A machine learning-based method for optimizing smelting process parameters

Abstract

This invention discloses a machine learning-based method for optimizing smelting process parameters, belonging to the field of machine learning technology. The method includes: obtaining historical furnace smelting process state sequences and current furnace smelting process state sequences; extracting process parameter adjustment events and constructing a parameter response state node sequence; executing the DENCLUE algorithm to form parameter response density attractors, parameter response transition edges, parameter response attraction path graphs, and target mass attractors; retrieving qualified parameter response attraction paths and combining candidate process parameter trajectory sets; constructing an improved MTGNN model to generate predicted attraction paths; forming the target process parameter trajectory; and sending the next parameter control segment or updating the target process parameter trajectory. This invention, by introducing the DENCLUE algorithm and the improved MTGNN model, achieves intelligent matching, continuous prediction, and execution feedback optimization of smelting process parameter adjustment paths.

Description

Technical Field

[0001] This invention relates to the field of machine learning technology, and in particular to a method for optimizing smelting process parameters based on machine learning. Background Technology

[0002] In the smelting process, process parameters such as oxygen flow rate, electrode power, fuel addition, reducing agent addition, and slagging agent addition directly affect furnace temperature, furnace pressure, flue gas composition, molten metal composition, and the final product composition. Current smelting process control typically relies on process experience, fixed process procedures, historical furnace statistics, and online monitoring data. Operators or control systems adjust process parameters based on temperature, composition, and equipment status. As the scale of smelting process data increases, some production systems are beginning to use data analysis models to predict furnace condition changes and final quality, and to provide parameter adjustment suggestions based on the prediction results, in order to reduce the impact of human experience differences on smelting stability.

[0003] Existing machine learning optimization methods typically divide historical furnace cycles into samples of similar operating conditions, then use clustering models, time series models, or graph neural network models to predict temperature changes, composition changes, or endpoint quality, and select parameter adjustment schemes based on the predicted values. While these methods can improve parameter recommendation efficiency in some scenarios, most still rely on variable correlation or single-step prediction results as the basis for adjustment. They fail to adequately characterize the continuous relationship between process parameter adjustment events, furnace condition response lags, equipment execution status, and endpoint composition compliance, leading to problems such as response misalignment, path discontinuity, and the inability to promptly correct post-execution deviations in the recommended trajectory.

[0004] Therefore, how to provide a machine learning-based method for optimizing smelting process parameters is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose a machine learning-based method for optimizing smelting process parameters. This invention utilizes the DENCLUE algorithm and an improved MTGNN model to jointly analyze process parameter adjustment events, furnace condition response status, equipment operating status, and final product composition data from furnace smelting operation data. It details the complete process from constructing parameter response density attractors, retrieving target mass attraction paths, matching current furnace condition attractors, predicting candidate process parameter trajectories, to execution feedback updates, achieving continuous optimization of smelting process parameter trajectories. Compared to methods that adjust parameters based solely on single-point predictions or empirical rules, this invention maintains the continuity of parameter adjustment paths, reduces prediction bias caused by furnace condition response lags, and improves the final component qualification rate, equipment execution safety, and smelting control stability.

[0006] A method for optimizing smelting process parameters based on machine learning according to an embodiment of the present invention includes: Collect furnace smelting operation data and preprocess the furnace smelting operation data to obtain the historical furnace smelting process status sequence and the current furnace smelting process status sequence. Extract process parameter adjustment events from historical furnace status sequences and construct parameter response status node sequences; The DENCLUE algorithm is executed based on the parameter response state node sequence to perform parameter action phase constraints, parameter response transition reachability constraints, and target mass path constraints. The density contribution is calculated to form parameter response density attractors, parameter response transition edges, parameter response attraction path graphs, and target mass attractors. Based on the current furnace state sequence, locate the current starting attractor, retrieve the qualified parameter response attraction path leading to the target mass attractor, extract the parameter response transfer edge and combine the candidate process parameter trajectory set; An improved MTGNN model is constructed based on the current furnace condition, the current starting attractor, and the set of candidate process parameter trajectories. Combined with parameter response transfer edge execution path constraint propagation, time delay alignment propagation, and attractor order preservation prediction, a predicted attraction path is generated. The predicted attraction path is matched and verified with the qualified parameter response attraction path, and the candidate process parameter trajectory set is screened by combining the target mass attractor arrival state, equipment operating state and product composition state to form the target process parameter trajectory. Send the current parameter control segment of the target process parameter trajectory, match the furnace condition response and termination attractor with the parameter response density attractor, and send the next parameter control segment or update the target process parameter trajectory according to the matching result.

[0007] Optionally, the furnace smelting operation data specifically includes furnace identifier, collection timestamp, oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, slagging agent addition data, furnace temperature data, furnace pressure data, flue gas composition data, molten metal composition data, equipment operating status data, and product final composition data.

[0008] Optionally, the preprocessing of the furnace smelting operation data includes... The smelting operation data of each furnace is grouped according to the furnace number identifier. Furnaces whose production status is completed and have product end-point component data are classified as historical furnaces, and furnaces whose production status is smelting in progress and have not yet formed product end-point component data are classified as current furnaces. The smelting operation data of the historical furnaces and the current furnaces are sorted, time-aligned, missing data is filled, anomalies are removed, and values ​​are normalized according to the collection timestamp, forming the historical furnace status sequence and the current furnace status sequence, respectively.

[0009] Optionally, the construction parameter response state node sequence includes: Based on the process parameter data and equipment operation status data in the historical furnace status sequence, the sampling locations where the process parameters changed and the equipment was in normal execution status were screened to determine the process parameter adjustment events. The furnace condition status before the adjustment, the process parameter change process, the response time and the furnace condition response status are extracted from the process parameter adjustment event to form a parameter response status node; Arrange the parameter response status nodes corresponding to the same process parameter adjustment event according to the collection timestamp, and construct the parameter response status node sequence.

[0010] Optionally, the formation of the parameter response density attractor, parameter response transition edge, parameter response attraction path graph, and target mass attractor includes: The DENCLUE algorithm is executed based on the parameter response state node sequence. Historical segments that match the furnace condition before adjustment and have not experienced the same type of process parameter adjustment event are retrieved according to the process parameter adjustment event, the furnace condition state before adjustment and the furnace condition response state. The actual response segment and the control response segment are obtained. Counterfactual parameter response ablation is performed on the actual response fragment and the control response fragment, and nodes caused by changes in oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data or slagging agent addition data are retained to obtain the effective parameter response state node sequence. The response time-delay manifold folding is performed according to the changes in flue gas composition data, furnace temperature data, furnace pressure data, and molten metal composition data. The phase constraint node sequence is obtained by preserving sequential phase nodes, suppressing reverse phase nodes, and skipping phase nodes through parameter-based phase constraints. The target quality attractor seed is determined from the historical furnaces where the final product composition data is qualified and the equipment operation status data is normal. The target quality backtracking growth is performed along the process parameter adjustment event and phase constraint node sequence. The reachable nodes are retained and the unreachable nodes are suppressed by the target quality path constraint to obtain the target quality backtracking node sequence. Parameter response transfer reachability constraints are applied to adjacent nodes in the target quality backtracking node sequence. Transfer reachable nodes are retained. Density contribution is calculated and parameter response density attractors are determined. Adjacent parameter response density attractors are connected to generate parameter response transfer edges and parameter response attraction path graphs. The path endpoints where the product endpoint component data is qualified and the equipment operating status data is normal are determined as the target quality attractors.

[0011] Optionally, the extraction of parameter response transition edges and the combination of candidate process parameter trajectory sets includes: Based on the process parameter status and furnace condition response status in the current furnace state sequence, the current furnace condition response status is matched with the parameter response density attractor to determine the current starting attractor; Using the current starting attractor as the path start point and the target mass attractor as the path end point, the parameter response attraction path is retrieved in the parameter response attraction path diagram, and then continuously filtered according to the normal equipment operating status, qualified product end component, and parameter response transfer edge to obtain qualified parameter response attraction paths. Based on the parameter response transfer edge in the qualified parameter response attraction path, the adjustment direction, adjustment magnitude and execution order of oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data and slagging agent addition data are extracted and combined to form a candidate process parameter trajectory set.

[0012] Optionally, generating the predicted attraction path includes: An improved MTGNN model is constructed, which includes a state encoding module, a transition graph construction module, a transition propagation module, and a path prediction module. The state coding module performs attractor state space reparameterization and feedback assimilation based on the current furnace state, the current starting attractor, the parameter response density attractor, and historical execution feedback records, mapping the current furnace state and the furnace response after execution to the current attractor state code. The transfer graph construction module performs event-driven construction based on process parameter adjustment events and parameter response transfer edges in the candidate process parameter trajectory set. It uses parameter response density attractors as graph nodes and parameter response transfer edges as event-triggered directed edges to generate candidate parameter response transfer graphs. The transfer propagation module performs path constraint propagation along the parameter response transfer edges defined by the candidate process parameter trajectory set, based on the current attractor state encoding and candidate parameter response transfer map. It also performs forward propagation along the direction from the current starting attractor to the target mass attractor, reverse traction along the direction from the target mass attractor to the current starting attractor, and performs counter-evidence blocking and time-delay alignment propagation along the paths where the product endpoint component data is unqualified or the equipment operating status data is abnormal, generating a bidirectional constraint transfer propagation sequence. The path prediction module performs attractor order preservation prediction based on the bidirectional constraint transfer propagation sequence, candidate process parameter trajectory set, and historical execution feedback records. It generates attractor transition residuals based on the matching results between the predicted target end parameter response density attractor and the corresponding target end parameter response density attractor in the candidate process parameter trajectory set. It then corrects the transfer propagation intensity according to the attractor transition residuals and historical execution feedback records to generate the predicted attraction path. The improved MTGNN model was trained using path matching error, target mass attractor arrival error, and attractor transition residual as optimization objectives. The parameters of the state encoding module, transition graph construction module, transition propagation module, and path prediction module were continuously optimized. Training was stopped when the total error decreased by less than 0.001 for 10 consecutive training rounds, and the improved MTGNN model was obtained after training.

[0013] Optionally, the execution of attractor state space reparameterization and feedback assimilation includes: Based on the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data in the current furnace condition, and by calculating the deviation between these data and the furnace condition response state corresponding to each parameter response density attractor, the attractor matching result is obtained. Using the parameter response density attractor as the state coordinate and the phase position and adjacent parameter response transition edges as the transition coordinate, an attractor state space is constructed. Based on the attractor matching results, the current furnace condition is projected onto the attractor state space to generate the current attractor state code. Based on the historical execution feedback record, read the parameter response density attractor corresponding to the furnace condition response after the execution of the previous parameter control segment and the parameter response density attractor at the target end of the previous parameter response transfer edge. According to the matching result of the two, correct the current attractor state code and the propagation strength of the corresponding parameter response transfer edge to obtain the current attractor state code after feedback assimilation.

[0014] Optionally, forming the target process parameter trajectory includes: Based on the order of attractors of parameter response density and the connection relationship of parameter response transition edges in the predicted attraction path and qualified parameter response attraction path, a matching verification is performed to obtain the path matching result; Based on the path matching results, the predicted attraction path is checked to see if it reaches the target mass attractor, and whether the corresponding equipment operating status data is normal and the product endpoint component data is qualified, thus obtaining the trajectory verification results; Based on the trajectory verification results, candidate process parameter trajectories that match the path, reach the target mass attractor, have normal equipment operation status, and have qualified product endpoint components are retained from the candidate process parameter trajectory set to form the target process parameter trajectory.

[0015] Optionally, sending the next parameter control segment or updating the target process parameter trajectory includes: Based on the current parameter control segment of the target process parameter trajectory, determine the adjustment direction, adjustment range, and execution sequence of oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, and slagging agent addition data, and send them to the smelting execution equipment; Collect furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data after the execution of the current parameter control segment. Determine the parameter response density attractor corresponding to the furnace condition response after execution and match it with the parameter response density attractor at the target end of the current parameter response transfer edge to generate an execution feedback record. When the execution feedback record shows a match, the next parameter control segment is sent. When the execution feedback record shows an inconsistency, the parameter response density attractor corresponding to the furnace condition response after execution is written into the current attractor state code, and the target process parameter trajectory is updated.

[0016] The beneficial effects of this invention are: This invention discloses a machine learning-based method for optimizing smelting process parameters. By collecting furnace operation data and uniformly processing process parameter adjustment events, furnace response status, equipment operating status, and final product composition data, the optimization of smelting process parameters no longer relies on single predicted values ​​or manual experience. The DENCLUE algorithm is used to form parameter response density attractors, parameter response transition edges, and parameter response attraction path graphs. This transforms effective parameter adjustment processes from historical furnace runs into searchable and verifiable response paths, improving the matching accuracy between current furnace conditions and historically effective process paths.

[0017] This invention improves the MTGNN model by jointly predicting the current furnace condition, the current starting attractor, the candidate process parameter trajectory set, and historical execution feedback records. This allows for the generation of predicted attraction paths while maintaining attractor order and continuous parameter response transition edges, reducing the impact of furnace response lag, equipment execution deviations, and abnormal historical paths on parameter recommendation results. By updating the target process parameter trajectory using execution feedback records, subsequent control segments can be promptly corrected after the current parameter control segment is executed, improving the final product composition qualification rate, smelting process stability, and equipment operational safety. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a machine learning-based method for optimizing smelting process parameters proposed in this invention. Figure 2 This is a data flow diagram of the DENCLUE algorithm, which is a machine learning-based method for optimizing smelting process parameters proposed in this invention. Figure 3 This is a schematic diagram of the structure of the improved MTGNN model for a machine learning-based method for optimizing smelting process parameters proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figure 1 , Figure 2 and Figure 3 A machine learning-based method for optimizing smelting process parameters includes: Collect furnace smelting operation data and preprocess the furnace smelting operation data to obtain the historical furnace smelting process status sequence and the current furnace smelting process status sequence. Extract process parameter adjustment events from historical furnace status sequences and construct parameter response status node sequences; The DENCLUE algorithm is executed based on the parameter response state node sequence to perform parameter action phase constraints, parameter response transition reachability constraints, and target mass path constraints. The density contribution is calculated to form parameter response density attractors, parameter response transition edges, parameter response attraction path graphs, and target mass attractors. Based on the current furnace state sequence, locate the current starting attractor, retrieve the qualified parameter response attraction path leading to the target mass attractor, extract the parameter response transfer edge and combine the candidate process parameter trajectory set; An improved MTGNN model is constructed based on the current furnace condition, the current starting attractor, and the set of candidate process parameter trajectories. Combined with parameter response transfer edge execution path constraint propagation, time delay alignment propagation, and attractor order preservation prediction, a predicted attraction path is generated. The predicted attraction path is matched and verified with the qualified parameter response attraction path, and the candidate process parameter trajectory set is screened by combining the target mass attractor arrival state, equipment operating state and product composition state to form the target process parameter trajectory. Send the current parameter control segment of the target process parameter trajectory, match the furnace condition response and termination attractor with the parameter response density attractor, and send the next parameter control segment or update the target process parameter trajectory according to the matching result.

[0021] In this embodiment, the furnace smelting operation data specifically includes furnace identifier, collection timestamp, oxygen supply flow rate data, electrode power data, fuel addition amount data, reducing agent addition amount data, slagging agent addition amount data, furnace temperature data, furnace pressure data, flue gas composition data, molten metal composition data, equipment operating status data, and product final composition data.

[0022] In this embodiment, the preprocessing of the furnace smelting operation data includes... The smelting operation data of each furnace is grouped according to the furnace number. Furnaces with completed production status and final product component data are classified as historical furnaces, while furnaces with ongoing smelting status and no final product component data are classified as current furnaces. The production status of a furnace batch refers to the furnace batch execution stage identifier recorded in the smelting production management system. The furnace batch execution stage identifier includes "smelting in progress" and "completed". "Smelting in progress" means that the furnace batch is still in the stage of feeding, oxygen supply, heating, reduction, slag formation or composition adjustment. "Completed" means that the furnace batch has completed smelting and formed the final product composition data. The smelting operation data of the historical furnaces and the current furnaces are sorted, time-aligned, missing data filled, anomaly removed, and normalized according to the collection timestamp, forming the historical furnace status sequence and the current furnace status sequence, respectively. The historical furnace state sequence and the current furnace state sequence are formed respectively, as follows: Sampling locations are established for the divided historical furnace batches and the current furnace batch at 5-second sampling intervals. Based on the collection timestamp, the oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, slagging agent addition data, furnace temperature data, furnace pressure data, flue gas composition data, molten metal composition data, and equipment operating status data within the same furnace batch are written into the corresponding sampling locations. When the duration of consecutive missing data items does not exceed 15 seconds, read the previous and next valid sample values ​​at the missing position. When one sampling position is consecutively missing, write the average of the previous and next valid sample values. When two or three sampling positions are consecutively missing, write the difference between the previous and next valid sample values ​​at equal intervals according to the collection timestamp order at the missing sampling position. When the duration of consecutive missing data exceeds 15 seconds, delete the corresponding missing sampling position. When the equipment operation status data is marked as sensor failure, actuator non-response, or communication interruption, delete the corresponding sampling position. When the numerical data exceeds three times the standard deviation of the average of the previous five valid sampling positions in the same furnace, delete the corresponding sampling position. Normalize the numerical data to the 0 to 1 range according to the minimum and maximum values ​​of the same data item in the historical furnace. Connect the sampling positions of the historical furnace in ascending order according to the collection timestamp, and write the product endpoint component data at the end of the historical furnace to form the historical furnace status sequence. Connect the sampling positions of the current furnace in ascending order according to the collection timestamp to form the current furnace status sequence.

[0023] In this embodiment, the construction of the parameter response state node sequence includes: Based on process parameter data and equipment operating status data from historical furnace status sequences, sampling locations where process parameters changed and equipment was in normal operating status were selected to determine process parameter adjustment events, where: Process parameter data refers to oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, and slagging agent addition data in the historical furnace state sequence; Equipment operation status data refers to the execution status indicators of oxygen supply equipment, electrode power supply equipment, fuel addition equipment, reducing agent addition equipment, and slagging agent addition equipment. The execution status indicators include normal execution, sensor failure, actuator non-response, and communication interruption. The process parameter adjustment event is identified as follows: According to the collection timestamp, the oxygen supply flow rate, electrode power, fuel addition, reducing agent addition, and slagging agent addition data of adjacent sampling locations are read. For the same process parameter data, the value of the previous sampling location is subtracted from the value of the next sampling location, the absolute value is taken, and then divided by the value of the previous sampling location to obtain the change ratio. When the value of the previous sampling location is 0 and the value of the next sampling location is greater than 0, the change ratio is recorded as 100%. When the values ​​of the previous and next sampling locations are both 0, the change ratio is recorded as 0. When the change ratio of any process parameter data reaches 2%, and the equipment operation status data of the same sampling location is normal execution, the corresponding sampling location is marked as a candidate adjustment location. Candidate adjustment positions with a collection time interval of no more than 10 seconds, the same adjustment object, and the same adjustment direction are merged into one process parameter adjustment event. The collection timestamp of the first candidate adjustment position is recorded as the start time of the process parameter adjustment event, and the collection timestamp of the last candidate adjustment position is recorded as the end time of the process parameter change. The process parameter adjustment event extracts the furnace condition status before adjustment, the process parameter change process, the response time, and the furnace condition response status, forming a parameter response status node, where: Furnace condition response status refers to the furnace condition change record within 60 seconds after the occurrence of a process parameter adjustment event, which is composed of furnace temperature data, furnace pressure data, flue gas composition data and molten metal composition data according to the collection timestamp. The furnace condition change record includes the furnace temperature value, furnace pressure value, flue gas composition value and molten metal composition value corresponding to each sampling location. The constituent parameter response state nodes are as follows: Based on the start time of the process parameter adjustment event, read the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data within 30 seconds before the start time. Calculate the average value of the same data item to form the furnace condition status before adjustment. Read the process parameter data from the start time to the end time of the process parameter change. Write the names of process parameters with a change rate of 2% into the adjustment object. Write the upward adjustment direction if the value of the next sampling position is greater than the value of the previous sampling position. Write the downward adjustment direction if the value of the next sampling position is less than the value of the previous sampling position. Write the change range by subtracting the value of the sampling position before the start time from the value of the end time of the process parameter change. The process parameter change process is composed of the adjustment object, adjustment direction, and change range. Read the furnace temperature data, furnace pressure data, flue gas composition data and molten metal composition data within 60 seconds after the start time to form the furnace condition response status. The difference between the time when the change ratio of any data item in the furnace condition response status relative to the furnace condition status before adjustment first reaches 1% and the start time is recorded as the response time. Write the furnace condition status before adjustment, the process parameter change process, the response time and the furnace condition response status into the same record to form a parameter response status node. Arrange the parameter response status nodes corresponding to the same process parameter adjustment event according to the collection timestamp, and construct a parameter response status node sequence, where: Construct the parameter response state node sequence as follows: The parameter response status nodes within the same historical furnace are grouped according to the furnace number identifier and arranged in ascending order according to the start time of the process parameter adjustment event. When the start time interval between adjacent parameter response status nodes is less than 10 seconds, and the adjustment object, adjustment direction, and furnace condition response status change direction are the same, the adjacent parameter response status nodes are merged into one parameter response status node. The sorted and merged parameter response status nodes are connected according to the collection timestamp to construct a parameter response status node sequence.

[0024] In this embodiment, the formation of the parameter response density attractor, parameter response transition edge, parameter response attraction path graph, and target mass attractor includes: The DENCLUE algorithm is executed based on the parameter response state node sequence. Historical segments matching the pre-adjustment furnace condition and not showing similar process parameter adjustment events are retrieved according to the process parameter adjustment event, the furnace condition status before adjustment, and the furnace condition response status. This yields the actual response segment and the reference response segment, where: A historical segment refers to a 90-second continuous record of furnace smelting operation data in the historical furnace status sequence. With a 5-second sampling interval, the historical segment includes 19 sampling locations from 30 seconds before the start time of the process parameter adjustment event to 60 seconds after the start time. Each sampling location includes the collection timestamp, oxygen supply flow data, electrode power data, fuel addition data, reducing agent addition data, slagging agent addition data, furnace temperature data, furnace pressure data, flue gas composition data, molten metal composition data, and equipment operating status data. The actual response fragment and the control response fragment are obtained as follows: Read the start time of the process parameter adjustment event, and determine the historical segment from 30 seconds before the start time to 60 seconds after the start time as the actual response segment. Select a continuous 90-second historical segment in the historical furnace state sequence by sliding in 5-second increments. Take the 7th sampling position as the comparison start position. Delete the historical segments in which the same process parameter adjustment event occurred within 60 seconds after the comparison start position. Calculate the average furnace temperature, average furnace pressure, average flue gas composition, and average molten metal composition of the actual response segment and the historical segment to be compared in the 6 sampling positions before the start position. Add the absolute values ​​of the differences between the four types of average values ​​corresponding to the two historical segments and divide by 4 to obtain the furnace condition deviation. Determine the historical segment with a furnace condition deviation of no more than 0.05 as the control response segment. When the number of control response segments exceeds 5, retain the 5 control response segments with the smallest furnace condition deviation. Counterfactual parameter response ablation is performed on the actual response fragment and the control response fragment. Nodes caused by changes in oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, or slagging agent addition data are retained to obtain the effective parameter response state node sequence, where: Counterfactual parameter response ablation refers to a comparative analysis process consisting of actual response segments and control response segments, used to distinguish between the response portion caused by process parameter adjustment events and the response portion caused by natural fluctuations in furnace conditions during furnace condition changes. The sequence of valid parameter response state nodes is obtained as follows: Furnace temperature, furnace pressure, flue gas composition, and molten metal composition data are read from the actual response segment and the control response segment within 60 seconds after the start time. The sampled value at the start time is taken as the value before the response, and the sampled value with the largest absolute difference from the value before the response within 60 seconds after the start time is taken as the value after the response. The response change amplitude is obtained by subtracting the response change amplitude from the value before the response. The counterfactual response difference is obtained by subtracting the response change amplitude of the control response segment from the response change amplitude of the actual response segment. When the counterfactual response difference is greater than 0, it is a positive difference; when it is less than 0, it is a negative difference. When the response change amplitude of the actual response segment is greater than 0, it is a positive furnace condition response; when it is less than 0, it is a negative furnace condition response. When the absolute value of the counterfactual response difference reaches 0.03 and the direction of the counterfactual response difference is consistent with the direction of the furnace condition response, the corresponding parameter response state node is retained. When the absolute value of the counterfactual response difference is less than 0.03, or the direction of the counterfactual response difference is opposite to the direction of the furnace condition response, the corresponding parameter response state node is deleted. The retained parameter response state nodes are connected according to the collection timestamp to obtain the effective parameter response state node sequence. Response time-delay manifold folding is performed based on the changes in flue gas composition data, furnace temperature data, furnace pressure data, and molten metal composition data. Phase constraints are applied using parameters to retain sequential phase nodes, suppress reverse phase nodes, and eliminate skipped phase nodes, resulting in a phase-constrained node sequence, where: Response time delay manifold folding refers to a node representation method that aligns parameter response state nodes with different response time lengths in different historical furnaces to the same response process coordinate system according to the initial response position, fast migration position, and component convergence position. The execution of response time-delayed manifold folding is specifically as follows: The furnace condition response status after the start time of the process parameter adjustment event is read from the effective parameter response status node sequence. The sampling position where the change ratio of flue gas composition data first reaches 1% is determined as the initial response position. The furnace temperature change rate and furnace pressure change rate between adjacent sampling positions are calculated respectively. The normalized value of the furnace temperature change rate is multiplied by 0.6 and the normalized value of the furnace pressure change rate is multiplied by 0.4 and then added to obtain the comprehensive change rate. The sampling position with the largest comprehensive change rate is determined as the fast migration position. The first sampling position where the change ratio of the molten metal composition data at three consecutive sampling positions is less than 0.5% is determined as the composition convergence position. The nodes between the start time and the initial response position are written into the initial response segment. The nodes between the initial response position and the fast migration position are written into the fast migration segment. The nodes between the fast migration position and the composition convergence position are written into the composition convergence segment. The folded parameter response status nodes are obtained. The parameter action phase constraint refers to the sequential restriction relationship between the parameter response state nodes in the initial response segment, the fast transition segment, and the component convergence segment. Sequential phase nodes are nodes in which the phase changes sequentially according to the initial response segment, the fast transition segment, and the component convergence segment. Reverse phase nodes are nodes in which the phase returns from the component convergence segment to the fast transition segment or from the fast transition segment to the initial response segment. Jump phase nodes are nodes in which the phase directly enters the component convergence segment from the initial response segment. The phase-constrained node sequence is obtained as follows: The initial density contribution of each parameter response state node after the response time-delay manifold folding is set to 1.00. Phase markers are written according to the response segment in which the parameter response state node is located. Nodes located in the initial response segment are marked with initial response phase markers, nodes located in the fast migration segment are marked with fast migration phase markers, and nodes located in the component convergence segment are marked with component convergence phase markers. When the phase markers of adjacent nodes change in the order of initial response phase, fast migration phase, and component convergence phase, the next node is marked as a sequential phase node, and the density contribution of the next node is kept at 1.00. When the phase marker of an adjacent node returns from the component convergence phase to the fast migration phase or from the fast migration phase to the initial response phase, the next node is marked as a reverse phase node, and the density contribution of the next node is adjusted to 0.20. When the phase marker of an adjacent node directly enters the component convergence phase from the initial response phase, the next node is marked as a jump phase node, and the density contribution of the next node is adjusted to 0.40. The parameter response state nodes with completed phase markers and density contribution assignments are connected according to the acquisition timestamp to obtain the phase-constrained node sequence. The target quality attractor seed is determined from historical furnace batches where the final product composition data is qualified and the equipment operating status data is normal. Target quality backtracking growth is then performed along the process parameter adjustment event and phase constraint node sequence. Reachable nodes are retained and unreachable nodes are suppressed through target quality path constraints, resulting in the target quality backtracking node sequence, where: The determination of the target mass attractor seed is as follows: For historical furnaces where the product endpoint component data falls between the lower and upper limits of the product endpoint component recorded in the current furnace production task record, and the equipment operation status data shows normal execution, the last parameter response status node of each eligible historical furnace is selected as a candidate endpoint node. When the number of candidate endpoint nodes is 1, the candidate endpoint node is determined as the seed of the target mass attractor. When the number of candidate endpoint nodes is greater than 1, the first ungrouped candidate endpoint node is selected as the endpoint group baseline node according to the ascending order of the collection timestamp. The node distance between the remaining ungrouped candidate endpoint nodes and the endpoint group baseline node is calculated. The node distance is the average of the normalized absolute values ​​of the differences between the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data. Candidate endpoint nodes with a node distance of no more than 0.04 are written into the same endpoint node group. After completing one endpoint node group, the candidate endpoint node with the earliest collection timestamp among the remaining ungrouped candidate endpoint nodes is selected as the new endpoint group reference node until all candidate endpoint nodes are grouped. The endpoint node group with the largest number of nodes is selected, and the average values ​​of the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data in the endpoint node group are used to form the target mass attractor seed. Target quality backtracking growth refers to a reverse hierarchical connection relationship consisting of parameter response state nodes prior to the target quality attractor seed, with the target quality attractor seed as the backtracking starting point and process parameter adjustment events as the reverse connection basis. The execution of target quality backtracking growth is as follows: Using the target mass attractor seed as the layer 0 node, the parameter response status nodes within the same historical furnace batch whose acquisition timestamps are earlier than the layer 0 node and are directly connected to the layer 0 node through process parameter adjustment events are used as the layer 1 backtracking nodes. The parameter response status nodes whose acquisition timestamps are earlier than the layer 1 backtracking nodes and are directly connected to the layer 1 backtracking nodes through process parameter adjustment events are used as the layer 2 backtracking nodes. The parameter response status nodes whose acquisition timestamps are earlier than the layer 2 backtracking nodes and are directly connected to the layer 2 backtracking nodes through process parameter adjustment events are used as the layer 3 backtracking nodes. The parameter response status nodes are connected in the order of layer 3 backtracking nodes, layer 2 backtracking nodes, layer 1 backtracking nodes and target mass attractor seed to obtain the target mass backtracking candidate sequence. The target quality path constraint refers to the restriction relationship between parameter response state nodes and the target quality attractor seed within the connection range of three process parameter adjustment events. The target quality path constraint includes reachable nodes and unreachable nodes. Reachable nodes are parameter response state nodes that can enter the target quality backtracking candidate sequence, and unreachable nodes are parameter response state nodes that cannot enter the target quality backtracking candidate sequence. The target quality backtracking node sequence is obtained as follows: Read the parameter response state nodes and their corresponding density contributions in the phase constraint node sequence, and determine whether each parameter response state node can enter the target quality backtracking candidate sequence within the connection range of 3 process parameter adjustment events. Parameter response state nodes that can enter the target quality backtracking candidate sequence are marked as reachable nodes, and parameter response state nodes that cannot enter the target quality backtracking candidate sequence are marked as unreachable nodes. The density contribution of reachable nodes remains unchanged, and the density contribution of unreachable nodes is multiplied by 0.30. The density contribution of parameter response state nodes whose product endpoint component data is unqualified or whose equipment operating status data is not executed normally is multiplied by 0.10. Connect the reachable nodes after density contribution adjustment according to the collection timestamp to obtain the target quality backtracking node sequence. For adjacent nodes in the target quality backtracking node sequence, parameter response transition reachability constraints are applied, and transition reachable nodes are retained. Density contributions are calculated, and parameter response density attractors are determined. Adjacent parameter response density attractors are connected to generate parameter response transition edges and parameter response attraction path graphs. The path endpoints where the product endpoint component data is qualified and the equipment operating status data is normal are determined as the target quality attractors, where: Parameter response transfer reachability constraints refer to the transfer restriction relationship between adjacent parameter response state nodes in the target quality backtracking node sequence. The transfer restriction relationship is determined by process parameter adjustment events and equipment operating status data. Parameter response transfer reachability constraints include transferable nodes and transfer-unreachable nodes. The execution of the parameter response transition reachability constraint is as follows: Read two parameter response status nodes with adjacent timestamps in the target quality backtracking node sequence, check whether there is a process parameter adjustment event between the two adjacent nodes where the change ratio of oxygen supply flow data, electrode power data, fuel addition data, reducing agent addition data, or slagging agent addition data reaches 2%, and check whether the corresponding equipment operation status data is executed normally. When both conditions are met, mark the latter parameter response status node as a transferable node and keep the density contribution unchanged. When either condition is not met, mark the latter parameter response status node as a transferable unreachable node and multiply the density contribution by 0.20. The density contribution is calculated and the parametric response density attractor is determined as follows: Read each parameter response state node in the target quality backtracking node sequence. Use the average of the normalized absolute values ​​of the differences in furnace temperature data, furnace pressure data, flue gas composition data, molten metal composition data, and process parameter variation amplitudes as the node distance. When the node distance between two parameter response state nodes does not exceed 0.08, the density contribution of one parameter response state node is added to the other. When the node distance exceeds 0.08, no density contribution is added. After adding the density contributions of all nodes, parameter response state nodes whose density contribution is greater than 1.2 times the average density contribution of the same historical furnace are identified as high-density nodes. For each node, the first unmerged high-density node is selected as the attractor reference node in descending order of density contribution. The node distance between the remaining unmerged high-density nodes and the attractor reference node is calculated. High-density nodes with a node distance of no more than 0.05 are written into the same attractor group. After completing one attractor group, the high-density node with the largest density contribution among the remaining unmerged high-density nodes is selected as the new attractor reference node until all high-density nodes are grouped. The average value of the furnace temperature data, furnace pressure data, flue gas composition data, molten metal composition data and process parameter variation amplitude in each attractor group is determined as the parameter response density attractor. The generation of parameter response transition edges and parameter response attraction path graphs is as follows: Read adjacent parameter response density attractors within the same historical furnace according to the acquisition timestamp, take the previous parameter response density attractor as the starting point of the transfer edge and the next parameter response density attractor as the ending point of the transfer edge, read the process parameter adjustment event between the two parameter response density attractors, write the adjustment object, adjustment direction, change range, start time, end time and equipment operating status data into the same parameter response transfer edge, take the parameter response density attractor as a graph node and the parameter response transfer edge as a directed connection relationship, and connect them according to the acquisition timestamp to form a parameter response attraction path graph; The determination of the target mass attractor is as follows: Read the path endpoints in the parameter response attraction path diagram, filter the path endpoints whose product endpoint component data is between the lower limit and upper limit of the product endpoint component recorded in the current furnace production task record, and whose equipment operation status data is normal execution. Read the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data of the path endpoints and the target mass attractor seed. Calculate the absolute value of the normalized difference in furnace temperature, the absolute value of the normalized difference in furnace pressure, the average value of the normalized difference in flue gas composition, and the average value of the normalized difference in molten metal composition. Add the four values ​​together and divide by 4 to obtain the node distance between the path endpoint and the target mass attractor seed. Path endpoints with a node distance of no more than 0.05 are identified as the target mass attractors.

[0025] In this embodiment, the step of extracting parameter response transition edges and combining candidate process parameter trajectory sets includes: Based on the process parameter states and furnace condition response states in the current furnace state sequence, the current furnace condition response state is matched with the parameter response density attractor to determine the current initial attractor, where: The current starting attractor is determined as follows: Read the furnace temperature, furnace pressure, flue gas composition, and molten metal composition data within the latest 60 seconds of the current furnace state sequence, calculate the average value of each, and form the current furnace condition response state. Read the furnace temperature, furnace pressure, flue gas composition, and molten metal composition data of each parameter response density attractor, calculate the average value of the normalized absolute value of the difference between the current furnace condition response state and each parameter response density attractor, obtain the node distance, and determine the parameter response density attractor with the smallest node distance not exceeding 0.08 as the current starting attractor. Using the current starting attractor as the path origin and the target mass attractor as the path endpoint, a parameter response attraction path is retrieved from the parameter response attraction path graph. Qualified parameter response attraction paths are then obtained by continuously filtering based on whether the equipment is operating normally, the final product composition is qualified, and the parameter response transition edge. Among these: Retrieving the parameter response attraction path in the parameter response attraction path graph is specifically as follows: Using the current starting attractor as the path start point and the target mass attractor as the path end point, the parameter response transfer edge is searched edge by edge along the direction of the parameter response attraction path graph. When the end point of the previous parameter response transfer edge is equal to the start point of the next parameter response transfer edge, the parameter response transfer edge is considered to be continuous. When the search path reaches the target mass attractor, the parameter response attraction path is obtained. The number of parameter response transfer edges in each parameter response attraction path does not exceed 6. The qualified parameter response attraction path is obtained as follows: Read the parameter response transfer edge, equipment operation status data and product endpoint component data in the parameter response attraction path, retain the parameter response attraction path with continuous parameter response transfer edge, normal equipment operation status data, and product endpoint component data located between the lower limit value and upper limit value of product endpoint component recorded in the current furnace production task sheet, and obtain a qualified parameter response attraction path. Based on the parameter response transfer edges in the qualified parameter response attraction path, the adjustment direction, adjustment magnitude, and execution order of oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, and slagging agent addition data are extracted and combined to form a candidate process parameter trajectory set, wherein: The candidate process parameter trajectory set is formed by combining the following: Read the parameter response transfer edges in the qualified parameter response attraction path, extract the adjustment object, adjustment direction, and adjustment magnitude from the process parameter changes recorded in the parameter response transfer edges, and determine the execution order according to the ascending order of the acquisition timestamps of the parameter response transfer edges. The adjustment object is the oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, or slagging agent addition data that have undergone numerical changes. The adjustment direction is either upward or downward adjustment. The adjustment magnitude is the process parameter value at the end position of the parameter response transfer edge minus the process parameter value at the sampling position before the start position of the parameter response transfer edge. Connect the adjustment object, adjustment direction, adjustment magnitude, and execution order within the same qualified parameter response attraction path to form a candidate process parameter trajectory. Write all candidate process parameter trajectories into the same set to form a candidate process parameter trajectory set.

[0026] In this embodiment, generating the predicted attraction path includes: An improved MTGNN model is constructed, comprising a state encoding module, a transition graph construction module, a transition propagation module, and a path prediction module, wherein: The improved MTGNN model is constructed as follows: The traditional MTGNN model is structurally optimized by adding attractor state space reparameterization processing to the input mapping layer and temporal convolution layer to obtain a state coding module. This module converts the current furnace condition state into the current attractor state code, reduces the deviation between the current furnace condition state and the historical parameter response path, and improves the matching stability of the current initial attractor. Based on the graph learning layer, a parameter response transition edge event-driven construction process is added to obtain the transition graph construction module. The graph connection relationship is determined by the parameter response density attractor and the parameter response transition edge, which reduces the interference of irrelevant variables and improves the accuracy of the expression of process parameter adjustment path. Based on the graph convolutional propagation layer and the temporal convolutional layer, path constraint propagation, time delay alignment propagation and target mass attractor reverse traction processing are added to obtain the transfer propagation module, which enables state propagation to proceed along the parameter response transfer edge defined by the candidate process parameter trajectory set, thereby reducing the impact of path jumps and response misalignments on the prediction results. Based on the prediction output layer, attractor sequence preservation prediction and attractor transition residual correction processing are added to obtain the path prediction module, which transforms the output result from time series prediction value to predicted attraction path, and corrects the propagation result according to attractor transition residual to improve the consistency between the predicted attraction path and the qualified parameter response attraction path. The improved MTGNN model is obtained by combining the state encoding module, transition graph construction module, transition propagation module and path prediction module. This model retains the ability to process multivariate time series data and integrates parameter response density attractors, parameter response transition edges and target mass attractors to improve the accuracy and feasibility of predicting smelting process parameter trajectories. The state coding module, based on the current furnace condition state, the current starting attractor, the parameter response density attractor, and historical execution feedback records, performs attractor state space reparameterization and feedback assimilation, mapping the current furnace condition state and the furnace condition response after execution to the current attractor state code, wherein: The status coding module includes: Attractor Matcher: Calculates the attractor distance between the current furnace condition and the parameter response density attractor; State space projector: Writes the current furnace condition state into the attractor state space position corresponding to the matched attractor or the nearest neighbor attractor; Feedback Assimilation: Adjusts the propagation strength of parameter response transition edges based on historical execution feedback records; Encoding buffer: Stores and outputs the current attractor state encoding; The transition graph construction module, based on process parameter adjustment events and parameter response transition edges in the candidate process parameter trajectory set, performs event-driven construction of transition edges, using parameter response density attractors as graph nodes and parameter response transition edges as event-triggered directed edges to generate a candidate parameter response transition graph, where: The transition graph construction module includes: Trajectory Reader: Reads candidate process parameter trajectories from the candidate process parameter trajectory set; Transfer edge register: stores the starting parameter response density attractor, ending parameter response density attractor, adjustment object, adjustment direction, adjustment magnitude, and execution order of parameter response transfer edges; Event-triggered connector: Adjust the event settings according to process parameters to establish event-triggered directed edges; Candidate graph cache: Stores and outputs candidate parameter response transition graphs; The event-driven construction of the transition edge is performed as follows: Read the candidate process parameter trajectory and parameter response transition edge in the candidate process parameter trajectory set. Take the starting parameter response density attractor and the ending parameter response density attractor of the parameter response transition edge as nodes to be connected. Take the adjustment object, adjustment direction, adjustment magnitude and execution order recorded by the parameter response transition edge as event trigger information. When there is corresponding event trigger information in the candidate process parameter trajectory, establish a directed connection from the starting parameter response density attractor to the ending parameter response density attractor. Generate the candidate parameter response transition graph, specifically as follows: Write the parameter response density attractors involved in the candidate process parameter trajectory set into the graph nodes, write the directed connections corresponding to the event triggering information into the event triggering directed edges, and write the adjustment object, adjustment direction, adjustment magnitude, execution order, equipment operating status data and propagation intensity into the event triggering directed edges. Connect the graph nodes and event triggering directed edges according to the execution order of the candidate process parameter trajectories to generate the candidate parameter response transfer graph. The transfer propagation module, based on the current attractor state encoding and candidate parameter response transfer map, performs path constraint propagation along the parameter response transfer edges defined by the candidate process parameter trajectory set, performs forward propagation along the direction from the current starting attractor to the target mass attractor, performs reverse traction along the direction from the target mass attractor to the current starting attractor, and performs counter-evidence blocking and time-delay alignment propagation along paths where the product endpoint component data is unqualified or the equipment operating status data is abnormal, generating a bidirectional constraint transfer propagation sequence, wherein: The transfer and propagation module includes: Path constraint propagator: limits the propagation process to be executed along the parameter response transfer edges in the candidate process parameter trajectory set; Forward propagator: Propagates from the current starting attractor to the target mass attractor, and calculates the forward propagation strength; Reverse traction device: Backtracking from the target mass attractor to the current starting attractor, calculate the reverse traction strength; Counter-evidence blocking and time delay aligner: Blocks paths where the final component data of the product is unqualified or the equipment operating status data is abnormal, and corrects parameter response transfer edges with a response time difference exceeding 15 seconds; Propagation sequence buffer: Stores and outputs bidirectional constraint transition propagation sequences; The execution of path constraint propagation is as follows: Read the current attractor state code and candidate parameter response transition graph after feedback assimilation, write the current attractor state code after feedback assimilation into the graph node corresponding to the current starting attractor, and propagate along the parameter response transition edges existing in the candidate process parameter trajectory set. Do not propagate along directed connections outside the candidate process parameter trajectory set. Continue propagation when the endpoint of the previous parameter response transition edge is equal to the starting point of the next parameter response transition edge, and stop propagating the corresponding path when the endpoint of the previous parameter response transition edge is not equal to the starting point of the next parameter response transition edge. The execution of positive propagation is as follows: Starting from the current initial attractor, propagate towards the target mass attractor according to the execution order of the parameter response transfer edges. For each parameter response transfer edge passed, multiply the current propagation intensity by the propagation intensity of the parameter response transfer edge to obtain the forward propagation intensity of the endpoint parameter response density attractor. When the propagation reaches the target mass attractor, record the parameter response density attractor, parameter response transfer edge passed, and forward propagation intensity. The execution of reverse traction is as follows: Starting from the target mass attractor, backtrack along the opposite direction of the parameter response transfer edge towards the current starting attractor. The traction strength of the target mass attractor is set to 1.00. Each time a reverse parameter response transfer edge is passed, the traction strength is multiplied by 0.90 and written to the previous parameter response density attractor. When backtracking to the current starting attractor, the parameter response density attractor, parameter response transfer edge and reverse traction strength are recorded. The execution of the disproving blocking and time-delayed aligned propagation is as follows: Read the parameter response transfer edges passed through by the forward propagation and reverse traction. When the product endpoint component data corresponding to the parameter response transfer edge is not located between the lower limit and upper limit of the product endpoint component recorded in the current furnace production task sheet, or when the equipment operation status data is not executed normally, set the propagation intensity of the parameter response transfer edge to 0.00 and stop the corresponding path propagation. Read the response time in the parameter response status nodes corresponding to the start and end points of the parameter response transfer edge, calculate the average value as the edge response time. When the number of parameter response transfer edges that have completed propagation on the same propagation path is 0, do not perform time delay alignment. When the number of parameter response transfer edges that have completed propagation on the same propagation path is greater than 0, calculate the average response time of the propagated edges. When the absolute value of the difference between the edge response time and the average response time of the propagated edges exceeds 15 seconds, multiply the propagation intensity of the current parameter response transfer edge by 0.80, and adjust the propagation time position of the current parameter response transfer edge according to the average response time of the propagated edges. Write the node order, edge order, forward propagation intensity, reverse traction intensity, and propagation time position into the same sequence to generate a bidirectional constraint transfer propagation sequence. The path prediction module, based on the bidirectional constraint transfer propagation sequence, the candidate process parameter trajectory set, and historical execution feedback records, performs attractor sequence preservation prediction. It generates attractor transition residuals based on the matching results between the predicted target-end parameter response density attractor and the corresponding target-end parameter response density attractor in the candidate process parameter trajectory set. Then, it corrects the transfer propagation strength according to the attractor transition residuals and historical execution feedback records to generate the predicted attraction path, where: The path prediction module includes: Sequence checker: Removes propagation sequences where node jumps or parameter response transitions are discontinuous; Predicted Passage Strength Calculator: Calculates predicted passage strength based on forward propagation strength and reverse traction strength; Attractor transition residual calculator: Calculates the attractor transition residual between the predicted target end parameter response density attractor and the corresponding target end parameter response density attractor; Path output buffer: Stores and outputs predicted attraction paths; Perform attractor order preservation prediction, specifically as follows: Read the node order, edge order, forward propagation intensity, and reverse traction intensity in the bidirectional constraint transfer propagation sequence, and read the parameter response transfer edge execution order in the candidate process parameter trajectory set. When the node order is consistent with the parameter response transfer edge execution order and adjacent parameter response transfer edges are continuous, retain the corresponding propagation sequence. When the node order jumps or adjacent parameter response transfer edges are not continuous, delete the corresponding propagation sequence. Add the forward propagation intensity and reverse traction intensity of the same parameter response transfer edge in the retained propagation sequence and divide by 2 to obtain the predicted pass intensity of the parameter response transfer edge. The generation of attractor transition residuals is as follows: Read each parameter response transition edge in the retained propagation sequence, and use the endpoint parameter response density attractor of the parameter response transition edge as the predicted target parameter response density attractor. Read the parameter response transition edges with the same execution order in the candidate process parameter trajectory set, and use the endpoint parameter response density attractor of the parameter response transition edge as the corresponding target parameter response density attractor. Calculate the average of the normalized absolute values ​​of the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data between the two to obtain the attractor transition residual, and write the attractor transition residual into the corresponding parameter response transition edge. Generate predicted attraction paths, specifically: Read the attractor transition residuals and historical execution feedback records. When the attractor transition residuals do not exceed 0.05, keep the predicted pass intensity of the corresponding parameter response transfer edge unchanged. When the attractor transition residuals exceed 0.05, multiply the predicted pass intensity of the corresponding parameter response transfer edge by 0.80. When the parameter response density attractor corresponding to the furnace condition response after execution in the historical execution feedback record is inconsistent with the parameter response density attractor at the target end of the previous parameter response transfer edge, multiply the predicted pass intensity of the corresponding parameter response transfer edge by 0.70 again. Calculate the average value of the corrected predicted pass intensity of all parameter response transfer edges in each retained propagation sequence. Select the propagation sequence with the highest average value that can continuously reach the target mass attractor from the current starting attractor. When the average values ​​are the same, select the propagation sequence with the smallest sum of attractor transition residuals. Connect the parameter response density attractors and parameter response transfer edges in the selected propagation sequence according to the execution order to generate a predicted attraction path. The improved MTGNN model was trained using path matching error, target mass attractor arrival error, and attractor transition residual as optimization objectives. The parameters of the state encoding module, transition graph construction module, transition propagation module, and path prediction module were continuously optimized. Training was stopped when the total error decreased by less than 0.001 over 10 consecutive training rounds, resulting in the completed improved MTGNN model. The improved MTGNN model is trained as follows: The current furnace condition, current starting attractor, parameter response density attractor, candidate process parameter trajectory set, historical execution feedback record, and qualified parameter response attraction path are read from the historical furnace state sequence. The current furnace condition, candidate process parameter trajectory set, and historical execution feedback record within the same historical furnace are used as training inputs, and the qualified parameter response attraction path and target quality attractor are used as supervision targets. The training inputs are sequentially fed into the state encoding module, transition graph construction module, transition propagation module, and path prediction module, and the predicted attraction path is output. The node order and parameter response transfer edge order in the predicted attraction path are compared item by item with the qualified parameter response attraction path. Matches at the same position are recorded as 0, and inconsistencies as 1. The average of all comparison results is calculated to obtain the path matching error. It is then determined whether the endpoint of the predicted attraction path is the target mass attractor; reaching the target mass attractor is recorded as 0, and not reaching it as 1, thus obtaining the target mass attractor arrival error. The predicted target-end parameter response density attractor for each parameter response transfer edge in the predicted attraction path is read, along with the target-end parameter response density attractors of the same execution order in the candidate process parameter trajectory set. The average of the normalized absolute values ​​of the differences between the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data is calculated to obtain the attractor transition residual. The path matching error is multiplied by 0.40, the target mass attractor arrival error is multiplied by 0.30, and the attractor transition residual is multiplied by 0.30, and then summed to obtain the total training error function. The parameters of the state encoding module, transition graph construction module, transition propagation module, and path prediction module are updated according to the gradient descent direction. When the total error function decreases by less than 0.001 for 10 consecutive rounds of training, training is stopped, and the improved MTGNN model is obtained after training.

[0027] In this embodiment, the process of performing attractor state space reparameterization and feedback assimilation includes: Based on the current furnace temperature, pressure, flue gas composition, and molten metal composition data, and by calculating the deviation between these data and the corresponding furnace response states of the attractors for each parameter's response density, the attractor matching results are obtained. The attractor matching results are as follows: Read the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data from the current furnace condition status, and read the furnace condition response status of each parameter response density attractor. Calculate the average of the normalized absolute values ​​of the differences between the current furnace condition status and each parameter response density attractor to obtain the attractor distance. When the minimum attractor distance does not exceed 0.08, the corresponding parameter response density attractor is determined as the matched attractor. When the minimum attractor distance exceeds 0.08, the corresponding parameter response density attractor is determined as the nearest neighbor attractor. The attractor distance is used as the deviation of the current furnace condition status. The attractor matching result is composed of the matched attractor or nearest neighbor attractor, the attractor distance, and the current furnace condition status. Using the parametric response density attractor as the state coordinate and the phase position and adjacent parametric response transition edges as the transition coordinate, an attractor state space is constructed. Based on the attractor matching results, the current furnace condition is projected onto the attractor state space to generate the current attractor state code, where: Constructing the attractor state space, specifically: Read all parameter response density attractors, use each parameter response density attractor as a state coordinate point, read the phase mark of the parameter response density attractor and the adjacent parameter response transition edge, use the phase mark as the phase position, and use the start point, end point, adjustment object, adjustment direction and adjustment magnitude of the adjacent parameter response transition edge as the transition coordinates. Connect the state coordinate point, phase position and transition coordinates according to the correspondence between the parameter response density attractor, phase mark and adjacent parameter response transition edge to construct the attractor state space; The current attractor state encoding is generated as follows: Read the attractor matching results and attractor state space. First, create a blank coded record. Write the attractor identifier of the matching attractor or nearest neighbor attractor into the first item of the coded record. When the attractor matching result is a matching attractor, write 1.00 into the second item; when the attractor matching result is a nearest neighbor attractor, write 0.00 into the second item. Write the attractor distance into the third item. Then, calculate the normalized values ​​of the furnace temperature, furnace pressure, flue gas composition, and molten metal composition in the current furnace condition. Write items 4 through 7, write the phase position of the matched attractor or nearest neighbor attractor in the attractor state space into item 8, write the initial response phase into 1.00, write the fast transition phase into 2.00, write the component convergence phase into 3.00, write the adjustment object, adjustment direction, and adjustment magnitude in the adjacent parameter response transition edge into items 9 through 11 in sequence, write 1.00 when the adjustment direction is upward, and write -1.00 when the adjustment direction is downward, and connect items 1 through 11 in the writing order to generate the current attractor state code; Based on historical execution feedback records, the parameter response density attractor corresponding to the furnace condition response after the execution of the previous parameter control segment and the parameter response density attractor at the target end of the previous parameter response transition edge are read. The current attractor state code and the propagation strength of the corresponding parameter response transition edge are corrected according to their matching results to obtain the current attractor state code after feedback assimilation, where: Historical execution feedback records refer to the records formed after the execution of the previous parameter control segment, including the previous parameter control segment, the furnace condition response status after the execution of the previous parameter control segment, the parameter response density attractor corresponding to the furnace condition response status after execution, the parameter response density attractor of the target end of the previous parameter response transfer edge, and the matching result of the two. The current attractor state encoding after feedback assimilation is obtained as follows: When the historical execution feedback record is empty, keep the first to eleventh items of the current attractor state code unchanged, set the propagation strength of the parameter response transfer edge connected to the current attractor state code to 1.00, and use the current attractor state code as the current attractor state code after feedback assimilation. When the historical execution feedback record is not empty, read the parameter response density attractor corresponding to the furnace condition response state after execution and the parameter response density attractor at the target end of the previous parameter response transfer edge. When the two are consistent, keep the first to eleventh items of the current attractor state code unchanged, multiply the propagation strength of the previous parameter response transfer edge by 1.10, and write the corrected propagation strength into the current attractor state code. The 12th item of the attractor state encoding is used to obtain the current attractor state encoding after feedback assimilation. When the two are inconsistent, the 1st item of the current attractor state encoding is updated to the attractor identifier of the parameter response density attractor corresponding to the furnace condition response state after execution, the 2nd item is updated to 0.00, the 3rd item is updated to the attractor distance between the furnace condition response state after execution and the parameter response density attractor at the target end of the previous parameter response transfer edge, the 8th item is updated to the phase position of the parameter response density attractor corresponding to the furnace condition response state after execution, the propagation intensity of the previous parameter response transfer edge is multiplied by 0.70 and written into the 12th item to obtain the current attractor state encoding after feedback assimilation.

[0028] In this embodiment, forming the target process parameter trajectory includes: Based on the order of attractors by parameter response density and the connection relationship of parameter response transition edges in the predicted attraction path and the qualified parameter response attraction path, a matching verification is performed to obtain the path matching result, where: The matching verification is performed as follows: Read the order of the parameter response density attractors in the predicted attraction path and the qualified parameter response attraction path, and compare the attractor identifiers according to their positions. Read the parameter response transfer edges at the same position in the two paths, and compare the starting parameter response density attractor, ending parameter response density attractor, adjustment object, adjustment direction, and adjustment magnitude of the parameter response transfer edges. When the attractor identifiers are consistent, the starting and ending points of the parameter response transfer edges are consistent, the adjustment object and adjustment direction are consistent, and the difference in adjustment magnitude does not exceed 5%, it is recorded as a path match. If any comparison item does not meet the requirements, it is recorded as a path mismatch, and the path matching result is obtained. Based on the path matching results, the predicted attraction path is verified to ensure it reaches the target mass attractor, and the corresponding equipment operating status data and product endpoint component data are normal. This yields the trajectory verification results, including: The trajectory verification results are as follows: Read the path matching result, the predicted attraction path endpoint, the equipment operation status data corresponding to the parameter response transfer edge in the predicted attraction path, and the product endpoint component data corresponding to the path endpoint. If the path matching result is a path match, the predicted attraction path endpoint is the target mass attractor, the equipment operation status data are all in normal execution, and the product endpoint component data is between the lower limit and upper limit of the product endpoint component recorded in the current furnace production task sheet, the trajectory verification is considered passed. If any condition is not met, the trajectory verification is considered failed, and the trajectory verification result is obtained. Based on the trajectory verification results, candidate process parameter trajectories that match the path, reach the target mass attractor, have normal equipment operation status, and whose final product components are qualified are retained from the candidate process parameter trajectory set to form the target process parameter trajectory, where: The target process parameter trajectory is formed as follows: Read the trajectory verification result corresponding to each candidate process parameter trajectory in the candidate process parameter trajectory set, and retain the candidate process parameter trajectories that pass the trajectory verification. When the number of retained candidate process parameter trajectories is 1, the retained candidate process parameter trajectory is determined as the target process parameter trajectory. When the number of retained candidate process parameter trajectories is greater than 1, the candidate process parameter trajectory with the highest predicted average intensity is selected as the target process parameter trajectory. When the predicted average intensity is the same, the candidate process parameter trajectory with the smallest sum of attractor transition residuals is selected as the target process parameter trajectory.

[0029] In this embodiment, sending the next parameter control segment or updating the target process parameter trajectory includes: Based on the current parameter control segment of the target process parameter trajectory, the adjustment direction, adjustment range, and execution sequence of oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, and slagging agent addition data are determined and sent to the smelting execution equipment, wherein: The adjustment direction, adjustment range, and execution sequence for oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, and slagging agent addition data are determined as follows: The first unsent parameter response transfer edge in the target process parameter trajectory is read as the current parameter control segment. The adjustment object, adjustment direction, and adjustment magnitude are read from the process parameter changes recorded in the current parameter control segment. The adjustment object is oxygen supply flow data, electrode power data, fuel addition data, reducing agent addition data, or slagging agent addition data. The adjustment direction is upward or downward. The adjustment magnitude is the process parameter value at the end position of the current parameter control segment minus the process parameter value at the sampling position before the start position of the current parameter control segment. The execution order is determined according to the acquisition timestamp order of the current parameter control segment in the target process parameter trajectory. The adjustment object, adjustment direction, adjustment magnitude, and execution order are sent to the corresponding smelting execution equipment. Collect furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data after the execution of the current parameter control segment. Determine the parameter response density attractor corresponding to the furnace condition response after execution, and match it with the parameter response density attractor at the target end of the current parameter response transfer edge to generate an execution feedback record, in which: The parameter response density attractor corresponding to the furnace condition response after execution is determined as follows: The time when the current parameter control segment sends a successful acknowledgment is taken as the start time of the post-execution response. Starting from the 5th second after the start time of the post-execution response, furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data are collected every 5 seconds. This collection continues until the 60th second, resulting in 12 post-execution sampling positions. The average values ​​of the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data in the 12 post-execution sampling positions are calculated to form the post-execution furnace condition response. The average value of the normalized absolute difference between the post-execution furnace condition response and each parameter response density attractor is calculated to obtain the post-execution attractor distance. The parameter response density attractor with the smallest post-execution attractor distance is determined as the parameter response density attractor corresponding to the post-execution furnace condition response. Generate execution feedback records, specifically: Read the current parameter control segment, the furnace condition response after execution, the parameter response density attractor corresponding to the furnace condition response after execution, and the parameter response density attractor at the target end of the current parameter response transfer edge. When the parameter response density attractor corresponding to the furnace condition response after execution is consistent with the parameter response density attractor at the target end of the current parameter response transfer edge, it is recorded as a consistent match. When they are inconsistent, it is recorded as an inconsistent match. Write the current parameter control segment, the furnace condition response after execution, the parameter response density attractor corresponding to the furnace condition response after execution, the parameter response density attractor at the target end of the current parameter response transfer edge, and the matching result into the same record to generate an execution feedback record. When the execution feedback record shows a match, the next parameter control segment is sent. When the execution feedback record shows an inconsistency, the parameter response density attractor corresponding to the furnace condition response after execution is written into the current attractor state code, and the target process parameter trajectory is updated, where: Update the target process parameter trajectory, specifically; Read the execution feedback record. When the execution feedback record matches, determine the first parameter response transfer edge after the current parameter control segment in the target process parameter trajectory as the next parameter control segment. Read the adjustment object, adjustment direction, adjustment range and execution sequence in the next parameter control segment and send them to the smelting execution equipment. When the execution feedback record shows a mismatch, the parameter response density attractor corresponding to the furnace condition response after execution is written into the current attractor state encoding. The transfer propagation module and path prediction module are re-executed to obtain the updated predicted attraction path. Based on the updated predicted attraction path, the candidate process parameter trajectory set is re-selected to form the updated target process parameter trajectory. The first parameter response transfer edge in the updated target process parameter trajectory, starting from the parameter response density attractor corresponding to the furnace condition response after execution, is read and determined as the new current parameter control segment. The adjustment object, adjustment direction, adjustment magnitude, and execution order in the new current parameter control segment are read and sent to the smelting execution equipment. After the transmission is completed, the furnace condition response after execution is collected again and a new execution feedback record is generated until the target process parameter trajectory reaches the target mass attractor.

[0030] Example 1: To verify the feasibility of this invention in practice, it was applied to a metal smelting electric furnace control system. The system is connected to a furnace front control cabinet, a furnace temperature detection device, a furnace pressure detection device, a flue gas analyzer, an online molten metal composition detection device, and smelting execution equipment. Traditional methods recommend parameters using statistical averages of similar furnace runs and fixed process thresholds. When the furnace temperature is lower than the process target or the molten metal composition is close to the boundary, the system adjusts the oxygen supply flow rate, electrode power, and reducing agent addition based on historical average experience. While traditional methods can complete basic control, they lack continuous assessment of the furnace condition response after each parameter adjustment. This can easily lead to situations where the next control segment is sent before the previous one has fully taken effect, resulting in fluctuations in the final composition and increased energy consumption.

[0031] This embodiment selects 300 completed historical furnace runs as samples, of which 240 runs are used for training and 60 runs are used for verification. For each furnace run, data on oxygen supply flow rate, electrode power, fuel addition, reducing agent addition, slagging agent addition, furnace temperature, furnace pressure, flue gas composition, molten metal composition, equipment operating status, and final product composition are recorded at 5-second sampling intervals. The final product requirements are: carbon content 0.16% to 0.20%, silicon content 0.28% to 0.36%, manganese content 0.65% to 0.78%, and final temperature 1605°C to 1625°C. After preprocessing, the system deletes 386 sensor communication interruption sampling locations and 82 actuator non-response sampling locations, and completes sampling locations with consecutive missing intervals not exceeding 15 seconds, ultimately forming 300 historical furnace run status sequences.

[0032] In historical furnace runs, the system identified 1684 process parameter adjustment events. For example, in a successful furnace run, the oxygen supply flow rate increased from 850 cubic meters per hour to 905 cubic meters per hour, the electrode power increased from 24.0 MW to 24.6 MW, the reducing agent addition decreased from 136 kg to 122 kg, and within 60 seconds after the adjustment, the furnace temperature rose from 1588°C to 1601°C, the carbon monoxide volume fraction decreased from 19.4% to 16.9%, and the molten metal carbon content decreased from 0.239% to 0.207%. The final carbon content was 0.184%, silicon content was 0.326%, and manganese content was 0.704%. In another unsuccessful furnace run, after the oxygen supply flow rate increased from 860 cubic meters per hour to 935 cubic meters per hour, the final carbon content decreased to 0.151%, below the lower limit. The system used the corresponding path as a counter-evidence sample. After counterfactual parameter response ablation, response time delay manifold folding, target mass backtracking growth, and transfer reachability constraints, the system retains 1216 valid parameter response state nodes, forming 76 parameter response density attractors, 188 parameter response transition edges, and 8 target mass attractors.

[0033] In a current furnace operation, the system collected the following furnace conditions: furnace temperature 1587 degrees Celsius, furnace pressure 3.8 kPa, carbon monoxide volume fraction 20.2%, carbon dioxide volume fraction 10.9%, molten metal carbon content 0.241%, silicon content 0.405%, and manganese content 0.641%. All equipment was operating normally. Traditional methods, based on the average of similar furnace operations, recommend increasing the oxygen supply flow rate from 855 cubic meters per hour to 930 cubic meters per hour and the electrode power from 24.1 MW to 25.4 MW. This invention first matches the current furnace condition with 76 parameter response density attractors. The attractor distance for parameter response density attractor number 23 is calculated to be 0.043, which is less than 0.08. Therefore, parameter response density attractor number 23 is determined as the current starting attractor.

[0034] The system uses the current starting attractor as the path origin and the target mass attractor as the path endpoint. Five parameter response attraction paths are retrieved from the parameter response attraction path graph. Two paths are deleted due to abnormal equipment operation or the predicted endpoint silicon content exceeding 0.36%. The remaining three paths form three candidate process parameter trajectories. An improved MTGNN model is used to predict these three trajectories. The first trajectory has an average predicted pass intensity of 0.805 and a total attractor transition residual of 0.047; the second trajectory has an average predicted pass intensity of 0.691 and a total attractor transition residual of 0.064; and the third trajectory has an average predicted pass intensity of 0.850 and a total attractor transition residual of 0.029. The system selects the third trajectory as the target process parameter trajectory. The current parameter control segment is that the oxygen supply flow rate increases from 855 cubic meters per hour to 895 cubic meters per hour, the electrode power increases from 24.1 MW to 24.5 MW, the amount of slagging agent added increases from 275 kg to 310 kg, and the amount of reducing agent added remains at 138 kg.

[0035] After the current parameter control segment is successfully sent, the system collects furnace condition data every 5 seconds starting from the 5th second, continuously until the 60th second, obtaining a total of 12 post-execution sampling positions. The average values ​​of the 12 sampling positions are: furnace temperature 1597 degrees Celsius, furnace pressure 3.5 kPa, carbon monoxide volume fraction 17.4%, carbon dioxide volume fraction 12.6%, molten metal carbon content 0.218%, silicon content 0.374%, and manganese content 0.662%. The system calculates the distance between the post-execution furnace condition response and the 76 parameter response density attractors. The distance of parameter response density attractor number 36 is the smallest, at 0.038, and parameter response density attractor number 36 is consistent with the target end of the current parameter response transition edge. Therefore, the execution feedback record shows a match, and the system continues to send the next parameter control segment, causing the furnace condition to enter the composition convergence stage.

[0036] In 30 simulated samples of current furnace runs, the conventional method and the method of this invention were compared using the same initial furnace condition data. The conventional method achieved a 73.3% pass rate for carbon content in 22 furnaces, while the method of this invention achieved a 93.3% pass rate for carbon content in 28 furnaces. The conventional method achieved a 76.7% pass rate for silicon content in 23 furnaces, while the method of this invention achieved a 90.0% pass rate for silicon content in 27 furnaces. The conventional method achieved an 80.0% pass rate for manganese content in 24 furnaces, while the method of this invention achieved a 93.3% pass rate for manganese content in 28 furnaces. The conventional method achieved a 66.7% pass rate for all three endpoint components in 20 furnaces, while the method of this invention achieved a 86.7% pass rate for all three endpoint components in 26 furnaces. The average deviation of the endpoint temperature was 13.6 degrees Celsius for the conventional method and 6.4 degrees Celsius for the method of this invention. The conventional method required an average of 4.5 parameter control segments to reach the acceptable endpoint range, while the method of this invention required an average of 3.2 parameter control segments. The average power consumption was 411.8 kWh per ton for the conventional method and 394.6 kWh per ton for the method of this invention. The above results demonstrate that the present invention can reduce the impact of furnace condition response misalignment and path discontinuity on the final quality by parameter response path matching, continuous prediction, and execution feedback updates, thereby improving the accuracy and feasibility of smelting process parameter optimization.

[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for optimizing smelting process parameters based on machine learning, characterized in that, include: Collect furnace smelting operation data and preprocess the furnace smelting operation data to obtain the historical furnace smelting process status sequence and the current furnace smelting process status sequence. Extract process parameter adjustment events from historical furnace status sequences and construct parameter response status node sequences; The DENCLUE algorithm is executed based on the parameter response state node sequence to perform parameter action phase constraints, parameter response transition reachability constraints, and target mass path constraints. The density contribution is calculated to form parameter response density attractors, parameter response transition edges, parameter response attraction path graphs, and target mass attractors. Based on the current furnace state sequence, locate the current starting attractor, retrieve the qualified parameter response attraction path leading to the target mass attractor, extract the parameter response transfer edge and combine the candidate process parameter trajectory set; An improved MTGNN model is constructed based on the current furnace condition, the current starting attractor, and the set of candidate process parameter trajectories. Combined with parameter response transfer edge execution path constraint propagation, time delay alignment propagation, and attractor order preservation prediction, a predicted attraction path is generated. The predicted attraction path is matched and verified with the qualified parameter response attraction path, and the candidate process parameter trajectory set is screened by combining the target mass attractor arrival state, equipment operating state and product composition state to form the target process parameter trajectory. Send the current parameter control segment of the target process parameter trajectory, match the furnace condition response and termination attractor with the parameter response density attractor, and send the next parameter control segment or update the target process parameter trajectory according to the matching result.

2. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, The specific furnace smelting operation data includes furnace identifier, collection timestamp, oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, slagging agent addition data, furnace temperature data, furnace pressure data, flue gas composition data, molten metal composition data, equipment operating status data, and product final composition data.

3. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, The preprocessing of furnace smelting operation data includes: The smelting operation data of each furnace is grouped according to the furnace number identifier. Furnaces whose production status is completed and have product end-point component data are classified as historical furnaces, and furnaces whose production status is smelting in progress and have not yet formed product end-point component data are classified as current furnaces. The smelting operation data of the historical furnaces and the current furnaces are sorted, time-aligned, missing data is filled, anomalies are removed, and values ​​are normalized according to the collection timestamp, forming the historical furnace status sequence and the current furnace status sequence, respectively.

4. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, The constructed parameter response state node sequence includes: Based on the process parameter data and equipment operation status data in the historical furnace status sequence, the sampling locations where the process parameters changed and the equipment was in normal execution status were screened to determine the process parameter adjustment events. The furnace condition status before the adjustment, the process parameter change process, the response time and the furnace condition response status are extracted from the process parameter adjustment event to form a parameter response status node; Arrange the parameter response status nodes corresponding to the same process parameter adjustment event according to the collection timestamp, and construct the parameter response status node sequence.

5. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, The formation of the parameter response density attractor, parameter response transition edge, parameter response attraction path graph, and target mass attractor includes: The DENCLUE algorithm is executed based on the parameter response state node sequence. Historical segments that match the furnace condition before adjustment and have not experienced the same type of process parameter adjustment event are retrieved according to the process parameter adjustment event, the furnace condition state before adjustment and the furnace condition response state. The actual response segment and the control response segment are obtained. Counterfactual parameter response ablation is performed on the actual response fragment and the control response fragment, and nodes caused by changes in oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data or slagging agent addition data are retained to obtain the effective parameter response state node sequence. The response time-delay manifold folding is performed according to the changes in flue gas composition data, furnace temperature data, furnace pressure data, and molten metal composition data. The phase constraint node sequence is obtained by preserving sequential phase nodes, suppressing reverse phase nodes, and skipping phase nodes through parameter-based phase constraints. The target quality attractor seed is determined from the historical furnaces where the final product composition data is qualified and the equipment operation status data is normal. The target quality backtracking growth is performed along the process parameter adjustment event and phase constraint node sequence. The reachable nodes are retained and the unreachable nodes are suppressed by the target quality path constraint to obtain the target quality backtracking node sequence. Parameter response transfer reachability constraints are applied to adjacent nodes in the target quality backtracking node sequence. Transfer reachable nodes are retained. Density contribution is calculated and parameter response density attractors are determined. Adjacent parameter response density attractors are connected to generate parameter response transfer edges and parameter response attraction path graphs. The path endpoints where the product endpoint component data is qualified and the equipment operating status data is normal are determined as the target quality attractors.

6. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, The extraction of parameter response transition edges and the combination of candidate process parameter trajectory sets include: Based on the process parameter status and furnace condition response status in the current furnace state sequence, the current furnace condition response status is matched with the parameter response density attractor to determine the current starting attractor; Using the current starting attractor as the path start point and the target mass attractor as the path end point, the parameter response attraction path is retrieved in the parameter response attraction path diagram, and then continuously filtered according to the normal equipment operating status, qualified product end component, and parameter response transfer edge to obtain qualified parameter response attraction paths. Based on the parameter response transfer edge in the qualified parameter response attraction path, the adjustment direction, adjustment magnitude and execution order of oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data and slagging agent addition data are extracted and combined to form a candidate process parameter trajectory set.

7. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, The generation of the predicted attraction path includes: An improved MTGNN model is constructed, which includes a state encoding module, a transition graph construction module, a transition propagation module, and a path prediction module. The state coding module performs attractor state space reparameterization and feedback assimilation based on the current furnace state, the current starting attractor, the parameter response density attractor, and historical execution feedback records, mapping the current furnace state and the furnace response after execution to the current attractor state code. The transfer graph construction module performs event-driven construction based on process parameter adjustment events and parameter response transfer edges in the candidate process parameter trajectory set. It uses parameter response density attractors as graph nodes and parameter response transfer edges as event-triggered directed edges to generate candidate parameter response transfer graphs. The transfer propagation module performs path constraint propagation along the parameter response transfer edges defined by the candidate process parameter trajectory set, based on the current attractor state encoding and candidate parameter response transfer map. It also performs forward propagation along the direction from the current starting attractor to the target mass attractor, reverse traction along the direction from the target mass attractor to the current starting attractor, and performs counter-evidence blocking and time-delay alignment propagation along the paths where the product endpoint component data is unqualified or the equipment operating status data is abnormal, generating a bidirectional constraint transfer propagation sequence. The path prediction module performs attractor order preservation prediction based on the bidirectional constraint transfer propagation sequence, candidate process parameter trajectory set, and historical execution feedback records. It generates attractor transition residuals based on the matching results between the predicted target end parameter response density attractor and the corresponding target end parameter response density attractor in the candidate process parameter trajectory set. It then corrects the transfer propagation intensity according to the attractor transition residuals and historical execution feedback records to generate the predicted attraction path. The improved MTGNN model was trained using path matching error, target mass attractor arrival error, and attractor transition residual as optimization objectives. The parameters of the state encoding module, transition graph construction module, transition propagation module, and path prediction module were continuously optimized. Training was stopped when the total error decreased by less than 0.001 for 10 consecutive training rounds, and the improved MTGNN model was obtained after training.

8. The method for optimizing smelting process parameters based on machine learning according to claim 7, characterized in that, The execution of attractor state space reparameterization and feedback assimilation includes: Based on the furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data in the current furnace condition, and by calculating the deviation between these data and the furnace condition response state corresponding to each parameter response density attractor, the attractor matching result is obtained. Using the parameter response density attractor as the state coordinate and the phase position and adjacent parameter response transition edges as the transition coordinate, an attractor state space is constructed. Based on the attractor matching results, the current furnace condition is projected onto the attractor state space to generate the current attractor state code. Based on the historical execution feedback record, read the parameter response density attractor corresponding to the furnace condition response after the execution of the previous parameter control segment and the parameter response density attractor at the target end of the previous parameter response transfer edge. According to the matching result of the two, correct the current attractor state code and the propagation strength of the corresponding parameter response transfer edge to obtain the current attractor state code after feedback assimilation.

9. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, The formation of the target process parameter trajectory includes: Based on the order of attractors of parameter response density and the connection relationship of parameter response transition edges in the predicted attraction path and qualified parameter response attraction path, a matching verification is performed to obtain the path matching result; Based on the path matching results, the predicted attraction path is checked to see if it reaches the target mass attractor, and whether the corresponding equipment operating status data is normal and the product endpoint component data is qualified, thus obtaining the trajectory verification results; Based on the trajectory verification results, candidate process parameter trajectories that match the path, reach the target mass attractor, have normal equipment operation status, and have qualified product endpoint components are retained from the candidate process parameter trajectory set to form the target process parameter trajectory.

10. The method for optimizing smelting process parameters based on machine learning according to claim 1, characterized in that, Sending the next parameter control segment or updating the target process parameter trajectory includes: Based on the current parameter control segment of the target process parameter trajectory, determine the adjustment direction, adjustment range, and execution sequence of oxygen supply flow rate data, electrode power data, fuel addition data, reducing agent addition data, and slagging agent addition data, and send them to the smelting execution equipment; Collect furnace temperature data, furnace pressure data, flue gas composition data, and molten metal composition data after the execution of the current parameter control segment. Determine the parameter response density attractor corresponding to the furnace condition response after execution and match it with the parameter response density attractor at the target end of the current parameter response transfer edge to generate an execution feedback record. When the execution feedback record shows a match, the next parameter control segment is sent. When the execution feedback record shows an inconsistency, the parameter response density attractor corresponding to the furnace condition response after execution is written into the current attractor state code, and the target process parameter trajectory is updated.

Images