Intelligent management and control and energy-saving optimization method and system for energy consumption of smelting process

Abstract

The present application relates to the technical field of data management optimization, in particular to a smelting process energy consumption intelligent management and control and energy saving optimization method and system. The method comprises the following steps: determining the reference energy consumption coefficient of each working condition cluster based on the regression relationship between the duration of each working condition cluster and the unit product energy consumption, generating the stable judgment threshold corresponding to each working condition cluster, and intercepting the continuous stable state time period as a stable activation segment; a time weight function is constructed to weight and correct the duration of each stable activation segment to obtain a weighted activation duration; taking the duration of each working condition cluster as a decision variable and taking the minimum unit product energy consumption as an optimization objective, an optimization model containing duration boundary constraint, working condition transfer constraint and process completion constraint is constructed to generate the control setting value sequence of the underlying device and execute it, which effectively reduces the energy consumption of the overall smelting process.

Description

Technical Field

[0001] This invention relates to the field of management data optimization technology, and in particular to intelligent control and energy-saving optimization methods and systems for smelting process energy consumption. Background Technology

[0002] Smelting is a crucial process in modern metallurgy and materials processing manufacturing, directly impacting product quality and performance, and is also one of the most energy-intensive steps in the production process. Currently, data-driven process parameter control and operating condition analysis have gradually become important technological directions for energy-saving optimization of smelting processes.

[0003] In existing energy management technologies for smelting processes, some solutions focus on the collection, prediction, and optimization control of energy consumption data. For example, Chinese patent application CN120373912A discloses an energy consumption prediction and optimization system for energy-saving management. Its core features include: a dynamic energy consumption prediction module that uses an improved time series decomposition model and a multimodal LSTM neural network to predict energy consumption trends in future periods based on historical energy consumption data and real-time operating status, and generates a baseline energy consumption curve; an intelligent optimization control module that generates a set of equipment control instructions based on the prediction results and preset energy-saving strategies, including equipment start-up and shutdown time optimization, power adjustment instructions, and equipment linkage rules, and achieves energy consumption regulation through time-of-use electricity price optimization algorithms and dynamic thresholds; and an adaptive calibration module that fuses multi-source data through Kalman filtering and uses an incremental learning mechanism to update model parameters online.

[0004] However, when dealing with the multidimensional process parameters generated during the smelting process, the aforementioned methods typically focus on predicting overall energy consumption trends and optimizing equipment-level control. They struggle to accurately quantify the energy consumption contribution of different operating conditions within the smelting process and rarely consider the differences in the frequency of occurrence and the intensity of energy consumption fluctuations of each operating condition during the production cycle, resulting in an insufficiently refined characterization of the relationship between operating conditions and energy consumption. Furthermore, existing optimization methods largely rely on prediction results to drive preset energy-saving strategies, lacking the ability to find global operating condition paths with the goal of minimizing energy consumption per unit of product.

[0005] Therefore, how to extract accurate energy consumption correlation characteristics from industrial operation data and establish a global energy-saving control strategy that takes into account both operating condition transfer rules and process completion constraints has become a key issue that urgently needs to be addressed in the intelligent upgrading of smelting processes. Summary of the Invention

[0006] To address the technical challenge of extracting accurate energy consumption correlation features from industrial operation data and establishing a global energy-saving control strategy that takes into account both operating condition transfer rules and process completion constraints, this invention provides a method and system for intelligent control and energy-saving optimization of smelting process energy consumption.

[0007] In a first aspect, the present invention provides a method for intelligent control and energy-saving optimization of energy consumption in smelting processes, employing the following technical solution:

[0008] Intelligent control and energy-saving optimization methods for smelting process energy consumption, including the following steps: S1. Cluster the multidimensional process parameter data points of the smelting process to obtain multiple working condition clusters and their corresponding durations; determine the baseline energy consumption coefficient of each working condition cluster based on the regression relationship between the duration of each working condition cluster and the energy consumption per unit product; perform normalized inverse proportional mapping on the baseline energy consumption coefficient to generate the stability judgment threshold corresponding to each working condition cluster; calculate the normalized dispersion of each data point relative to its respective working condition cluster; determine the data points with normalized dispersion below the corresponding stability judgment threshold as stable states, and extract continuous stable state time periods as stable activation segments; S2. Statistically analyze the working condition clusters in different process periods. The frequency of occurrence within the segment and its corresponding energy consumption deviation are used to construct a time weighting function to weight and correct the duration of each stable activation segment, thus obtaining the weighted activation duration. Based on the correlation between the weighted activation duration and the energy consumption per unit product, the accurate energy consumption coefficient of each operating condition cluster is determined. S3, with the duration of each operating condition cluster as the decision variable and the minimum energy consumption per unit product as the optimization objective, an optimization model containing duration boundary constraints, operating condition transition constraints, and process completion constraints is constructed. The operating condition cluster transition path and the target duration of each operating condition cluster are solved, and the control setpoint sequence of the underlying equipment is generated and sent out for execution.

[0009] This invention provides an intelligent control and energy-saving optimization method for smelting processes. It identifies different operating conditions by clustering multi-dimensional process parameters and generates a dynamic stability threshold using a benchmark energy consumption coefficient inverse proportional mapping, accurately capturing stable activation segments that exclude transitional interference. Furthermore, it constructs a time weighting function by combining the frequency of occurrence of each operating condition with energy consumption deviations at different process stages to correct the duration, thereby deriving an accurate energy consumption coefficient. Finally, under the premise of meeting real process constraints, it performs global optimization with the goal of minimizing energy consumption per unit product, generating and issuing control sequences. This method optimizes traditional extensive preset energy-saving strategies, achieving closed-loop control from detailed characterization of complex operating conditions to global energy consumption optimization, reducing energy consumption per unit product in the actual smelting process and ensuring the reliable completion of production tasks.

[0010] According to the intelligent energy consumption control and energy-saving optimization method for smelting process provided by the present invention, the method of clustering multi-dimensional process parameter data points of the smelting process to obtain multiple operating condition clusters includes: collecting the temperature of the smelting furnace, the fan pressure and the power supply at a preset sampling frequency to obtain each data point in the multi-dimensional time series; removing outliers from the multi-dimensional time series using the isolated forest algorithm and normalizing the cleaned data using the Z-score normalization method; and clustering the normalized data based on the K-Means++ algorithm, determining the number of cluster centers by the silhouette coefficient, and dividing each cluster center and its subordinate data into multiple operating condition clusters.

[0011] According to the intelligent control and energy-saving optimization method for smelting process provided by the present invention, the determination of the benchmark energy consumption coefficient of each working condition cluster based on the regression relationship between the duration of each working condition cluster and the energy consumption per unit product includes: counting the number of data points of each working condition cluster in each batch of smelting process in chronological order, multiplying by the sampling period to obtain the duration of each working condition cluster; using the vector formed by the duration of each working condition cluster in each batch of smelting process as the independent variable, and the corresponding energy consumption per unit product in each batch as the dependent variable; using a multiple linear regression algorithm with non-negative constraints to fit the linear mapping relationship between the independent variable and the dependent variable, constructing a preliminary correlation model, and using the partial regression coefficients of each independent variable in the preliminary correlation model as the benchmark energy consumption coefficients of the corresponding working condition cluster.

[0012] This invention constructs a multiple linear regression model with non-negative constraints, using batch duration as the independent variable and unit product energy consumption as the dependent variable. Compared with simple local energy consumption rough calculation, this invention directly extracts the true unit-time marginal contribution of each smelting condition to the total energy consumption from macro-historical data, ensuring that the determined benchmark energy consumption coefficient has clear physical dimensions and engineering guidance significance, and avoiding evaluation distortion caused by fluctuation of a single parameter.

[0013] According to the intelligent energy consumption control and energy-saving optimization method for smelting processes provided by the present invention, the step of calculating the normalized dispersion of each data point relative to its respective working condition cluster, determining data points with normalized dispersion lower than the corresponding stability determination threshold as stable states, and extracting continuous stable state time periods as stable activation segments includes: calculating the Euclidean distance from each data point in the working condition cluster to the centroid of the working condition cluster, dividing the Euclidean distance by the maximum Euclidean distance from the training sample in the working condition cluster to the centroid of the working condition cluster, and then performing interval truncation processing to obtain normalized dispersion values ​​in the range of 0 to 1; traversing the working condition cluster sequence, when the normalized dispersion of N consecutive data points is less than the stability determination threshold, determining the N consecutive data points as stable states, where N is a preset positive integer greater than or equal to 5; merging adjacent stable states belonging to the same working condition cluster and extracting them to form continuous stable activation segments.

[0014] This invention calculates the normalized dispersion of real-time multidimensional spatial coordinates to the centroid of the working condition, and combines a dynamic sliding window and merging mechanism to extract continuous and stable active segments. Compared with the traditional approach of rigidly cutting the working condition by relying on fixed parameter thresholds, this method can intelligently filter out transient transition noise data that are prone to control oscillations in the non-steady metallurgical production cycle, and accurately capture the stable operating range that can truly characterize the essence of the working condition.

[0015] According to the intelligent control and energy-saving optimization method for smelting process provided by the present invention, the step of constructing a time weight function by statistically analyzing the occurrence frequency of each working condition cluster in different process time periods and its corresponding energy consumption deviation includes: dividing the smelting process cycle into multiple process time segments of equal length, statistically analyzing the occurrence frequency of each working condition cluster in each process time segment in all batches to obtain the corresponding occurrence frequency; calculating the absolute value of the difference between the average unit product energy consumption of the batch in which each working condition cluster occurs in each process time segment and the overall average unit product energy consumption of all batches, as the energy consumption deviation; and normalizing the product of the occurrence frequency and the energy consumption deviation to fit and generate a time weight function that varies with process time.

[0016] This invention slices the time axis of the smelting process and extracts the frequency of occurrence of specific working conditions and the corresponding absolute energy consumption deviation to generate a continuous time weighting function. Compared with the traditional static evaluation of the whole furnace duration, this invention restores the completely different energy consumption fluctuation evolution law caused by the same operation at different stages of the smelting life cycle, making the weighted correction of the working condition duration closer to the complex industrial physical mechanism.

[0017] According to the intelligent control and energy-saving optimization method for smelting process energy consumption provided by the present invention, the method for determining the precise energy consumption coefficient of each working condition cluster based on the correlation between weighted activation duration and unit product energy consumption includes: based on the operating data of each batch, using the weighted activation duration of each working condition cluster as the input variable and the unit product energy consumption of the corresponding batch as the output variable, establishing a refined correlation model using a random forest regression algorithm, extracting the feature importance of each working condition cluster, and obtaining the precise energy consumption coefficient by proportionally correcting the benchmark energy consumption coefficient using the feature importance, wherein the feature importance is used to characterize the strength of the contribution of the corresponding working condition cluster to the unit product energy consumption prediction result.

[0018] This invention introduces a random forest regression algorithm to calculate the nonlinear relationship between weighted activation duration and energy consumption, and uses the output feature importance to proportionally correct the baseline energy consumption coefficient. Compared with the limitations of a single linear model, this scheme achieves secondary calibration of complex thermal-engineering relationships while preserving physical dimensions, thereby improving the system's resolution in identifying the energy consumption contribution of heavy-load, high-energy-consuming operating conditions and fine-tuning auxiliary operating conditions.

[0019] According to the intelligent control and energy-saving optimization method for smelting process provided by the present invention, the duration boundary constraint is that the duration of each working condition cluster must be within the upper and lower limits of historical statistical values; the working condition transition constraint is that the transition between adjacent working condition clusters must satisfy that the Markov transition probability in the historical working condition sequence is greater than a preset lower limit; the process completion constraint includes at least one of the following: planned total duration constraint, endpoint temperature constraint, output constraint, and termination working condition constraint.

[0020] The intelligent control and energy-saving optimization method for smelting process energy consumption provided by the present invention uses a genetic algorithm to solve the optimization model, including: encoding candidate solutions into chromosomes containing working condition paths and durations; penalizing candidate solutions that violate constraints in fitness calculation; and outputting the combination of target durations for each working condition cluster and the corresponding working condition cluster transfer path after iteration, which minimizes the energy consumption per unit product.

[0021] According to the intelligent control and energy-saving optimization method for smelting process provided by the present invention, the step of generating and issuing the control setpoint sequence of the underlying equipment includes: establishing a mapping relationship between the operating condition cluster and the control parameters based on the statistical values ​​of the control parameters corresponding to each operating condition cluster in the historical stable active segment; sequentially reading the control parameters of each operating condition cluster according to the operating condition cluster transfer path, and generating a segmented control plan in combination with the target duration; inserting a transition segment for the section where the parameter jump between adjacent operating condition clusters exceeds the equipment ramping capacity; and writing the final control setpoint sequence into the programmable logic controller.

[0022] Secondly, this invention provides an intelligent energy consumption control and energy-saving optimization system for smelting processes, employing the following technical solution: A smart energy consumption control and energy-saving optimization system for smelting processes includes a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned smart energy consumption control and energy-saving optimization method for smelting processes is implemented.

[0023] By adopting the above technical solution, the above-mentioned intelligent energy consumption control and energy-saving optimization method for smelting process is generated into a computer program and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and processor for convenient use.

[0024] The present invention has the following technical effects: Based on the above technical solutions, the intelligent control and energy-saving optimization method and system for smelting process energy consumption provided by this invention obtains the baseline energy consumption coefficient by clustering and preliminary correlation analysis of multi-dimensional process parameters, and accurately extracts continuous stable activation segments using the normalized dispersion and the energy consumption adaptive stability threshold obtained by normalized inverse proportional mapping, reducing the interference of abnormal fluctuation data and improving the accuracy of operating condition feature identification. On this basis, a time weight function is constructed by combining the frequency of operating conditions and energy consumption deviation to weight the duration. The established refined correlation model can more accurately represent the contribution strength of each operating condition cluster to the unit product energy consumption, and obtains accurate energy consumption coefficients while maintaining the dimensions of the baseline energy consumption coefficient and non-negative constraints, deeply characterizing the real impact of complex operating conditions on energy consumption. Furthermore, by constructing and solving an optimization model with the goal of minimizing unit product energy consumption, the optimal operating condition cluster transfer path and target duration are obtained. Combined with the process completion constraints and the mapping relationship of the underlying control parameters corresponding to the operating condition clusters, a setpoint sequence that can be written to the controller is generated, realizing the scientific control of the underlying equipment, reducing the ineffective residence of high-energy-consuming operating conditions, and reducing the overall energy consumption of the smelting process. Attached Figure Description

[0025] Figure 1 A flowchart illustrating the intelligent energy consumption control and energy-saving optimization method for smelting processes provided in this embodiment of the invention; Figure 2 This is a schematic diagram illustrating the variation of the global contour coefficient with the number of cluster centers provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the change of normalized dispersion over time, as provided in an embodiment of the present invention. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] Example 1 This invention discloses a method for intelligent control and energy-saving optimization of energy consumption in smelting processes. For details, please refer to [link to relevant documentation]. Figure 1 As shown, Figure 1 The flowchart illustrates the intelligent energy consumption control and energy-saving optimization method for smelting processes provided in this embodiment of the invention. The method specifically includes the following steps: S1: Determine the stability threshold and extract stable activation segments.

[0028] For example, in this embodiment of the invention, the multidimensional process parameter data points of the smelting process are clustered to obtain multiple working condition clusters and their corresponding durations; the baseline energy consumption coefficient of each working condition cluster is determined based on the regression relationship between the duration of each working condition cluster and the energy consumption per unit product; the baseline energy consumption coefficient is normalized and inversely proportionally mapped to generate the stability judgment threshold corresponding to each working condition cluster; the normalized dispersion of each data point relative to its working condition cluster is calculated; data points with normalized dispersion lower than the corresponding stability judgment threshold are judged as stable states, and continuous stable state time periods are extracted as stable activation segments.

[0029] Multidimensional process parameters, including furnace temperature, fan pressure, and power supply, are collected by sensors deployed on the smelting furnace. Further parameters can be included, such as gas flow rate, cooling water pressure, and power supply frequency. Multidimensional time-series data is read, and a standardization module performs zero-mean standardization on the multidimensional process parameters to reduce the influence of dimensions. The K-Means algorithm is used to perform cluster analysis on the standardized data, and the cluster centroids are initialized using K-Means++. The number of clusters is set as the total number of operating condition categories. The `fit_predict` method is used to obtain the operating condition cluster label to which each data point belongs and the centroid coordinates of each operating condition cluster. The time-series data of each batch of smelting cycles is traversed, recording the start and end times of consecutive entries belonging to the same operating condition cluster label. The time difference and sampling period are calculated to extract the duration of each operating condition cluster.

[0030] In one embodiment, the acquired multidimensional process parameters are clustered to obtain multiple operating condition clusters, including: Temperature sensor data, fan pressure data, and power data of the underlying equipment are collected at a preset sampling frequency to form a multidimensional time series. The isolated forest algorithm is used to remove outliers from the multidimensional time series, and the Z-score normalization method is used to normalize the cleaned data. The normalized data is clustered based on the K-Means++ algorithm, and the number of cluster centers is determined by the silhouette coefficient. Each cluster center and its subordinate data are divided into multiple operating condition clusters.

[0031] During the multidimensional process parameter acquisition phase, the sampling frequency of the underlying equipment is preferably set to 1Hz to 5Hz, for example, 2Hz, meaning data is collected every 0.5s. This allows for real-time acquisition of the furnace internal temperature (range 1200℃ to 1600℃), the fan supply pressure (range 10kPa to 50kPa), and the active power of the induction power supply (range 500kW to 2500kW), forming a multidimensional time-series dataset. For anomaly detection, the hyperparameters of the Isolation Forest algorithm are set as follows: the number of trees is set to 100 to 200, the maximum number of samples is set to 256, and the contamination rate is preset to between 0.02 and 0.05, such as 0.03. Anomaly scores are calculated for each sample, and outliers are identified and removed based on the preset contamination rate or preset anomaly score threshold. The Z-score normalization formula is then applied. ,in and The mean and standard deviation of the cleaned data are given, and each parameter is converted into standardized data with a mean of 0 and a standard deviation of 1. The initial cluster center selection probability of the K-Means++ clustering algorithm is proportional to the square of the nearest distance from the sample point to the selected cluster center. The traversal range of the number of candidate cluster centers K is set from 3 to 12. For each K value, the silhouette coefficient of the corresponding model is calculated, ranging from -1 to 1. When a certain K value, such as K=5, reaches its maximum global average silhouette coefficient (e.g., above 0.65), K is determined as the optimal number of cluster centers. The standardized multidimensional time series data points are then divided into these 5 specific smelting condition clusters according to the principle of closest proximity to the centroid. The change in silhouette coefficient with the number of clusters is illustrated in the diagram below. Figure 2 As shown, Figure 2 This is a schematic diagram illustrating the variation of the global contour coefficient with the number of cluster centers, provided in an embodiment of the present invention.

[0032] After data collection based on the above steps, the duration of each operating condition cluster is used as the independent variable, and the unit product energy consumption of the corresponding batch is used as the dependent variable. The unit product energy consumption is the quotient of the total power consumption of the corresponding batch and the output of qualified products, in kWh / t. A multiple linear regression algorithm with non-negative constraints is called to perform regression analysis, and the non-negative partial regression coefficients of the regression equation are obtained. The linear mapping relationship between the independent and dependent variables is fitted, and the non-negative partial regression coefficients are used as the benchmark energy consumption coefficients for each operating condition cluster. Since the duration is preferably uniformly converted to hours, the physical meaning of the benchmark energy consumption coefficient is the unit time marginal contribution of unit product energy consumption relative to the duration of the corresponding operating condition cluster, in kWh / (t·h). Thus, the product of the duration and the precise energy consumption coefficient in the subsequent objective function has the meaning of unit product energy consumption.

[0033] For each multidimensional time series data point, the `euclidean` function is called to calculate the Euclidean distance between the data point and the centroid coordinates of the operating condition cluster to which the data point belongs. The maximum Euclidean distance from the training samples within the operating condition cluster to the centroid is used as the normalized denominator. If the maximum Euclidean distance is zero, it is replaced with a preset minimum positive number to obtain the initial dispersion. The initial dispersion is then truncated to an upper limit, so that values ​​greater than 1 are set to 1, resulting in the normalized dispersion with values ​​between 0 and 1. Using pandas' conditional filtering function, each row is compared. If the normalized dispersion of a certain time node is less than the energy consumption stability judgment threshold, the node is marked as a stable activation state. The sliding window algorithm is used to identify and extract adjacent data intervals on the time axis that are both in a stable activation state as continuous stable activation segments.

[0034] In one embodiment, the operating condition cluster sequence and duration of each batch of smelting process are extracted, and a preliminary correlation model is constructed through regression analysis to obtain the baseline energy consumption coefficient of each operating condition cluster, including: The number of data points for each working condition cluster in each batch of smelting process is counted in chronological order, and the duration of each working condition cluster is obtained by multiplying the data points by the sampling period. The vector formed by the duration of each working condition cluster in each batch of smelting process is used as the independent variable, and the energy consumption per unit product of each batch is used as the dependent variable. A multiple linear regression algorithm with non-negative constraints is used to fit the linear mapping relationship between the independent and dependent variables to construct a preliminary correlation model. The partial regression coefficients of each independent variable in the preliminary correlation model are used as the benchmark energy consumption coefficients of the corresponding working condition clusters.

[0035] For any i-th smelting batch, the total number of data points belonging to a specific operating condition cluster is counted sequentially according to the timestamp, assuming a fixed sampling period for the data acquisition system. If the duration is set to 0.5 seconds, the actual duration of this operating condition cluster in the batch is calculated as the product of the total number of data points and 0.5, in seconds. Converting the duration to hours allows us to construct a row vector of independent variables with a duration dimension of K, for example, K=5. Simultaneously, we obtain the actual total weight of smelted metal in the batch, such as 10 tons, and the total power consumption, such as 5500 kWh. The quotient of these two is used as the dependent variable, for example, the unit product energy consumption of this batch is 550 kWh / t. Based on this, we collect M samples, preferably an integer greater than 500, of consecutive historical smelting batches, thereby constructing a feature matrix of size 500×5 and a continuous label vector of size 500×1. Based on this, the non-negative least squares algorithm NNLS or the L-BFGS-B optimization algorithm with boundary constraints is used to solve the model. The objective function is to minimize the sum of squared residuals, and all parameter vectors are restricted to be greater than or equal to zero. After convergence, the output is a partial regression coefficient vector containing five non-negative floating-point sequences, for example, 12.5, 45.2, 108.6, 75.3, 20.1, which correspond to the marginal contribution per unit time of energy consumption of unit product under conditions such as heating, holding, and melting, respectively. These five partial regression coefficients are used as the benchmark energy consumption coefficients for each condition cluster.

[0036] Continuing with the above normalized inverse proportional mapping formula, the baseline energy consumption coefficient is mapped to the energy consumption stability judgment threshold. For example, when the preset range of the stability threshold is 0.30 to 0.45, the melting condition with the highest baseline energy consumption coefficient can be classified as a high-energy-consumption-contribution condition, corresponding to a lower stability threshold in the range of 0.30 to 0.35; the auxiliary condition with a lower baseline energy consumption coefficient can be classified as a low-energy-consumption-contribution condition, corresponding to a higher stability threshold in the range of 0.40 to 0.45; and the remaining conditions use an intermediate stability threshold in the range of 0.35 to 0.40. Through this sorting and grading method, the thresholds of different condition clusters all fall within the preset range, avoiding threshold out-of-control due to differences in the order of magnitude of energy consumption coefficients.

[0037] When setting the energy consumption stability judgment threshold, instead of directly using a single constant scaling factor divided by the baseline energy consumption coefficient, the threshold is sorted and categorized based on the baseline energy consumption coefficient of each operating condition cluster. Specifically, the operating condition clusters are sorted from high to low according to the baseline energy consumption coefficient and divided into high energy consumption contribution operating conditions, medium energy consumption contribution operating conditions, and low energy consumption contribution operating conditions. For high energy consumption contribution operating conditions, the energy consumption stability judgment threshold is set to a lower value, such as 0.30 to 0.35; for medium energy consumption contribution operating conditions, the threshold is set to an intermediate value, such as 0.35 to 0.40; and for low energy consumption contribution operating conditions, the threshold is set to a higher value, such as 0.40 to 0.45. Thus, operating condition clusters with high baseline energy consumption coefficients correspond to stricter stability screening conditions, while operating condition clusters with low baseline energy consumption coefficients correspond to relatively lenient stability screening conditions, while ensuring that the energy consumption stability judgment threshold of each operating condition cluster falls within the preset threshold range.

[0038] In one embodiment, the normalized dispersion of multidimensional time series data points relative to the centroid of the corresponding operating condition cluster is calculated. When the normalized dispersion is less than the energy consumption stability determination threshold, it is determined to be in a stable activation state. Continuous stable activation segments are extracted, including: Calculate the Euclidean distance from each data point within the working condition cluster to the centroid of the working condition cluster, and then divide this Euclidean distance by the maximum Euclidean distance from the training samples within the working condition cluster to the centroid of the working condition cluster. After interval truncation, the normalized dispersion with a value between 0 and 1 is obtained. Traverse the working condition cluster sequence. When the normalized dispersion of N consecutive data points is less than the stability determination threshold, the N consecutive data points are determined to be in a stable state, where N is a preset positive integer greater than or equal to 5. Merge adjacent stable states that belong to the same working condition cluster and truncate them to form a continuous stable activation segment.

[0039] For real-time multidimensional data points that have been determined to belong to the current working condition cluster, the Euclidean distance is calculated by using the multidimensional spatial coordinates of the data points and the corresponding centroid coordinates. All data points belonging to the working condition cluster in the historical training sample library are traversed to find the maximum Euclidean distance from the data points to the centroid of the corresponding working condition cluster. For example, the maximum Euclidean distance is 8.45. After performing a division operation, the initial dispersion is obtained, and the initial dispersion greater than 1 is truncated to 1 to generate a normalized dispersion sequence with values ​​in the interval between 0 and 1. For example, the dispersion is calculated to be 0.25 at a certain moment. Then, a fixed-length sliding time window is opened, and the preset positive integer N is configured to be a preset positive integer greater than or equal to 5, for example, N=10, corresponding to a time span of 5 seconds at a frequency of 2Hz. The normalized discreteness sequence is scanned point by point through the sliding window. When the discreteness values ​​of 10 consecutive data points in the window are all less than the energy consumption stability judgment threshold obtained by mapping the baseline energy consumption coefficient of the current operating condition cluster, this dynamic threshold range is generally between 0.30 and 0.45. For example, if the output at the current moment is 0.32, the control unit writes a stable activation state flag signal with a Boolean value of True to the memory. After the state flag is completed, the Boolean flag sequence is scanned forward. For adjacent True state segments that belong to the same operating condition cluster and are not interrupted by other operating conditions in between, or whose disconnection time is less than the tolerance limit, preferably configured to be 1 second, i.e., missing only 2 sampling points, they are merged. This extracts a continuous stable activation segment from the process cycle, such as a time stamp from 10:15:20 to 10:28:45, with a duration of 805 seconds. Figure 3 As shown, Figure 3 This is a schematic diagram illustrating the change of normalized dispersion over time, as provided in an embodiment of the present invention.

[0040] Thus, by performing cluster analysis on multidimensional process parameters and extracting stable activation fragments, the embodiments of the present invention can effectively eliminate noise interference in the transition phase, improve the accuracy of operating condition feature identification, and provide an accurate data foundation for the subsequent construction of a refined energy consumption correlation model.

[0041] S2: Determine the precise energy consumption coefficient for each operating condition cluster.

[0042] For example, in this embodiment of the invention, the frequency of occurrence of each operating condition cluster in different process periods and the corresponding energy consumption deviation are statistically analyzed to construct a time weighting function, so as to weight and correct the duration of each stable activation segment to obtain the weighted activation duration; the accurate energy consumption coefficient of each operating condition cluster is determined based on the correlation between the weighted activation duration and the energy consumption per unit product.

[0043] After dividing the smelting process cycle into multiple equal-length process time segments, the frequency of occurrence of each operating condition cluster in each batch within each process time segment is statistically analyzed using grouping statistics. The frequency of occurrence, ranging from 0 to 1, is calculated using the total number of valid batches or the total number of valid occurrences within the same time segment as the denominator, avoiding inconsistencies in dimensions caused by dividing the total number of data points by the number of operating condition switches. The absolute value of the difference between the average unit product energy consumption of the batch in which the operating condition cluster occurs within the corresponding time segment and the overall average unit product energy consumption of all batches is calculated as the energy consumption deviation. The energy consumption deviation is used to represent the intensity of the impact of the operating condition cluster on energy consumption fluctuations within that time segment, and is not used to separately determine the direction of increased or decreased energy consumption; the direction and dimensions of energy consumption contribution are maintained by the baseline energy consumption coefficient and its non-negativity constraint. After multiplying the occurrence frequency by the energy consumption deviation and normalizing, a time weighting function is constructed. This function iterates through all continuous and stable activation segments extracted in the above steps, multiplying the duration of each segment by the average weight within the corresponding time interval of the operating condition cluster to which the segment belongs, to calculate the weighted activation duration.

[0044] In one embodiment, the frequency of occurrence of each operating condition cluster in different process periods and its corresponding energy consumption deviation are statistically analyzed to construct a time weighting function, resulting in a weighted activation duration, including: The smelting process cycle is divided into multiple equal-length process time segments. The frequency of occurrence of each condition cluster in each batch within each process time segment is statistically analyzed to obtain the corresponding occurrence frequency. The absolute value of the difference between the average unit product energy consumption of the batch in which each condition cluster occurs within each process time segment and the overall average unit product energy consumption of all batches is calculated as the energy consumption deviation. The product of the occurrence frequency and the energy consumption deviation is normalized, and a time weighting function that varies with process time is fitted to generate the result. The continuous stable activation segment is discretized on the time axis. The average weight is calculated by integrating the time weighting function over the corresponding time interval and dividing by the length of the time interval. The average weight is then multiplied by the actual duration of the continuous stable activation segment to obtain the weighted activation duration.

[0045] Specifically, the smelting cycle, with a total duration of approximately 120 minutes, is divided into 12 continuous process time segments, each 10 minutes in length, and defined as follows: , to In the accumulated feature library of 1000 historical smelting batches, the frequency of occurrence of specific working condition clusters was obtained by counting the aggregations in each segment. For example, the frequency of occurrence of a certain smelting working condition cluster in... The occurrence occurred 850 times within this time period. When there are 1000 valid batch records in this time period, 850 is divided by 1000 to obtain the occurrence frequency of 0.85, instead of dividing by the total number of operating condition switching times. Further, related batches are extracted, and the average unit product energy consumption is calculated. Assuming the measured average is 570 kWh / t, the baseline average unit product energy consumption of all 1000 historical batches is extracted, assumed to be 550 kWh / t. The absolute value of the subtraction is calculated, outputting an absolute energy consumption deviation of 20 kWh / t. The occurrence frequency of 0.85 is multiplied by the energy consumption deviation of 20 to generate a weighted original scalar of 17. This weighted original scalar is then normalized using Min-Max in the set of weighted original scalars for all operating condition clusters and all time periods. During normalization, a very small positive number can be added to the denominator to prevent the maximum value from equaling the minimum value and causing division by zero. The very small positive number is only used as a numerical stabilization term and does not change the meaning of the weight as an independent smoothing term. The scalar values ​​for the entire time interval are mapped to coefficients between 0.1 and 1.0 using the Min-Max scaling method. A continuous-time weighting function model corresponding to the j-th operating condition cluster, with the process seconds t as the abscissa, is then approximated and fitted using Gaussian kernel smoothing or a 6th-order spline curve. When implementing weighted switching of operating conditions, if a certain segment of continuous stable activation falls within the 12th minute, with an absolute time start point between 720 seconds and 25th minute, and an absolute time end point within the 1500-second interval, the actual duration is 780 seconds. A definite integral operation is performed between the upper and lower bounds of 720 and 1500. After obtaining the integral value, it is divided by the interval length of 780 to obtain the interval average weight. Assuming the calculation result is 0.85, the product operation of 0.85 and the actual duration of 780 seconds is performed to obtain 663 seconds as the weighted activation duration for subsequent optimization.

[0046] The vector composed of the weighted activation durations of all operating condition clusters is used as the feature input matrix. The unit product energy consumption (i.e., total energy consumption) of the corresponding batch smelting cycle divided by the qualified product output is used as the target output vector. A refined correlation model is constructed by calling the random forest regression algorithm. The feature importance score is calculated by the reduction in mean squared error caused by the splitting of regression tree nodes or by the permutation importance. The feature importance score is normalized to obtain the normalized importance score, which is further converted into a proportional correction factor that fluctuates around 1, without replacing the physical dimension of the baseline energy consumption coefficient. The baseline energy consumption coefficient is then used as the energy consumption contribution direction and dimension benchmark. The baseline energy consumption coefficient is proportionally corrected using the proportional correction factor, and the correction result is subjected to non-negativity constraint processing to obtain the accurate energy consumption coefficient of each operating condition cluster.

[0047] In one embodiment, the precise energy consumption coefficient for each operating condition cluster is determined based on the correlation between weighted activation duration and unit product energy consumption, including: Based on the operational data of each batch, the weighted activation duration of each operating condition cluster is used as the input variable, and the unit product energy consumption of the corresponding batch is used as the output variable. A refined correlation model is established using the random forest regression algorithm, and the feature importance score of the refined correlation model is output. The feature importance score is normalized to obtain the feature importance of each operating condition cluster. The final feature importance is a proportional correction factor that fluctuates around 1. The benchmark energy consumption coefficient is used as the energy consumption contribution direction and dimension benchmark. The benchmark energy consumption coefficient is proportionally corrected using the feature importance, and the correction result is subjected to non-negativity constraint processing to obtain the accurate energy consumption coefficient. The feature importance is used to characterize the strength of the contribution of the corresponding operating condition cluster to the unit product energy consumption prediction result, and is not used alone to determine the direction of energy consumption increase or decrease.

[0048] The constraints set for the optimization model include: duration boundary constraints, which require the duration of each working condition cluster to be within the upper and lower limits of historical statistical values; working condition transition constraints, which require the transition between adjacent working condition clusters to satisfy that the Markov transition probability in the historical working condition sequence is greater than a preset lower limit; and process completion constraints, which include at least one of the following: total planned duration constraint, endpoint temperature constraint, output constraint, and termination working condition constraint.

[0049] By using a genetic algorithm to solve the optimization model, the combination of target durations for each operating condition cluster that minimizes energy consumption per unit product and the corresponding optimal operating condition cluster transfer path are obtained.

[0050] For the weighted activation durations of the five generated smelting condition clusters, time is constructed as a five-dimensional input feature column vector, and the actual unit product energy consumption of the corresponding batch, such as 552.4 kWh / t, is extracted as the regression target. The model is trained using a random forest regression algorithm, with the number of decision trees set to 300, the maximum split depth set to 15 layers, and bootstrap enabled (Bootstrap=True). After cross-validation convergence, the output is a score of feature importance obtained by accumulating the reduction in mean squared error caused by splitting each regression tree node, or a score of permutation importance obtained by observing the increment of the mean squared error of the validation set after shuffling a single feature, instead of using the Gini impurity reduction method in classification tasks. For example, the output consists of five floating-point numbers: 0.05, 0.15, 0.45, 0.25, and 0.10, with a sum of 1.

[0051] To ensure that the accurate energy consumption coefficient and the baseline energy consumption coefficient are on the same physical dimension, the operating condition clusters are first sorted and categorized according to their feature importance scores, and preset proportional correction factors are set for different categorizations. Specifically, operating condition clusters with high feature importance are classified into the high contribution categorization, and their proportional correction factors are set to values ​​greater than 1, such as 1.10 to 1.30; operating condition clusters with feature importance in the middle range are classified into the medium contribution categorization, and their proportional correction factors are set to values ​​close to 1, such as 0.95 to 1.05; operating condition clusters with low feature importance are classified into the low contribution categorization, and their proportional correction factors are set to values ​​less than 1 but greater than 0, such as 0.70 to 0.95. Subsequently, the baseline energy consumption coefficient of each operating condition cluster is multiplied by the proportional correction factor of the corresponding categorization to obtain the corresponding accurate energy consumption coefficient. Thus, the coefficient corresponding to the operating condition cluster with high feature importance can be appropriately increased, and the coefficient corresponding to the operating condition cluster with low feature importance can be appropriately decreased, while maintaining the physical dimension and non-negative properties of the accurate energy consumption coefficient.

[0052] When solving the optimization model, the following constraints are set: a 5×5 first-order Markov state transition matrix is ​​constructed by analyzing the time-series state records; the probability of a transition must be higher than 0.01, and transitions with physical limitations or extremely low probabilities are removed; the duration of each case cluster must fall within ±3 of a normal distribution. Within the window, for example, the refining condition cluster is limited to a minimum dwell time of 600 seconds and a maximum of 1200 seconds. This is based on the sample standard deviation of the duration of each operating condition cluster, obtained from historical operational data. Furthermore, to prevent the optimization results from collapsing only towards the lower limit of the duration of each operating condition cluster, process completion constraints are also set, including at least one of the following: total planned duration constraint, target temperature constraint, target output constraint, and termination condition constraint. For example, the sum of the target durations of all operating condition clusters must fall within the total planned duration. Within the allowable deviation range, the predicted endpoint temperature calculated based on the historical average heating contribution or effective heat input contribution of each operating condition cluster must reach the target furnace exit temperature range, the batch output must meet the lower limit of the production plan, and the terminal operating condition of the optimal path must be a termination operating condition that allows furnace exit or allows heat preservation. The initial population size of the genetic algorithm is set to 200, and iterative optimization is performed for up to 500 generations with a crossover probability of 0.8 and an adaptive dynamic mutation rate decreasing from 0.1 to 0.01, aiming to minimize the energy consumption per unit product. For example, the target energy consumption per unit product is minimized to 480 kWh / t. The output is the duration configuration scheme that minimizes the unit energy consumption, such as 900 seconds for the heating section and 750 seconds for the refining section, as well as the corresponding optimal operating condition cluster transfer path, such as the transfer path from heating cluster 1 to melting cluster 3 and then to heat preservation cluster 2.

[0053] Thus, by using a weighted calculation of the frequency of operating conditions and the energy consumption deviation, this embodiment of the invention can effectively reflect the time-varying energy consumption characteristics of complex operating conditions, improve the fitting accuracy of the precise energy consumption coefficient, and thus provide an accurate data foundation for the global optimization of the subsequent optimization model.

[0054] S3: Generate the control setting value sequence of the underlying device and send it down for execution.

[0055] For example, in this embodiment of the invention, the duration of each operating condition cluster is used as the decision variable, and the minimum energy consumption per unit product is used as the optimization objective. An optimization model is constructed that includes duration boundary constraints, operating condition transition constraints, and process completion constraints. The operating condition cluster transition path and the target duration of each operating condition cluster are solved, and the control setpoint sequence of the underlying equipment is generated and sent out for execution.

[0056] The objective function for the mathematical optimization problem is constructed as the sum of the dot product of the target duration and the corresponding precise energy consumption coefficient for each set of decision variables. Before calculating the objective function, the target duration of each set of working conditions is uniformly converted to hours so that the product of the target duration and the precise energy consumption coefficient has the meaning of energy consumption per unit product. A genetic algorithm is preferably used to solve the optimization model. Specifically, candidate solutions are encoded as chromosomes containing integer genes for working condition paths and real genes for durations. The integer genes for working condition paths represent the working condition cluster number selected for the l-th process segment, and the real genes for durations represent the target duration of that process segment. If the same working condition cluster appears multiple times in the path, the durations of each segment are summed to obtain the target duration of that working condition cluster.

[0057] In the genetic algorithm fitness calculation, firstly, the transition probability between integer genes in adjacent working condition clusters is checked against the Markov state transition matrix to see if it exceeds a preset lower limit. If not, a penalty term is applied to the candidate solution or path repair is performed. Secondly, it is checked whether the target duration of each working condition cluster is within the historical statistical upper and lower limits, and whether the sum of the durations of all segments meets the planned total duration range. Thirdly, based on historical data or process mechanisms, effective heat input contribution, heating contribution, heat preservation contribution, or output contribution parameters corresponding to each working condition cluster are pre-established, and the predicted endpoint temperature, predicted output, or process completion index corresponding to the candidate path is calculated. If the predicted endpoint temperature does not reach the target temperature range, the predicted output is lower than the planned output lower limit, or the process completion index is lower than the completion threshold, the candidate solution is determined not to meet the process completion constraint and a penalty is applied. Thus, the optimization model does not merely obtain a formally low energy consumption result by compressing the duration of each working condition cluster, but rather seeks the path and duration with the lowest energy consumption per unit product while completing the smelting production task.

[0058] In a specific implementation, the fitness function of the genetic algorithm can be composed of a target energy consumption per unit product and constraint penalty terms. The target energy consumption per unit product is the sum of the products of the target duration and the precise energy consumption coefficient for each operating condition cluster. The constraint penalty terms include penalties for illegal transfers, duration exceeding limits, total duration deviation, insufficient target temperature, insufficient target output, and illegal termination of operating conditions. The genetic algorithm sequentially performs population initialization, fitness calculation, selection, crossover, mutation, and elite retention operations, and finally outputs the optimal operating condition cluster transfer path and the target duration for each operating condition segment or each operating condition cluster.

[0059] Before sending the optimization results to the underlying equipment, a mapping table between operating condition clusters and underlying control parameters is established. This mapping table can be obtained by statistically analyzing the control setpoints corresponding to each operating condition cluster in historical stable active segments, including parameters such as set temperature, heating slope, fan pressure, gas flow rate, power supply, cooling water valve opening, and holding time. Specifically, the median, mean, or centroid inverse normalized value of each control variable within the stable active segment of the corresponding operating condition cluster can be used as the basic setpoint parameters for that operating condition cluster. After the genetic algorithm outputs the optimal transition path, the control unit reads the basic setpoint parameters of each operating condition cluster in sequence according to this transition path and generates a segmented control plan based on the target duration. If the set temperature, flow rate, or power change corresponding to adjacent operating condition clusters exceeds the equipment's allowable ramp rate, a linear ramp or speed-limiting transition segment is inserted between adjacent segments, thus forming an executable sequence of control setpoints.

[0060] A network connection is established with the field-programmable logic controller (PLC) using the OPCUA industrial communication protocol. The control setpoint sequence is written to the PLC's target node according to the register address mapping relationship. For example, the set temperature is written to the temperature control register, the fan pressure or gas flow rate to the flow control register, the valve opening to the valve execution register, and the heating power to the power control register. During execution, the field controller continues to read real-time temperature, pressure, power, and output feedback signals, and uses PID control or model predictive control to make small corrections to the segmented setpoints based on deviations, ensuring that the actual operating trajectory follows the optimal operating condition cluster transfer path and target duration. Thus, the path and duration output by the optimization model can be explicitly translated into control parameters such as temperature, flow rate, valve opening, and heating power that can be executed by the underlying equipment, completing a closed loop from optimization results to equipment control.

[0061] The ablation experiment used 3000 consecutive historical smelting batches of data from a large smelter as experimental conditions. Temperature, fan pressure, and power consumption of the underlying equipment were extracted to form a multidimensional time series, with the hardware sampling frequency uniformly configured at 2Hz. All batch data were divided into a training set containing 2500 batches and a test set containing 500 batches. The experiment set up three comparison dimensions: the basic control group used only conventional clustering and multiple linear regression for optimization; the partial ablation group added a stable activation state extraction mechanism based on normalized discreteness to the basic control group; and the complete scheme group covered all implementation steps, including time-weighted processing, random forest accuracy coefficient correction, and global optimization using a genetic algorithm with Markov state transition constraints, target duration boundary constraints, and process completion constraints.

[0062] The average absolute error of the unit product energy consumption prediction for the basic control group model was 35.2 kWh / t, and the actual average energy consumption of the test batches guided by this strategy was 556.4 kWh / t. For the partial ablation group, by eliminating abnormal fluctuations during the transition phase, the average absolute error of prediction was reduced to 24.5 kWh / t, and the average energy consumption after guiding production was reduced to 538.1 kWh / t. The average absolute error of prediction for the complete scheme group was reduced to 11.3 kWh / t. After implementing control based on the optimal transfer path and target duration output by the genetic algorithm in the test set, the measured average energy consumption of the batches reached a minimum of 482.6 kWh / t, achieving a total energy saving of over 30,000 kWh across 500 test batches.

[0063] Compared to the baseline control group, the complete solution reduced energy consumption per unit product by 73.8 kWh / t, demonstrating a significant improvement in energy efficiency. The core reason for this technological leap lies in the fact that this invention utilizes a time-weighting function to accurately map the nonlinear correlation between the frequency of occurrence of specific operating conditions in each process segment and the intensity of energy consumption fluctuations. Furthermore, it employs random forest feature importance to rigorously calibrate the coefficients of the energy consumption correlation model. The lower bound of Markov state transition probabilities, statistical boundaries of dwell time, and process completion constraints applied during the optimization process ensure that the output operating condition evolution path avoids invalid or high-energy-consuming illegal jumps. Simultaneously, it prevents the optimization result from failing to complete the smelting production task due to simply compressing the duration, thus guaranteeing the executability of the control strategy at the workshop level.

[0064] Thus, by collaboratively optimizing and constraining the evolution path of global operating conditions, the embodiments of the present invention can effectively avoid the ineffective dwelling of high-energy-consuming operating conditions, improve the physical feasibility of control commands, and thereby provide an accurate data foundation for the automated regulation of underlying equipment.

[0065] Example 2 Embodiment 2 of the present invention also discloses an intelligent energy consumption control and energy-saving optimization system for smelting processes, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement the intelligent energy consumption control and energy-saving optimization method for smelting processes provided by the present invention. Specifically, it includes the following modules: The calculation module is used to cluster the multidimensional process parameter data points of the smelting process to obtain multiple working condition clusters and their corresponding durations; based on the regression relationship between the duration of each working condition cluster and the energy consumption per unit product, the benchmark energy consumption coefficient of each working condition cluster is determined; the benchmark energy consumption coefficient is normalized and inversely proportionally mapped to generate the stability judgment threshold corresponding to each working condition cluster; the normalized dispersion of each data point relative to its working condition cluster is calculated; data points with normalized dispersion lower than the corresponding stability judgment threshold are judged as stable states, and continuous stable state time periods are extracted as stable activation segments. The extraction module is used to statistically analyze the frequency of occurrence of each operating condition cluster in different process periods and construct a time weighting function based on the corresponding energy consumption deviation. This function is used to weight and correct the duration of each stable activation segment to obtain the weighted activation duration. Based on the correlation between the weighted activation duration and the energy consumption per unit product, the accurate energy consumption coefficient of each operating condition cluster is determined. The optimization module is used to construct an optimization model that includes duration boundary constraints, operating condition transition constraints, and process completion constraints, with the duration of each operating condition cluster as the decision variable and the minimum energy consumption per unit product as the optimization objective. It solves for the operating condition cluster transition path and the target duration of each operating condition cluster, generates the control setpoint sequence of the underlying equipment, and issues it for execution.

[0066] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0067] The above are all preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent management and control of energy consumption and energy saving optimization of smelting process, characterized in that, include: S1, cluster the multidimensional process parameter data points of the smelting process to obtain multiple working condition clusters and their corresponding durations; Based on the regression relationship between the duration of each operating condition cluster and the energy consumption per unit product, the baseline energy consumption coefficient of each operating condition cluster is determined. The baseline energy consumption coefficient is then normalized and inversely proportionally mapped to generate the stability judgment threshold corresponding to each operating condition cluster. Calculate the normalized dispersion of each data point relative to its respective working condition cluster. Data points with normalized dispersion lower than the corresponding stability determination threshold are determined to be in a stable state, and continuous stable state time periods are extracted as stable activation segments. S2, Calculate the frequency of occurrence of each working condition cluster in different process periods and the corresponding energy consumption deviation to construct a time weight function, so as to weight and correct the duration of each stable activation segment to obtain the weighted activation duration; The precise energy consumption coefficient of each operating condition cluster is determined based on the correlation between weighted activation duration and unit product energy consumption; S3 uses the duration of each operating condition cluster as the decision variable and the minimum energy consumption per unit product as the optimization objective. It constructs an optimization model that includes duration boundary constraints, operating condition transition constraints, and process completion constraints. It solves for the operating condition cluster transition path and the target duration of each operating condition cluster, generates the control setpoint sequence of the underlying equipment, and issues it for execution.

2. The intelligent energy consumption control and energy-saving optimization method for smelting process according to claim 1, characterized in that, The multidimensional process parameter data points of the smelting process are clustered to obtain multiple operating condition clusters, including: Temperature, blower pressure, and power consumption of the smelting furnace were collected at a preset sampling frequency to obtain data points in the multidimensional time series. Outlier removal was performed on the multidimensional time series using the isolated forest algorithm, and the cleaned data was normalized using the Z-score normalization method. The normalized data was clustered based on the K-Means++ algorithm, and the number of cluster centers was determined by the silhouette coefficient. Each cluster center and its subordinate data were divided into multiple operating condition clusters.

3. The intelligent control and energy-saving optimization method for smelting process energy consumption according to claim 1, characterized in that, The determination of the baseline energy consumption coefficient for each operating condition cluster based on the regression relationship between the duration of each operating condition cluster and the energy consumption per unit product includes: The number of data points for each working condition cluster in each batch of smelting process is counted in chronological order, and the duration of each working condition cluster is obtained by multiplying the data points by the sampling period. The vector formed by the duration of each working condition cluster in each batch of smelting process is used as the independent variable, and the energy consumption per unit product of each batch is used as the dependent variable. A multiple linear regression algorithm with non-negative constraints is used to fit the linear mapping relationship between the independent and dependent variables to construct a preliminary correlation model. The partial regression coefficients of each independent variable in the preliminary correlation model are used as the benchmark energy consumption coefficients of the corresponding working condition clusters.

4. The intelligent energy consumption control and energy-saving optimization method for smelting process according to claim 3, characterized in that, The process involves calculating the normalized dispersion of each data point relative to its respective operating condition cluster, classifying data points with normalized dispersion below the corresponding stability threshold as stable states, and extracting continuous stable state time periods as stable activation segments, including: Calculate the Euclidean distance from each data point within the working condition cluster to the centroid of the working condition cluster, and then divide this Euclidean distance by the maximum Euclidean distance from the training samples within the working condition cluster to the centroid of the working condition cluster. After interval truncation, the normalized dispersion with a value between 0 and 1 is obtained. Traverse the working condition cluster sequence. When the normalized dispersion of N consecutive data points is less than the stability determination threshold, the N consecutive data points are determined to be in a stable state, where N is a preset positive integer greater than or equal to 5. Merge adjacent stable states that belong to the same working condition cluster and truncate them to form a continuous stable activation segment.

5. The intelligent energy consumption control and energy-saving optimization method for smelting process according to claim 1, characterized in that, The method of statistically analyzing the frequency of occurrence of each operating condition cluster in different process periods and constructing a time weighting function based on its corresponding energy consumption deviation includes: The smelting process cycle is divided into multiple equal-length process time segments. The frequency of occurrence of each condition cluster in each process time segment is counted in all batches to obtain the corresponding occurrence frequency. The absolute value of the difference between the average unit product energy consumption of the batch in which each condition cluster occurs in each process time segment and the overall average unit product energy consumption of all batches is calculated as the energy consumption deviation. The product of the occurrence frequency and the energy consumption deviation is normalized and fitted to generate a time weight function that varies with the process time.

6. The intelligent energy consumption control and energy-saving optimization method for smelting process according to claim 1, characterized in that, The method of determining the precise energy consumption coefficient of each operating condition cluster based on the correlation between weighted activation duration and unit product energy consumption includes: based on the operating data of each batch, using the weighted activation duration of each operating condition cluster as the input variable and the unit product energy consumption of the corresponding batch as the output variable, establishing a refined correlation model using the random forest regression algorithm, extracting the feature importance of each operating condition cluster, and obtaining the precise energy consumption coefficient by proportionally correcting the benchmark energy consumption coefficient using the feature importance, wherein the feature importance is used to characterize the strength of the contribution of the corresponding operating condition cluster to the unit product energy consumption prediction result.

7. The intelligent energy consumption control and energy-saving optimization method for smelting process according to claim 1, characterized in that, The duration boundary constraint requires that the duration of each working condition cluster be within the upper and lower limits of historical statistical values; the working condition transition constraint requires that the transition between adjacent working condition clusters satisfy that the Markov transition probability in the historical working condition sequence is greater than a preset lower limit; the process completion constraint includes at least one of the following: planned total duration constraint, endpoint temperature constraint, output constraint, and termination working condition constraint.

8. The method for intelligent control and energy-saving optimization of smelting process energy consumption according to claim 1, characterized in that, The optimization model is solved using a genetic algorithm, including: Candidate solutions are encoded as chromosomes containing operating condition paths and durations. In fitness calculations, candidates that violate constraints are penalized. After iteration, the combination of target durations for each operating condition cluster that minimizes energy consumption per unit product and the corresponding operating condition cluster transition paths are output.

9. The intelligent energy consumption control and energy-saving optimization method for smelting process according to claim 1, characterized in that, The generation and execution of the control setting value sequence for the underlying device includes: Based on the statistical values ​​of control parameters corresponding to each operating condition cluster in the historical stable activation segment, establish the mapping relationship between operating condition clusters and control parameters; read the control parameters of each operating condition cluster sequentially according to the operating condition cluster transfer path, and generate a segmented control plan in combination with the target duration; insert a transition segment for the section where the parameter jump between adjacent operating condition clusters exceeds the equipment's ramping capacity; and write the final control setpoint sequence into the programmable logic controller.

10. A smart energy consumption control and energy-saving optimization system for smelting processes, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the intelligent control and energy-saving optimization method for smelting process energy consumption according to any one of claims 1-9 is implemented.

Images