Production rhythm self-adaptive converter steelmaking oxygen consumption prediction method and system

Abstract

The invention provides a production rhythm self-adaptive converter steelmaking oxygen consumption prediction method and system, and belongs to the technical field of information.The method comprises the steps that only data of a converter steelmaking blowing period is adopted to train a Transform time sequence prediction model, and an initial oxygen consumption prediction model is obtained; a Bayesian optimization algorithm is adopted to optimize hyper-parameters of the initial oxygen consumption prediction model, and a pre-trained oxygen consumption prediction model is obtained; and finally, performing converter steelmaking oxygen consumption prediction by using the pre-trained oxygen consumption prediction model. By means of the method, the starting moment and the ending moment of blowing can be accurately predicted. Aiming at the problem that the prediction precision is reduced due to data quality change caused by steelmaking plan change, the method is automatically aligned with the steelmaking plan, the precision and aging requirements when the converter steelmaking oxygen consumption production rhythm changes are met, oxygen production can be guided, and the oxygen production cost can be reduced.

Description

Technical Field

[0001] This invention belongs to the field of information technology, specifically relating to a method and system for predicting oxygen consumption in converter steelmaking with adaptive production rhythm. Background Technology

[0002] Oxygen is an essential energy medium in steel production, mainly involved in oxygen-enriched combustion in blast furnaces and top blowing in converters. During normal production, the oxygen demand in blast furnaces remains almost constant, while converter steelmaking is an intermittent, quasi-cyclical production process. The production rhythm of converter steelmaking is easily affected by actual conditions such as steelmaking plans, process operations, and fault repairs.

[0003] Steelmaking processes can be divided into blowing and non-blowing periods. The blowing period has a relatively fixed length, ranging from 18 to 25 minutes depending on the converter model. Compared to the blowing period, the non-blowing period is longer and less variable, making it difficult to accurately predict the oxygen consumption of converter steelmaking, especially at the beginning and end of the blowing period. Due to the variable production rhythm of converter steelmaking, the total oxygen demand when multiple converters are operating simultaneously fluctuates greatly, making it difficult for dispatchers to adjust the production load of oxygen-generating equipment in a timely and effective manner, leading to insufficient oxygen supply or oxygen release.

[0004] Currently, the relevant research in domestic and international literature and patents on predicting oxygen consumption in steel production is as follows:

[0005] Zhao Jun et al. (A method for predicting oxygen load in steel enterprises based on production plan, application number: 202010205754.6) extracted the main features affecting oxygen consumption from actual industrial operation data, and trained a neural network prediction model for each converter based on this. The prediction results of a blowing stage output by the model were matched with the converter production plan to obtain the total oxygen load prediction value within a planned time period.

[0006] Wang Jianjun et al. (A method for controlling blast furnace oxygen consumption based on least squares support vector machine, application number: 201310117694.2) used the least squares support vector machine method to establish a blast furnace oxygen consumption prediction model and adjusted the blast furnace oxygen supply according to the prediction results of the model.

[0007] Qin Bo et al. (Qin Bo, Wu Qingchao, Zhang Juanjuan, Wang Jianguo, Zhang Wenxing. Research on oxygen consumption prediction in converter steelmaking based on PSO-optimized SVM [J]. Measurement and Control Technology, 2014, 33(12):121-124.) established a model for predicting oxygen consumption in converter steelmaking by using the penalty coefficient, insensitivity loss coefficient and Gaussian kernel width coefficient of PSO-optimized SVM. Experimental results show that the model has high accuracy and generalization ability.

[0008] Zhang Ziyang et al. (Zhang Ziyang, Sun Yangguang. Prediction of converter oxygen blowing amount based on grey Elman neural network [J]. Computer Applications and Software, 2018, 35(11):103-107.) established a converter oxygen blowing amount prediction model using grey Elman neural network, and solved the problem of the network easily getting trapped in local optima and overfitting by optimizing the weights and thresholds.

[0009] Among the aforementioned patents, there are numerous methods for predicting oxygen demand in blast furnaces, but fewer methods for predicting oxygen consumption in converter steelmaking. In applied technologies, the prediction models all use entire historical data periods as training data; however, when production rhythms change, the model's prediction accuracy decreases. Therefore, analyzing the converter blowing cycle is beneficial for improving prediction accuracy. Developing an adaptive oxygen consumption prediction method for converter steelmaking can provide predictive data support for scientific oxygen scheduling. Summary of the Invention

[0010] To address the shortcomings of existing technologies, this application proposes a method and system for predicting oxygen consumption in converter steelmaking with an adaptive production rhythm. Utilizing actual industrial production data, the method analyzes the converter blowing cycle, eliminating data from non-blowing periods, and only predicts the time-series oxygen consumption during the converter blowing period. The model is entirely based on an attention mechanism and employs Bayesian methods for hyperparameter optimization, enabling accurate prediction of the start and end stages of the blowing period. Aligning the prediction results with the steelmaking plan allows for accurate prediction of the oxygen demand of each converter over a future period, providing scientific data guidance for energy managers in developing scheduling plans.

[0011] Firstly, this application proposes a method for predicting oxygen consumption in converter steelmaking in steel enterprises with an adaptive production rhythm, including:

[0012] Obtain the oxygen consumption during the converter blowing period at the current moment;

[0013] Take the first I oxygen consumption data of the current converter blowing period and input it into the pre-trained oxygen consumption prediction model to obtain a prediction sequence of a specified length. The establishment process of the pre-trained oxygen consumption prediction model includes: using the historical converter steelmaking plan of each converter, the oxygen consumption data of the corresponding time period and the Transformer time series prediction model to establish an initial oxygen consumption prediction model, and using the Bayesian optimization algorithm to optimize the hyperparameters of the initial oxygen consumption prediction model to obtain the pre-trained oxygen consumption prediction model.

[0014] The predicted sequence is inversely normalized to obtain the predicted oxygen consumption of the corresponding converter in the next blowing cycle.

[0015] Align the predicted oxygen consumption for the next blowing cycle with the steelmaking plan to obtain the predicted oxygen consumption for the next blowing cycle.

[0016] The initial oxygen consumption prediction model is established using historical converter steelmaking plans for each converter, oxygen consumption data for corresponding time periods, and the Transformer time-series prediction model. This includes:

[0017] Obtain the historical steelmaking plans and oxygen consumption data for each converter during the corresponding time period, and perform timestamp alignment on the historical steelmaking plans and oxygen consumption data for each converter during the corresponding time period to obtain the aligned data.

[0018] The aligned data is preprocessed, retaining only the data from the blowing stage;

[0019] The preprocessed data is divided into a training set and a test set. The preprocessed data includes the input sequence and the output sequence at the corresponding time.

[0020] The Transformer time series prediction model was trained using the training set to obtain the initial oxygen consumption prediction model.

[0021] The Transformer temporal prediction model includes an encoder and a decoder, wherein the encoder includes a first input layer, N encoding layers and a first positional encoding layer;

[0022] The first input layer is connected to the first position coding layer, and the first position coding layer is connected to each coding layer in sequence;

[0023] Each coding layer includes: a first self-attention layer, a first normalization layer, a first feedforward connection layer, and a second normalization layer; the first self-attention layer is connected to the first normalization layer, the first normalization layer is connected to the first feedforward connection layer, and the first feedforward connection layer is connected to the second normalization layer.

[0024] The first input layer is used to extend the input sequence to two-dimensional data using a broadcast method;

[0025] The first position encoding layer is used to add position biases to each of the two-dimensional data after the input sequence is expanded;

[0026] Each coding layer is used to encode the output data of the first position coding layer using an attention mechanism;

[0027] The decoder includes: a second input layer, M decoding layers, a second positional coding layer, and an output layer;

[0028] The second input layer is connected to the second position coding layer, and the second position coding layer is connected to each coding layer in turn;

[0029] Each decoding layer includes: a second self-attention layer, a third normalization layer, a decoding attention layer, a fourth normalization layer, a second feedforward connection layer, and a fifth normalization layer; the second self-attention layer is connected to the third normalization layer, the third normalization layer is connected to the decoding attention layer, the decoding attention layer is connected to the fourth normalization layer, the fourth normalization layer is connected to the second feedforward connection layer, and the second feedforward connection layer is connected to the fifth normalization layer;

[0030] The second input layer is used to extend the output sequence to two-dimensional data using a broadcast method;

[0031] The second position encoding layer is used to add position biases to each of the two-dimensional data after the output sequence is expanded;

[0032] Each decoding layer is used to decode the output data of the second positional coding layer using an attention mechanism.

[0033] The output layer is used to output the decoded data.

[0034] The hyperparameters of the initial oxygen consumption prediction model include: the number of encoding layers, the number of decoding layers, the length of the input sequence, and the length of the output sequence.

[0035] Furthermore, the hyperparameters of the initial oxygen consumption prediction model are optimized using a Bayesian optimization algorithm to obtain a pre-trained oxygen consumption prediction model, including:

[0036] Step S2.5: Determine the search range of hyperparameters, wherein the search range of the number of encoding layers and the search range of the number of decoding layers are determined to be 1 to 10 layers, and the length of the input sequence and the length of the output sequence are determined to be an integer multiple of the longest blowing cycle;

[0037] Step S2.6: Define the objective function using the current combination of hyperparameters;

[0038] Step S2.7: Based on the Gaussian process, establish the probability distribution model of the objective function;

[0039] Step S2.8: Select the next set of hyperparameters using the acquisition function;

[0040] Step S2.9: Evaluate the next set of hyperparameters selected using the validation set;

[0041] Step S2.10: Update the probability distribution model of the objective function based on the current evaluation results, return to step S2.7 for iteration until the preset maximum number of iterations is met, and output the optimized hyperparameters.

[0042] The objective function is calculated as follows:

[0043]

[0044] Where θ is the current combination of hyperparameters, L val (θ) represents the error value of the oxygen consumption prediction model for each converter trained using the current hyperparameter combination on the validation set. Let be the predicted value of the oxygen consumption prediction model for the i-th sample. Let N be the actual oxygen consumption data corresponding to the i-th sample, and N be the number of samples in the validation set.

[0045] The probability distribution model of the objective function is calculated as follows:

[0046] L val (θ)|H~GP(m(θ),k(θ,θ'))

[0047] Among them, L val (θ) represents the error value of the oxygen consumption prediction model for each converter trained using the current hyperparameter combination on the validation set, m(θ) is the mean function of the Gaussian process of the current hyperparameter combination, θ' is the existing hyperparameter combination, k(θ,θ') is the covariance function between the current hyperparameter combination and the existing hyperparameter combination, H is the historical tuning result of the existing hyperparameter combination, and GP is the Gaussian process.

[0048] The acquisition function is calculated as follows:

[0049]

[0050] Where y represents the possible output value of the objective function under the hyperparameter combination θ. * p is the currently known optimal objective function value. M (y|θ) represents the probability distribution of the objective function at the hyperparameter combination θ. Let θ be the desired improvement value at the current hyperparameter combination θ.

[0051] The next set of selected hyperparameters is calculated as follows:

[0052]

[0053] Where Θ is the set of all possible combinations of hyperparameters, θ * For the next set of hyperparameters selected, EI y *(θ) represents the expected improvement at the current hyperparameter combination θ, and arg max is the value to be taken. The hyperparameter at its maximum value.

[0054] Secondly, this application proposes a production rhythm adaptive oxygen consumption prediction system for converter steelmaking in steel enterprises, comprising:

[0055] The data acquisition module is used to obtain the oxygen consumption during the converter blowing period at the current moment;

[0056] The oxygen consumption prediction module is used to take the first I oxygen consumption data of the current converter blowing period and input them into the pre-trained oxygen consumption prediction model to obtain a prediction sequence of a specified length. The pre-trained oxygen consumption prediction model is established by: using the historical converter steelmaking plan of each converter, the oxygen consumption data of the corresponding time period and the Transformer time series prediction model to establish an initial oxygen consumption prediction model, and using the Bayesian optimization algorithm to optimize the hyperparameters of the initial oxygen consumption prediction model to obtain the pre-trained oxygen consumption prediction model.

[0057] The prediction result output module is used to inversely normalize the prediction sequence to obtain the predicted oxygen consumption value for the corresponding converter in the next blowing cycle; and to align the predicted oxygen consumption value for the next blowing cycle with the steelmaking plan to obtain the predicted oxygen consumption result for the next blowing cycle.

[0058] Thirdly, this application proposes an electronic device, comprising: one or more processors, and a memory for storing instructions, which, when executed by the one or more processors, cause the one or more processors to execute the aforementioned method for predicting oxygen consumption in converter steelmaking for steelmaking with adaptive production rhythm.

[0059] Fourthly, this application proposes a computer-readable storage medium storing executable instructions that, when executed, cause a processor to perform the aforementioned method for predicting oxygen consumption in converter steelmaking for steelmaking with adaptive production rhythm.

[0060] Fifthly, this application proposes a computer program product, including a computer program or instructions, which, when executed by a processor, implements the aforementioned method for predicting oxygen consumption in converter steelmaking for steelmaking with adaptive production rhythm.

[0061] Beneficial effects:

[0062] This application proposes a production rhythm-adaptive method and system for predicting oxygen consumption in converter steelmaking. This method, based on attention mechanisms and Bayesian optimization, is a production rhythm-adaptive approach for predicting oxygen consumption in steel enterprises' converter steelmaking processes. It trains the time-series prediction model using only data from the converter steelmaking blowing period. The key hyperparameters of the prediction model are optimized using a Bayesian algorithm, enabling accurate prediction of the start and end times of blowing. To address the issue of reduced prediction accuracy due to data quality variations caused by changes in the steelmaking plan, this method automatically aligns with the steelmaking plan, ensuring that the model output matches the current production rhythm. This method meets the accuracy and timeliness requirements for oxygen consumption prediction in converter steelmaking when the production rhythm changes, and can guide oxygen production and reduce oxygen production costs. Attached Figure Description

[0063] Figure 1 Flowchart of a method for predicting oxygen consumption in converter steelmaking in an iron and steel enterprise with adaptive production rhythm according to an embodiment of this application.

[0064] Figure 2 A schematic diagram of the Transformer time-series prediction model structure in this application embodiment;

[0065] Figure 3 A schematic diagram illustrating the prediction results of an embodiment of this application;

[0066] Figure 4 This application presents a schematic diagram of a production rhythm adaptive oxygen consumption prediction system for converter steelmaking in steel enterprises. Detailed Implementation

[0067] The specific implementation methods of this application will be further described in detail below with reference to the accompanying drawings and embodiments.

[0068] To address the limitations of current prediction technologies, this application proposes a method and system for predicting oxygen consumption in converter steelmaking with an adaptive production rhythm, aiming to achieve accurate and reliable prediction of oxygen consumption in converter steelmaking. This method can provide data support for energy managers to formulate scheduling plans, thereby reducing oxygen emissions, stabilizing production supply, and lowering enterprise energy consumption.

[0069] Example 1:

[0070] This embodiment proposes a method for predicting oxygen consumption in converter steelmaking in steel enterprises with adaptive production rhythm, such as... Figure 1 As shown, it includes:

[0071] Step S1: Obtain the oxygen consumption during the converter blowing period at the current moment;

[0072] Step S2: Take the first I oxygen consumption data of the current converter blowing period and input it into the pre-trained oxygen consumption prediction model to obtain a prediction sequence of a specified length; the pre-trained oxygen consumption prediction model is established by: using the historical converter steelmaking plan of each converter, the oxygen consumption data of the corresponding time period and the Transformer time series prediction model to establish an initial oxygen consumption prediction model, and using the Bayesian optimization algorithm to optimize the hyperparameters of the initial oxygen consumption prediction model to obtain the pre-trained oxygen consumption prediction model;

[0073] Step S3: The predicted sequence is inversely normalized to obtain the predicted oxygen consumption value for the next blowing cycle of the corresponding converter;

[0074] Step S4: Align the predicted oxygen consumption for the next blowing cycle with the steelmaking plan to obtain the predicted oxygen consumption for the next blowing cycle.

[0075] The initial oxygen consumption prediction model is established using historical converter steelmaking plans for each converter, oxygen consumption data for corresponding time periods, and the Transformer time-series prediction model. This includes:

[0076] Step S2.1: Obtain the historical steelmaking plan and oxygen consumption data for each converter during the corresponding period, and perform timestamp alignment on the historical steelmaking plan and oxygen consumption data for each converter during the corresponding period to obtain the aligned data.

[0077] In this embodiment, the historical steelmaking plans of each converter and the oxygen consumption data of the corresponding time period are obtained. The original data is timestamped to ensure that each steelmaking plan corresponds to the oxygen consumption data of its actual occurrence stage, providing a reliable data foundation for subsequent modeling.

[0078] Step S2.2: Preprocess the aligned data, retaining only the data from the blowing stage;

[0079] In this embodiment, the aligned data is cleaned and integrated, completing preprocessing steps such as blowing cycle analysis, furnace division, and data removal from non-blowing stages. Input-output pairs are constructed based on the preprocessed data, and training, testing, and validation sets are created. The specific steps are as follows:

[0080] Step S2.2.1: Consistency check. Data outside the blowing period is set to 0. Data within the blowing period is processed as follows:

[0081] Step S2.2.2: Outlier Correction. Based on the maximum oxygen consumption of the converter, larger outliers are corrected. Data exceeding the upper limit of oxygen consumption is considered to be caused by instrument drift and is directly corrected to the upper limit of oxygen consumption.

[0082] Step S2.2.3: Missing value completion. For zero values ​​during the blowing period, fill in the missing values ​​by averaging the data before and after that value;

[0083] Step S2.2.4: Data Integration. Average the oxygen consumption per minute, transforming the data granularity from seconds to minutes.

[0084] Step S2.2.5: Data Normalization. Perform max-min normalization on the processed data;

[0085] Step S2.2.6: Dataset Partitioning. Based on the temporal characteristics of the converter steelmaking process, the sliding time window algorithm is used to map the data sequence into input-output pairs and divide it into training, test, and validation sets.

[0086] in:

[0087] (1) In step S2.2.6, before constructing the input-output pair, it is necessary to analyze the blowing time of each heat in the steelmaking plan. Assuming the maximum blowing time is T, the length O of the model output sequence is selected as T, and the length I of the input sequence is selected as 3-4 times the length of the output sequence.

[0088] (2) The specific implementation steps of constructing input-output pairs (input-output pairs include: input sequence and output sequence at corresponding time) using the sliding window algorithm in step S2.2.6 are explained as follows: Let the normalized data sequence be X = [x1, x2, ..., x N ], where N is the length of the sequence. The length L of the sliding window is set to L = 1 + 0. The sliding window is used to slice the data sequence X, and then the first 1 elements and the last 0 elements of the resulting sequence are sliced ​​to obtain N-L+1 input-output pairs, namely (X1, Y1), (X2, Y2), ... (X... N-L+1 ,Y N-L+1 ), where X i =[x i ,x i+1 ,...x i+I-1 ], (i = 1, 2, ..., N-L+1), Y i =[x i+I ,x i+I+1 ,...,x i+I+O-1 ], (i = 1, 2, ..., N-L+1);

[0089] (3) In step S2.2.6, all input-output pairs need to be integrated into batches of size 32, and all batches need to be divided into training, testing, and validation sets in a ratio of 8:1:1 for model training, testing, and validation. The model structure will then be described using the input-output pair (X1, Y1), and (X1, Y1) will be denoted as (X, Y), where X = [x1, x2, ..., x...]. I ], Y = [y1, y2, ..., y O ].

[0090] Step S2.3: Divide the preprocessed data into a training set and a test set. The preprocessed data includes the input sequence and the output sequence at the corresponding time.

[0091] Step S2.4: Train the Transformer time series prediction model using the training set to obtain the initial oxygen consumption prediction model.

[0092] In this embodiment, the structural parameters of the Transformer time-series prediction model corresponding to each converter are set and initialized, including defining core modules such as the encoder structure, decoder structure, and attention mechanism. The hyperparameters that need to be set in this embodiment include: the number of encoding layers E, the number of decoding layers D, the number of attention layers H, the dimension d of the positional encoding layer, the learning rate γ, the number of dataset iterations P, and the dropout rate θ. In this embodiment, when optimizing using the Bayesian algorithm later, it is not necessary to optimize all hyperparameters; only the number of encoding layers and decoding layers need to be optimized, as well as the length of the input sequence and the length of the output sequence.

[0093] Specifically, the Transformer time-series prediction model, such as Figure 2 As shown, it includes: an encoder and a decoder, wherein the encoder includes: a first input layer, N encoding layers and a first position encoding layer;

[0094] The first input layer is connected to the first position coding layer, and the first position coding layer is connected to each coding layer in sequence;

[0095] Each coding layer includes: a first self-attention layer, a first normalization layer, a first feedforward connection layer, and a second normalization layer; the first self-attention layer is connected to the first normalization layer, the first normalization layer is connected to the first feedforward connection layer, and the first feedforward connection layer is connected to the second normalization layer.

[0096] The first input layer is used to extend the input sequence to two-dimensional data using a broadcast method;

[0097] In this embodiment, the input sequence of the input layer is X = [x1, x2, ..., x...]. IAfter broadcasting, the output sequence of the input layer is [[x1,x1,...,x1],[x2,x2,...,x2],...[x I ,x I ,...,x I ]], where the length of the one-dimensional sequence is d.

[0098] The first position encoding layer is used to add position biases to each of the two-dimensional data after the input sequence is expanded;

[0099] In this embodiment, to enable the model to learn the positional information of data in the sequence, a bias is added to each sequence in the positional encoding layer; wherein, the formula for calculating the bias in the positional encoding layer is:

[0100] PE(p,2i)=sin(p / 10000 2i / d )

[0101] PE(p,2i+1)=cos(p / 10000 2i / d )

[0102] The input sequence of the positional coding layer is X = [[x1,x1,...,x1],[x2,x2,...,x2],...[x I ,x I ,...,x I The output sequence after positional encoding is Xp = [[xp1,xp1,...,xp1],[xp2,xp2,...,xp2],...[xp I XP I ,...,xp I ]).

[0103] Each encoding layer is used to encode the output data of the first position encoding layer using an attention mechanism; each decoding layer includes: a second self-attention layer, a third normalization layer, a decoding attention layer, a fourth normalization layer, a second feedforward connection layer, and a fifth normalization layer; the second self-attention layer is connected to the third normalization layer, the third normalization layer is connected to the decoding attention layer, the decoding attention layer is connected to the fourth normalization layer, the fourth normalization layer is connected to the second feedforward connection layer, and the second feedforward connection layer is connected to the fifth normalization layer;

[0104] In this embodiment, the attention layer encodes the input sequence, and the scaling dot product attention calculation formula is as follows:

[0105]

[0106] Where the values ​​of Q, K, and V are all [[xp1,xp1,...,xp1],[xp2,xp2,...,xp2],...[xpI XP I ,...,xp I The softmax calculation formula is as follows:

[0107]

[0108] Input sequence Xp = [[xp1,xp1,...,xe1],[xp2,xp2,...,xp2],...[xp I XP I ,...,xp I The encoded output sequence is Xe = [[xe] 11 xe 12 ,...,xe 1d ],[xe 21 ,xe 22 ,...,xe 2d ],...[xe I1 ,xe I2 ,...,xe Id ]).

[0109] The output sequence of the first attention layer is input into the first normalization layer. The output of the first normalized layer is used as the input of the first feedforward connection layer. The output of the first feedforward connection layer is used as the input of the second normalization layer. The first feedforward connection layer is a fully connected network, and its calculation formula is as follows:

[0110] FFN(x) = max(0, xW1+b1)W2+b2

[0111] Where W1 is the first weight, W2 is the second weight, b1 is the first bias vector, and b2 is the second bias vector.

[0112] The normalized output sequence of the second layer is fed into the next coding layer, and so on, up to the Nth coding layer. The network structure of all coding layers is exactly the same as that of the first coding layer. Let the output sequence of the Nth coding layer be z = [[z] 11 ,z 12 ,...z 1d ],[z 21 ,z 22 ,...z 2d ],...,[z I1 ,z I2 ,...z Id ]).

[0113] The decoder includes: a second input layer, M decoding layers, a second positional coding layer, and an output layer;

[0114] The second input layer is connected to the second positional encoding layer, and the second positional encoding layer is sequentially connected to each encoding layer. In this embodiment, a decoder structure is defined in the Transformer converter steelmaking oxygen consumption prediction model, and the output of each sub-layer in the decoding layer is calculated as follows:

[0115] The second input layer is used to extend the output sequence to two-dimensional data using a broadcast method;

[0116] The input sequence of the input layer is Y = [y1, y2, ..., y]. O-1 After broadcasting, the output sequence of the input layer is [[y1,y1,...,y1],[y2,y2,...,y2],...[y I ,y I ,...,y I ]], where the length of the one-dimensional sequence is d.

[0117] The second position encoding layer is used to add position biases to each of the two-dimensional data after the output sequence is expanded;

[0118] In this embodiment, to enable the model to learn the positional information of data in the sequence, a bias is added to each sequence in the positional encoding layer. The calculation formula for the bias in the second positional encoding layer is the same as that in the first positional encoding layer. The encoded output sequence is Ye = [[ye 11 ye 12 ,...,ye 1d ],[ye 21 ,ye 22 ,...,ye 2d ],...[ye I1 ,ye I2 ,...,ye Id ]).

[0119] Each decoding layer is used to decode the output data of the second position coding layer using an attention mechanism;

[0120] The second attention layer decodes the input sequence, and its calculation formula is the same as that of the first attention layer.

[0121] Where, the values ​​of Q and V are both Ye = [[ye 11 ye 12 ,...,ye 1d ],[ye 21 ,ye 22 ,...,ye 2d ],...[ye I1 ,ye I2 ,...,ye IdThe value of vector K is [[z]]. 11 ,z 12 ,...z 1d ],[z 21 ,z 22 ,...z 2d ],...,[z I1 ,z I2 ,...z Id ]).

[0122] The output sequence of the second attention layer is input to the third normalization layer. The output of the third normalization layer is used as the input of the decoding attention layer. The output of the decoding attention layer is used as the input of the fourth normalization layer. The output of the fourth normalization layer is used as the input of the second feedforward connection layer. The output of the second feedforward connection layer is used as the input of the fifth normalization layer. The second feedforward connection layer is a fully connected network, and its calculation formula is the same as that of the first feedforward connection layer.

[0123] The normalized output sequence of the fifth layer is fed into the next decoding layer, and so on, up to the Mth decoding layer. The network structure of all decoding layers is exactly the same as that of the first decoding layer.

[0124] The output of the Mth decoding layer is fed into the output layer (which is also a fully connected layer) to obtain the predicted value. The calculation formula is the same as that of the first feedforward connection layer.

[0125] The output layer is used to output the decoded data.

[0126] Furthermore, the hyperparameters of the initial oxygen consumption prediction model are optimized using a Bayesian optimization algorithm to obtain a pre-trained oxygen consumption prediction model, including:

[0127] Step S2.5: Determine the search range of hyperparameters, wherein the search range of the number of encoding layers and the search range of the number of decoding layers are determined to be 1 to 10 layers, and the length of the input sequence and the length of the output sequence are determined to be an integer multiple of the longest blowing cycle;

[0128] In this embodiment, the Bayesian optimization algorithm is used to optimize the key hyperparameters of the Transformer time series prediction model, including the number of encoding layers, the number of decoding layers, the input sequence length, and the output sequence length. Based on the optimized hyperparameters, the models corresponding to each converter are trained, tested, and validated. First, the search space for the hyperparameters to be optimized is determined. The search range for the number of encoding and decoding layers is set to 1 to 10 layers, and the input and output sequence lengths are set to integer multiples of the longest blowing cycle T. These hyperparameters are defined as optimization variables, and the search space is constructed as the input to the Bayesian optimization algorithm.

[0129] Step S2.6: Define the objective function using the current combination of hyperparameters;

[0130] In this embodiment, the root mean square error (RMSE) of the prediction model on the validation set is set as the evaluation metric. The objective function is calculated as follows:

[0131]

[0132] Where θ is the current combination of hyperparameters, L val (θ) represents the error value of the oxygen consumption prediction model for each converter trained using the current hyperparameter combination on the validation set. Let be the predicted value of the oxygen consumption prediction model for the i-th sample. Let N be the actual oxygen consumption data corresponding to the i-th sample, and N be the number of samples in the validation set.

[0133] Step S2.7: Based on the Gaussian process, establish the probability distribution model of the objective function;

[0134] In this embodiment, based on the existing tuning history H of hyperparameter combinations, Gaussian process (GP) regression is introduced to establish the objective function L. val The probability distribution model M of (θ) is used to predict the objective function value of unknown hyperparameter combinations.

[0135] The probability distribution model of the objective function is calculated as follows:

[0136] L val (θ)|H~GP(m(θ),k(θ,θ'))

[0137] Among them, L val (θ) represents the error value of the oxygen consumption prediction model for each converter trained using the current hyperparameter combination on the validation set, m(θ) is the mean function of the Gaussian process of the current hyperparameter combination, assumed to be 0, θ' is the existing hyperparameter combination, k(θ,θ') is the covariance function between the current hyperparameter combination and the existing hyperparameter combination, H is the historical tuning result of the existing hyperparameter combination, and GP is the Gaussian process.

[0138] k(θ,θ') is the covariance function, which in this embodiment is the radial basis function kernel function. The corresponding calculation formula is as follows:

[0139]

[0140] in, Let l represent the signal variance, and l represent the length of the kernel function.

[0141] Step S2.8: Select the next set of hyperparameters using the acquisition function;

[0142] In this embodiment, the next set of hyperparameters is selected using an acquisition function. Based on the prediction results of the probability distribution model of the current objective function, the expected improvement (EI) acquisition function is used to select the combination of hyperparameters with the greatest potential benefit.

[0143] The acquisition function is calculated as follows:

[0144]

[0145] Where y represents the possible output value of the objective function under the hyperparameter combination θ. * p is the currently known optimal objective function value. M (y|θ) represents the probability distribution of the objective function at the hyperparameter combination θ. Let θ be the desired improvement value at the current hyperparameter combination θ.

[0146] The next set of selected hyperparameters is calculated as follows:

[0147]

[0148] Where Θ is the set of all possible combinations of hyperparameters, θ * For the next set of hyperparameters selected, Let arg max be the desired improvement value at the current hyperparameter combination θ, and let arg max be the value to be taken. The hyperparameter at its maximum value.

[0149] Step S2.9: Evaluate the next set of hyperparameters selected using the validation set;

[0150] In this embodiment, the currently selected hyperparameter combination is evaluated. The selected hyperparameter combination is applied to the construction and training process of the Transformer model, the model performance is evaluated on the validation set, and the corresponding error index, i.e., the objective function value, is recorded.

[0151] Step S2.10: Update the probability distribution model of the objective function based on the current evaluation results, return to step S2.7 for iteration until the preset maximum number of iterations is met, and output the optimized hyperparameters.

[0152] In this embodiment, the probability distribution model is updated. Based on the current evaluation results, the probability distribution model is updated by adding new observation points (hyperparameter combinations and corresponding error values) to the tuning history set H, forming an iterative optimization closed loop. This continues until the preset maximum number of iterations is met. The determined optimal hyperparameter combination is output. All hyperparameter combinations and performance indicators recorded during the Bayesian optimization process are sorted, and the set of hyperparameters with the smallest prediction error is selected as the final configuration.

[0153] In this embodiment, the models corresponding to each converter are trained, tested, and validated based on the optimized hyperparameters.

[0154] Finally, the trained oxygen consumption prediction models for each converter are used, and the latest oxygen consumption data is extracted from the real-time database at fixed time intervals through a data dumping program, serving as input for the Transformer time-series prediction model. Simultaneously, the model is deployed to the energy management system.

[0155] The model is fed with the first I data points of the current moment to obtain a prediction sequence of length O. After inverse normalization, the predicted oxygen consumption of the converter in the next blowing cycle is obtained. The prediction results are aligned with the steelmaking plan and sent to the front-end interface of the energy management system in the form of a JSON string, providing scientific data guidance for scheduling plan decision-making and formulation.

[0156] This embodiment uses actual production data from a converter in a large steel enterprise as an example to perform a 25-minute real-time prediction of the converter's oxygen consumption during the steelmaking process. The average prediction accuracy is over 97%, and the prediction results are shown in the figure below. Figure 3 As shown, the method can effectively track data change trends. This embodiment can accurately learn the start and end stages of the converter blowing process, obtaining the oxygen consumption values ​​at the corresponding moments, providing data support for energy management personnel.

[0157] Example 2:

[0158] This embodiment proposes a production rhythm adaptive oxygen consumption prediction system for converter steelmaking in steel enterprises, such as... Figure 4 As shown, it includes: a data acquisition module, an oxygen consumption prediction module, and a prediction result output module;

[0159] The data acquisition module is connected to the oxygen consumption prediction module, and the oxygen consumption prediction module is connected to the prediction result output module.

[0160] The data acquisition module is used to obtain the oxygen consumption during the converter blowing period at the current moment;

[0161] The oxygen consumption prediction module is used to take the first I oxygen consumption data of the current converter blowing period and input them into the pre-trained oxygen consumption prediction model to obtain a prediction sequence of a specified length. The pre-trained oxygen consumption prediction model is established by: using the historical converter steelmaking plan of each converter, the oxygen consumption data of the corresponding time period and the Transformer time series prediction model to establish an initial oxygen consumption prediction model, and using the Bayesian optimization algorithm to optimize the hyperparameters of the initial oxygen consumption prediction model to obtain the pre-trained oxygen consumption prediction model.

[0162] The prediction result output module is used to inversely normalize the prediction sequence to obtain the predicted oxygen consumption value for the corresponding converter in the next blowing cycle; and to align the predicted oxygen consumption value for the next blowing cycle with the steelmaking plan to obtain the predicted oxygen consumption result for the next blowing cycle.

[0163] The various embodiments in this application are described in a progressive manner. The same or similar parts between the various embodiments can be referred to each other. Each embodiment focuses on describing the differences from other embodiments.

[0164] The scope of protection of this application is not limited to the embodiments described above. Obviously, those skilled in the art can make various modifications and variations to this disclosure without departing from the scope and spirit of this disclosure. If such modifications and variations fall within the scope of equivalent technology of this disclosure, then the intent of this disclosure also includes such modifications and variations.

Claims

1. A method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm, characterized in that, include: Obtain the oxygen consumption during the converter blowing period at the current moment; Take the first I oxygen consumption data of the current converter blowing period and input them into the pre-trained oxygen consumption prediction model to obtain a prediction sequence of a specified length. The process of establishing the pre-trained oxygen consumption prediction model includes: using the historical converter steelmaking plan of each converter, the oxygen consumption data of the corresponding time period and the Transformer time series prediction model to establish an initial oxygen consumption prediction model, and using the Bayesian optimization algorithm to optimize the hyperparameters of the initial oxygen consumption prediction model to obtain the pre-trained oxygen consumption prediction model. The predicted sequence is inversely normalized to obtain the predicted oxygen consumption of the corresponding converter in the next blowing cycle. Align the predicted oxygen consumption for the next blowing cycle with the steelmaking plan to obtain the predicted oxygen consumption for the next blowing cycle.

2. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 1, characterized in that, The initial oxygen consumption prediction model is established using historical converter steelmaking plans for each converter, oxygen consumption data for corresponding time periods, and the Transformer time-series prediction model. This includes: Obtain the historical steelmaking plans and oxygen consumption data for each converter during the corresponding time period, and perform timestamp alignment on the historical steelmaking plans and oxygen consumption data for each converter during the corresponding time period to obtain the aligned data. The aligned data is preprocessed, retaining only the data from the blowing stage; The preprocessed data is divided into a training set and a test set. The preprocessed data includes the input sequence and the output sequence at the corresponding time point. The Transformer time series prediction model was trained using the training set to obtain the initial oxygen consumption prediction model.

3. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 1, characterized in that, The Transformer temporal prediction model includes an encoder and a decoder, wherein the encoder includes a first input layer, N encoding layers and a first positional encoding layer; The first input layer is connected to the first position coding layer, and the first position coding layer is connected to each coding layer in sequence; Each coding layer includes: a first self-attention layer, a first normalization layer, a first feedforward connection layer, and a second normalization layer; the first self-attention layer is connected to the first normalization layer, the first normalization layer is connected to the first feedforward connection layer, and the first feedforward connection layer is connected to the second normalization layer. The first input layer is used to extend the input sequence to two-dimensional data using a broadcast method; The first position encoding layer is used to add position biases to each of the two-dimensional data after the input sequence is expanded; Each coding layer is used to encode the output data of the first position coding layer using an attention mechanism; The decoder includes: a second input layer, M decoding layers, a second positional coding layer, and an output layer; The second input layer is connected to the second position coding layer, and the second position coding layer is connected to each coding layer in turn; Each decoding layer includes: a second self-attention layer, a third normalization layer, a decoding attention layer, a fourth normalization layer, a second feedforward connection layer, and a fifth normalization layer; the second self-attention layer is connected to the third normalization layer, the third normalization layer is connected to the decoding attention layer, the decoding attention layer is connected to the fourth normalization layer, the fourth normalization layer is connected to the second feedforward connection layer, and the second feedforward connection layer is connected to the fifth normalization layer; The second input layer is used to extend the output sequence to two-dimensional data using a broadcast method; The second position encoding layer is used to add position biases to each of the two-dimensional data after the output sequence is expanded; Each decoding layer is used to decode the output data of the second positional coding layer using an attention mechanism. The output layer is used to output the decoded data.

4. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 1, characterized in that, The hyperparameters of the initial oxygen consumption prediction model include: the number of encoding layers, the number of decoding layers, the length of the input sequence, and the length of the output sequence.

5. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 1, characterized in that, Furthermore, the hyperparameters of the initial oxygen consumption prediction model are optimized using a Bayesian optimization algorithm to obtain a pre-trained oxygen consumption prediction model, including: Step S2.5: Determine the search range of hyperparameters, wherein the search range of the number of encoding layers and the search range of the number of decoding layers are determined to be 1 to 10 layers, and the length of the input sequence and the length of the output sequence are determined to be an integer multiple of the longest blowing cycle; Step S2.6: Define the objective function using the current combination of hyperparameters; Step S2.7: Based on the Gaussian process, establish the probability distribution model of the objective function; Step S2.8: Select the next set of hyperparameters using the acquisition function; Step S2.9: Evaluate the next set of hyperparameters selected using the validation set; Step S2.10: Update the probability distribution model of the objective function based on the current evaluation results, return to step S2.7 for iteration until the preset maximum number of iterations is met, and output the optimized hyperparameters.

6. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 5, characterized in that, The objective function is calculated as follows: Where θ is the current combination of hyperparameters, L val (θ) represents the error value of the oxygen consumption prediction model for each converter trained using the current hyperparameter combination on the validation set. Let be the predicted value of the oxygen consumption prediction model for the i-th sample. Let N be the actual oxygen consumption data corresponding to the i-th sample, and N be the number of samples in the validation set.

7. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 5, characterized in that, The probability distribution model of the objective function is calculated as follows: L val (θ)|H~GP(m(θ),k(θ,θ')) Among them, L val (θ) represents the error value of the oxygen consumption prediction model for each converter trained using the current hyperparameter combination on the validation set, m(θ) is the mean function of the Gaussian process of the current hyperparameter combination, θ' is the existing hyperparameter combination, k(θ,θ') is the covariance function between the current hyperparameter combination and the existing hyperparameter combination, H is the historical tuning result of the existing hyperparameter combination, and GP is the Gaussian process.

8. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 5, characterized in that, The acquisition function is calculated as follows: Where y represents the possible output value of the objective function under the hyperparameter combination θ. * p is the currently known optimal objective function value. M (y|θ) represents the probability distribution of the objective function at the hyperparameter combination θ. Let θ be the desired improvement value at the current hyperparameter combination θ.

9. The method for predicting oxygen consumption in converter steelmaking with adaptive production rhythm according to claim 5, characterized in that, The next set of selected hyperparameters is calculated as follows: Where Θ is the set of all possible combinations of hyperparameters, θ * For the next set of hyperparameters selected, Let arg max be the desired improvement value at the current hyperparameter combination θ, and let arg max be the value to be taken. The hyperparameter at its maximum value.

10. A converter steelmaking oxygen consumption prediction system with adaptive production rhythm, characterized in that, include: The data acquisition module is used to obtain the oxygen consumption during the converter blowing period at the current moment; The oxygen consumption prediction module is used to take the first I oxygen consumption data of the converter blowing period at the current moment and input them into the pre-trained oxygen consumption prediction model to obtain a prediction sequence of a specified length. The process of establishing the pre-trained oxygen consumption prediction model includes: using the historical converter steelmaking plan of each converter, the oxygen consumption data of the corresponding time period and the Transformer time series prediction model to establish an initial oxygen consumption prediction model, and using the Bayesian optimization algorithm to optimize the hyperparameters of the initial oxygen consumption prediction model to obtain the pre-trained oxygen consumption prediction model. The prediction result output module is used to inversely normalize the prediction sequence to obtain the predicted oxygen consumption value for the corresponding converter in the next blowing cycle; and to align the predicted oxygen consumption value for the next blowing cycle with the steelmaking plan to obtain the predicted oxygen consumption result for the next blowing cycle.

Images