bias of the input gate, the W C is the weight of the candidate gate, the b C is the bias of the candidate gate, and the current unit is updated as follows:

[0032]

[0033] Among them, the C t The range is between (0,1), and the input of the output gate at time t is the output h of the previous timestamp. t-1 and the input x at the current time t , output o t Use the following formula to calculate:

[0034] o t =σ(W o ·[h t-1 ,x t ]+b o ).

[0035] Further defining, the output of the LSTM unit is obtained by calculating the output of the output gate and the output of the unit state, as shown below:

[0036] h t =o t *tanh(C t ).

[0037] Further defined, the final output result of the fully connected layer of the matching model is as follows:

[0038]

[0039] In order to obtain the relationship between different ratios over a long period of time, the intermediate segment data is hidden during calculation. The x represents the number of hidden units when output from the LSTM unit. represents the hidden state skipped under different batch ratios, where x is the hidden state of the skip connection from timestamp t-x+1 to t, expressed as This is the final prediction result of the hybrid CNN-LSTM model.

[0040] Beneficial effects of the present invention:

[0041] 1. The present invention provides a coking coal blending method based on convolutional neural network, which aims to help coking plants select suitable coal types for production operations in coal blending work to obtain maximum revenue per ton of coke; the present invention also optimizes some problems existing in the prior art, such as the large deviation of calculation results caused by relying solely on data-driven, the scientificity and reliability of coal blending results caused by relying solely on coal blending experience, and the complex chemical reaction process is not fully considered in coal blending based on the establishment of a coke production mechanism model, and the versatility of the coal blending in different furnace types is poor.

[0042] 2. The overall architecture of the algorithm model proposed in the present invention adopts a combination of CNN (convolutional neural network) and LSTM (long short-term memory network) to perform ratio prediction. CNN can extract effective features and patterns from multi-variable inputs, LSTM captures complex long-term dependencies, and automatically selects the most suitable related time series data, so as to achieve high-precision coking coal blending prediction.

[0043] 3. After clarifying the objective function and various constraints, the present invention establishes a convolutional neural network and a long short-term memory network model. After selecting various activation functions, learning rates, and loss functions, the work experience and professional business knowledge of coal blending experts are structured and trained in combination with model bias items, fully retaining the influence of the professional knowledge of business personnel and enhancing the reliability and accuracy of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The present invention can be further illustrated by means of non-limiting examples given in the accompanying drawings;

[0045] Figure 1 This is an overall architecture diagram of the algorithm model of an embodiment of a coking coal blending method based on a convolutional neural network of the present invention. DETAILED DESCRIPTION

[0046] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments. The technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present invention.

[0047] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0048] In addition, the descriptions of "first", "second", etc. in the present invention are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" and "second" may explicitly or implicitly include at least one of the features. In addition, the technical solutions between the various embodiments can be combined with each other, but it must be based on the ability of ordinary technicians in the field to implement. When the combination of technical solutions is contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0049] A coking coal blending method based on a convolutional neural network of the present invention comprises the following steps:

[0050] By collecting historical coal blending production data provided by a large coking plant of a listed company, according to different indicators such as production line, furnace type, coke type, coking time, furnace temperature, coal type, etc., the data is cleaned and made into a data set, which is divided into a training set and a validation set. Then, the original model is obtained by using the training set for preliminary training of the model, and the original model is verified and fine-tuned according to the validation set to obtain the target model.

[0051] After obtaining the proportioning model suitable for the coking plant, the coal information is collected as input (inventory information or purchaseable information), and the optimized coal proportioning information for the furnace can be output after analysis and calculation of the optimization model.

[0052] like Figure 1 As shown, the specific steps include the following:

[0053] S1: Collecting original data, which is the historical coal blending data of the coking plant for many years, and using actual production data as the data source;

[0054] S2: Data cleaning and data set production, cleaning the raw data, removing irrelevant data and outliers, and generating a data set;

[0055] S3: feature extraction and evaluation, extracting features from the raw data;

[0056] Specifically, this feature is a feature that helps model learning, and screens out features that have a greater impact on the performance of the matching model, and analyzes the contribution of the feature to the prediction results of the matching model.

[0057] S4: Algorithm construction, based on a deep neural network model, uses coal blending data as input for feature training and optimizes the blending model, which contains multiple convolutional layers and LSTM units, and is combined with a fully connected layer to generate output.

[0058] In the practical application of this embodiment, the data cleaning and data set preparation in S2 specifically include selecting a method of deleting, interpolating, filling the mean or median with respect to missing values ​​in the data, and the data set includes a training set and a validation set.

[0059] Specifically, when cleaning data, the consistency of data format, unit, and scope should be ensured to avoid analytical deviations caused by inconsistent data.

[0060] In the practical application of this embodiment, the objective function of the optimization ratio model in S4 is as follows:

[0061] max P=P C +P P -W T XR T +B

[0062] Where P is the total profit per ton of coke, C For coke yield, the P P is the product income, X is the mass of a single type of coal, and W T is the unit price of a single type of coal, the R T is the coal-to-coke ratio of the coal type, and B is the adjustment parameter of the proportioning model.

[0063] In the practical application of this embodiment, the chemical product yield P P Including tar, crude benzene, sulfur, coal gas and ammonium sulfate.

[0064] In the actual application of this embodiment, the adjustment parameter B of the proportioning model includes the business experience of the coal blending engineer, the historical coal blending data learning experience and the model self-learning experience.

[0065] In the practical application of this embodiment, after determining the core objective function, CNN is selected as the first layer of the hybrid network for data operation. The goal is to extract the short-term patterns and dependencies between different input variables of coal type information. The convolution layer is composed of multiple convolution filters with a width of m and a height of n, denoted as t. The tth filter scans the input matrix X to generate:

[0066] g t =tanh(a t *Y+b t )

[0067] In the formula, g t represents the output function, the a t is the input vector, Y represents the weight, and b t Indicates deviation.

[0068] In the practical application of this embodiment, an output matrix is ​​also included, and the output matrix is ​​represented by q c , the output is input as an input variable into the LSTM unit for calculation. The calculation process includes the following steps:

[0069] f t =σ(x t Y f +h t-1 W hf +b r )

[0070] Among them, the long short-term memory network can solve the problem of gradient disappearance. It can use hidden vector m and memory vector n to store information for a long time. The memory vector allows interaction with the output from the previous state and the next input state to choose to update or keep the input vector.

[0071] The LSTM unit includes a forget gate f t , input gate i t and output gate o t , the σ is the activation function, the f t The range is between (0,1), the w hf is the weight of the forget gate, the x t is the input of the current layer at time t, the h t-1 is the output of the previous layer, and combining it as input yields:

[0072] i t =σ(W i ·[h t-1 ,x t ]+b i )

[0073]

[0074] Among them, i t The range is between (0,1), the W i is the weight of the input gate, the b i is the bias of the input gate, the W C is the weight of the candidate gate, the b C is the bias of the candidate gate, and the current unit is updated as follows:

[0075]

[0076] Among them, the C t The range is between (0,1), and the input of the output gate at time t is the output h of the previous timestamp. t-1 and the input x at the current time t , output o t Use the following formula to calculate:

[0077] o t =σ(W o ·[h t-1 ,x t ]+b o ).

[0078] In the practical application of this embodiment, the output of the LSTM unit is obtained by calculating the output of the output gate and the output of the unit state, as shown below:

[0079] h t =o t *tanh(C t ).

[0080] like Figure 1 As shown in the overall structure diagram of the algorithm model, this model contains a series of convolutional layers and a series of LSTM blocks, which are combined together through a fully connected layer in the final stage to produce output. The input data first passes through the convolutional layer and the pooling layer, and the corresponding features are extracted. The obtained features are used as input to the LSTM block, which is designed to remember historical data and relatively long-term dependencies. In the actual application of this embodiment, the final output result of the fully connected layer of the matching model is as follows:

[0081]

[0082] In order to obtain the relationship between different ratios over a long period of time, the intermediate segment data is hidden during calculation. The x represents the number of hidden units when output from the LSTM unit. represents the hidden state skipped under different batch ratios, where x is the hidden state of the skip connection from timestamp t-x+1 to t, expressed as This is the final prediction result of the hybrid CNN-LSTM model.

[0083] The key technical points of this solution are:

[0084] 1. Use convolutional neural networks and long short-term memory networks to build an optimized coal blending model to assist coal coking. Modeling is done based on the historical coal blending data of the coking plant, and a model belonging to a specific plant area is established to reflect the coal blending characteristics of the plant area.

[0085] 2. According to the coal blending optimization model, the coking plant is guided to carry out coal blending production. Within the range of coal input indicators, combined with the prices of crude benzene, tar, sulfur, coal gas, ammonium sulfate and other chemical products, the raw coal ratio with the highest comprehensive profit is output.

[0086] 3. The optimization ratio model will predict the thermal strength of the produced coke (for example, M40, M25, M10 and other indicators), and can limit related indicators to make the coal ratio entering the furnace more feasible and ensure product quality.

[0087] In summary, the present invention establishes a convolutional neural network and a long short-term memory network model after clarifying the objective function and various constraints. After various activation functions are selected, the learning rate is selected, and the loss function is determined, the work experience and professional business knowledge of the coal blending expert are structured, and the model bias item is combined for training to fully retain the influence of the professional knowledge of the business personnel and enhance the reliability and accuracy of the model.

[0088] The present invention calculates coal blending with the goal of maximizing the comprehensive coke benefit, and has been put into practical application in a coking plant of a certain enterprise, achieving the effect of saving a large amount of production costs.

[0089] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A coking coal blending method based on convolutional neural network, characterized in that: The following steps are included: S1: Collecting original data, which is the historical coal blending data of the coking plant for many years, and using actual production data as the data source; S2: Data cleaning and data set production, cleaning the raw data, removing irrelevant data and outliers, and generating a data set; S3: feature extraction and evaluation, extracting features from the raw data; S4: Algorithm construction, based on a deep neural network model, uses coal blending data as input for feature training and optimizes the blending model, which contains multiple convolutional layers and LSTM units, and is combined with a fully connected layer to generate output.

2. A method for coking coal blending based on convolutional neural network according to claim 1, characterized in that: The data cleaning and data set preparation in S2 specifically include selecting a method of deleting, interpolating, filling the mean or median with respect to missing values ​​in the data, and the data set includes a training set and a validation set.

3. A method for coking coal blending based on convolutional neural network according to claim 1, characterized in that: The objective function of the optimization ratio model in S4 is as follows: max P=P C +P P -W T XR T +B Where P is the total profit per ton of coke, C For coke yield, the P P is the product income, X is the mass of a single type of coal, and W T is the unit price of a single type of coal, the R T is the coal-to-coke ratio of the coal type, and B is the adjustment parameter of the proportioning model.

4. A method for coking coal blending based on convolutional neural network according to claim 3, characterized in that: The product income P P Including tar, crude benzene, sulfur, coal gas and ammonium sulfate.

5. A method for coking coal blending based on convolutional neural network according to claim 3, characterized in that: The adjustment parameter B of the proportioning model includes the business experience of coal blending engineers, historical coal blending data learning experience and model self-learning experience.

6. A method for coking coal blending based on convolutional neural network according to claim 3, characterized in that: After determining the core objective function, CNN is selected as the first layer of the hybrid network for data operation. The goal is to extract the short-term patterns and dependencies between different input variables of coal information. The convolution layer consists of multiple convolution filters with a width of m and a height of n, denoted as t. The tth filter scans the input matrix X to produce: g t =tanh(a t *Y+b t ) In the formula, g t represents the output function, the a t is the input vector, Y represents the weight, and b t Indicates deviation.

7. A method for coking coal blending based on convolutional neural network according to claim 6, characterized in that: Also included is an output matrix, which is denoted as q c , the output is input as an input variable into the LSTM unit for calculation. The calculation process includes the following steps: f t =σ(x t Yx f +h t-1 W hf +b r ) Wherein, the LSTM unit includes a forget gate f t , input gate i t and output gate o t , the σ is the activation function, the f t The range is between (0,1), the w hf is the weight of the forget gate, the x t is the input of the current layer at time t, the h t-1 is the output of the previous layer, and combining it as input yields: i t =σ(W i ·[h t-1 ,x t ]+b i ) Among them, i t The range is between (0,1), the W i is the weight of the input gate, the b i is the bias of the input gate, the W C is the weight of the candidate gate, the b C is the bias of the candidate gate, and the current unit is updated as follows: Among them, the C t The range is between (0,1), and the input of the output gate at time t is the output h of the previous timestamp. t-1 and the input x at the current time t , output o t Use the following formula to calculate: the t =σ(W o ·[h t-1 ,x t ]+b o )。 8. A method for coking coal blending based on convolutional neural network according to claim 7, characterized in that: The output of the LSTM unit is obtained by calculating the output of the output gate and the output of the unit state, as shown below: h t =o t *tanh(C t )。 9. A method for coking coal blending based on convolutional neural network according to claim 8, characterized in that: The final output of the fully connected layer of the matching model is as follows: In order to obtain the relationship between different ratios over a long period of time, the intermediate segment data is hidden during calculation. The x represents the number of hidden units when output from the LSTM unit. represents the hidden state skipped under different batch ratios, where x is the hidden state of the skip connection from timestamp t-x+1 to t, expressed as This is the final prediction result of the hybrid CNN-LSTM model.

Images