Reinforcement learning-based parameter collaborative optimization control method for green electrode roasting process

Abstract

This invention provides a collaborative optimization control method for parameters in the green electrode roasting process based on reinforcement learning, belonging to the field of green electrode roasting technology. The method includes: collecting multi-source state data of the green electrode roasting process and extracting features; constructing a reinforcement learning model based on deep deterministic policy gradient to evaluate the operational efficiency of the roasting process; constructing a dynamic association network for the multi-source data; establishing a multi-objective optimization model, solving it using a non-dominated sorting genetic algorithm to obtain the parameter optimization solution set, and selecting the optimal control parameters for the current operating condition using the analytic hierarchy process (AHP); and converting the optimal control parameters into execution instructions through an industrial controller. This invention, through the combination of multi-source data fusion and dynamic association network construction, reinforcement learning and multi-objective optimization, accurately adapts to the complex roasting process, achieving collaborative optimization of quality, energy consumption, and equipment lifespan, improving control accuracy and reliability, and ensuring stable and efficient production.

Description

Technical Field

[0001] This invention belongs to the field of green electrode roasting technology, specifically a collaborative optimization control method for green electrode roasting process parameters based on reinforcement learning. Background Technology

[0002] Calcination of green electrodes is a critical step in electrode production, directly determining the quality and performance of the finished electrodes, production energy consumption, and equipment lifespan, and having a decisive impact on the subsequent application of the electrode products. Currently, the green electrode calcination process faces multiple technical bottlenecks and industry pain points. First, the calcination process is a complex industrial process involving multiple coupled factors, including process control, environmental changes, and material properties. Traditional control methods are difficult to accurately adapt to the complex characteristics of this process, resulting in a lack of scientific basis for parameter adjustment.

[0003] Secondly, current roasting parameter control relies heavily on manual experience, which is highly subjective and difficult to adapt to dynamic disturbances such as fluctuations in raw material properties and changes in grid voltage. This makes it impossible to achieve multi-objective synergistic optimization of quality, energy consumption, and equipment lifespan. Furthermore, current roasting control in the industry often focuses solely on quality indicators, neglecting the balance between energy costs and equipment lifespan. This leads to excessive energy consumption and rapid equipment wear during production, increasing production costs and hindering the industry's green and sustainable development.

[0004] Meanwhile, the utilization efficiency of various production data during the roasting process is low, and there is a lack of effective data processing and value mining methods, which fails to fully leverage the guiding role of data in production control, making it difficult for parameter optimization to align with actual production conditions. Therefore, developing a method for controlling green electrode roasting parameters that can accurately adapt to complex roasting processes and achieve multi-objective synergistic optimization has become a pressing technical problem for the current electrode manufacturing industry. Summary of the Invention

[0005] This invention provides a collaborative optimization control method for parameters in the green electrode roasting process based on reinforcement learning, in order to overcome the deficiencies in the prior art.

[0006] This invention provides a method for collaborative optimization control of parameters in the green electrode calcination process based on reinforcement learning, comprising:

[0007] Real-time acquisition of multi-source state data during the green electrode roasting process, including roasting process data, furnace environment data, and green electrode characteristic data.

[0008] Empirical mode decomposition (EMD) was used to denoise the multi-source state data, and factor analysis was employed to extract features, yielding process features, environmental features, and electrode characteristic features.

[0009] A reinforcement learning model based on deep deterministic policy gradients is constructed, using process characteristics, environmental characteristics, and electrode characteristics as state inputs to evaluate the operational efficiency of the calcination process in real time, and to make preliminary parameter adjustments based on the evaluation results.

[0010] A Bayesian filtering algorithm is used to fuse multi-source state data, and a dynamic association network of multi-source data is constructed by combining mutual information and complex network clustering.

[0011] A multi-objective optimization model is established based on the dynamic correlation network to comprehensively optimize calcination quality, energy consumption cost, and equipment life. The non-dominated sorting genetic algorithm III is used to solve the multi-objective optimization model to obtain the parameter optimization solution set. The optimal control parameters for the current operating condition are selected by the analytic hierarchy process.

[0012] By utilizing fuzzy PID control technology, the optimal control parameters are converted into execution instructions through an industrial controller, and the parameters of the roasting process are adjusted by combining feedforward-feedback composite control.

[0013] According to the reinforcement learning-based collaborative optimization control method for green electrode roasting process parameters provided by this invention, the roasting process data includes roasting temperature curves, heating rate, holding time, and atmosphere flow rate. Furnace environment data includes oxygen content, carbon monoxide concentration, and furnace wall temperature distribution within the furnace. Green electrode characteristic data includes the density, porosity, resistivity, and geometric dimensions of the green electrode.

[0014] The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning provided by the present invention includes the following process for denoising multi-source state data using empirical mode decomposition:

[0015] Empirical mode decomposition is performed on each dimension of the multi-source state data to obtain several intrinsic mode functions and a residual component.

[0016] Calculate the correlation coefficient for each intrinsic mode function.

[0017] Set a correlation coefficient threshold and retain the intrinsic mode functions and residual components whose correlation coefficients reach the threshold.

[0018] Reconstruct the denoised data to complete the denoising process for multi-source state data.

[0019] The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning provided by the present invention includes a feature extraction process using factor analysis, comprising:

[0020] Construct a multi-source state data matrix and calculate the standardized value of each indicator in the data matrix.

[0021] Calculate the correlation coefficient matrix of the standardized data, and perform eigenvalue decomposition on the correlation coefficient matrix to obtain eigenvalues ​​and corresponding eigenvectors.

[0022] The number of common factors is determined based on the cumulative contribution rate of eigenvalues, and the factor loading matrix is ​​calculated.

[0023] The rotated factor loading matrix is ​​obtained by maximizing variance orthogonal rotation. Based on the absolute value of the loading of each index on the common factor, the common factors are defined as process characteristics, environmental characteristics, and electrode characteristics, respectively.

[0024] The process of constructing a reinforcement learning model based on deep deterministic policy gradients, according to the reinforcement learning-based collaborative optimization control method for the green electrode calcination process provided by this invention, includes:

[0025] Define the state space and action space. The state space consists of process characteristics, environmental characteristics, and electrode characteristics. The action space includes the adjustment ranges for calcination temperature, heating rate, and atmosphere flow rate.

[0026] Design a reward function and construct a Critic network and an Actor network. The Critic network is a fully connected neural network with two hidden layers, outputting the action value. The Actor network is a fully connected neural network with two hidden layers, outputting the action.

[0027] The network parameters are updated using a policy gradient, and a soft update strategy for the target network is adopted.

[0028] The reinforcement learning model is trained iteratively until the loss function converges, resulting in a reinforcement learning evaluation model.

[0029] The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning provided by the present invention includes the following process for constructing a dynamic association network of multi-source data by combining mutual information and complex network clustering:

[0030] The multi-source data, after being fused by Bayesian filtering, is divided into time series categories. There are 1 data window, each containing 1 data window. A window data matrix is ​​obtained by sampling 10 consecutive points.

[0031] Calculate the mutual information values ​​between the data in each dimension within each data window.

[0032] The probability density function is calculated using the kernel density estimation method to obtain the mutual information matrix.

[0033] Set a mutual information threshold. When the mutual information value between two dimensions reaches the mutual information threshold, it is defined that there is a strong correlation between the two dimensions, and a weighted correlation matrix is ​​constructed for each window.

[0034] The spectral clustering algorithm is used to perform cluster analysis on the correlation matrix of each window.

[0035] The clustering results are obtained by solving the eigenvalue problem of the Laplacian matrix, and each cluster is used as a node in a dynamic association network.

[0036] Calculate the similarity between clusters in different windows. When the similarity reaches a preset similarity threshold, establish a connection between the nodes of the two clusters, with the edge weight being the similarity value, thus forming a dynamic association network of multi-source data.

[0037] The process of establishing a multi-objective optimization model that integrates roasting quality, energy consumption cost, and equipment lifespan, according to the reinforcement learning-based collaborative optimization control method for green electrode roasting process parameters provided by the present invention, includes:

[0038] Based on the window-weighted association matrix weights in dynamic association networks, an association influence factor is defined.

[0039] The objective function for roasting quality is defined as follows:

[0040]

[0041] In the formula, This represents the objective function for roasting quality. This indicates the overall score for roasting quality. Indicates the baseline quality score. This represents the adjustment coefficient. Indicates the related influencing factors.

[0042] The energy consumption cost objective function is defined by the following formula:

[0043]

[0044] In the formula, This represents the objective function for energy consumption cost. Indicates the actual energy consumption per unit of product. This indicates the maximum permissible energy consumption per unit of product. Indicates the related influencing factors.

[0045] The objective function for determining the equipment's lifespan is expressed by the following formula:

[0046]

[0047] In the formula, Represents the objective function for equipment lifespan. Indicates the equipment's baseline lifespan. Indicates the number of critical and easily damaged components. Indicates the first The cumulative wear and tear of each component. Indicates the first The design life of each component Indicates the related influencing factors.

[0048] By integrating the objective functions of calcination quality, energy consumption cost, and equipment lifespan, and setting constraints, a multi-objective optimization model is obtained.

[0049] The method for collaborative optimization control of parameters in the live electrode roasting process based on reinforcement learning provided by this invention, which uses a non-dominated sorting genetic algorithm III to solve a multi-objective optimization model and obtain the parameter optimization solution set, includes the following steps:

[0050] Step 1: Initialize the population, where each individual represents a set of roasting parameter vectors. Set the population size to [value missing]. Initial individuals are randomly generated based on constraints.

[0051] Step two: Calculate the objective function value for each individual.

[0052] Step 3: Based on the non-dominance relationship, stratify and sort the population. Assign individuals not dominated by other individuals to the first frontier layer, the remaining individuals not dominated to the second frontier layer, and so on, until all frontier layers are obtained. Calculate the crowding distance for each frontier layer individual. If an individual is a boundary individual of a frontier layer, set the crowding distance component of the corresponding objective function to 1.

[0053] Step 4: Select parent individuals based on front-end layer sorting and crowding distance.

[0054] Step 5: Use simulated binary crossover to generate offspring individuals, and use polynomial mutation to generate offspring individuals.

[0055] Step six: Merge the parent and child individuals to obtain a size of For the mixed population, the non-dominated ranking and crowding distance calculations were re-performed, and the top [population] was selected. Individuals make up the next generation of the population.

[0056] Repeat steps one through six in an iterative process until the preset maximum number of iterations is reached, and finally optimize the parameter set of all non-dominated individuals in the population.

[0057] The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning provided by the present invention includes the following process for selecting the optimal control parameters for the current operating condition using the analytic hierarchy process:

[0058] A hierarchical structure is established, including an objective layer, a criterion layer, and a solution layer. The objective layer is used to select the optimal calcination control parameters, the criterion layer includes calcination quality, energy consumption cost, and equipment lifespan, and the solution layer includes candidate solutions from the parameter optimization solution set.

[0059] Construct the criterion layer judgment matrix, and calculate the maximum eigenvalue and corresponding eigenvector of the criterion layer judgment matrix.

[0060] The eigenvectors are normalized to obtain the criterion weight vector.

[0061] Perform a consistency check, calculate the consistency index, find the average random consistency index, and when the consistency ratio reaches the preset condition, the judgment matrix meets the consistency requirement; otherwise, reconstruct the judgment matrix.

[0062] Construct a decision matrix for each scheme layer, calculate the maximum eigenvalue and normalized eigenvector of each decision matrix, and perform a consistency check.

[0063] Calculate the overall weight of each scheme, select the scheme with the largest overall weight, and the corresponding parameter vector is the optimal control parameter for the current working condition.

[0064] The reinforcement learning-based collaborative optimization control method for the parameters of the green electrode roasting process provided by the present invention, which combines feedforward-feedback composite control to adjust the roasting process parameters, includes the following steps:

[0065] Key parameters in the optimal control parameters are extracted, including the target calcination temperature, target heating rate, and target atmosphere flow rate.

[0066] Establish a feedforward control model and calculate the feedforward control quantity.

[0067] A fuzzy PID feedback controller is constructed, with the inputs being the calcination quality deviation and the rate of change of the deviation.

[0068] Design a fuzzy control rule table and formulate control rules.

[0069] The PID parameter correction is obtained by clarifying the data using the center of gravity method. The PID controller parameters are then adjusted in real time, and the PID feedback control quantity is calculated.

[0070] The composite control quantity is calculated based on the feedforward control quantity and the feedback control quantity, and then converted into an execution instruction by an industrial controller to adjust the parameters of the roasting process.

[0071] The reinforcement learning-based collaborative optimization control method for green electrode roasting process parameters provided by this invention, by employing multi-source data fusion and dynamic correlation network construction technology, can deeply explore the dynamic correlation between multi-dimensional data such as process parameters, furnace environment state and green electrode characteristics during roasting, capture the coupling law of each factor as the operating conditions change, and provide comprehensive and reliable data support for subsequent parameter optimization.

[0072] By introducing a technical approach combining reinforcement learning and multi-objective optimization, a collaborative optimization system was constructed that considers calcination quality, energy consumption costs, and equipment lifespan, breaking through the limitations of traditional control methods that focus solely on quality indicators while neglecting energy consumption and equipment wear. Reinforcement learning technology can perceive the operational status of the calcination process in real time and dynamically optimize adjustment strategies, while the multi-objective optimization model achieves a balanced consideration of multi-dimensional performance indicators. This ensures the stability and consistency of the finished electrode quality while effectively reducing the production energy consumption per unit product, decreasing the wear rate of key components such as furnace refractory bricks and heating elements, extending equipment maintenance cycles, and improving the company's economic efficiency and green sustainable production level. By adopting a feedforward-feedback composite control architecture, the impact of external disturbances such as raw material characteristic fluctuations and grid voltage changes on the calcination process can be predicted and offset in advance, while simultaneously correcting process deviations in real time. This effectively solves the problems of regulation lag, overshoot, and steady-state error commonly found in traditional control, improving the accuracy and stability of parameter control and ensuring that the entire calcination process remains under stable and controllable optimal conditions.

[0073] By using the analytic hierarchy process (AHP) to scientifically select optimal parameters, the importance weight of various performance indicators under different working conditions can be comprehensively weighed. The optimal solution that best suits the current production conditions can be selected from multiple sets of optimization parameter solutions, thus avoiding the one-sidedness of parameter selection under a single optimization objective. Attached Figure Description

[0074] The invention will now be further described with reference to the accompanying drawings.

[0075] Figure 1 This is a flowchart illustrating the collaborative optimization control method for parameters in the green electrode roasting process based on reinforcement learning in this invention.

[0076] Figure 2 This is a schematic diagram of the feature extraction process using factor analysis in this invention;

[0077] Figure 3 This is a schematic diagram of the process for constructing a dynamic correlation network of multi-source data in this invention. Detailed Implementation

[0078] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0079] like Figures 1 to 3 As shown in the embodiment of the present invention, the method for collaborative optimization control of parameters in the green electrode calcination process based on reinforcement learning includes:

[0080] Real-time acquisition of multi-source state data during the green electrode roasting process, including roasting process data, furnace environment data, and green electrode characteristic data.

[0081] The roasting process data includes the roasting temperature profile, heating rate, holding time, and atmosphere flow rate. The furnace environment data includes the oxygen content, carbon monoxide concentration, and furnace wall temperature distribution within the furnace. The green electrode characteristic data includes the density, porosity, resistivity, and geometric dimensions of the green electrode.

[0082] Empirical mode decomposition (EMD) was used to denoise the multi-source state data, and factor analysis was employed to extract features, yielding process features, environmental features, and electrode characteristic features.

[0083] The process of denoising multi-source state data using empirical mode decomposition methods includes:

[0084] For each dimension of the multi-source state data Empirical mode decomposition is performed to obtain several eigenmode functions. and a residual component ,satisfy , This indicates the number of intrinsic mode functions. Let represent the k-th intrinsic mode function.

[0085] Calculate the correlation coefficient for each intrinsic mode function. ,in For covariance, Let Variance be the variance.

[0086] Set correlation coefficient threshold Retain correlation coefficient The intrinsic mode functions and residual components.

[0087] Reconstruct the denoised data This completes the denoising process for multi-source state data.

[0088] The process of feature extraction using factor analysis includes:

[0089] Construct a multi-source state data matrix, and calculate the standardized value of each index in the data matrix, expressed by the formula:

[0090]

[0091] In the formula, For the first The first sample The original values ​​of each indicator The range of values ​​is , The total number of samples, The range of values ​​is , For the total number of indicators, For the first The average of the indicators, , For the first The standard deviation of each indicator , For the first The first sample Standardized data for each indicator.

[0092] Calculate the correlation coefficient matrix of standardized data The formula is expressed as:

[0093]

[0094] in, This represents the correlation coefficient between the i-th indicator and the j-th indicator. , This represents the standardized data of the i-th indicator in the k-th sample. This represents the standardized data of the j-th indicator in the k-th sample. For the first The average of the standardized indicators.

[0095] For the correlation coefficient matrix Perform eigenvalue decomposition to obtain eigenvalues. and corresponding feature vectors ,in , This is the eigenvector index.

[0096] Based on the cumulative contribution rate of eigenvalues Determine the number of common factors The factor loading matrix is ​​calculated using the following formula:

[0097]

[0098] In the formula, For the front Composed of eigenvectors matrix, For the front Composed of eigenvalues Diagonal matrix.

[0099] The rotated factor loading matrix is ​​obtained by maximizing variance orthogonal rotation. Based on the absolute values ​​of the loadings of each indicator on the common factors, the common factors are defined as process characteristics. Environmental characteristics and electrode characteristics ,in , , , These are the elements of the rotated factor loading matrix.

[0100] A reinforcement learning model based on deep deterministic policy gradients is constructed, using process characteristics, environmental characteristics, and electrode characteristics as state inputs to evaluate the operational efficiency of the calcination process in real time, and to make preliminary parameter adjustments based on the evaluation results.

[0101] The process of constructing a reinforcement learning model based on deep deterministic policy gradients includes:

[0102] Define the state space It consists of process characteristics, environmental characteristics, and electrode characteristic characteristics, namely... ,in For process feature vectors, For environmental feature vectors, Let be the electrode characteristic feature vector, and let the state space dimension be . , The number of common factors.

[0103] Define action space Including the amount of roasting temperature adjustment Heating rate adjustment amount Atmosphere flow rate adjustment ,Right now ,in , , .

[0104] Design reward function The formula is expressed as:

[0105]

[0106] In the formula, The roasting quality score is calculated by weighting the density compliance rate, porosity compliance rate, and resistivity uniformity of the finished electrode, with a value range of [range missing]. . Indicates energy consumption per unit of product. , for Real-time power at any given moment For roasting time, This represents the total mass of the live electrode. This represents the equipment loss coefficient, calculated based on the cumulative loss of high-temperature components in the furnace body, and its value range is... . , , Denotes the weight coefficients of the reward function, satisfying ,and Adjustments can be made based on actual working conditions and requirements.

[0107] Construct a Critic network and an Actor network. The Critic network is a fully connected neural network with two hidden layers. Its input dimensions are the state space dimension plus the action space dimension, and its output dimension is 1, outputting the action value. The network loss function is:

[0108]

[0109] In the formula, Indicates the number of training samples in a batch. This indicates that the current Critic network is for the th Sample Action value estimation, This represents the state of the i-th sample. This represents the action of the i-th sample. Indicates the first The instant reward corresponding to each sample This represents the discount factor, used to weigh immediate rewards against future rewards. This indicates the target Critic network's action pair for the next state. Action value estimation, Indicates the target Actor network's next state The output action.

[0110] The Actor network is a fully connected neural network with two hidden layers. Its input dimension is the state space dimension, and its output dimension is the action space dimension, outputting the action. , For the Actor network parameters, the network parameters are updated using policy gradients, and the gradient formula is:

[0111]

[0112] In the formula, This represents the gradient of the objective function of the Actor network. This represents the gradient of the Critic network with respect to the action at the current output action of the Actor network. This represents the gradient of the Actor network output with respect to the network parameters.

[0113] Set up an experience replay buffer to store the data generated from each interaction. The sample is stored in the buffer.

[0114] Using a target network soft update strategy, the parameter update formulas for the target Critic network and the target Actor network are as follows:

[0115]

[0116]

[0117] In the formula, , These represent the parameters of the current Critic network and Actor network, respectively. , These represent the parameters of the target Critic network and the target Actor network, respectively. This represents the soft update coefficient, with a value range of... .

[0118] Iteratively train the model until the loss function converges (the mean of the loss function for 100 consecutive training batches is less than 100%). ), thus obtaining the reinforcement learning evaluation model.

[0119] A Bayesian filtering algorithm is used to fuse multi-source state data, and a dynamic association network of the multi-source data is constructed using a combination of mutual information and complex network clustering techniques. The process includes:

[0120] The multi-source data, after being fused by Bayesian filtering, is divided into time series categories. There are 1 data window, each containing 1 data window. A window data matrix is ​​obtained by sampling consecutive points. ,in , , , For multi-source data dimensions, .

[0121] Calculate the mutual information value between the data in each dimension within each data window. The formula for calculating mutual information is:

[0122]

[0123] In the formula, , These represent the first and second parts of the multi-source data, respectively. Wei, Di Dimensional data, , , express and The joint probability density function, , They represent , The marginal probability density function.

[0124] The probability density function is calculated using the kernel density estimation method, and the mutual information matrix is ​​obtained. , The window number is where For the first The first window Wei and Di Mutual information values ​​of dimensional data.

[0125] Set mutual information threshold (Take the 75th percentile of all mutual information values), when At that time, it was believed that the first Wei and Di Dimensional data in the first The windows are strongly correlated, and the edge weights are... ,otherwise Construct a weighted correlation matrix for each window. .

[0126] The spectral clustering algorithm is used to perform cluster analysis on the correlation matrix of each window. The objective function of spectral clustering is:

[0127]

[0128] In the formula, Represents the clustering indicator matrix. For the first Clustering label vectors of dimensional data, For multi-source data dimensions, This represents the weight of the edge connecting the p-th and q-th dimensions of data in the t-th window. Representation of the correlation matrix The diagonal elements of the degree matrix, , This represents the sum of the correlation strengths between the q-th dimension data and all other dimensions data in the t-th data window. , This represents the clustering label vector of the q-th dimension data.

[0129] The clustering results are obtained by solving the eigenvalue problem of the Laplacian matrix, and each cluster is used as a node in a dynamic association network.

[0130] The similarity between clusters from different windows is calculated using the following formula:

[0131]

[0132] In the formula, Indicates the first The first window A cluster, Indicates the first The first window A cluster, , Representing clusters , The number of data dimensions included. , They represent the first , The first window Wei and Di The weights of the edges connecting the dimensional data.

[0133] when ( When the similarity threshold is set to 0.6, a connection is established between two cluster nodes, with the edge weight being the similarity value, ultimately forming a dynamic association network of multi-source data.

[0134] A multi-objective optimization model is established based on the dynamic correlation network to comprehensively optimize calcination quality, energy consumption cost, and equipment life. The non-dominated sorting genetic algorithm III is used to solve the multi-objective optimization model to obtain the parameter optimization solution set. The optimal control parameters for the current operating condition are selected by the analytic hierarchy process.

[0135] The process of establishing a multi-objective optimization model includes:

[0136] Based on the window-weighted association matrix weights in dynamic association networks, we define association influence factors. The correlation between multiple sources at time t is represented by the following formula:

[0137]

[0138] In the formula, Let the weights be the edge weights connecting the p-th and q-th dimensions of the data in the t-th window. is the maximum weight value in the correlation matrix of the t-th window, and d is the dimension of the multi-source data;

[0139] The objective function for roasting quality is defined as follows:

[0140]

[0141] In the formula, This represents the objective function for roasting quality. This indicates the overall score for roasting quality. , This indicates the rate at which the finished electrode density meets the standard. , This indicates the pass rate of porosity in the finished electrode. , This indicates the uniformity of resistivity of the finished electrode. , For the resistivity of each sample, The average resistivity, This represents the weighting coefficient of the quality indicator, satisfying... ,and , This represents the baseline quality score, which is 80% of the best historical quality score. This represents the adjustment coefficient, and its value range is... This is used to adjust the sensitivity of the mass function.

[0142] The energy consumption cost objective function is defined by the following formula:

[0143]

[0144] In the formula, This represents the objective function for energy consumption cost. Indicates the actual energy consumption per unit of product. , for Total power of the calcination system at all times. Total roasting time The total mass of the live electrode. This indicates the maximum allowable energy consumption per unit of product, determined based on industry standards and the rated power of the equipment.

[0145] The objective function for determining the equipment's lifespan is expressed by the following formula:

[0146]

[0147] In the formula, Represents the objective function for equipment lifespan. This represents the equipment's baseline lifespan, taken as 80% of the design lifespan. This indicates the number of critical and easily damaged components (including furnace refractory bricks, heating elements, etc.). Indicates the first The cumulative wear and tear of each component is calculated based on operating time, operating temperature, and load. Indicates the first The design life of each component.

[0148] By combining the objective functions of calcination quality, energy consumption cost, and equipment lifespan, and setting constraints, a multi-objective optimization model is obtained, expressed by the formula:

[0149]

[0150] The constraints include:

[0151] ,in The roasting temperature, The lowest roasting temperature, This is the highest roasting temperature.

[0152] ,in For the heating rate, For the lowest heating rate, This represents the highest heating rate.

[0153] ,in For atmosphere flow, For minimum atmosphere flow rate, This represents the maximum atmosphere flow rate.

[0154] The process of solving a multi-objective optimization model and obtaining the parameter optimization solution set includes:

[0155] Step 1: Initialize the population, each individual Represents a set of calcination parameter vectors ,in The roasting temperature, For the heating rate, For atmospheric flow rate.

[0156] Population size set to Initial individuals are randomly generated based on constraints, using the following formula:

[0157]

[0158]

[0159]

[0160] In the formula, Indicates the value in Random numbers within an interval.

[0161] Step 2: Calculate the objective function value for each individual. .

[0162] Step 3: Stratify and rank the population based on non-dominance relationships, for individuals... and ,like Non-inferior on all objective functions And at least one objective function is strictly better than ,but Dominate Individuals in the population that are not dominated by other individuals are classified as the first frontal layer. Remove The remaining individuals that are not dominated are classified into the second frontier layer. This process continues to derive all frontier layers.

[0163] Calculate the crowding distance for each individual in the frontal layer. The formula is:

[0164]

[0165] In the formula, , They respectively represent the first in the same frontier layer The neighboring individuals of the first individual are in the second... The values ​​that can be taken on the objective function , They represent the first The maximum and minimum values ​​of the objective function in the current population.

[0166] If an individual is a boundary individual in the front layer (without left or right neighbors), then the crowding distance component of the corresponding objective function is set to 1.

[0167] Step 4: Select parent individuals based on the frontier layer ranking and crowding distance. The selection strategy is as follows: prioritize individuals with smaller frontier layer indices; within the same frontier layer, prioritize individuals with larger crowding distances. The number of individuals selected is... .

[0168] Step 5: Generate offspring individuals using simulated binary crossover. The crossover formula is as follows:

[0169]

[0170]

[0171] In the formula, , Representing the parent individuals , In the The values ​​that can be taken on each parameter , Representing offspring individuals , In the The values ​​that can be taken on each parameter This represents the parameters related to the cross-distribution index. , for Interval random numbers, This is the cross-distribution index.

[0172] The offspring are generated using polynomial mutation, and the mutation formula is as follows:

[0173]

[0174] In the formula, Indicates the parent individual in the first generation The values ​​that can be taken on each parameter This indicates that the individual after mutation is on the 1st... The values ​​that can be taken from each parameter. , , for Interval random numbers, This is the variation distribution index.

[0175] Crossover probability Probability of mutation , For parameter dimensions, here .

[0176] Step six: Merge the parent and child individuals to obtain a size of For the mixed population, the non-dominated ranking and crowding distance calculations were re-performed, and the top [population] was selected. Individuals make up the next generation of the population.

[0177] Repeat steps one through six in the iterative process until the preset maximum number of iterations is reached, and optimize the solution set of parameters for all non-dominant individuals in the final population. , This represents the number of non-dominated individuals.

[0178] The process of selecting the optimal control parameters for the current operating condition using the analytic hierarchy process includes:

[0179] Establish a hierarchical structure, target layer To "select the optimal calcination control parameters", the criterion layer , For roasting quality, For energy consumption costs, For equipment lifespan, solution layer , For the first parameter in the optimal solution set One candidate solution.

[0180] Constructing the criterion layer judgment matrix ,in Representation Criteria Relative to criteria The degree of importance is categorized by value. These represent equal importance, slightly important, obviously important, strongly important, and extremely important, respectively.

[0181] Calculate the judgment matrix Maximum eigenvalue and corresponding feature vectors .

[0182] For eigenvectors Normalization is performed to obtain the criterion weight vector. The normalization formula is , For feature vectors The Each element.

[0183] Perform a consistency check and calculate the consistency index. Find the average random consistency index When the consistency ratio If the condition is met, the judgment matrix satisfies the consistency requirement; otherwise, the judgment matrix is ​​reconstructed.

[0184] Constructing the scheme layer judgment matrix , ,in Representation scheme Compared to the plan In the guidelines The degree of superiority or inferiority is determined by the same rules as the criterion layer judgment matrix.

[0185] Calculate each judgment matrix separately Maximum eigenvalue and normalized eigenvectors And perform a consistency check.

[0186] Calculate the overall weight of each option. , .

[0187] Choose the option with the highest overall weight. Its corresponding parameter vector These are the optimal control parameters for the current operating conditions.

[0188] Using fuzzy PID control technology, the optimal control parameters are converted into execution commands through an industrial controller. Combined with feedforward-feedback composite control, the parameters of the calcination process are adjusted. The process includes:

[0189] Extract the key parameter from the optimal control parameters, namely the target calcination temperature. Target heating rate Target atmosphere flow rate .

[0190] A feedforward control model was established, and the main disturbance factors were identified based on the mechanism analysis of the calcination process. (Including fluctuations in raw material characteristics, grid voltage fluctuations, etc.), the formula for calculating the feedforward control quantity is:

[0191]

[0192] In the formula, This refers to feedforward control variables, including temperature feedforward, rate feedforward, and flow rate feedforward. Represents the feedforward gain matrix. , For temperature feedforward gain, For rate feedforward gain, For flow feedforward gain, This indicates the current disturbance value, which is collected in real time by the sensor. This represents the baseline disturbance value, which is the average disturbance value under normal operating conditions.

[0193] Construct a fuzzy PID feedback controller, with the input being the calcination quality deviation. and the rate of change of deviation ,in , Score the target quality. .

[0194] Will and It is divided into 5 fuzzy subsets: negative large (NL), negative small (NS), zero (ZO), positive small (PS), and positive large (PL).

[0195] Design a fuzzy control rule table and formulate control rules: If For NL and If it is NL, then For PL, For NL, For PS.

[0196] The PID parameter correction is obtained by clarifying the data using the center of gravity method. Adjust PID controller parameters in real time:

[0197]

[0198]

[0199]

[0200] In the formula, The initial parameters for the PID controller are determined using engineering tuning methods.

[0201] Calculate the PID feedback control quantity The formula is:

[0202]

[0203] In the formula, This refers to feedback control variables, including temperature feedback variables, rate feedback variables, and flow feedback variables.

[0204] Composite control quantity The industrial controller converts control quantities into execution commands, driving actuators such as temperature regulators, heating rate controllers, and atmosphere flow control valves.

[0205] Real-time acquisition of status data during the roasting process, and updating of current disturbance values. and quality deviation By dynamically adjusting the feedforward control quantity and the feedback control quantity, the parameters of the roasting process can be precisely adjusted.

[0206] In summary, this embodiment provides a collaborative optimization control method for parameters in the green electrode roasting process based on reinforcement learning. By employing multi-source data fusion and dynamic correlation network construction technology, it can deeply explore the dynamic correlation between multi-dimensional data such as process parameters, furnace environment status, and green electrode characteristics during the roasting process, capture the coupling law of each factor as the operating conditions change, and provide comprehensive and reliable data support for subsequent parameter optimization.

[0207] By introducing a technical approach combining reinforcement learning and multi-objective optimization, a collaborative optimization system was constructed that considers calcination quality, energy consumption costs, and equipment lifespan, breaking through the limitations of traditional control methods that focus solely on quality indicators while neglecting energy consumption and equipment wear. Reinforcement learning technology can perceive the operational status of the calcination process in real time and dynamically optimize adjustment strategies, while the multi-objective optimization model achieves a balanced consideration of multi-dimensional performance indicators. This ensures the stability and consistency of the finished electrode quality while effectively reducing the production energy consumption per unit product, decreasing the wear rate of key components such as furnace refractory bricks and heating elements, extending equipment maintenance cycles, and improving the company's economic efficiency and green sustainable production level. By adopting a feedforward-feedback composite control architecture, the impact of external disturbances such as raw material characteristic fluctuations and grid voltage changes on the calcination process can be predicted and offset in advance, while simultaneously correcting process deviations in real time. This effectively solves the problems of regulation lag, overshoot, and steady-state error commonly found in traditional control, improving the accuracy and stability of parameter control and ensuring that the entire calcination process remains under stable and controllable optimal conditions.

[0208] By using the analytic hierarchy process (AHP) to scientifically select optimal parameters, the importance weight of various performance indicators under different working conditions can be comprehensively weighed. The optimal solution that best suits the current production conditions can be selected from multiple sets of optimization parameter solutions, thus avoiding the one-sidedness of parameter selection under a single optimization objective.

[0209] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0210] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning, characterized in that, include: Real-time acquisition of multi-source state data during the green electrode roasting process, including roasting process data, furnace environment data, and green electrode characteristic data; The multi-source state data was denoised using the empirical mode decomposition method, and feature extraction was performed using factor analysis to obtain process features, environmental features, and electrode characteristic features. A reinforcement learning model based on deep deterministic policy gradient is constructed, using the process characteristics, environmental characteristics, and electrode characteristics as state inputs, to evaluate the operational efficiency of the calcination process in real time, and to make preliminary parameter adjustments based on the evaluation results. The Bayesian filtering algorithm is used to fuse the multi-source state data, and a dynamic association network of the multi-source data is constructed by combining mutual information and complex network clustering techniques. A multi-objective optimization model is established based on the dynamic correlation network, which integrates calcination quality, energy consumption cost and equipment life. The multi-objective optimization model is solved by non-dominated sorting genetic algorithm III to obtain the parameter optimization solution set. The optimal control parameters for the current working condition are selected by the analytic hierarchy process. Using fuzzy PID control technology, the optimal control parameters are converted into execution instructions by an industrial controller, and the parameters of the roasting process are adjusted by combining feedforward-feedback composite control.

2. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The roasting process data includes roasting temperature curve, heating rate, holding time, and atmosphere flow rate; the furnace environment data includes oxygen content, carbon monoxide concentration, and furnace wall temperature distribution; and the green electrode characteristic data includes the density, porosity, resistivity, and geometric dimensions of the green electrode.

3. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of denoising the multi-source state data using the empirical mode decomposition method includes: Empirical mode decomposition is performed on each dimension of the multi-source state data to obtain several intrinsic mode functions and one residual component; Calculate the correlation coefficient for each intrinsic mode function; Set a correlation coefficient threshold and retain the intrinsic mode functions and residual components whose correlation coefficients reach the threshold. Reconstruct the denoised data to complete the denoising process for multi-source state data.

4. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of feature extraction using factor analysis includes: Construct a multi-source state data matrix, and calculate the standardized value of each indicator in the data matrix; Calculate the correlation coefficient matrix of standardized data, and perform eigenvalue decomposition on the correlation coefficient matrix to obtain eigenvalues ​​and corresponding eigenvectors; The number of common factors is determined based on the cumulative contribution rate of eigenvalues, and the factor loading matrix is ​​calculated. The rotated factor loading matrix is ​​obtained by maximizing variance orthogonal rotation. Based on the absolute value of the loading of each index on the common factor, the common factors are defined as process characteristics, environmental characteristics, and electrode characteristics, respectively.

5. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of constructing a reinforcement learning model based on deep deterministic policy gradients includes: Define a state space and an action space. The state space consists of process characteristics, environmental characteristics, and electrode characteristic characteristics. The action space includes the calcination temperature adjustment, the heating rate adjustment, and the atmosphere flow rate adjustment. Design a reward function and construct a Critic network and an Actor network. The Critic network is a fully connected neural network with two hidden layers, which outputs the action value; the Actor network is a fully connected neural network with two hidden layers, which outputs the action. The network parameters are updated using a policy gradient, and a soft update strategy for the target network is adopted. The reinforcement learning model is trained iteratively until the loss function converges, thus obtaining the reinforcement learning evaluation model.

6. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of constructing dynamic association networks of multi-source data by combining mutual information and complex network clustering includes: The multi-source data, after being fused by Bayesian filtering, is divided into time series categories. There are 1 data window, each containing 1 data window. A window data matrix is ​​obtained by sampling several consecutive points; Calculate the mutual information values ​​between the data in each dimension within each data window; The probability density function is calculated using the kernel density estimation method to obtain the mutual information matrix; Set a mutual information threshold. When the mutual information value between two dimensions of data reaches the mutual information threshold, it is defined that there is a strong correlation between the two dimensions of data, and a weighted correlation matrix is ​​constructed for each window. The correlation matrix of each window is clustered using a spectral clustering algorithm. The clustering results are obtained by solving the eigenvalue problem of the Laplacian matrix, and each cluster is used as a node in a dynamic association network. Calculate the similarity between clusters in different windows. When the similarity reaches a preset similarity threshold, establish a connection between the nodes of the two clusters, with the edge weight being the similarity value, thus forming a dynamic association network of multi-source data.

7. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of establishing a multi-objective optimization model that integrates calcination quality, energy consumption cost, and equipment life includes: Based on the window-weighted association matrix weights in dynamic association networks, an association influence factor is defined. The objective function for roasting quality is defined as follows: In the formula, This represents the objective function for roasting quality. This indicates the overall score for roasting quality. Indicates the baseline quality score. This represents the adjustment coefficient. Indicates the correlation influencing factors; The energy consumption cost objective function is defined by the following formula: In the formula, This represents the objective function for energy consumption cost. Indicates the actual energy consumption per unit of product. This indicates the maximum permissible energy consumption per unit of product. Indicates the correlation influencing factors; The objective function for determining the equipment's lifespan is expressed by the following formula: In the formula, Represents the objective function for equipment lifespan. Indicates the equipment's baseline lifespan. Indicates the number of critical and easily damaged components. Indicates the first The cumulative wear and tear of each component. Indicates the first The design life of each component Indicates the correlation influencing factors; By combining the aforementioned calcination quality objective function, energy consumption cost objective function, and equipment life objective function, and setting constraints, a multi-objective optimization model is obtained.

8. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of using the non-dominated sorting genetic algorithm III to solve the multi-objective optimization model and obtain the parameter optimization solution set includes: Step 1: Initialize the population, with each individual representing a set of roasting parameter vectors; set the population size to [value missing]. Initial individuals are randomly generated based on constraints; Step two: Calculate the objective function value for each individual; Step 3: Sort the population into layers based on non-dominance relationships. Assign individuals that are not dominated by other individuals to the first front layer, and assign the remaining individuals that are not dominated to the second front layer, and so on to obtain all front layers. Calculate the crowding distance of each front layer individual. If an individual is a boundary individual of the front layer, set the crowding distance component of the corresponding objective function to 1. Step 4: Select parent individuals based on frontier layer sorting and crowding distance; Step 5: Use simulated binary crossover to generate offspring individuals, and use polynomial mutation to generate offspring individuals; Step six: Merge the parent and child individuals to obtain a size of For the mixed population, the non-dominated ranking and crowding distance calculations were re-performed, and the top [population] was selected. Individuals form the next generation population; Repeat steps one through six in an iterative process until the preset maximum number of iterations is reached, and finally optimize the parameter set of all non-dominated individuals in the population.

9. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of selecting the optimal control parameters for the current operating condition using the analytic hierarchy process includes: A hierarchical structure is established, including an objective layer, a criterion layer, and a scheme layer. The objective layer is used to select the optimal calcination control parameters, the criterion layer includes calcination quality, energy consumption cost, and equipment lifespan, and the scheme layer includes candidate schemes in the parameter optimization solution set. Construct the criterion layer judgment matrix, and calculate the maximum eigenvalue and corresponding eigenvector of the criterion layer judgment matrix; The feature vectors are normalized to obtain the criterion weight vectors; Perform a consistency check, calculate the consistency index, find the average random consistency index, and when the consistency ratio reaches the preset condition, the judgment matrix satisfies the consistency requirement; otherwise, reconstruct the judgment matrix. Construct the decision matrix for each scheme layer, calculate the maximum eigenvalue and normalized eigenvector of each decision matrix, and perform a consistency check. Calculate the overall weight of each scheme, select the scheme with the largest overall weight, and the corresponding parameter vector is the optimal control parameter for the current working condition.

10. The method for collaborative optimization control of parameters in the green electrode roasting process based on reinforcement learning according to claim 1, characterized in that, The process of adjusting the calcination process parameters using a combined feedforward-feedback control includes: Extract the key parameters from the optimal control parameters, including the target calcination temperature, target heating rate, and target atmosphere flow rate; Establish a feedforward control model and calculate the feedforward control quantity; Construct a fuzzy PID feedback controller, with the inputs being the calcination quality deviation and the rate of change of the deviation; Design a fuzzy control rule table and formulate control rules; The PID parameter correction is obtained by clarifying the process using the center of gravity method, and the PID controller parameters are adjusted in real time to calculate the PID feedback control quantity. The composite control quantity is calculated based on the feedforward control quantity and the feedback control quantity, and the composite control quantity is converted into an execution instruction by the industrial controller to adjust the parameters of the roasting process.

Images