Converter steam prediction method and system based on time domain segmentation

Abstract

The invention discloses a converter steam prediction method and system based on time domain segmentation, and the method comprises the following steps: collecting production data of a steel mill, and carrying out the preprocessing; based on the process production condition and the time period of original production data, converter smelting is divided into a planning stage and a time sequence production stage; in the planning stage, in combination with steelmaking production planning information, multi-heat interval prediction is conducted on the steam yield at the time t before converter smelting; in the time sequence production stage, the change trend of the steam flow in different smelting modes and different production stages is captured, real-time dynamic prediction is conducted on the steam flow in combination with furnace time real-time data, and the global dependency relationship of the steam flow in the long-time range is obtained. By adopting the technical scheme, the staged data characteristics in the prediction of the steam recovery amount of the converter are fully excavated, and the comprehensiveness and perspectiveness of the prediction are improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of converter steelmaking, and relates to a converter steam prediction method and system based on time domain segmentation. BACKGROUND

[0002] Steelmaking is one of the largest industrial energy consumption sectors in the world, and it is highly dependent on energy and resources. As an important secondary energy source for steel enterprises, steam is an important medium for energy conversion such as electricity, power, and heat. As an important secondary energy source, recovered steam can not only be used to drive power equipment in the converter steelmaking process, but also can be used for heating or driving steam turbines to generate electricity, thereby achieving efficient utilization of waste heat. Converter steam waste heat recovery and reuse can significantly reduce production costs and improve energy utilization efficiency.

[0003] In the converter steelmaking process, steam is mainly generated through high-temperature reactions in the converter and a high-temperature flue gas cooling system. High-concentration oxygen is blown into the converter through an oxygen lance, causing oxidation reactions in the converter to release a large amount of heat energy. This heat is conducted to the cooling water through the water-cooled walls of the furnace body, causing some of the cooling water to evaporate and generate steam. At the same time, high-temperature flue gas discharged from the converter enters the cooling system through a pipeline, and the flue gas exchanges heat with the cooling water in the cooling system. After the cooling water absorbs the heat from the flue gas, some of the cooling water evaporates to form steam. The steam generated by each converter flows through the flue gas pipeline to the corresponding steam drum for temporary storage. The steam in the steam drum is adjusted to ensure that the temperature and pressure meet the requirements of the heat accumulator, and then sent to the heat accumulator for storage.

[0004] However, since the converter steelmaking process includes multiple stages, each stage has its specific process parameters and requirements, and the steelmaking processes for different steel grades also differ, resulting in complex steelmaking processes. In addition, the rhythm of steelmaking production is affected by factors such as equipment status and raw material supply, resulting in unstable production rhythm, further increasing the difficulty of converter steam prediction. Existing research starts from historical steam data and does not consider the stage-by-stage nature of converter steelmaking or combine with the production plan of the steelmaking plant. The method of directly predicting from historical data cannot fully capture the dynamic rules of steam generation throughout the process, limiting the accuracy and practical application value of the prediction model. Moreover, the rhythm of converter production is not considered, the complexity of steam fluctuations under different modes is not fully analyzed, and the effective capture of time series changes is lacking, thereby affecting the prediction accuracy. SUMMARY

[0005] The purpose of the present application is to provide a converter steam prediction method and system based on time domain segmentation, which fully excavates the stage-by-stage data features in converter steam recovery amount prediction, and improves the comprehensiveness and forward-looking nature of the prediction.

[0006] In order to achieve the above object, the basic scheme of the present application is: a converter steam prediction method based on time domain segmentation, comprising the following steps:

[0007] Collecting production data of the steel plant and preprocessing;

[0008] Based on the time period of process production and original production data, the converter smelting is divided into a planning stage and a time sequence production stage;

[0009] In the planning stage, the steam production of t time before the converter smelting is interval predicted for multiple heats in combination with the steelmaking production plan information;

[0010] In the time sequence production stage, the change trend of steam flow under different smelting modes and different production stages is captured, and the steam flow is dynamically predicted in real time in combination with the real-time data of the heat, so as to obtain the global dependence relationship of the steam flow in a long time range.

[0011] The working principle and beneficial effects of the basic scheme are that: based on the time period of process production and original data, the converter smelting is divided into a planning stage and a time sequence production stage, wherein the planning stage provides prospective guidance for steam scheduling by predicting the steam production in the next two hours. The time sequence prediction focuses on the steam flow at the current time and adjusts the steam supply strategy in real time.

[0012] In the planning stage, the steam of two hours before the converter smelting is interval predicted in combination with the steelmaking production plan information. In the time sequence production stage, considering that there are various smelting modes in the converter smelting process, the steam fluctuation characteristics under different modes are significantly different, the local feature extraction capability of the steam flow under different smelting modes is enhanced, the change trend of the steam flow under different modes is accurately captured, the global dependence relationship of the steam flow in a long time range is effectively captured, and accurate prediction of the steam flow is realized.

[0013] The technical scheme can fully exploit the stage data features in the converter steam recovery prediction, and improves the comprehensiveness and prospectiveness of the prediction. It provides more accurate data support and decision basis for the scheduling optimization and energy efficiency improvement of the converter steam system.

[0014] Further, in the planning stage, the Black Kite Algorithm BKA is improved by using Latin Hypercube Sampling, Reverse Difference Mutation and Gaussian Cuckoo Mutation Distribution, and in combination with the prediction method, the steam production in the future t time is predicted to provide prospective guidance for steam scheduling;

[0015] In the time sequence production stage, dynamic snake convolution DSC and Transformer model are introduced for collaborative modeling, the steam flow at the current time is focused on, and the steam supply strategy is adjusted in real time.

[0016] The IBKA algorithm improves search comprehensiveness and reduces blind spots by introducing Latin hypercube sampling to ensure a uniform population distribution. In the local search, inverse differential mutation is used to enhance population diversity and avoid premature convergence. In the later stages of convergence, Gaussian cuckoo mutation is introduced to enhance global search capabilities and ensure the quality of the optimal solution.

[0017] In the time-series production stage, considering that there are multiple smelting modes in the converter steelmaking process, and that the steam fluctuation characteristics of different modes are significantly different, Dynamic Snake Convolution (DSC) and Transformer model are introduced for collaborative modeling. This enhances the ability to extract local features of steam flow under different smelting modes and accurately captures the changing trend of steam flow under different modes. Combined with Transformer model, the global dependency of steam flow over a long period of time is effectively captured, and accurate prediction of steam flow is achieved.

[0018] Furthermore, in the initialization population phase of the Black-winged Kite Algorithm (BKA), Latin hypercube sampling is used instead of random initialization. The specific method is as follows:

[0019] The population is randomly initialized by sampling from a uniformly distributed Latin hypercube in the interval (0,1), and the fitness value is calculated. :

[0020] ,

[0021] Where ub and lb are the upper and lower boundaries of the search range, respectively; lhs is based on the Latin hypercube sampling function and returns a 1×d matrix, where d is the dimension of the variable.

[0022] The IBKA algorithm improves the comprehensiveness of the search and reduces blind spots by introducing Latin hypercube sampling to ensure a uniform distribution of the population.

[0023] Furthermore, a reverse differential mutation is introduced into the prey attack phase of the Blackwing Kite Algorithm (BKA), specifically:

[0024] Introducing reverse individuals For the current individual Its reverse individual for:

[0025] ,

[0026] in, and These represent the maximum and minimum values ​​of the current population in each dimension, respectively.

[0027] Differential mutation is introduced before the attack, and the search direction is adjusted by using information from multiple random individuals to calculate new candidate solutions. :

[0028] ,

[0029] in, , , These are three different individuals randomly selected from the population, where F is the mutation scaling factor, which controls the intensity of individual mutations.

[0030] Employing a fitness-driven update mechanism, for the current individual Reverse individuals and mutated individuals The evaluation is performed, and the optimal individual is selected for updating based on the fitness function:

[0031] .

[0032] In local searches, inverse differential mutation is used to enhance population diversity and avoid premature convergence.

[0033] Furthermore, a Gaussian cuckoo mutation distribution is introduced in the later convergence stage of the Black-winged Kite Algorithm (BKA). The specific steps are as follows:

[0034] Gaussian mutation optimizes the quality of the final solution by applying small perturbations around the individual, enabling the individual to perform a fine search in a local region.

[0035] ,

[0036] in, For the mutated individual, For the current individual, Let N(0,1) be the intensity of variation, and let N(0,1) be a random variable that follows a standard normal distribution.

[0037] Introduce the Lévy flight variant to enhance the population's global search capabilities:

[0038] ,

[0039] in, It follows a Lévy distribution, which can generate long jump steps, allowing individuals to escape local optima.

[0040] In the later stages of convergence, a Gaussian cuckoo mutation is introduced to enhance the global search capability and ensure the quality of the optimal solution.

[0041] Furthermore, in the time-series production stage, dynamic serpentine convolution DSC is introduced, with the following specific steps:

[0042] An offset is introduced. For input features X∈R^B×C×L, an offset is generated through convolution. The offset adjusts the sampling position of the convolution operation, where R represents the real number field, B represents the number of input samples at one time, C represents the number of channels, and L represents the number of time steps or spatial locations in each sample.

[0043] ,

[0044] Where k is the kernel size, and dynamic convolution is:

[0045] ,

[0046] Where i represents the output position, m represents the convolution kernel position, K represents the dynamic weights, and Δi m Y represents the offset of the convolution kernel at position m; Y(i) represents the result after dynamic convolution calculation. Represents the weight values ​​of the dynamic convolution kernel;

[0047] In each stage of steam generation, DSC captures the complex dynamic features of the data through multi-view feature extraction. DSC uses a small convolution kernel or a small range of offset to capture short-term fluctuations, while DSC uses a larger convolution kernel or adaptive offset to capture features that reflect long-term trends.

[0048] By introducing dynamic serpentine convolution and adaptively adjusting the kernel size, the ability to extract local features of steam flow under different smelting modes is enhanced, and the changing trend of steam flow under different modes is accurately captured.

[0049] Furthermore, dynamic serpentine convolution DSC is introduced to co-model with the Transformer model, resulting in the DSC-Transformer model. The specific steps are as follows:

[0050] By utilizing the DSC layer, local dynamic changes in data can be captured at different time scales, adapting to frequent fluctuations and trends in time series.

[0051] By dynamically adjusting the convolution kernel, the DSC layer abstracts from low frequency to high frequency layer by layer, extracts features from the sequence from coarse-grained to fine-grained, identifies dynamic change features in time series data, and integrates these features into a fused feature vector.

[0052] The encoder is configured to convert dynamic features into hidden representations, capturing temporal context information;

[0053] The encoder includes multiple feature extraction layers, with the number of layers decreasing one by one, to abstract the input features layer by layer. Each layer includes several convolutional networks, pooling layers and residual modules. By setting different convolutional kernel size and convolutional stride parameters, data features are gradually abstracted and extracted.

[0054] The decoder recovers the original data features and location information to solve the problems of localization and detail restoration in time series. The decoder consists of multiple layers with the layer size increasing layer by layer. The size of the encoder output is restored by upsampling and directional convolution layer by layer.

[0055] Multi-view feature fusion is performed on the features output by the encoder and decoder to integrate dynamic information at different levels and scales.

[0056] The ReLU activation function is used to perform nonlinear processing on the fused features, and the predicted value at each time step is calculated.

[0057] By introducing Dynamic Snake Convolution (DSC) and the Transformer model, the global dependency of steam flow over a long period of time can be effectively captured, enabling accurate prediction of steam flow.

[0058] The present invention also provides a converter steam prediction system based on time-domain segmentation, including a data acquisition unit and a processing unit. The data acquisition unit is used to collect production data from the steel plant and transmit it to the processing unit.

[0059] The processing unit includes a planning stage prediction model and a time-series prediction model. The processing unit executes the method described in this invention. The planning stage prediction model is used to perform interval prediction of steam output for multiple furnace cycles at time t before converter smelting. The time-series prediction model is used to obtain the global dependency of steam flow over a long period of time, thereby realizing converter steam prediction.

[0060] The predictive model constructed by this system can accurately predict the fluctuation trend of steam recovery in each furnace over a period of time, and can also dynamically predict the steam flow rate in each furnace production process in real time. This provides data support for the dynamic balance of the steelmaking steam network, ensures the safe operation of the system, and improves energy utilization efficiency.

[0061] Furthermore, the prediction model for the planning phase includes an IBKA layer, an optimization model hyperparameter layer, an SVM layer, and an interval prediction layer connected in sequence.

[0062] After the planned data is input, it first passes through the IBKA layer, which includes operations such as Latin hypercube sampling, Latin hypercube mutation, reverse differential evolution, and Gaussian cuckoo, used to preprocess the planned data and perform feature extraction.

[0063] The hyperparameter optimization layer optimizes the hyperparameters of the model to improve model performance.

[0064] The optimized data is passed to the SVM layer, where the support vector machine algorithm is used for learning and modeling.

[0065] The output of the SVM layer is fed into the interval prediction layer to obtain the final prediction interval.

[0066] The planning-stage prediction model can accurately predict the fluctuation trend of steam recovery volume per furnace over a period of time in the future, which is beneficial to use.

[0067] Furthermore, the time-series prediction model includes a DSC layer, an encoder, a decoder, and a ReLU activation function layer connected in sequence;

[0068] The DSC layer is used to capture local dynamic changes in data at different time scales, adapting to frequent fluctuations and trends in time series.

[0069] The encoder converts dynamic features into hidden representations and captures temporal context information. The encoder includes multiple feature extraction layers, with the number of layers decreasing one by one, to abstract the input features layer by layer. Each layer includes several convolutional networks, pooling layers and residual modules. By setting different convolutional kernel sizes and convolutional stride parameters, data features are gradually abstracted and extracted.

[0070] The decoder is used to recover the original data features and location information. The decoder consists of multiple layers, with the layer size increasing layer by layer. The size of the encoder output is restored by layer-by-layer upsampling and directional convolution.

[0071] The ReLU activation function is used to perform nonlinear processing on the fused features and calculate the predicted value at each time step.

[0072] The time-series prediction model can dynamically predict the steam flow rate of each furnace production process in real time, ensuring long-term prediction accuracy. Attached Figure Description

[0073] Figure 1 This is a flowchart illustrating the converter steam prediction method based on time-domain segmentation of the present invention.

[0074] Figure 2 This is a flowchart illustrating the improved BKA algorithm of the converter steam prediction method based on time-domain segmentation in this invention.

[0075] Figure 3 This is the dynamic adaptation graph of the dynamic serpentine convolution DSC of the converter steam prediction method based on time-domain segmentation in this invention;

[0076] Figure 4 This is a flowchart of the multi-view feature fusion of the dynamic serpentine convolution DSC based on the converter steam prediction method of the present invention.

[0077] Figure 5This is a schematic diagram of the DSC-Transformer model of the converter steam prediction method based on time-domain segmentation of the present invention;

[0078] Figure 6 This is a schematic diagram of the planning stage prediction model of the converter steam prediction system based on time-domain segmentation of the present invention. Detailed Implementation

[0079] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0080] In the description of this invention, it should be understood that the terms "longitudinal", "lateral", "up", "down", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0081] In the description of this invention, unless otherwise specified and limited, it should be noted that the terms "installation", "connection" and "linking" should be interpreted broadly. For example, they can refer to mechanical or electrical connections, or internal connections between two components. They can be direct connections or indirect connections through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0082] To address the complexity and dynamism of steam demand changes during converter production, this invention discloses a converter steam forecasting method based on time-domain segmentation. It fully considers the role of forecasting in enterprise regulation during the planning stage, as well as the impact of different smelting modes and steelmaking stages on steam fluctuations during the time-series production stage. Different deep learning methods are used at different stages to adapt to different data characteristics and forecasting needs, thereby improving the accuracy and reliability of steam demand forecasting.

[0083] In converter steelmaking, steam generation is concentrated in the oxygen blowing stage, accounting for over 95% of the total steam volume. Only a small amount of steam is generated in the slag splashing stage, and almost no steam is released in the remaining stages. Therefore, when forecasting steam, modeling and predicting only the steam during the oxygen blowing stage can effectively characterize the steam variation trend throughout the entire converter steelmaking process, meeting the practical needs of steam forecasting. This approach improves prediction accuracy, reduces noise interference, and makes the model more targeted, thereby enhancing the accuracy and stability of the predictions. It also reduces computational costs, minimizes redundant calculations and data processing, improves computational efficiency, and enables the prediction model to respond more quickly to actual production needs.

[0084] like Figure 1 As shown, the converter steam prediction method based on time-domain segmentation includes the following steps:

[0085] Production data from steel mills is collected and preprocessed. However, production data obtained from steel mill sites is often subject to quality issues such as missing data, inconsistent data, duplicate data, and inconsistent features due to limitations in data collection equipment, transmission links, and other factors.

[0086] Data quality has a significant impact on model training. Therefore, it is necessary to preprocess data with poor quality before model training, such as cleaning up missing values, outliers, and duplicate values, to improve the quality of the dataset.

[0087] In addition to common data problems, industrial field data such as steel plants also suffer from data inconsistency issues. This means that the data features required by the model are not in the same table, and industrial data from different stages are stored in different tables. However, when training the model, it is necessary to align the data from different time dimensions.

[0088] Therefore, it is necessary to preprocess the data to obtain reliable model learning samples. The historical data of the converter contains detailed records of raw material composition, process parameters, and other data before and after steelmaking at different stages, providing rich data support for model training.

[0089] Based on the process production situation and the time-dependent nature of the original production data, converter smelting is divided into the planning stage and the time-series production stage;

[0090] During the planning phase, combined with steelmaking production plan information, the improved BKA algorithm (IBKA) and support vector machine (SVM) model are used to make interval predictions of steam output at time t before converter smelting for multiple heats, providing forward-looking guidance for steam scheduling.

[0091] During the time-series production phase, the changing trends of steam flow under different smelting modes and production stages are captured. Combined with real-time furnace data, real-time dynamic prediction of steam flow is performed to obtain the global dependency of steam flow over a long period. The steam supply strategy is adjusted in real-time based on the current steam flow rate.

[0092] Compared to traditional single-stage data prediction methods, this invention can fully exploit the stage-specific data characteristics in converter steam recovery prediction, improving the comprehensiveness and foresight of the prediction. This provides more accurate data support and decision-making basis for the scheduling optimization and energy efficiency improvement of the converter steam system.

[0093] In a preferred embodiment of the present invention, during the planning stage (2 hours before production), such as Figure 2 As shown, the Black-winged Kite Algorithm (BKA) is improved using Latin hypercube sampling, inverse difference mutation, and Gaussian cuckoo mutation distribution. Combined with prediction methods (such as Support Vector Machine (SVM) or neural networks; taking SVM as an example, the improved BKA algorithm is first used to optimize the hyperparameters of the SVM (such as the regularization parameter C and the Gaussian kernel (RBF) parameter gamma), and then the optimized parameter combination is passed to the SVM model, which then uses the input parameters to predict the data), predicting the steam production within the next time interval t. This provides forward-looking guidance for steam scheduling, optimizes production planning and energy scheduling, and ensures that steam supply matches demand.

[0094] The input features of the SVM model include production planning information such as steel grade (STEEL_GRADE), molten steel quantity (IRON_WEIGHT), oxygen blowing time (OXYGEN_TIME), and converter processing number (STATION_ID).

[0095] In the planning stage of converter steelmaking, the data input features only include four characteristics: converter processing number, steel grade, planned oxygen blowing time, and planned molten iron weight. The data features have low dimensionality and exhibit stable changes. These characteristics provide ideal conditions for the application of Support Vector Machines (SVM).

[0096] SVM exhibits significant advantages on low-dimensional data, enabling effective learning on smaller training datasets and avoiding overfitting issues that can arise from highly complex models. Secondly, by introducing a tolerance band, SVM enhances the model's tolerance to small errors, demonstrating excellent handling capabilities for occasional outliers during the converter planning phase. Furthermore, although the data features during the converter planning phase are relatively simple, implicit nonlinear relationships may still exist. The SVM kernel function can map low-dimensional data to a high-dimensional feature space, thereby capturing potential nonlinear relationships. This capability allows SVM to effectively reveal potentially complex patterns even when processing simple low-dimensional data, and further improves model accuracy through prediction using the optimal regression hyperplane in the high-dimensional space. Therefore, SVM demonstrates high efficiency, stability, and strong nonlinear handling capabilities in low-dimensional data regression prediction during the converter steelmaking planning phase, fully showcasing its unique advantages.

[0097] In the time-series production stage, dynamic serpentine convolutional DSC is introduced in collaboration with the Transformer model to focus on the current steam flow rate and adjust the steam supply strategy in real time. Through the deformable convolution kernel of the dynamic serpentine convolution, the changes in time-series data under different smelting modes and production stages are accurately and dynamically learned. Combined with real-time furnace data, real-time dynamic prediction of steam flow rate is achieved. This enables precise adjustment of steam supply, optimization of parameter control (temperature, pressure, etc.) in the steelmaking process, timely warnings of anomalies (such as excessive steam recovery exceeding the accumulator's storage capacity, requiring advance steam delivery), reduction of energy waste, and improvement of overall production efficiency.

[0098] Experimental results show that the prediction accuracy (R²) of the IBKA-SVM model reaches 92.3% in the planning phase, and the prediction accuracy of the DSC-Transformer model is further improved to 98.7% in the time-series production phase. This invention overcomes the limitation of existing technologies where planning forecasting and time-series forecasting are disconnected, and constructs a dynamic correlation mechanism between the two.

[0099] Specifically, the macroscopic, coarse-grained forecasts in the planning phase provide a reliable initial framework and constraints for the microscopic, fine-grained forecasts in the time-series phase; while the time-series phase utilizes its high-frequency real-time data to dynamically correct and calibrate the forecast results from the planning phase. This two-way interactive mechanism enables the overall forecasting system to possess not only the stability of macroscopic planning but also the sensitivity of microscopic response, thereby achieving a qualitative leap in the accuracy of forecasting time-series data (such as instantaneous steam production and energy demand fluctuations).

[0100] Thanks to the significant improvement in prediction accuracy during the timing phase, it is possible to accurately anticipate real-time trends in the near future, laying a solid foundation for closed-loop online systems. Existing technologies often suffer from gaps between prediction and control, making true adaptive optimization difficult. This invention, based on high-precision input curves of steam pressure, temperature, and liquid level at the second level, outputs steam flow rates at the second level for the next two minutes in real time, achieving forward-looking prediction and providing reliable, advanced decision-making support for online control.

[0101] In a preferred embodiment of the present invention, the Black-winged Kite Algorithm (BKA) mainly includes four stages: population initialization, cruising search (global exploration), prey attack (local exploitation), and information sharing (accelerated convergence). In the population initialization stage of the Black-winged Kite Algorithm (BKA), Latin hypercube sampling is used instead of random initialization. Specifically, the method is as follows:

[0102] The population is randomly initialized by sampling from a uniformly distributed Latin hypercube in the interval (0,1), and the fitness value is calculated. :

[0103] ,

[0104] Where ub and lb are the upper and lower boundaries of the search range, respectively; lhs is based on the Latin hypercube sampling function and returns a 1×d matrix, where d is the dimension of the variable.

[0105] Latin hypercube sampling (LHS) possesses excellent full coverage and equal probability sampling properties, resulting in a more uniform sample distribution compared to random sampling. Therefore, using LHS for population initialization in optimization algorithms helps reduce the imbalance in the initial population distribution, thereby decreasing the likelihood of the algorithm getting trapped in local optima in the early stages and improving global search capabilities. Furthermore, the broad distribution of the population accelerates the exploration of optimal solution regions in the search space, improving the efficiency of the algorithm's convergence to the optimal solution.

[0106] In a preferred embodiment of the present invention, a reverse differential mutation is introduced in the prey attack phase of the Blackwing Kite Algorithm (BKA), specifically as follows:

[0107] To ensure that individuals can explore unexplored areas of the search space, reverse individuals are introduced. For the current individual Its reverse individual for:

[0108] ,

[0109] in, and These represent the maximum and minimum values ​​of the current population in each dimension, respectively; this formula ensures that each individual can explore its reverse position in the search space, thus enhancing the coverage of the population.

[0110] To further enhance information exchange within the population, differential mutation is introduced before the attack behavior. This involves adjusting the search direction based on information from multiple random individuals and calculating new candidate solutions. :

[0111] ,

[0112] in, , , These are three different individuals randomly selected from the population. F is the mutation scaling factor, which controls the intensity of individual mutation. This allows individuals to no longer converge solely towards the global optimum, but to learn from the search information of other individuals, thereby expanding the search range and avoiding premature convergence to a local optimum.

[0113] To further improve the adaptability of the search, a fitness-driven update mechanism is adopted to update the current individual. Reverse individuals and mutated individuals The evaluation is performed, and the optimal individual is selected for updating based on the fitness function:

[0114] .

[0115] In a preferred embodiment of the present invention, in order to further improve the global search capability of the algorithm and prevent individuals from getting trapped in local optima in the later stages of optimization, a Gaussian cuckoo mutation distribution is introduced in the later convergence stage of the Black-winged Kite Algorithm (BKA). The specific steps are as follows:

[0116] Gaussian mutation optimizes the quality of the final solution by applying small perturbations around the individual, enabling the individual to perform a fine search in a local region.

[0117] ,

[0118] in, For the mutated individual, For the current individual, The variation intensity is N(0,1), which is a random variable that follows a standard normal distribution; this can prevent individuals from stagnating in the later stages of the search and improve the search ability of the population.

[0119] Relying solely on local perturbations may not be sufficient to effectively escape local optima; therefore, the Lévy flight mutation is introduced to enhance the population's search capability globally.

[0120] ,

[0121] in, The Lévy distribution generates long jump steps, allowing individuals to escape local optima. The synergistic effect of Gaussian and Lévy mutations enables the algorithm to maintain search activity in the later stages of convergence, enhancing global search capabilities and effectively addressing the problem of optimization stagnation in the later stages, thus improving the quality of the final solution.

[0122] In the final stage of the algorithm, Gaussian cuckoo mutation is introduced, which is combined with Gaussian perturbation mutation and Lévy flight mutation. These are used for local search fine-tuning and large-scale jumps, respectively, to enhance search diversity, improve the ability to escape local optima, and improve global search efficiency.

[0123] In a preferred embodiment of the present invention, dynamic serpentine convolution allows the shape and size of the convolution kernel to be dynamically adjusted according to the characteristics of the input data, and can flexibly adjust the sensing range according to different stages of steam data change in order to capture information at different scales.

[0124] At different stages of smelting, the dynamic convolution kernel will adaptively adjust according to the data fluctuation characteristics, ensuring that the model uses a larger range of convolution kernels to capture trends in the high steam fluctuation stage, while using a small scale convolution kernel to improve resolution in the low fluctuation stage, paying more attention to small changes, meeting the prediction needs of different stages, and enabling it to better adapt to the characteristics of the input data, and flexibly capture complex structures and diverse features.

[0125] In the sequential production stage, such as Figure 3 As shown, the dynamic snake-shaped convolution DSC is introduced, and the specific steps are as follows:

[0126] To allow the convolution kernel to more flexibly focus on the complex geometric features of the target, a deformation offset is introduced. For input features X∈R^B×C×L, the offset is generated through convolution. The offset adjusts the sampling position of the convolution operation, where R represents the real number field, B represents the number of input samples at one time, C represents the number of channels, and L represents the number of time steps or spatial locations in each sample.

[0127] ,

[0128] Where k is the kernel size, and dynamic convolution is:

[0129] ,

[0130] Where i represents the output position, m represents the convolution kernel position, K represents the dynamic weights, and Δi m Y represents the offset of the convolution kernel at position m; Y(i) represents the result after dynamic convolution calculation. Represents the weight values ​​of the dynamic convolution kernel;

[0131] In each stage of steam generation, DSC captures the complex dynamic features of the data through multi-view feature extraction. DSC uses small convolutional kernels or small-range offsets to capture short-term fluctuations, while using larger convolutional kernels or adaptive offsets to capture features reflecting long-term trends, such as... Figure 4 As shown.

[0132] By dynamically adjusting the convolution kernel, DSC can better adapt to each stage of the converter steam generation process, enabling detailed learning of the steam generation characteristics at different stages. This improves the model's understanding and prediction capabilities of complex steam dynamics, providing more accurate data support for the converter production process.

[0133] Throughout the various stages of steam generation, DSC captures the complex dynamic features of the data through multi-view feature extraction. For short-term fluctuations, DSC uses small convolution kernels or small-range offsets to accurately capture them, while for features reflecting long-term trends, DSC uses larger convolution kernels or adaptive offsets.

[0134] This allows DSC to flexibly adjust the position of the sensing field, observe steam flow data from multiple angles, and dynamically capture local and global changes in steam generation. By mapping time-series data to features from different perspectives, the model can more comprehensively extract different characteristics of the steam generation process, including local generation patterns, features at different time scales, and their dynamic changes, forming a rich feature representation.

[0135] To avoid network load and feature redundancy caused by multi-view feature extraction, random discarding and grouping strategies are introduced during the fusion process to prevent model overfitting, thereby effectively improving the accuracy and stability of steam time series prediction.

[0136] In a preferred embodiment of the present invention, dynamic serpentine convolution DSC is introduced to perform collaborative modeling with the Transformer model, such as... Figure 5 As shown, the specific steps to obtain the DSC-Transformer model are as follows:

[0137] Dynamic features in time series data can be extracted by using DSC. The DSC layer can capture local dynamic changes in data at different time scales to adapt to frequent fluctuations and trends in time series.

[0138] By dynamically adjusting the convolution kernel, the DSC layer abstracts from low frequency to high frequency layer by layer, extracts features from the sequence from coarse-grained to fine-grained, identifies dynamic change features in time series data, and integrates these features into a fused feature vector, providing rich time series information for subsequent encoder layers;

[0139] The encoder is configured to convert dynamic features into hidden representations, capturing temporal context information;

[0140] The encoder includes multiple feature extraction layers, with the number of layers decreasing one by one, to abstract the input features layer by layer. Each layer includes several convolutional networks, pooling layers, and residual modules. By setting different convolutional kernel sizes and convolutional stride parameters, data features are gradually abstracted and extracted. The encoder's structural design can ensure that the model gradually abstracts and integrates the local and global information of the data from the bottom layer to the top layer, making full use of the dynamic features provided by DSC.

[0141] To address the challenges of localization and detail restoration in time-series data, a decoder is used to recover the original data features and location information. The decoder consists of multiple layers, with each layer increasing in size. It restores the size of the encoder output through layer-by-layer upsampling and directional convolution. After upsampling, the data information is gradually recovered, and details are preserved. However, some location information may have been lost or shifted during the encoding process. Therefore, the decoder combines the dynamic features extracted by the encoder and DSC to enhance the fidelity and accuracy of location information, ensuring that the feature information of each location in the time-series data is restored and accurately located.

[0142] Multi-view feature fusion is performed on the features output by the encoder and decoder to integrate dynamic information at different levels and scales.

[0143] The ReLU activation function is used to perform nonlinear processing on the fused features, and the predicted value for each time step is calculated. Multi-view feature fusion ensures the model's balance in capturing short-term fluctuations and long-term trends, enabling the model to accurately predict dynamic changes in steam flow and providing more comprehensive and stable support for the final steam time series forecast.

[0144] After extracting features from multiple perspectives, the DSC integrates these features through a weighted fusion method, combining short-term fluctuation features with long-term trend features to form a complete feature representation. Furthermore, the weighted fusion strategy highlights and enhances important features, ensuring that the final feature representation more accurately reflects the dynamic changes in steam flow. This multi-perspective feature extraction and fusion mechanism enables the DSC to more comprehensively and accurately capture steam flow fluctuation patterns at various stages, enhancing the model's understanding of the complex dynamic characteristics of steam flow.

[0145] The overall model is mainly divided into two symmetrical parts: an encoder and a decoder, each containing multiple layers. In the encoder, dynamic serpentine convolutions extract local features to capture dynamic trends in the time series. As the network depth increases, the model can capture a wider range of temporal feature information. Shallow convolutions focus on local features and short-term changes, while deeper network layers extract global features and long-term trends. A multi-view feature fusion layer integrates features from different levels to ensure full utilization of various types of information, achieving optimal prediction results.

[0146] To comprehensively evaluate the model's performance, this invention employs multiple evaluation metrics, including the coefficient of determination (R-Square), mean absolute error, root mean square error, and mean absolute percentage error.

[0147] R², MAPE, MAE, and RMSE are commonly used metrics for evaluating the performance of regression models. R² measures the model's ability to explain the target value, ranging from 0 to 1; a value closer to 1 indicates a better fit to the data. MAPE is a metric for evaluating the accuracy of predictive models, particularly suitable for time series data prediction. It measures the model's predictive performance by calculating the percentage of absolute difference between predicted and actual values. MAE quantifies the mean absolute error between predicted and actual values, reflecting the degree of deviation in the model's predictions; a smaller MAE value indicates more accurate predictions. RMSE retains the original units of error, more intuitively representing the average difference between predicted and actual values. Combining these metrics allows for a comprehensive evaluation of the model's fit and predictive accuracy, helping to determine the model's applicability in different scenarios.

[0148] ,

[0149] ,

[0150] ,

[0151] ,

[0152] Through the above training process and evaluation methods, we can gain a detailed understanding of the model's learning and prediction performance when processing data in different time windows, thereby providing a scientific basis for the prediction and optimization of the converter steelmaking process.

[0153] To verify the performance of the IBKA algorithm, four standard test functions were selected from the 23 CEC tests, as shown in Table 1. The experiment was compared with the Snake Optimization Algorithm (SO), Particle Swarm Optimization (PSO), Goose Optimization Algorithm (GOOSE), Black-winged Kite Algorithm (BKA), and Subtractive Average Optimization Algorithm (SABO). The maximum number of iterations T was set to 500, the population size N to 30, and the dimension D to 30.

[0154] Table 1. Test Functions

[0155]

[0156] Table 2 presents the statistical results of the six algorithms after 10 runs, with bold indicating the optimal result. F2 and F4 are unimodal functions used to test the convergence speed and local search capability of the algorithms. Table 2 shows that, on the unimodal function, IBKA significantly outperforms the five swarm intelligence algorithms compared, and shows a significant improvement over BKA. Multimodal test functions are used to test the performance of the algorithms in solving complex multimodal optimization problems. Functions F8 and F10 are used to examine the algorithm's global search and ability to escape local optima. According to Table 2, compared to other algorithms, IBKA has the best parameters on both F8 and F10. This result indicates that IBKA has good global search capability and the ability to escape local optima.

[0157] Table 2. Optimization results of 6 algorithms

[0158]

[0159] The present invention also provides a converter steam prediction system based on time-domain segmentation, including a data acquisition unit and a processing unit. The data acquisition unit is used to collect production data from the steel plant and transmit it to the processing unit.

[0160] The processing unit includes a planning stage prediction model and a time-series prediction model. The processing unit executes the method described in this invention. The planning stage prediction model is used to perform interval prediction of steam output at time t before converter smelting for multiple furnace cycles. The time-series prediction model is used to obtain the global dependency of steam flow over a long period of time, thereby realizing converter steam prediction.

[0161] This invention combines DSC and Transformer for modeling steam time series data, integrating their respective advantages: the dynamic feature extraction of DSC and the powerful sequence modeling capabilities of Transformer. By utilizing the multi-scale features extracted by DSC at each time step, Transformer can further capture the complex dependencies in time series data, providing more comprehensive and accurate feature support for time series prediction.

[0162] In a preferred embodiment of the present invention, such asFigure 6 As shown, the prediction model in the planning phase includes an IBKA layer, an optimization model hyperparameter layer, an SVM layer, and an interval prediction layer connected in sequence.

[0163] After the planned data is input, it first passes through the IBKA layer, which includes operations such as Latin hypercube sampling, Latin hypercube mutation, reverse differential evolution, and Gaussian cuckoo, used to preprocess the planned data and perform feature extraction.

[0164] The hyperparameter optimization layer optimizes the hyperparameters of the model to improve model performance.

[0165] The optimized data is passed to the SVM layer, where the support vector machine algorithm is used for learning and modeling.

[0166] The output of the SVM layer is fed into the interval prediction layer to obtain the final prediction interval.

[0167] In a preferred embodiment of the present invention, the time-series prediction model includes a DSC layer, an encoder, a decoder, and a ReLU activation function layer connected in sequence.

[0168] The DSC layer is used to capture local dynamic changes in data at different time scales, adapting to frequent fluctuations and trends in time series.

[0169] The encoder converts dynamic features into hidden representations and captures temporal context information. The encoder includes multiple feature extraction layers, with the number of layers decreasing one by one, to abstract the input features layer by layer. Each layer includes several convolutional networks, pooling layers and residual modules. By setting different convolutional kernel size and convolutional stride parameters, data features are gradually abstracted and extracted.

[0170] The decoder is used to recover the original data features and location information. The decoder consists of multiple layers, with the layer size increasing layer by layer. The size of the encoder output is restored by layer-by-layer upsampling and directional convolution.

[0171] The ReLU activation function is used to perform non-linear processing on the fused features to calculate the predicted value at each time step.

[0172] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0173] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

[0174] The specific embodiments described herein are merely illustrative of the spirit of the invention. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to substitute them, without departing from the spirit of the invention or exceeding the scope defined by the appended claims.

Claims

1. A converter steam prediction method based on time-domain segmentation, characterized in that, Includes the following steps: Collect production data from steel mills and preprocess it; Based on the process production situation and the time-dependent nature of the original production data, converter smelting is divided into the planning stage and the time-series production stage; During the planning phase, based on steelmaking production plan information, the steam output at time t before converter smelting is predicted over multiple heats. During the time-series production phase, the changing trends of steam flow under different smelting modes and production stages are captured. Combined with real-time data of each furnace, the steam flow is dynamically predicted in real time, and the global dependence of steam flow over a long period of time is obtained.

2. The converter steam prediction method based on time-domain segmentation as described in claim 1, characterized in that, During the planning phase, the Black-winged Kite Algorithm (BKA) is improved using Latin hypercube sampling, inverse difference mutation, and Gaussian cuckoo mutation distribution. Combined with prediction methods, the steam production in the next time interval t is predicted, providing forward-looking guidance for steam scheduling. In the time-series production stage, dynamic snake convolution DSC is introduced to work in conjunction with the Transformer model to focus on the current steam flow rate and adjust the steam supply strategy in real time.

3. The converter steam prediction method based on time-domain segmentation as described in claim 2, characterized in that, In the initialization population phase of the Black-winged Kite Algorithm (BKA), Latin hypercube sampling is used instead of random initialization. The specific method is as follows: The population is randomly initialized by sampling from a uniformly distributed Latin hypercube in the interval (0,1), and the fitness value is calculated. : ， Where ub and lb are the upper and lower boundaries of the search range, respectively; lhs is based on the Latin hypercube sampling function and returns a 1×d matrix, where d is the dimension of the variable.

4. The converter steam prediction method based on time-domain segmentation as described in claim 3, characterized in that, In the prey attack phase of the Blackwing Kite Algorithm (BKA), a reverse differential mutation is introduced, specifically as follows: Introducing reverse individuals For the current individual Its reverse individual for: ， in, and These represent the maximum and minimum values ​​of the current population in each dimension, respectively. Differential mutation is introduced before the attack, and the search direction is adjusted by using information from multiple random individuals to calculate new candidate solutions. : ， in, , , These are three different individuals randomly selected from the population, where F is the mutation scaling factor, which controls the intensity of individual mutations. A fitness-driven update mechanism is adopted for the current individual. Reverse individuals and mutated individuals The evaluation is performed, and the optimal individual is selected for updating based on the fitness function: 。 5. The converter steam prediction method based on time-domain segmentation as described in claim 4, characterized in that, In the later stage of convergence of the Black-winged Kite Algorithm (BKA), a Gaussian cuckoo mutation distribution is introduced. The specific steps are as follows: Gaussian mutation optimizes the quality of the final solution by applying small perturbations around the individual, enabling the individual to perform a fine search in a local region. ， in, For the mutated individual, For the current individual, Let N(0,1) be the intensity of variation, and let N(0,1) be a random variable that follows a standard normal distribution. Introduce the Lévy flight variant to enhance the population's global search capabilities: ， in, It follows a Lévy distribution, which can generate long jump steps, allowing individuals to escape local optima.

6. The converter steam prediction method based on time-domain segmentation as described in claim 1, characterized in that, In the time-series production stage, dynamic snake-shaped convolution DSC is introduced, and the specific steps are as follows: An offset is introduced. For input features X∈R^B×C×L, an offset is generated through convolution. The offset adjusts the sampling position of the convolution operation, where R represents the real number field, B represents the number of input samples at one time, C represents the number of channels, and L represents the number of time steps or spatial locations in each sample. ， Where k is the kernel size, and dynamic convolution is: ， Where i represents the output position, m represents the convolution kernel position, K represents the dynamic weights, and Δi m Y represents the offset of the convolution kernel at position m; Y(i) represents the result after dynamic convolution calculation. Represents the weight values ​​of the dynamic convolution kernel; In each stage of steam generation, DSC captures the complex dynamic features of the data through multi-view feature extraction. DSC uses a small convolution kernel or a small range of offset to capture short-term fluctuations, while DSC uses a larger convolution kernel or adaptive offset to capture features that reflect long-term trends.

7. The converter steam prediction method based on time-domain segmentation as described in claim 6, characterized in that, The specific steps for introducing dynamic serpentine convolution DSC and the Transformer model for collaborative modeling to obtain the DSC-Transformer model are as follows: By utilizing the DSC layer, local dynamic changes in data can be captured at different time scales, adapting to frequent fluctuations and trends in time series. By dynamically adjusting the convolution kernel, the DSC layer abstracts from low frequency to high frequency layer by layer, extracts features from the sequence from coarse-grained to fine-grained, identifies dynamic change features in time series data, and integrates these features into a fused feature vector. The encoder is configured to convert dynamic features into hidden representations, capturing temporal context information; The encoder includes multiple feature extraction layers, with the number of layers decreasing one by one, to abstract the input features layer by layer. Each layer includes several convolutional networks, pooling layers and residual modules. By setting different convolutional kernel size and convolutional stride parameters, data features are gradually abstracted and extracted. The decoder recovers the original data features and location information to solve the problems of localization and detail restoration in time series. The decoder consists of multiple layers with the layer size increasing layer by layer. The size of the encoder output is restored by upsampling and directional convolution layer by layer. Multi-view feature fusion is performed on the features output by the encoder and decoder to integrate dynamic information at different levels and scales. The ReLU activation function is used to perform nonlinear processing on the fused features, and the predicted value at each time step is calculated.

8. A converter steam prediction system based on time-domain segmentation, characterized in that, It includes a data acquisition unit and a processing unit. The data acquisition unit is used to collect production data from the steel plant and transmit it to the processing unit. The processing unit includes a planning stage prediction model and a time-series prediction model. The processing unit executes the method described in any one of claims 1-7. The planning stage prediction model is used to perform interval prediction of steam output at time t before converter smelting for multiple furnace cycles. The time-series prediction model is used to obtain the global dependency of steam flow over a long period of time, thereby realizing converter steam prediction.

9. The converter steam prediction system based on time-domain segmentation as described in claim 8, characterized in that, The prediction model for the planning phase includes an IBKA layer, an optimization model hyperparameter layer, an SVM layer, and an interval prediction layer connected in sequence. After the planned data is input, it first passes through the IBKA layer, which includes operations such as Latin hypercube sampling, Latin hypercube mutation, reverse differential evolution, and Gaussian cuckoo, used to preprocess the planned data and perform feature extraction. The hyperparameter optimization layer optimizes the hyperparameters of the model to improve model performance. The optimized data is passed to the SVM layer, where the support vector machine algorithm is used for learning and modeling. The output of the SVM layer is fed into the interval prediction layer to obtain the final prediction interval.

10. The converter steam prediction system based on time-domain segmentation as described in claim 8, characterized in that, The time-series prediction model includes a DSC layer, an encoder, a decoder, and a ReLU activation function layer connected in sequence; The DSC layer is used to capture local dynamic changes in data at different time scales, adapting to frequent fluctuations and trends in time series. The encoder converts dynamic features into hidden representations and captures temporal context information. The encoder includes multiple feature extraction layers, with the number of layers decreasing one by one, to abstract the input features layer by layer. Each layer includes several convolutional networks, pooling layers and residual modules. By setting different convolutional kernel sizes and convolutional stride parameters, data features are gradually abstracted and extracted. The decoder is used to recover the original data features and location information. The decoder consists of multiple layers, with the layer size increasing layer by layer. The size of the encoder output is restored by layer-by-layer upsampling and directional convolution. The ReLU activation function is used to perform nonlinear processing on the fused features and calculate the predicted value at each time step.

Images