Method and system for dynamic allocation of heating power based on reinforcement learning

Abstract

The application belongs to the technical field of heating cooperative control of smelting furnaces, and particularly relates to a heating power dynamic allocation method and system based on enhanced learning, which comprises the following steps: collecting and decoupling furnace temperature operation data of multiple smelting furnaces in real time, extracting multi-domain fluctuation characteristics, dividing process stages through a smelting process deduction model to obtain heating reference parameters; constructing a coupling physical constraint equation based on the heating reference parameters, establishing a fuzzy programming model through interval two-type fuzzy sets, dividing a phased power sub-feasible region based on the constraint boundary of the inverse solution power through section decomposition, building upper and lower layer deep Q networks and reasoning to generate a multi-zone initial power allocation scheme, constructing a multi-furnace non-cooperative evolutionary game model to solve an equilibrium strategy set, screening an optimal scheduling scheme based on cross-efficiency evaluation and robustness analysis, calculating a safety confidence through D-S evidence theory, and automatically triggering power degradation and alarming in an abnormal situation. The application realizes fine dynamic scheduling of multi-furnace multi-zone heating power and strengthens operation safety.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for heating in smelting furnaces, specifically to a method and system for dynamic allocation of heating power based on reinforcement learning. Background Technology

[0002] Regenerative melting furnaces are thermal equipment used in the smelting of non-ferrous metals such as recycled aluminum. With the expansion of the industry's multi-furnace parallel production scale and the requirements for energy conservation and consumption reduction, the existing heating power control can no longer meet the overall smelting control needs. Existing studies have not decoupled the coupling characteristics of temperature rise trend, heat storage reversal period and random disturbance in furnace temperature signals, lack in-depth mining of fluctuation characteristics, and cannot provide a basis for phased power control. The power constraints are mostly based on deterministic thermal mechanism derivation, without systematically considering the coupled effects of multiple uncertainties such as ambient temperature fluctuations, batch differences in gas calorific value and unstable heat absorption rate of scrap aluminum. The constraint boundaries are generally too loose or too tight. Moreover, existing studies mostly focus on single furnace optimization, which cannot meet the heating power allocation problem of multiple furnaces. Therefore, a dynamic heating power allocation method and system based on reinforcement learning is needed to solve the above problems. Summary of the Invention

[0003] To address the technical problems mentioned in the background, this invention provides a method and system for dynamic allocation of heating power based on reinforcement learning.

[0004] The objective of this invention can be achieved through the following technical solutions: The first aspect of this invention provides a method for dynamic allocation of heating power based on reinforcement learning, the specific steps of which are as follows: Step 1: Collect furnace temperature operation data of each regenerator burner zone of multiple smelting furnaces in real time, decouple them into temperature rise trend, commutation period and random disturbance, decompose the commutation period term to extract multi-domain fluctuation characteristics, divide the process stages through the smelting process deduction model, and then statistically analyze the feature distribution center and fit the power response elastic coefficient to obtain the heating reference parameters of each stage. Step 2: Construct coupled physical constraint equations based on heating reference parameters at each process stage; determine uncertainty parameters and establish a fuzzy programming model through interval type II fuzzy sets; decompose the power boundary constraint based on cut set decomposition; and divide the power sub-feasible region after traversing feasible solutions. Step 3: Divide the action set based on the power sub-feasible region of each process stage, build an upper and lower deep Q network, obtain a hierarchical decision model through iterative training, infer and output the stage power strategy and partition power value, and generate a multi-zone initial power allocation scheme. Step 4: Construct a non-cooperative evolutionary game model with each smelting furnace as the main players, solve for the set of equilibrium power allocation strategies, select the optimal power scheduling scheme through cross-efficiency evaluation and robustness analysis, calculate the safety confidence based on DS evidence theory, and automatically trigger power degradation and alarms in case of anomalies.

[0005] In this application, based on step one, the furnace temperature data of each regenerable burner zone of multiple smelting furnaces is collected in real time, decoupled into temperature rise trend, commutation period, and random disturbance, the commutation period term is decomposed to extract multi-domain fluctuation characteristics, the process stages are divided through the smelting process derivation model, the feature distribution center is statistically analyzed and the power response elastic coefficient is fitted, and the heating reference parameters for each stage are combined to obtain the specific steps: Real-time collection of furnace temperature operation data for each regenerator burner zone of several smelting furnaces; decoupling of furnace temperature operation data to obtain temperature rise trend term, regenerator reversal cycle term and random disturbance term; The thermal storage commutation periodic term is decomposed into several narrowband intrinsic mode functions by multi-scale decomposition. The calculation logic is as follows: , Where TS is the thermal storage commutation period term, Here, p represents the narrowband intrinsic mode function, and p is the center frequency of each component. It is a first-order partial differential operator. Here, is the Dirac delta function, jc is the imaginary unit, t is the time point, * is the convolution operator, K is the total number of modal decompositions, and k is the modal function number. For the complex exponential demodulation term, fractional Fourier transforms are performed on several narrowband intrinsic mode functions to obtain the transform results. The fractional domain fuzzy entropy is calculated based on the transform results. Then, the instantaneous frequencies are calculated on several narrowband intrinsic mode functions. The marginal spectrum is obtained by integrating the instantaneous frequencies. The marginal spectrum energy proportion ER is calculated based on the marginal spectrum. The calculation logic is as follows: Where hc is the marginal spectrum, This is the lower limit angular frequency of the main frequency range of the component. The upper angular frequency of the corresponding component's main frequency range. This is the maximum number of points. For definite integral calculation; taking the burner commutation trigger signal as the initial time point, temperature response segments of the commutation interval are extracted periodically from the corresponding narrowband intrinsic mode function. The commutation interval is... , To preset the steady-state duration before commutation, the response segments of several consecutive commutation cycles under the same operating condition are aligned with equal length and arithmetically averaged to obtain the step response sequence of the inherent mode function. Based on the same step response sequence, the relative temperature rise peak within a single commutation cycle is taken as the impact peak. The impact peak is compared with the standard steady-state temperature rise amplitude in the database to obtain the commutation impact attenuation coefficient. The commutation impact attenuation coefficient, fractional-order domain fuzzy entropy, and marginal spectrum energy proportion are integrated into a set of fluctuation characteristics. A smelting process simulation model is established. The fluctuation feature set and furnace temperature operation data are input into the smelting process simulation model. The smelting process simulation model outputs the smelting process hidden state transition matrix and feature observation probability matrix. Based on the smelting process hidden state transition matrix and feature observation probability matrix, Viterbi decoding is performed on the furnace temperature operation data to output time series segmentation results, including charging, primary refining, secondary refining and settling. Gaussian kernel density estimation is performed on the fluctuation feature set of each segment to obtain the feature distribution center. The furnace temperature operation data is matched with the time interval of the corresponding process stage. The heating power time series data within this time period is extracted. Then, the furnace temperature change rate of each sampling point is calculated for the furnace temperature time series data within the same segment. The natural logarithmic transformation is performed on the heating power time series data and the furnace temperature change rate to obtain the first linear value and the second linear value. A univariate linear regression equation is constructed using the first linear value and the second linear value. The calculation logic is as follows: ,in The regression intercept term reflects the basic heat loss at this stage of the process. The regression slope represents the power response elasticity coefficient for this stage of the process. As a random error term, the power response elastic coefficient is obtained by solving. The characteristic distribution center and the power response elastic coefficient are integrated to obtain the heating reference parameters for each process stage.

[0006] In this application, based on step two, coupled physical constraint equations are constructed based on heating reference parameters at each process stage. Uncertainty parameters are determined through interval type II fuzzy sets, and a fuzzy programming model is established. Based on the inverse solution of the constraint power boundary using cut set decomposition, the feasible solutions are traversed, and the staged power sub-feasible regions are divided. The specific steps are as follows: Based on the heating reference parameters of each process stage, a coupled physical constraint equation for the temperature field and flow field is constructed, and its calculation logic is as follows: ,in The heat exchange efficiency of the regenerator under the current operating conditions. Based on the baseline thermal storage and heat exchange efficiency, The flow rate influence coefficient represents the heat exchange efficiency. The coefficient for flue gas generation per unit power. The actual heating output power of the regenerative burner zone. The calorific value of the reference gas, This represents the actual calorific value of the gas. As the baseline flue gas flow rate, The coefficient representing the temperature effect of heat exchange efficiency is given, where T is the actual operating temperature of the furnace. Using the reference furnace temperature, the coupled physical constraint equations are analyzed to obtain the physical constraint boundaries; Extract running samples with uncertain parameters from the database. Uncertain parameters include ambient temperature, calorific value of fuel gas and heat absorption rate of scrap aluminum. Analyze the running samples and obtain the upper membership function and lower membership function. Generate a set of interval type II fuzzy parameters based on the upper membership function and lower membership function. The interval type II fuzzy parameter set and physical constraint boundary are transformed to obtain fuzzy opportunity constraints. All fuzzy opportunity constraints are integrated to generate an interval type II fuzzy programming model. A clear analysis of the interval type II fuzzy programming model is performed. Parameter value intervals at different confidence levels are obtained through cut-set decomposition. These intervals are then imported into coupled physical constraint equations. Inverse calculations are performed on the furnace temperature compliance constraint, furnace pressure safety constraint, and heat storage and heat exchange efficiency constraint to obtain the upper and lower limits of each parameter, including the upper and lower limits of furnace temperature power, furnace pressure power, and heat exchange efficiency power. The intersection of the upper and lower limits of each parameter at the same confidence level is taken to obtain the power feasible interval. A power constraint space is constructed based on the power feasible intervals of each heat storage burner zone. Each dimension of this power constraint space corresponds to a heating power decision variable for a zone, and each set of constraint faces corresponds to a process or equipment constraint condition. Any coordinate point within the space corresponds to a multi-zone heating power allocation scheme that satisfies all process requirements, safety restrictions, and equipment rating conditions at the corresponding confidence level. Points outside the space violate at least one constraint and are considered infeasible power allocation schemes. The feasible solution traversal of the power constraint space is performed to obtain the initial feasible solution set. The initial feasible solution set and the heating reference parameters of each process stage are mapped and divided to obtain the power sub-feasible domains corresponding to each process stage. Each sub-feasible domain directly serves as the action space boundary of the reinforcement learning model under the corresponding process stage, so that the power decision always falls within the range of process compliance and equipment safety.

[0007] In this application, based on step three, which involves dividing the action set according to the power sub-feasible region of each process stage, an upper and lower depth Q-network is constructed. A hierarchical decision model is obtained through iterative training, and the stage power strategy and partition power values ​​are inferred and output to generate a multi-zone initial power allocation scheme. The specific steps are as follows: The power sub-feasible domains of each process stage are divided into stage granularity and partition granularity to obtain the action set of the upper process stage and the action set of the lower partition power level. The overall fluctuation characteristics of the smelting furnace, furnace temperature operation data, and heating reference parameters are concatenated into a global operating condition feature vector. The global operating condition feature vector is then input into the upper-level deep Q-network. The global operating condition feature vector and the upper-level process stage action set are analyzed, and the total power command and stage command are output. The total power command, stage command, and real-time operating condition feature vectors of each regenerator burner zone are input into the lower-level deep Q network. The lower-level deep Q network matches the command with the vector and the lower-level zone power level action set to obtain the power adjustment value of each regenerator burner zone. The upper-layer deep Q-network and the lower-layer deep Q-network are iteratively trained to obtain a hierarchical decision model. The hierarchical decision model is used to reason about the current smelting conditions and output the optimal stage power strategy and partition power value in sequence to obtain the initial power allocation scheme for multiple zones.

[0008] In this application, based on step four, a non-cooperative evolutionary game model is constructed with each smelting furnace as the main player. The equilibrium power allocation strategy set is obtained by solving the model. The optimal power scheduling scheme is selected through cross-efficiency evaluation and robustness analysis. The safety confidence level is calculated based on DS evidence theory. Power degradation and alarms are automatically triggered in case of anomalies. The specific steps are as follows: Each smelting furnace is set as the main player in the game. The initial power allocation scheme of the multi-zone is used as the game strategy set. Energy consumption cost, production efficiency and heat storage ball loss are used as the payoff functions to construct a non-cooperative evolutionary game model for multiple furnaces. The evolutionary stable strategy of the multi-furnace non-cooperative evolutionary game model is solved by replicating the dynamic equation. Based on the evolutionary stable strategy, the Nash equilibrium state of multi-furnace power scheduling is analyzed, and the power allocation strategy set under the game equilibrium is obtained. A cross-efficiency evaluation model is constructed, with the gas power input of each zone as the input index and the temperature compliance rate, heating rate and heat storage and heat exchange efficiency as the output index. The cross-efficiency is calculated for each scheme in the power allocation strategy set to obtain the cross-efficiency value. The power allocation strategy set is sorted according to the cross-efficiency value, and the top-ranked schemes are selected as candidate schemes. Robustness index analysis is performed on the candidate scheme set to obtain the robustness index. The scheme with the highest score is selected and marked as the power dynamic scheduling optimization scheme based on the cross-efficiency value and the robustness index. Based on the power dynamic scheduling optimization scheme, the power target value, adjustment step size and adjustment timing of each heating zone are analyzed to obtain the power adjustment parameter matrix; The power regulation parameter matrix is ​​sent to the actuators of each regenerable burner, and the power is adjusted in a closed loop by a fuzzy PID controller. Real-time feedback data of zone temperature and furnace pressure are collected and the adjustment amount is dynamically corrected. The evidence bodies for temperature anomalies, pressure anomalies, and equipment failures are constructed using the DS evidence theory. The safety confidence level is then calculated by integrating the safety confidence level of the current heating state. When the safety confidence level is lower than the preset confidence threshold, the heating power of the corresponding zone is automatically reduced to the safe range, and an abnormal alarm record is generated.

[0009] Please refer to Figure 2 As shown, the second aspect of the present invention provides a dynamic heating power allocation system based on reinforcement learning, including a feature extraction module, a constraint modeling module, a decision modeling module, and a game optimization module; The feature extraction module collects furnace temperature operation data of each regenerator burner zone of multiple smelting furnaces in real time, decouples them into temperature rise trend, commutation period and random disturbance, decomposes the commutation period term to extract multi-domain fluctuation features, divides the process stages through the smelting process deduction model, and then statistically analyzes the feature distribution center and fits the power response elastic coefficient to obtain the heating reference parameters of each stage. The constraint modeling module constructs coupled physical constraint equations based on heating reference parameters at each process stage, determines uncertainty parameters and establishes a fuzzy programming model through interval type II fuzzy sets, and decomposes the constraint power boundary based on cut set decomposition, and divides the power sub-feasible region after traversing feasible solutions. The decision modeling module divides the action set based on the power sub-feasible domain of each process stage, builds upper and lower deep Q networks, obtains a hierarchical decision model through iterative training, infers and outputs the stage power strategy and partition power value, and generates a multi-zone initial power allocation scheme. The game optimization module constructs a non-cooperative evolutionary game model with each smelting furnace as the main players, solves the equilibrium power allocation strategy set, selects the optimal power scheduling scheme through cross-efficiency evaluation and robustness analysis, calculates the safety confidence based on DS evidence theory, and triggers power degradation and alarms in case of anomalies.

[0010] Compared with the prior art, the beneficial effects of the present invention are: By decoupling multiple components of furnace temperature operation data, we can deeply extract multi-domain fluctuation features such as reversing impact attenuation coefficient, fractional-order domain fuzzy entropy, and marginal spectrum energy ratio. Based on the smelting process deduction model and Viterbi decoding, we can automatically divide the process stages and obtain the heating reference parameters composed of the feature distribution center and power response elastic coefficient of each stage, thus solving the problems of fuzzy empirical division boundaries and reference parameters deviating from actual working conditions. Based on heating reference parameters, a temperature field-flow field coupled physical constraint equation is constructed. The multi-source uncertainty of ambient temperature, gas calorific value and waste aluminum heat absorption rate is quantified by interval type II fuzzy set. The power constraint boundary with multiple confidence levels is obtained by inverse solution of cut set decomposition. The divided phased power sub-feasible domain can dynamically adapt to operating condition fluctuations, ensuring the compliance and safety of power schemes under fluctuating scenarios. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. The following drawings are not drawn to scale according to the actual size, but are intended to illustrate the main idea of ​​the present invention.

[0012] Figure 1 This is a diagram illustrating the method steps of the present invention.

[0013] Figure 2 This is a system module connection diagram of the present invention. Detailed Implementation

[0014] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are also within the scope of protection of the present invention.

[0015] Please refer to Figure 1 As shown, the first aspect of the present invention provides a method for dynamic allocation of heating power based on reinforcement learning, the specific steps of which are as follows: Step 1: Collect furnace temperature operation data of each regenerator burner zone of multiple smelting furnaces in real time, decouple them into temperature rise trend, commutation period and random disturbance, decompose the commutation period term to extract multi-domain fluctuation characteristics, divide the process stages through the smelting process deduction model, and then statistically analyze the feature distribution center and fit the power response elastic coefficient to obtain the heating reference parameters of each stage. Step 2: Construct coupled physical constraint equations based on heating reference parameters at each process stage; determine uncertainty parameters and establish a fuzzy programming model through interval type II fuzzy sets; decompose the power boundary constraint based on cut set decomposition; and divide the power sub-feasible region after traversing feasible solutions. Step 3: Divide the action set based on the power sub-feasible region of each process stage, build an upper and lower deep Q network, obtain a hierarchical decision model through iterative training, infer and output the stage power strategy and partition power value, and generate a multi-zone initial power allocation scheme. Step 4: Construct a non-cooperative evolutionary game model with each smelting furnace as the main players, solve for the set of equilibrium power allocation strategies, select the optimal power scheduling scheme through cross-efficiency evaluation and robustness analysis, calculate the safety confidence based on DS evidence theory, and automatically trigger power degradation and alarms in case of anomalies.

[0016] In this application, based on step one, the furnace temperature data of each regenerable burner zone of multiple smelting furnaces is collected in real time, decoupled into temperature rise trend, commutation period, and random disturbance, the commutation period term is decomposed to extract multi-domain fluctuation characteristics, the process stages are divided through the smelting process derivation model, the feature distribution center is statistically analyzed and the power response elastic coefficient is fitted, and the heating reference parameters for each stage are combined to obtain the specific steps: Real-time furnace temperature operation data from each regenerator burner zone of several smelting furnaces is collected. The furnace temperature operation data is decoupled to obtain a temperature rise trend term, a regenerator reversal cycle term, and a random disturbance term. It should be noted that the temperature rise trend term is extracted through morphological opening operations, constructed by creating a three-layer morphological structure element adapted to different process time scales. These three layers correspond to three time scales in the smelting process: minute-level process temperature rise trend, ten-second-level regenerator reversal cycle, and second-level random noise. The execution logic of the morphological opening operation is as follows: first, an erosion operation is performed on the original signal, then an expansion operation is performed on the erosion result to obtain the temperature rise trend term. The temperature rise trend term reflects the overall heat accumulation state of the smelting furnace. The furnace temperature operation data and the temperature rise trend term... Differential operations are performed to obtain the residual time series. Within the reversing delay interval specified by the process, the normalized autocorrelation function is calculated on the residual time series. The corresponding delays with coefficients exceeding the preset autocorrelation threshold are screened out, and the standard regenerator reversing cycle is obtained. The standard temperature cycle template is extracted from the database, the reversing cycle is segmented, and the matching correlation coefficient between the segment waveform and the standard cycle template is calculated for each segment. For segments with matching correlation coefficients that meet the standard, the reversing temperature waveform is directly restored. For low-matching segments with superimposed feeding and slag removal disturbances, the interference is removed by least squares fitting. Then, all segment extraction results are spliced ​​together to obtain the regenerator reversing cycle term. The furnace temperature operation data is subtracted sequentially from the temperature rise trend term and the regenerator reversing cycle term to obtain the random disturbance term. The thermal storage commutation periodic term is decomposed into several narrowband intrinsic mode functions by multi-scale decomposition. The calculation logic is as follows: , Where TS is the thermal storage commutation period term, Here, p represents the narrowband intrinsic mode function, and p is the center frequency of each component. It is a first-order partial differential operator. Here, is the Dirac delta function, jc is the imaginary unit, t is the time point, * is the convolution operator, K is the total number of modal decompositions, and k is the modal function number. For the complex exponential demodulation term, fractional Fourier transforms are performed on several narrowband intrinsic mode functions to obtain the transform results. The fractional-order fuzzy entropy is then calculated based on these results. It should be noted that the order p of the fractional Fourier transform is selected within the range of 0.5-1.5. In this embodiment, through order optimization, the optimal order p=0.8 is chosen, maximizing the energy concentration of the transformed signal and minimizing the fractional-order fuzzy entropy. The transform result is a fractional-order amplitude sequence. Phase space reconstruction is performed based on this fractional-order amplitude sequence to obtain several vectors. The distance between any two vectors is calculated, and this distance is input into an exponential fuzzy membership function to obtain the fuzzy similarity. Its calculation logic is as follows: Where exp is the natural exponential function, dc is the distance, r is the similarity tolerance for fuzzy similarity calculation, and n is the fuzzy boundary gradient coefficient. The global average similarity is obtained by averaging the similarities of all vectors. The corresponding fractional-order domain fuzzy entropy is then calculated based on the global average similarity. The calculation logic is as follows: ,in The global average similarity is given by m, where m is the embedding dimension of the phase space reconstruction, and IN is the natural logarithm operation. Then, the instantaneous frequencies are calculated for several narrowband intrinsic mode functions. Integrating these instantaneous frequencies yields the marginal spectrum. Based on this marginal spectrum, the marginal spectrum energy percentage (ER) is calculated. The calculation logic is as follows: Where hc is the marginal spectrum, This is the lower limit angular frequency of the main frequency range of the component. The upper angular frequency of the corresponding component's main frequency range. This is the maximum number of points. For definite integral calculation; taking the burner commutation trigger signal as the initial time point, temperature response segments of the commutation interval are extracted periodically from the corresponding narrowband intrinsic mode function. The commutation interval is... , To preset the steady-state duration before commutation, the response segments of several consecutive commutation cycles under the same operating condition are aligned with equal length and arithmetically averaged to obtain the step response sequence of the inherent mode function. Based on the same step response sequence, the relative temperature rise peak within a single commutation cycle is taken as the impact peak. The impact peak is compared with the standard steady-state temperature rise amplitude in the database to obtain the commutation impact attenuation coefficient. The commutation impact attenuation coefficient, fractional-order domain fuzzy entropy, and marginal spectrum energy proportion are integrated into a set of fluctuation characteristics. A smelting process simulation model is established. The fluctuation feature set and furnace temperature operation data are input into the smelting process simulation model. The smelting process simulation model outputs the smelting process hidden state transition matrix and feature observation probability matrix. Based on the smelting process hidden state transition matrix and feature observation probability matrix, Viterbi decoding is performed on the furnace temperature operation data to output time series segmentation results, including charging, primary refining, secondary refining and settling. Gaussian kernel density estimation is performed on the fluctuation feature set of each segment to obtain the feature distribution center. The furnace temperature operation data is matched with the time interval of the corresponding process stage. The heating power time series data within this time period is extracted. Then, the furnace temperature change rate of each sampling point is calculated for the furnace temperature time series data within the same segment. The natural logarithmic transformation is performed on the heating power time series data and the furnace temperature change rate to obtain the first linear value and the second linear value. A univariate linear regression equation is constructed using the first linear value and the second linear value. The calculation logic is as follows: ,in The regression intercept term reflects the basic heat loss at this stage of the process. The regression slope represents the power response elasticity coefficient for this stage of the process. As a random error term, the power response elastic coefficient is obtained by solving. The characteristic distribution center and the power response elastic coefficient are integrated to obtain the heating reference parameters for each process stage. It should be noted that the smelting process deduction model is designed to automatically identify the boundaries of smelting process stages and provide a basis for process division for staged heating power control. The model is trained using the Baum-Welch iterative algorithm. During the iteration process, the joint probability of matching the current observed features with each process hidden state from the start of the sequence to each time step is calculated using forward probability recursion. Then, the conditional probability of matching all observed sequences starting from each process hidden state at each time step is calculated using backward probability recursion. The forward and backward probability results are fused, and the hidden state occupancy probability and state transition occupancy probability at each time step are calculated. The model parameters are updated in this way. Multiple iterations are performed until the model can accurately measure the observed features. After the log-likelihood of the sequence converges, the output is the hidden state transition matrix and the feature observation probability matrix of the smelting process. The observation vector of the model is a normalized feature vector, which consists of the partitioned average furnace temperature, the reversal impact attenuation coefficient, the fractional domain fuzzy entropy, the marginal spectrum energy ratio, and the furnace temperature change rate. The number of hidden states is set to 4, corresponding to the four process stages of charging, primary refining, secondary refining, and settling. The observation probability follows a Gaussian mixture distribution. The model convergence criterion is that the change in the log-likelihood of the observation sequence for three consecutive iterations is less than a preset change. Based on the smelting furnace process identification scenario, the relative change value is taken as 10⁻³, and the smaller the value, the better the model parameters. Higher accuracy comes with more training iterations and longer processing time. The dimensions of the hidden state transition matrix of the smelting process correspond to each process stage. A probability mapping relationship between the state and fluctuation feature values ​​of each process stage is established, representing the distribution probability of the corresponding feature value under a certain process stage. This reflects the differences in furnace temperature fluctuation and power change characteristics at different stages. Then, based on the Viterbi optimal path decoding algorithm, the hidden state path is solved time-by-time for the furnace temperature operation sequence data. By dynamically programming, all possible combinations of process stage sequences are traversed, and the hidden state sequence that maximizes the probability of the entire observation sequence is selected as the optimal partitioning result. The time-series segmentation result of the entire smelting cycle is output. The process is divided into time intervals and boundary nodes for each stage, and feature samples of each operating condition are collected. Gaussian kernel density estimation is performed on all feature samples of each operating condition within the segment. The probability density distribution curves of each dimension of features are fitted by an adaptive bandwidth kernel function. The feature value corresponding to the density peak is taken as the feature distribution center of the process stage, which represents the typical steady-state operating level of furnace temperature, power and other operating parameters under the stage. The feature distribution center of each stage and the power response elasticity coefficient are two types of statistical indicators that constitute the heating benchmark parameters of each process stage in the whole smelting process, providing a constraint benchmark and response parameter basis at the process mechanism level for the dynamic allocation of heating power in each stage.

[0017] In this application, based on step two, coupled physical constraint equations are constructed based on heating reference parameters at each process stage. Uncertainty parameters are determined through interval type II fuzzy sets, and a fuzzy programming model is established. Based on the inverse solution of the constraint power boundary using cut set decomposition, the feasible solutions are traversed, and the staged power sub-feasible regions are divided. The specific steps are as follows: Based on the heating reference parameters of each process stage, a coupled physical constraint equation for the temperature field and flow field is constructed, and its calculation logic is as follows: ,in The heat exchange efficiency of the regenerator under the current operating conditions. Based on the baseline thermal storage and heat exchange efficiency, The flow rate influence coefficient represents the heat exchange efficiency. The coefficient for flue gas generation per unit power. The actual heating output power of the regenerative burner zone. The calorific value of the reference gas, This represents the actual calorific value of the gas. As the baseline flue gas flow rate, The coefficient representing the temperature effect of heat exchange efficiency is given, where T is the actual operating temperature of the furnace. Using the reference furnace temperature, the coupled physical constraint equations are analyzed to obtain the physical constraint boundaries; Extract running samples with uncertain parameters from the database. Uncertain parameters include ambient temperature, calorific value of fuel gas and heat absorption rate of scrap aluminum. Analyze the running samples and obtain the upper membership function and lower membership function. Generate a set of interval type II fuzzy parameters based on the upper membership function and lower membership function. The interval type II fuzzy parameter set and physical constraint boundary are transformed to obtain fuzzy opportunity constraints. All fuzzy opportunity constraints are integrated to generate an interval type II fuzzy programming model. A clear analysis of the interval type II fuzzy programming model is performed. Parameter value intervals at different confidence levels are obtained through cut-set decomposition. These intervals are then imported into coupled physical constraint equations. Inverse calculations are performed on the furnace temperature compliance constraint, furnace pressure safety constraint, and heat storage and heat exchange efficiency constraint to obtain the upper and lower limits of each parameter, including the upper and lower limits of furnace temperature power, furnace pressure power, and heat exchange efficiency power. The intersection of the upper and lower limits of each parameter at the same confidence level is taken to obtain the power feasible interval. A power constraint space is constructed based on the power feasible intervals of each heat storage burner zone. Each dimension of this power constraint space corresponds to a heating power decision variable for a zone, and each set of constraint faces corresponds to a process or equipment constraint condition. Any coordinate point within the space corresponds to a multi-zone heating power allocation scheme that satisfies all process requirements, safety restrictions, and equipment rating conditions at the corresponding confidence level. Points outside the space violate at least one constraint and are considered infeasible power allocation schemes. The feasible solution traversal of the power constraint space is performed to obtain the initial feasible solution set. The initial feasible solution set and the heating reference parameters of each process stage are mapped and divided to obtain the power sub-feasible domains corresponding to each process stage. Each sub-feasible domain is directly used as the action space boundary of the reinforcement learning model under the corresponding process stage, so that the power decision always falls within the range of process compliance and equipment safety. It should be noted that steady-state analysis of the coupled physical constraint equations yields a quantitative mapping relationship between heating power and furnace temperature, furnace pressure, and heat storage efficiency under deterministic operating conditions, generating physical constraint boundaries. Long-term operating samples of uncertain parameters are extracted from the database, and statistical analysis is used to obtain the extreme value fluctuation range, most probable distribution interval, and discrete distribution characteristics of each parameter. Based on this, upper and lower membership functions are constructed for each parameter. The upper membership function covers all fluctuation boundaries of the parameter, describing the maximum possible range of parameter values, while the lower membership function characterizes the parameter's most concentrated value. The distribution interval describes the core confidence range of parameter values. It is defined by two membership functions, one above the other, defining the interval type II fuzzy number for each parameter. This is then integrated to generate an interval type II fuzzy parameter set. Based on this set and the resolved physical constraint boundaries, the rigid process constraints under deterministic operating conditions are transformed into fuzzy chance constraints, thus completing the construction of the interval type II fuzzy programming model. Specifically, it no longer requires that the three types of constraints—furnace temperature compliance, furnace pressure safety, and heat storage and heat exchange efficiency compliance—be absolutely true under all operating conditions. Instead, it transforms the requirement that the probability of these constraints being true is not lower than a pre-set confidence level. The opportunity constraint is expressed horizontally, and a complete interval type-two fuzzy programming model is constructed. This transforms the power constraint problem with uncertainty into a quantifiable mathematical programming problem. Subsequently, the interval type-two fuzzy programming model is clarified through a set decomposition method, converting the fuzzy parameters into clear value intervals at different confidence levels. Specifically, for each set confidence level, the upper and lower membership functions of the three types of interval type-two fuzzy parameters are extracted to obtain the parameter value intervals corresponding to ambient temperature, gas calorific value, and waste aluminum heat absorption rate at that confidence level. Higher confidence levels result in narrower parameter value intervals and stricter constraints. The parameter value ranges at each confidence level are successively substituted into the calibrated coupled physical constraint equations to perform inverse calculations for the furnace temperature compliance constraint, furnace pressure safety constraint, and heat storage and heat exchange efficiency constraint. The inverse solution logic for the three types of constraints is as follows: In the inverse solution of the furnace temperature compliance constraint, the most unfavorable combination of the upper limit of ambient temperature, the lower limit of gas calorific value, and the upper limit of waste aluminum heat absorption rate is taken when solving for the lower limit of power; the most unfavorable combination of the lower limit of ambient temperature, the upper limit of gas calorific value, and the lower limit of waste aluminum heat absorption rate is taken when solving for the upper limit of power; in the inverse solution of the furnace pressure safety constraint, the lower limit of gas calorific value is taken as the most unfavorable parameter to solve for the upper limit of power, and the upper limit of gas calorific value is taken to solve for the lower limit of power.In the inverse kinematics solution constrained by thermal storage and heat exchange efficiency, the lower limit of the gas calorific value is taken as the most unfavorable parameter. The upper limit of power is obtained by solving a simultaneous equation. Based on the same confidence level, the intersection of the upper and lower limits of power corresponding to all constraint categories is taken. The comprehensive lower limit of power is the maximum value of the lower limits of each constraint, and the comprehensive upper limit of power is the minimum value of the upper limits of each constraint. This ensures that the obtained power range satisfies all constraint requirements. By substituting the most unfavorable parameter combination, the allowable power range corresponding to each type of constraint is derived, namely the upper and lower limits of power for furnace temperature, furnace pressure, and heat exchange efficiency. The most unfavorable parameter combination is used for solving interval uncertainties. The conservative principle refers to finding the parameter value combination that makes it most difficult to satisfy the corresponding constraint conditions and the corresponding power feasible boundary within the parameter cut-off range of a given confidence level. The power boundary obtained by solving this combination can guarantee that the constraint conditions are valid throughout the entire parameter fluctuation range. For the same confidence level, the intersection of the upper and lower limits of the corresponding power for all constraint categories is taken, where the lower limit of the comprehensive power is the maximum value of the lower limits of each constraint and the upper limit of the comprehensive power is the minimum value of the upper limits of each constraint. This ensures that the obtained power range can satisfy all constraint requirements, thus obtaining the single-partition power feasible range under this confidence level. In this application, based on step three, which involves dividing the action set according to the power sub-feasible region of each process stage, an upper and lower depth Q-network is constructed. A hierarchical decision model is obtained through iterative training, and the stage power strategy and partition power values ​​are inferred and output to generate a multi-zone initial power allocation scheme. The specific steps are as follows: The power sub-feasible domains of each process stage are divided into stage granularity and partition granularity to obtain the action set of the upper process stage and the action set of the lower partition power level. The overall fluctuation characteristics of the smelting furnace, furnace temperature operation data, and heating reference parameters are concatenated into a global operating condition feature vector. The global operating condition feature vector is then input into the upper-level deep Q-network. The global operating condition feature vector and the upper-level process stage action set are analyzed, and the total power command and stage command are output. The total power command, stage command, and real-time operating condition feature vectors of each regenerator burner zone are input into the lower-level deep Q network. The lower-level deep Q network matches the command with the vector and the lower-level zone power level action set to obtain the power adjustment value of each regenerator burner zone. The upper-layer deep Q-network and the lower-layer deep Q-network are iteratively trained to obtain a hierarchical decision model. The hierarchical decision model is used to reason about the current smelting conditions and output the optimal stage power strategy and partition power value in sequence to obtain the initial power allocation scheme for multiple zones. It should be noted that when the smelting process simulation model outputs the time-series segmented results, it triggers the upper-level DQN decision, such as the transition to primary refining after feeding or the transition to static setting after refining meets the standards, and updates the total power command and stage command. The lower-level DQN decision triggering cycle is aligned with the regenerator burner reversing cycle. Specifically, each reversing cycle triggers a partition power decision, and the power update is performed in the interval between each regenerator burner reversing, such as the transition interval before the reversing valve is switched and the flue gas flow direction is stabilized. The process cycle time compliance rate refers to the proportion of smelting furnaces in the total number of production furnaces where the deviation between the actual duration and the standard process cycle time of each process stage in a single furnace is within the preset allowable range during the statistical period. The lower-layer deep Q-network uses a multi-branch fully connected structure with shared feature layers and independent branches in each region: the input feature vector is processed through two shared fully connected layers to extract common operating condition features, and then processed through independent branches equal to the number of regenerative burner regions. Each branch contains two fully connected layers, and the parameters of each branch are trained independently, outputting the Q-values ​​of each power level for the corresponding region. The convergence criterion for the alternating training of the two-layer network is that, in 20 consecutive iterations, the fluctuation amplitude of the upper-layer network loss and the fluctuation amplitude of the lower-layer network loss are both less than the preset fluctuation amplitude. The convergence threshold for loss fluctuation during DQN training generally ranges from 10⁻³ to 10⁻¹. In this application, the layered architecture uses the following values: a typical value of 5×10⁻³ for the upper-layer global scheduling network and a typical value of 10⁻² for the lower-layer partitioned power network. The action space is discretized based on the power sub-feasible domain corresponding to each process stage. The upper-layer process stage action set uses the upper and lower limits of the total power in each stage's sub-feasible domain as boundaries, and discretizes the total heating power of a single stage into several standard power levels by adjusting the step size at equal intervals. Timing control actions for stage holding and forward process switching are set. The total power value corresponding to all actions is strictly constrained within the feasible domain of the corresponding process stage. The lower-layer partitioned power level action set... Using the total power level of the upper layer as the overall constraint, for each regenerable burner zone, several sub-power levels are discretized with the feasible power range of that zone as the boundary and the minimum single-step adjustment accuracy as the step size. The upper-layer deep Q-network uses a three-layer fully connected hidden layer structure for feature extraction and value fitting. The output corresponds to the Q value of all actions in the action set of the upper-layer process stage. The action with the largest Q value is selected to output the optimal total power command and process stage command, undertaking the functions of controlling the heating rhythm of the entire melting process, deciding the timing of process advancement, and scheduling the total power. The lower-layer deep Q-network is adapted to a multi-branch fully connected structure of the zone dimension. The output corresponds to the action Q value of all selectable power levels in each zone, and the optimal value is selected zone by zone. The power adjustment level outputs the zone power setting value. The upper-level reward is calculated based on the whole process gas energy consumption, process cycle time compliance rate, and global constraint compliance. The lower-level reward is calculated based on the zone temperature tracking accuracy, furnace temperature field uniformity, and local constraint compliance. During the training process of the hierarchical reinforcement learning decision model, the target network separation mechanism and the priority experience replay strategy are configured simultaneously. A two-layer joint experience replay pool is constructed to store complete state-action-reward transition samples. High-value samples are weighted and sampled according to the priority of temporal differential error. First, the upper-layer network parameters are fixed to train the lower-layer zone allocation network until convergence. Then, the lower-layer network parameters are fixed to train the upper-layer global scheduling network. Multiple rounds of alternating iterations are performed until the loss functions of both layers tend to stabilize.

[0018] In this application, based on step four, a non-cooperative evolutionary game model is constructed with each smelting furnace as the main player. The equilibrium power allocation strategy set is obtained by solving the model. The optimal power scheduling scheme is selected through cross-efficiency evaluation and robustness analysis. The safety confidence level is calculated based on DS evidence theory. Power degradation and alarms are automatically triggered in case of anomalies. The specific steps are as follows: Each smelting furnace is set as the main player in the game. The initial power allocation scheme of the multi-zone is used as the game strategy set. Energy consumption cost, production efficiency and heat storage ball loss are used as the payoff functions to construct a non-cooperative evolutionary game model for multiple furnaces. The evolutionary stable strategy of the multi-furnace non-cooperative evolutionary game model is solved by replicating the dynamic equation. Based on the evolutionary stable strategy, the Nash equilibrium state of multi-furnace power scheduling is analyzed, and the power allocation strategy set under the game equilibrium is obtained. A cross-efficiency evaluation model is constructed, with the gas power input of each zone as the input index and the temperature compliance rate, heating rate and heat storage and heat exchange efficiency as the output index. The cross-efficiency is calculated for each scheme in the power allocation strategy set to obtain the cross-efficiency value. The power allocation strategy set is sorted according to the cross-efficiency value, and the top-ranked schemes are selected as candidate schemes. Robustness index analysis is performed on the candidate scheme set to obtain the robustness index. The scheme with the highest score is selected and marked as the power dynamic scheduling optimization scheme based on the cross-efficiency value and the robustness index. Based on the power dynamic scheduling optimization scheme, the power target value, adjustment step size and adjustment timing of each heating zone are analyzed to obtain the power adjustment parameter matrix; The power regulation parameter matrix is ​​sent to the actuators of each regenerable burner, and the power is adjusted in a closed loop by a fuzzy PID controller. Real-time feedback data of zone temperature and furnace pressure are collected and the adjustment amount is dynamically corrected. The evidence bodies for temperature anomalies, pressure anomalies, and equipment failures are constructed using the DS evidence theory. The safety confidence level is then calculated by integrating the safety confidence level of the current heating state. When the safety confidence level is lower than the preset confidence threshold, the heating power of the corresponding zone is automatically reduced to the safe range, and an abnormal alarm record is generated. It should be noted that the game involves each regenerative smelting furnace operating in parallel within the workshop. Each furnace can independently adjust its power allocation strategy to optimize its own operational benefits, and there is no centralized forced scheduling. Therefore, it is a non-cooperative game. The set of optional strategies for each furnace consists of all feasible initial power allocation schemes for the current process stage. Each strategy corresponds to a set of power settings for each regenerative burner zone. A comprehensive benefit function is constructed that includes both positive benefits and negative costs. The comprehensive benefit U of a single furnace is calculated by coupling the production efficiency benefit RP, energy consumption cost CE, and regenerative ball loss cost CW. The calculation logic is as follows: Set the percentage of smelting furnaces using the j-th strategy at the initial time. The replication dynamic equation is obtained based on the proportion of smelting furnaces, and its calculation logic is as follows: ,in The total profit per furnace is given by US, where US is the average profit of the game population, and mc is the percentage of furnaces in the total number of furnaces. To determine the rate of change of strategy proportion over time, the replication dynamic equations for all strategies are set to 0. Solving the system of equations yields the equilibrium points for all strategy proportions. The equilibrium points are then verified using the Jacobian matrix local stability criterion. If the real parts of all eigenvalues ​​of the Jacobian matrix corresponding to an equilibrium point are less than 0, then that equilibrium point represents an evolutionarily stable strategy. In this case, any unilateral change in power strategy by any smelting furnace will not improve its own revenue, and the power allocation across multiple furnaces reaches a stable Nash equilibrium. The power allocation strategy set under the game equilibrium is output. Each power allocation strategy set is set as a decision-making unit, and an input-output evaluation index is constructed. The input index is the gas heating power of each regenerator burner zone, representing the scale of power resource input. The output indexes are the zone temperature compliance rate, furnace temperature rise rate, and regenerator ball heat exchange efficiency, corresponding to... The calculation logic for cross-efficiency, including temperature control quality, production efficiency, and heat exchange performance, is as follows: For any decision unit, the cross-efficiency evaluation model aims to maximize its own efficiency. It solves for the optimal input and output weights using linear programming to obtain the self-evaluation efficiency value of that unit. The optimal weights of each decision unit are then used as common evaluation weights to calculate the efficiency values ​​of all other decision units, resulting in the other-evaluation efficiency values ​​for any two units. Finally, the arithmetic mean of all other-evaluation efficiency values ​​and self-evaluation efficiency values ​​for all decision units is taken to obtain the cross-efficiency value. The robustness index calculation uses random fluctuations of the uncertainty parameter as the test condition. Based on its interval type-II fuzzy distribution characteristics, Monte Carlo random sampling is performed to generate several sets of fluctuating operating condition samples. Each set of samples is substituted into each candidate scheme to calculate the robustness index of the scheme. The calculation logic is as follows: Where FG is the mean of the single-furnace performance index, KM is the standard deviation of the performance index, and RI is the robustness index. In the comprehensive scoring calculation, the cross-efficiency value is weighted at 0.6, and the robustness index is weighted at 0.4. The weights can be adjusted according to production needs and must sum to 1. The three indicators of single-furnace comprehensive benefits are all converted to the monetary unit of yuan / furnace through economic value conversion: production efficiency benefits are calculated from the smelting time deviation and the unit time capacity benefit coefficient; energy consumption costs are calculated from the total gas consumption and the gas unit price; and regenerator ball loss costs are calculated from the power fluctuation amplitude, running time, and unit loss conversion coefficient. The power adjustment parameter matrix is ​​arranged with each regenerator burner zone as the row and the core as the column. The adjustment parameters are listed in the column. The power target value is the steady-state optimal heating power setpoint for each zone in the corresponding scheme. The adjustment step size is the maximum allowable amplitude of a single power adjustment. The adjustment timing is the trigger time node of the power adjustment, aligned with the regenerator burner reversing cycle. Power adjustment is performed during the reversing interval. The fuzzy PID controller takes the deviation and rate of change of the zone temperature target value from the measured value as input. First, the numerical values ​​are converted into fuzzy linguistic variables through a fuzzification interface. Then, fuzzy inference is performed based on the preset temperature adjustment fuzzy rule base. The proportional, integral, and derivative coefficients of the PID controller are dynamically corrected in real time. The real-time power adjustment is output through defuzzification. The DS evidence theory sets the recognition framework as follows. ={safe, abnormal, fault}, corresponding to three heating operation states, three independent evidence bodies are constructed for each state, and their basic probability assignments are calculated. The specific function for the temperature abnormality evidence body is: ; The specific function for the evidence body of abnormal stress is: The inputs for equipment failure evidence are valve opening deviation, gas flow fluctuation coefficient, and cumulative equipment runtime. Based on the health lookup table rule: a pre-established equipment health rating table is created, divided into three levels—healthy, sub-healthy, and faulty—based on three dimensions, with corresponding baseline probability values ​​output. The basic probability distributions of the three types of evidence are then successively fused using the Dempster combination rule to obtain the fused state probability distribution. , In the above formula For zone temperature deviation, For the rate of temperature change, and The calibration thresholds for temperature deviation and rate of temperature change are 15℃ and 2℃, respectively. , and These represent the probabilities of temperature-based evidence supporting safe, abnormal, and fault states, respectively. The uncertainty of temperature evidence is used to cover aspects such as measurement errors and minor operating condition disturbances that cannot be clearly attributed. This represents the difference between the current measured furnace pressure and the standard furnace pressure. and The furnace pressure safety allowable threshold and furnace pressure alarm threshold are 50 Pa and 150 Pa, respectively. , and , respectively, represent the support probabilities of temperature evidence for safe, abnormal, and fault states; KL is the evidence conflict coefficient, characterizing the degree of contradiction among the three types of evidence, with a value range of ([0,1]); A1, A2, and A3 represent the state propositions supported by temperature evidence, pressure evidence, and equipment fault evidence, respectively. For the intersection operation of sets, As evidence of equipment failure, the fusion probability value corresponding to the safety status in the identification framework is taken and marked as the safety confidence level of the current heating status.

[0019] Please refer to Figure 2 As shown, the second aspect of the present invention provides a dynamic heating power allocation system based on reinforcement learning, including a feature extraction module, a constraint modeling module, a decision modeling module, and a game optimization module; The feature extraction module collects furnace temperature operation data of each regenerator burner zone of multiple smelting furnaces in real time, decouples them into temperature rise trend, commutation period and random disturbance, decomposes the commutation period term to extract multi-domain fluctuation features, divides the process stages through the smelting process deduction model, and then statistically analyzes the feature distribution center and fits the power response elastic coefficient to obtain the heating reference parameters of each stage. The constraint modeling module constructs coupled physical constraint equations based on heating reference parameters at each process stage, determines uncertainty parameters and establishes a fuzzy programming model through interval type II fuzzy sets, and decomposes the constraint power boundary based on cut set decomposition, and divides the power sub-feasible region after traversing feasible solutions. The decision modeling module divides the action set based on the power sub-feasible domain of each process stage, builds upper and lower deep Q networks, obtains a hierarchical decision model through iterative training, infers and outputs the stage power strategy and partition power value, and generates a multi-zone initial power allocation scheme. The game optimization module constructs a non-cooperative evolutionary game model with each smelting furnace as the main players, solves the equilibrium power allocation strategy set, selects the optimal power scheduling scheme through cross-efficiency evaluation and robustness analysis, calculates the safety confidence based on DS evidence theory, and triggers power degradation and alarms in case of anomalies.

[0020] The above formulas all use dimensionless numerical calculations, which can be achieved through standardization, normalization, etc., and will not be elaborated here. The relevant expressions are obtained by collecting a large amount of actual production and operation data of regenerative smelting furnaces and fitting them through simulation, which closely approximates the actual thermal operation state of the smelting furnace. The preset parameters in the formulas can be set by those skilled in the art according to the smelting furnace type and specifications, raw material operating characteristics, production process requirements and scheduling control accuracy.

[0021] The above embodiments can be implemented in whole or in part through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented as a computer program product, which contains one or more computer-executable instructions. When the instructions are loaded and run on a computing device, they fully or partially implement the dynamic heating power allocation process and function described in this invention. The device includes a workshop-level heating control host, an edge computing terminal, a regenerative burner controller, a PLC industrial control unit, and a cloud-based production operation and maintenance platform. The instructions can be stored on a computer-readable storage medium or transmitted between devices via industrial Ethernet, fieldbus, or wireless industrial communication networks. The readable storage medium includes industrial-grade storage media such as solid-state drives, industrial-grade USB flash drives, memory cards, disks, and optical discs.

[0022] The execution order of each power scheduling process in this invention is based on functional and logical relationships, and is not limited by the step number. Those skilled in the art will understand that the system modules and algorithm steps described in this invention can be implemented using electronic hardware or a combination of hardware and software. Appropriate implementation methods can be selected for different furnace types, production capacity scales, and on-site management architectures, all of which are within the scope of protection of this invention.

[0023] The reinforcement learning-based dynamic heating power allocation system disclosed in this invention can also be implemented in other non-illustrative ways. The module division is only a logical functional division; other division methods can be used in practice. Multiple modules or components can be combined or integrated into the same hardware device, and some functional features can be tailored according to on-site requirements. The coupling and connection between modules can be indirect, such as electrical connection, industrial bus communication, industrial Ethernet communication, or other forms.

[0024] The modules described as separate components can be physically separated or not. The components displayed as modules can be physical units or distributed across multiple industrial control nodes. Some or all of the modules can be selected to achieve the purpose of the solution according to actual production scheduling needs.

[0025] When the functions of this invention are implemented as software functional units and sold or used as products, they can be stored in a computer-readable storage medium. This software product includes several instructions for driving industrial control hosts, workshop production management systems, thermal regenerative combustion control devices, programmable logic controllers, and other equipment to execute all or part of the power allocation and scheduling steps of this invention. The storage medium includes industrial-grade media capable of storing program code, such as solid-state drives, flash drives, portable hard drives, read-only memory, and random access memory.

[0026] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the claims.

Claims

1. A method for dynamic allocation of heating power based on reinforcement learning, characterized in that, Includes the following steps: Step 1: Collect furnace temperature operation data of each regenerator burner zone of multiple smelting furnaces in real time, decouple them into temperature rise trend, commutation period and random disturbance, decompose the commutation period term to extract multi-domain fluctuation characteristics, divide the process stages through the smelting process derivation model, and analyze to obtain the heating reference parameters of each stage. Step 2: Construct coupled physical constraint equations based on heating reference parameters, determine uncertainty parameters and establish a fuzzy programming model through interval type II fuzzy sets, decompose the inverse solution of the constraint power boundary based on the cut set, and divide the power sub-feasible region after traversing the feasible solution. Step 3: Divide the action set based on the power sub-feasible region of each process stage, build an upper and lower deep Q network, obtain a hierarchical decision model through iterative training, infer and output the stage power strategy and partition power value, and generate a multi-zone initial power allocation scheme. Step 4: Construct a non-cooperative evolutionary game model with each smelting furnace as the main players, solve for the set of equilibrium power allocation strategies, select the optimal power scheduling scheme through cross-efficiency evaluation and robustness analysis, calculate the safety confidence based on DS evidence theory, and automatically trigger power degradation and alarms in case of anomalies.

2. The method for dynamic allocation of heating power based on reinforcement learning according to claim 1, characterized in that, In step one, furnace temperature data for each regenerable burner zone of multiple smelting furnaces is collected in real time, decoupled into temperature rise trend, commutation period, and random disturbance, and multi-domain fluctuation characteristics are extracted by decomposing the commutation period term. The specific steps are as follows: Real-time collection of furnace temperature operation data for each regenerator burner zone of several smelting furnaces; decoupling of furnace temperature operation data to obtain temperature rise trend term, regenerator reversal cycle term and random disturbance term; The heat storage commutation periodic term is decomposed into several narrowband intrinsic mode functions (IMFs). Fractional Fourier transforms are performed on each IMF to obtain the transform results. The fractional domain fuzzy entropy is calculated based on the transform results. The instantaneous frequency is then calculated for each IMF. The marginal spectrum is obtained by integrating the instantaneous frequency. The marginal spectrum energy ratio is calculated based on the marginal spectrum. Taking the burner commutation trigger signal as the initial time point, temperature response segments of the commutation interval are extracted periodically from the corresponding narrowband IMF. The response segments of several consecutive commutation cycles under the same operating condition are aligned to the same length and arithmetically averaged to obtain the step response sequence of the IMF. The relative temperature rise peak within a single commutation cycle is taken as the impact peak based on the same step response sequence. The impact peak is compared with the standard steady-state temperature rise amplitude in the database to obtain the commutation impact attenuation coefficient. The commutation impact attenuation coefficient, fractional domain fuzzy entropy, and marginal spectrum energy ratio are integrated into a set of fluctuation characteristics.

3. The method for dynamic allocation of heating power based on reinforcement learning according to claim 2, characterized in that, In step one, the process stages are divided using a smelting process deduction model, the characteristic distribution center is statistically analyzed, and the power response elastic coefficient is fitted to obtain the heating reference parameters for each stage. The specific steps are as follows: A smelting process simulation model is established. The fluctuation feature set and furnace temperature operation data are input into the smelting process simulation model. The smelting process simulation model outputs the smelting process hidden state transition matrix and feature observation probability matrix. Based on the smelting process hidden state transition matrix and feature observation probability matrix, Viterbi decoding is performed on the furnace temperature operation data to output time series segmentation results. Gaussian kernel density estimation is performed on the fluctuation feature set of each segment to obtain the feature distribution center. Match the furnace temperature operation data to the time interval of the corresponding process stage, extract the heating power time series data within that time period, and then calculate the furnace temperature change rate of each furnace temperature sampling point for the furnace temperature time series data in the same segment. Perform natural logarithmic transformation on the heating power time series data and each furnace temperature change rate to obtain the first linear value and the second linear value. A univariate linear regression equation is constructed using the first and second linear values, and the power response elastic coefficient is obtained by solving it. The characteristic distribution center and the power response elastic coefficient are then integrated to obtain the heating reference parameters for each process stage.

4. The method for dynamic allocation of heating power based on reinforcement learning according to claim 1, characterized in that, In step two, coupled physical constraint equations are constructed based on the heating reference parameters of each process stage. Uncertainty parameters are determined through interval type II fuzzy sets, and a fuzzy programming model is established. The specific steps are as follows: Based on the heating reference parameters of each process stage, a coupled physical constraint equation for the temperature field and flow field is constructed. The coupled physical constraint equation is analyzed to obtain the physical constraint boundary. Extract running samples with uncertain parameters from the database. Uncertain parameters include ambient temperature, calorific value of fuel gas and heat absorption rate of scrap aluminum. Analyze the running samples and obtain the upper membership function and lower membership function. Generate a set of interval type II fuzzy parameters based on the upper membership function and lower membership function. The interval type II fuzzy parameter set and physical constraint boundary are transformed to obtain fuzzy opportunity constraints. All fuzzy opportunity constraints are integrated to generate an interval type II fuzzy programming model.

5. The method for dynamic allocation of heating power based on reinforcement learning according to claim 4, characterized in that, In step two, based on the inverse solution of the truncated set decomposition constrained power boundary, after traversing feasible solutions, the power sub-feasible region is divided into stages. The specific steps are as follows: The interval type II fuzzy programming model is analyzed for clarity. The parameter value intervals under different confidence levels are obtained by truncation decomposition. The parameter value intervals are then imported into the coupled physical constraint equations. The furnace temperature compliance constraint, furnace pressure safety constraint, and heat storage and heat exchange efficiency constraint are calculated by inverse solution to obtain the upper and lower limits of each parameter. The intersection of the upper and lower limits of each parameter under the same confidence level is taken to obtain the power feasible interval. The power constraint space is constructed based on the power feasible intervals of each heat storage burner zone. The feasible solution traversal of the power constraint space is performed to obtain the initial feasible solution set. The initial feasible solution set and the heating reference parameters of each process stage are mapped and partitioned to obtain the power sub-feasible region corresponding to each process stage.

6. The method for dynamic allocation of heating power based on reinforcement learning according to claim 1, characterized in that, In step three, action sets are divided based on the power sub-feasible regions of each process stage, upper and lower depth Q-networks are built, and a hierarchical decision model is obtained through iterative training. The stage power strategy and partition power value are inferred and output to generate a multi-zone initial power allocation scheme. The specific steps are as follows: The power sub-feasible domains of each process stage are divided into stage granularity and partition granularity to obtain the action set of the upper process stage and the action set of the lower partition power level. The overall fluctuation characteristics of the smelting furnace, furnace temperature operation data, and heating reference parameters are concatenated into a global operating condition feature vector. The global operating condition feature vector is then input into the upper-level deep Q-network. The global operating condition feature vector and the upper-level process stage action set are analyzed, and the total power command and stage command are output. The total power command, stage command, and real-time operating condition feature vectors of each regenerator burner zone are input into the lower-level deep Q network. The lower-level deep Q network matches the command with the vector and the lower-level zone power level action set to obtain the power adjustment value of each regenerator burner zone. The upper-layer deep Q-network and the lower-layer deep Q-network are iteratively trained to obtain a hierarchical decision model. The hierarchical decision model is used to reason about the current smelting conditions and output the optimal stage power strategy and partition power value in sequence to obtain the initial power allocation scheme for multiple zones.

7. The method for dynamic allocation of heating power based on reinforcement learning according to claim 1, characterized in that, In step four, a non-cooperative evolutionary game model is constructed with each smelting furnace as the main players. The equilibrium power allocation strategy set is obtained by solving the model. The optimal power scheduling scheme is then selected through cross-efficiency evaluation and robustness analysis. The specific steps are as follows: Each smelting furnace is set as the main player in the game. The initial power allocation scheme of the multi-zone is used as the game strategy set. Energy consumption cost, production efficiency and heat storage ball loss are used as the payoff functions to construct a non-cooperative evolutionary game model for multiple furnaces. The evolutionary stable strategy of the multi-furnace non-cooperative evolutionary game model is solved by replicating the dynamic equation. Based on the evolutionary stable strategy, the Nash equilibrium state of multi-furnace power scheduling is analyzed, and the power allocation strategy set under the game equilibrium is obtained. A cross-efficiency evaluation model is constructed, with the gas power input of each zone as the input index and the temperature compliance rate, heating rate and heat storage and heat exchange efficiency as the output index. The cross-efficiency is calculated for each scheme in the power allocation strategy set to obtain the cross-efficiency value. The power allocation strategy set is sorted according to the cross-efficiency value, and the top-ranked schemes are selected as candidate schemes. Robustness index analysis is performed on the candidate scheme set to obtain the robustness index. The scheme with the highest score is selected and marked as the power dynamic scheduling optimization scheme based on the cross-efficiency value and the robustness index.

8. The method for dynamic allocation of heating power based on reinforcement learning according to claim 7, characterized in that, In step four, the security confidence level is calculated based on the DS evidence theory, and power degradation and alarms are automatically triggered in case of anomalies. The specific steps are as follows: Based on the power dynamic scheduling optimization scheme, the power target value, adjustment step size and adjustment timing of each heating zone are analyzed to obtain the power adjustment parameter matrix; The power regulation parameter matrix is ​​sent to the actuators of each regenerable burner, and the power is adjusted in a closed loop by a fuzzy PID controller. Real-time feedback data of zone temperature and furnace pressure are collected and the adjustment amount is dynamically corrected. The evidence bodies for temperature anomalies, pressure anomalies, and equipment failures are constructed using the DS evidence theory. The safety confidence level is then calculated by integrating the safety confidence level of the current heating state. When the safety confidence level is lower than the preset confidence threshold, the heating power of the corresponding zone is automatically reduced to the safe range, and an abnormal alarm record is generated.

9. A dynamic heating power allocation system based on reinforcement learning, characterized in that, The method for dynamic allocation of heating power based on reinforcement learning as described in any one of claims 1-8 includes: The feature extraction module collects furnace temperature operation data of each regenerator burner zone of multiple smelting furnaces in real time, decouples them into temperature rise trend, commutation period and random disturbance, decomposes the commutation period term to extract multi-domain fluctuation features, divides the process stages through the smelting process deduction model, and then statistically analyzes the feature distribution center and fits the power response elastic coefficient to obtain the heating reference parameters of each stage. The constraint modeling module constructs coupled physical constraint equations based on heating reference parameters at each process stage, determines uncertainty parameters and establishes a fuzzy programming model through interval type II fuzzy sets, and decomposes the constraint power boundary based on cut set decomposition, and divides the power sub-feasible region after traversing feasible solutions. The decision modeling module divides the action set based on the power sub-feasible domain of each process stage, builds upper and lower deep Q networks, obtains a hierarchical decision model through iterative training, infers and outputs the stage power strategy and partition power value, and generates a multi-zone initial power allocation scheme. The game optimization module constructs a non-cooperative evolutionary game model with each smelting furnace as the main players, solves the equilibrium power allocation strategy set, selects the optimal power scheduling scheme through cross-efficiency evaluation and robustness analysis, calculates the safety confidence based on DS evidence theory, and triggers power degradation and alarms in case of anomalies.

Images