Intelligent coal blending coking optimization method and system based on deep learning
By combining deep learning and multi-objective genetic algorithms, the problems of large prediction errors, neglect of environmental constraints, and insufficient dynamic adaptability in coking coal blending methods are solved, achieving high-precision, low-cost, and environmentally friendly optimization of coke production.
Patent Information
- Application Number
- CN202511065720.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-31
- Publication Date
- 2025-11-21
AI Technical Summary
Existing coking coal blending methods rely on manual experience or simple models, which are difficult to adapt to changes in coal type and market demand. This results in large prediction errors, neglect of environmental constraints, and insufficient dynamic adaptability. Furthermore, multi-objective optimization algorithms have slow convergence speeds, making it difficult to achieve a dynamic balance between quality, cost, and environmental protection.
A deep learning-based intelligent coal blending method is adopted, which combines deep neural networks and multi-objective genetic algorithms. Through data preprocessing, coke quality prediction and coal blending optimization, nonlinear fitting and dynamic adjustment are achieved to optimize the ratio of high-sulfur coal and low-volatile coal, and real-time optimization is carried out in combination with coking process parameters.
To improve the accuracy of coke quality prediction, reduce quality fluctuations, achieve optimal cost ratios, reduce the use of high-priced coal types, improve production efficiency, reduce sulfur dioxide emissions, and ensure that coke quality consistently meets standards.
Smart Images
Figure CN120996430A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of coking in the metallurgical industry, and in particular to an intelligent coal blending and coking optimization method and system based on deep learning. BACKGROUND
[0002] Coking is a key link in steel production, and coke quality directly affects blast furnace smelting efficiency and pig iron quality. Traditional coal blending methods mainly rely on manual experience or simple linear regression models to calculate the blending ratio according to coal quality indicators (such as volatile matter, ash content, sulfur content, and caking index G value).
[0003] Currently, domestic and foreign coking enterprises generally use the following coal blending methods. First, the experience-based coal blending method, which is adjusted by experienced technical personnel based on historical data, but is limited by individual experience and is difficult to adapt to changes in coal types and fluctuations in market demand. Second, the statistical regression method, which establishes a linear regression model based on historical data to predict coke quality, but is difficult to handle nonlinear relationships and has poor adaptability to changes in coal types. Third, the expert system, which is a coal blending recommendation system based on a rule base, but the rules are updated with a lag and cannot achieve dynamic optimization. Finally, traditional optimization algorithms such as linear programming and genetic algorithms, which can optimize a single objective (such as the lowest cost or the best quality), but are difficult to achieve dynamic balance of multiple objectives.
[0004] In recent years, with the development of artificial intelligence technology, machine learning has been applied to coke quality prediction, and some enterprises have begun to try to introduce machine learning models (such as neural networks) into coal blending optimization. For example, some literature proposes a method of training a DNN model to predict coke quality based on historical data, but it does not consider the cost of coal blending and environmental constraints, and the model has limited generalization ability. In addition, the application of multi-objective optimization algorithms (such as genetic algorithms) in existing technologies is still in its early stages, and there are problems such as slow convergence speed and easy to fall into local optimum.
[0005] Experience-based models rely on manual experience or simple addition and are difficult to capture the complex nonlinear relationship between coal quality indicators and coke quality, resulting in large prediction errors (e.g., M 40 The prediction bias is often more than 5%); single-objective optimization algorithms usually only focus on a single dimension of quality or cost, ignoring key environmental constraints such as sulfur content, resulting in potential environmental non-compliance or cost out of control in actual production; traditional genetic algorithms use fixed crossover and mutation rates, which have insufficient dynamic adaptability when coal types fluctuate, are prone to premature convergence, and cause the blending ratio scheme to fall into local optimum; at the same time, mechanism models and data-driven models are not closely combined, hybrid models are missing, and explicit relationships (such as the linear relationship between volatile matter and M 40 ) are not effectively utilized, and only DNN is relied on to handle the nonlinear part, resulting in poor overall model interpretability.
[0006] To this end, the present application aims to provide an intelligent coal blending and coking optimization method and system based on deep learning to solve the above problems. SUMMARY
[0007] The present application aims to solve the above problems by providing an intelligent coal blending and coking optimization method and system based on deep learning, which predicts the coke quality (CSR, CRI, etc.) under different coal blending schemes through a deep learning model, and then combines market coal prices and inventory conditions to achieve optimal cost ratio, and uses the nonlinear fitting capability of deep learning to optimize the blending ratio of high-sulfur coal and low-volatile coal, while dynamically adjusting according to coking process parameters (such as furnace temperature and coking time).
[0008] To achieve the above purpose, the technical solution of the present application is as follows:
[0009] The present application provides an intelligent coal blending and coking optimization method based on deep learning, which comprises the following steps:
[0010] S1, data preprocessing: normalizing single coal parameters and calculating the weighted characteristics of blended coal using an additive model;
[0011] The formula of the additive model is:
[0012] Data normalization is performed using Min-Max standardization, and the formula is:
[0013] where X mix is a certain characteristic of blended coal, x i is the single coal characteristic value, and p i is the blending ratio;
[0014] S2, coke quality prediction: input the weighted characteristics of blended coal into a deep neural network (DNN) or hybrid model to predict the mechanical strength (M 40 , M 10 ) and reactivity strength (CRI, CSR) of coke;
[0015] S3, coal blending optimization: calculate the constraint conditions of maximum quality and minimum cost using a multi-objective genetic algorithm, and generate an optimal coal blending scheme according to the constraint conditions and the mechanical strength (M 40 , M 10 ) and reactivity strength (CRI, CSR) of coke;
[0016] The multi-objective genetic algorithm (MOGA) includes the following target functions:
[0017] Maximum quality: f1 = max(M40, min(CRI)),
[0018] Cost minimization: f2 = min(∑p i ·c i ),
[0019] Environmental constraints: sulfur content s mix ≤s max ,
[0020] Fitness function: Fitness = ω1·f1+ω2·f2+λ·Penalty(S mix ),
[0021] Where ω1, ω2 are weights, and λ is a penalty coefficient.
[0022] The single coal parameters include indexes of single coal and matching weight.
[0023] The deep neural network (DNN) outputs coke mechanical strength (M 40 , M 10 ) and reactivity strength by inputting indexes of single coal.
[0024] The deep neural network hidden layer activation function selects ReLU, the output layer is a linear activation function, the loss function adopts mean square error (MSE), and the optimizer adopts Adam algorithm.
[0025]
[0026] The mixed model (mechanism + data driven): the relationship between coke strength and volatile matter and caking index is counted by linear regression, and the relationship between the thickness of the colloid layer and the ash content is counted by the deep neural network.
[0027] The linear regression model is:
[0028] M 40 = β0+ β1·v mix + β z ·G mix + ∈.
[0029] The application also provides an intelligent coal blending and coking optimization system based on deep learning, which comprises a data preprocessing module, a coke quality prediction module and a coal blending optimization module.
[0030] The data preprocessing module normalizes single coal parameters and calculates the weighted characteristics of blended coal by using an additive model.
[0031] The formula of the additive model is:
[0032] Data normalization is performed by using Min-Max standardization, and the formula is:
[0033] The coke quality prediction module inputs the blending coal weighted characteristics into a deep neural network (DNN) or a hybrid model to predict the coke mechanical strength (M 40 , M 10 ) and reactivity strength (CRI, CSR);
[0034] The coal blending optimization module generates an optimal blending scheme that meets the quality, cost and environmental protection constraints by using a multi-objective genetic algorithm.
[0035] Compared with the prior art, the present application has the following advantages:
[0036] The present application accurately predicts the coke quality (CSR, CRI, etc.) under different blending schemes by using a deep learning model, reduces the dependence on artificial experience and reduces quality fluctuations; the cost-optimal blending is realized by combining market coal prices and inventory conditions while meeting the coke quality requirements; the strong non-linear fitting capability of the deep learning model is used to adapt to the blending optimization of different coal qualities (such as high-sulfur coal and low-volatile coal); at the same time, the coking process parameters (such as furnace temperature and coking time) are dynamically adjusted to improve the coke yield; the blending scheme is dynamically adjusted based on online detection data (such as coke microstructure and thermal strength) to form a closed-loop optimization system;
[0037] The intelligent coal blending and coking optimization method and system of the present application can improve the prediction accuracy by at least five percentage points and simultaneously reduce the overall coke quality fluctuation amplitude by thirty percent, significantly enhancing the quality stability; at the same time, by deeply optimizing the blending scheme of raw coal, it is expected to effectively reduce the proportion of high-priced coal, thereby achieving the goal of reducing the comprehensive cost of coal blending by five to ten percent; when facing a complex and variable raw coal environment, the final coke quality can still be stable and meet the standards; the method / system can quickly adjust the blending ratio based on real-time data in the production process, ensuring the quality, improving the production efficiency, and effectively reducing the emission of sulfur dioxide (SO2) in the production process. BRIEF DESCRIPTION OF DRAWINGS
[0038] Figure 1 is a flowchart of the intelligent coal blending and coking method in the embodiments of the present application;
[0039] Figure 2 is a system architecture diagram of the intelligent coal blending and coking system in the embodiments of the present application;
[0040] Figure 3 is a DNN model structure diagram in the embodiments of the present application;
[0041] Figure 4 is a multi-objective genetic algorithm (MOGA) schematic diagram in the embodiments of the present application. DETAILED DESCRIPTION
[0042] In order to make the person skilled in the art better understand the present application, the technical scheme of the present application will be further described in detail below in combination with the embodiments of the present application and the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.
[0043] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below in combination with the embodiments.
[0044] Embodiments:
[0045] Coal blending optimization of a coking plant:
[0046] I. Overall technical process
[0047] The core of the present application is to realize intelligent coal blending through the cooperation of three modules of data preprocessing, coke quality prediction and coal blending optimization. The specific steps are as follows:
[0048] 1. Input single coal parameters: including volatile matter (V daf ), caking index (G value), ash content (A d ), sulfur content (S t,d ) and unit cost (c i ).
[0049] 2. Data preprocessing module: normalize the original data and calculate the weighted characteristics of blended coal.
[0050] Coke quality prediction module: predict the mechanical strength (M 40 , M 10 ) and reactivity (CRI, CSR) of coke through a hybrid model.
[0051] 3. Coal blending optimization module: adopt improved genetic algorithm to generate an optimal blending scheme meeting the quality, cost and environmental protection constraints.
[0052] II. Specific implementation of the data preprocessing module
[0053] Step 1: Weighted calculation of blended coal characteristics
[0054] For the blending scheme P={p1, p2,..., p n} (satisfying ∑P i =1), the blended coal indicators are calculated:
[0055]
[0056] Where x iXi is the value of a certain property (such as Vdaf, G value) of the ith coal. mix Xi is the value of a certain property (such as Vdaf, G value) of the ith coal.
[0057] Step 2: Data normalization processing
[0058] Using Min-Max standardization to eliminate dimensional differences:
[0059]
[0060] Implementation parameters:
[0061] Input range: Vdaf ∈ [15%, 35%], G value ∈ [40, 100], Ad ∈ [5%, 15%], St,d ∈ [0.3%, 1.5%].
[0062] Normalized output values are mapped to the [0, 1] interval.
[0063] Three, specific implementation of the coke quality prediction module
[0064] Model architecture selection
[0065] A hybrid model of mechanism and data-driven is used, composed of linear regression sub-model and DNN sub-model in parallel:
[0066] Linear regression part: capture the explicit relationship between volatile matter, caking index and coke strength
[0067] M 40 = β0+ β1· v mix + β z · G mix + ∈
[0068] Coefficients are fitted by least squares method (example values of β1, β2, β3: -25.3, 0.82, 0.15) to historical data.
[0069] DNN part: handle the nonlinear effect of ash and sulfur.
[0070] Final prediction value = linear output × weight 1 + DNN output × weight 2.
[0071] Table 1 Model training details
[0072]
[0073] Four, specific implementation of the coal blending optimization module
[0074] Modeling of multi-objective optimization problem
[0075] Objective function:
[0076] Multi-Objective Genetic Algorithm (MOGA), objective functions include:
[0077] (1) Mass maximization: f1 = max(M40, min(CRI))
[0078] (2) Cost minimization: f2 = min(∑p i ·c i )
[0079] (3) Environmental constraint: sulfur content s mix ≤s max
[0080] Key improvements:
[0081] 1. Dynamic crossover rate (Pc) and mutation rate (Pm):
[0082]
[0083] where gen is the current generation, gen max = 500.
[0084] 2. Elite preservation strategy: the top 10% individuals of each generation directly enter the next generation.
[0085] 3. Constraint penalty function:
[0086]
[0087] Five, examples and verification
[0088] Table 2 Implementation case data
[0089]
[0090] Table 3 Optimization results
[0091]
[0092] Table 4 Model verification indicators
[0093]
[0094] Six, system iteration and deployment
[0095] 1. Real-time feedback mechanism:
[0096] (1) The actual mass data of each batch of coke is returned to the database
[0097] (2) When the prediction error is greater than 1.5 for 5 consecutive batches, trigger model retraining
[0098] 2. Parameter update rules:
[0099] Linear regression coefficients are updated monthly.
[0100] The DNN model is fully trained every quarter.
[0101] 9. Innovations to be protected
[0102] (1) Hybrid prediction architecture
[0103] The pioneering dual-pathway prediction framework of "linear mechanism model + deep neural network" is used to process V in the linear part. daf The dominant relationship between / G value and DNN captures the nonlinear effect of ash / sulfur content, significantly improving M 40 CRI prediction accuracy;
[0104] (2) Adaptive Multi-Objective Optimization Algorithm
[0105] Improved MOGA algorithm: Introduce Hamming distance to dynamically adjust crossover / mutation rate, combined with elitist retention strategy to avoid premature convergence;
[0106] Design a fitness function with a penalty term:
[0107] Fitness=ω1·f1+ω2·f2+λ·Penalty(S mix )
[0108] Achieve synergistic optimization of the three objectives of quality, cost, and environmental protection;
[0109] (3) Closed-loop iteration mechanism
[0110] Based on real-time production data feedback, using R 2 The parameters of the model are automatically updated based on the values >0.95 and RMSE <1.0 to adapt to fluctuations in coal quality.
[0111] (4) System Integration Innovation
[0112] Build an intelligent system that integrates data preprocessing, quality prediction, multi-objective optimization, and online feedback to replace human experience-based decision-making.
[0113] The above specific embodiments are merely explanations of the present invention and are not intended to limit the present invention. After reading this specification, those skilled in the art can make modifications to these embodiments without contributing any inventive step, but as long as they are within the scope of the claims of the present invention, they are protected by patent law.
Claims
1. A deep learning-based intelligent coal blending and coking optimization method, characterized by: The method comprises the following steps: S1, data preprocessing: normalizing single coal parameters and calculating blended coal weighted characteristics using an additive model; The formula of the additive model is: Data normalization was performed using Min-Max standardization, whose formula is: where X mix is a certain property of the blended coal, x i is a property value of the single coal, p i is a blending ratio; S2, coke quality prediction: inputting the blended coal weighted characteristics into a deep neural network or a hybrid model to predict coke mechanical strength and reactivity strength; S3, coal blending optimization: calculating constraint conditions for maximizing quality and minimizing cost by a multi-objective genetic algorithm, and generating an optimal coal blending scheme according to the constraint conditions and the coke mechanical strength and reactivity strength; The multi-objective genetic algorithm (MOGA) includes the following objective functions: Maximizing quality: f1 = max (M40, min (CRI)), Cost minimization: f2 = min(∑p i ·c i ), Environmental constraints: sulfur content s mix ≤ s max , Fitness = ω1 · f1 + ω2 · f2 + λ · Penalty(S mix ), Where ω1, ω2 are weights, and λ is a penalty coefficient.
2. The intelligent coal blending and coking optimization method based on deep learning according to claim 1, characterized in that: The single coal parameters include indexes of single coals and blending weights.
3. The intelligent coal blending and coking optimization method based on deep learning according to claim 1, characterized in that: The deep neural network outputs coke mechanical strength and reactivity strength by inputting the indexes of single coals; The deep neural network uses ReLU as the activation function of the hidden layer, linear activation function for the output layer, mean square error (MSE) as the loss function, and Adam algorithm as the optimizer; 4. The intelligent coal blending and coking optimization method based on deep learning according to claim 1, characterized in that: The hybrid model: statistically analyzing the relationship between coke strength and volatile matter and caking index by linear regression, and statistically analyzing the relationship between the thickness of the colloid layer and ash content by a deep neural network; The linear regression model is: M 40 = β0+ β1· v mix + β z · G mix + ∈.
5. The intelligent coal blending and coking optimization system based on deep learning according to claim 1, characterized in that: The system comprises a data preprocessing module, a coke quality prediction module, and a coal blending optimization module; The data preprocessing module normalizes single coal parameters and calculates blended coal weighted characteristics using an additive model; The formula of the additive model is: Data normalization was performed using Min-Max standardization, whose formula is: The coke quality prediction module inputs the blended coal weighted characteristics into a deep neural network or a hybrid model to predict coke mechanical strength and reactivity strength; The coal blending optimization module generates an optimal blending scheme that meets the quality, cost, and environmental protection constraints by using a multi-objective genetic algorithm.