Blast furnace molten iron silicon content flexible prediction method based on data fusion and time embedding

Abstract

The invention discloses a blast furnace molten iron silicon content flexible prediction method based on data fusion and time embedding, and relates to the field of industrial production. According to the method, time label information of data is utilized, a time domain relation between the data is mined through a learnable mapping method, and a sequence with non-uniform time intervals is mapped into a uniform sequence; through the structural design of the neural network, high-frequency uniformly sampled industrial data and low-frequency non-uniformly sampled data can be simultaneously input into a prediction model, and data fusion is performed, so that the problem of effective information loss caused by downsampling is avoided; the complete sequence-to-sequence architecture based on the attention mechanism is utilized, the prediction model can output molten iron silicon content prediction data of any length, and the requirement for any prediction duration is met. The prediction precision of the silicon content of the molten iron is improved, the prediction flexibility and usability are improved, and the method has more reference value for field production.

Description

Technical Field

[0001] The invention relates to the field of industrial production, and in particular to a flexible prediction method for silicon content in molten iron of a blast furnace based on data fusion and time embedding. Background Art

[0002] The silicon content of molten iron is an important indicator of molten iron quality and can indirectly reflect the thermal state of the furnace. Therefore, it is of great significance for the quality control of molten iron and the operation control of blast furnaces. The silicon content of molten iron at the tap hole cannot be obtained in real time at industrial sites. It needs to be sampled and manually tested, so there is a time lag. In addition, since the blast furnace is a large inertia system, the impact of production operation adjustments on quality indicators also has a large lag time. Therefore, the early prediction of the silicon content of molten iron is of great significance for the optimization of blast furnace production.

[0003] The blast furnace is a typical large-inertia, high-dimensional, complex nonlinear industrial production system. Its interior is relatively closed, and sensors can only be installed on the furnace top, furnace surface, and furnace bottom for data collection. Therefore, it is very difficult to model the blast furnace from the perspective of physical and chemical mechanisms. The data-driven modeling method avoids the study of complex mechanisms and only uses observational data to establish a black box model to predict quality indicators. In the data modeling method, the neural network model can characterize the nonlinear relationship in the system and learn autonomously through data collected from the industrial site. Among them, the sequence-to-sequence model architecture based on recurrent neural networks has a strong modeling ability for time series data commonly found in industry, and therefore has become one of the important models for predicting the silicon content of molten iron in blast furnaces.

[0004] There are two types of industrial data in the blast furnace production process. One type is data collected automatically by sensors, such as oxygen enrichment flow, furnace top pressure, hot air temperature, actual coal injection amount, etc. This type of data is uniformly sampled data, and the sampling time interval is 10s. The second type is manual test data, such as the silicon content of molten iron. The time label of this type of data is determined by manual sampling, and the time interval is approximately 20 to 40 minutes. The existence of heterogeneous industrial data and non-uniform time labels makes it difficult for recurrent neural networks that calculate according to unit time steps to directly process industrial raw data. Therefore, it is difficult to make full use of industrial raw data and build an accurate prediction model for the silicon content of molten iron.

[0005] The traditional approach is to perform mechanical preprocessing operations on the original data, average or downsample the high-frequency data in the time domain, reduce the magnitude of the data, and correspond one-to-one with the low-frequency data on the time label, and process the corresponding data in chronological order by step, ignoring the non-uniform sampling characteristics of the original data; or further perform linear interpolation on the original non-uniform data in the time domain, convert the non-uniform data into uniform data, and then conduct the subsequent modeling research. The fusion problem of heterogeneous industrial data and the processing problem of non-uniform sampling data are not considered. The interpolation processing of the silicon content of molten iron will cause the loss of original information; because the blast furnace is a large inertia industrial process, half an hour of input data is not enough to predict the silicon content.

[0006] Therefore, technicians in this field are committed to developing a flexible prediction method for silicon content in blast furnace hot metal based on data fusion and time embedding. Modeling is performed on heterogeneous uniformly sampled and non-uniformly sampled raw industrial data, making full use of the effective information contained in the time label to improve the prediction accuracy of hot metal silicon content; outputting the predicted data of hot metal silicon content at any future time point improves the flexibility and availability of the prediction, and is more valuable for on-site production. Summary of the invention

[0007] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is the fusion problem of heterogeneous industrial data and the processing problem of non-uniform sampling data, so as to improve the prediction accuracy of molten iron silicon content.

[0008] To achieve the above object, the present invention provides a flexible prediction method for silicon content in molten iron of a blast furnace based on data fusion and time embedding, comprising the following steps:

[0009] Step 1: Preprocess the original data;

[0010] Step 2: perform correlation analysis to obtain sensor data most relevant to the silicon content of molten iron;

[0011] Step 3: Construct a data set, where each piece of data consists of a time series of silicon content in molten iron and sensor data in a corresponding time range;

[0012] Step 4: The high-frequency sensor data is fused with the low-frequency and non-uniform historical molten iron silicon content data through the data fusion layer of the encoder, and a representation sequence corresponding to the input time step of the historical molten iron silicon content is extracted (the representation sequence refers to the vector combination at each time step output by the data fusion layer);

[0013] Step 5: The encoder uniform mapping layer uses the non-uniform time tag information to map the non-uniform representation sequence to a uniform time point, and obtains a high-level representation sequence with uniform time intervals (the high-level representation sequence refers to the vector combination at each time step output by the uniform mapping layer);

[0014] Step 6: Send the high-level representation sequence output by the uniform mapping layer and the decoder state vector (decoder state vector, which refers to the state of the hidden layer neurons of the decoder network) of the previous prediction time step to the attention mechanism layer, calculate the attention weight of each high-level representation (high-level representation, which refers to the vector at a certain time step output by the uniform mapping layer) of the current prediction time step, and further calculate the context feature vector (context feature vector, which refers to the vector output by the attention mechanism layer) of the current time step;

[0015] Step 7: Send the context feature vector of the current time step and the predicted value of silicon content in molten iron at the previous time step to the decoder, and calculate the decoder state vector and the predicted value of silicon content in molten iron at the current time step;

[0016] Step 8: Calculate the loss function based on the predicted value and true value of the molten iron silicon content, and then perform model back propagation and parameter optimization.

[0017] Furthermore, in step 1, data preprocessing includes abnormal data processing, missing value filling, filtering, and normalization.

[0018] Furthermore, in step 2, the time tag of the low-frequency data is used as a reference, and the high-frequency data is sampled according to a certain lag time, so as to achieve the correspondence of heterogeneous data under different lag time settings.

[0019] Furthermore, in step 2, correlation analysis is performed by comprehensively using the Pearson correlation coefficient, the Spearman correlation coefficient, and the maximum mutual information coefficient.

[0020] Furthermore, in step 3, a time series of silicon content in molten iron of a specific length is selected, and the latest silicon content data in the sequence is removed, and the remaining data constitutes a historical silicon content sequence in molten iron; according to the time range corresponding to the historical silicon content sequence in molten iron, a time series of input variables such as oxygen enrichment flow rate and top pressure of a certain length is selected; all the data selected above constitute one piece of data in the data set; in each piece of data, the latest silicon content data is used as the prediction data of the model, and all other data are used as the input data of the model.

[0021] Furthermore, in step 3, the entire data set is divided into a training set, a validation set, and a test set; the training set is used for model training and parameter optimization, the validation set is used for hyperparameter adjustment, and the test set is used to test the actual performance of the model.

[0022] Furthermore, in step 4, the original high-frequency uniform sensor data and the low-frequency non-uniform historical molten iron silicon content data are directly processed to achieve data fusion.

[0023] Furthermore, the step 4 indicates that the sequence includes sensor data and historical molten iron silicon content information.

[0024] Furthermore, in step 5, a time tag embedding mechanism is used to map the representation sequence of the non-uniform time intervals to uniform time points.

[0025] Furthermore, in step 5, the time label information of the data is used to mine the time domain relationship between the data through a learnable mapping method, and the sequence of non-uniform time intervals is mapped into a uniform sequence.

[0026] Furthermore, in step 5, the uniform mapping layer uses a learnable time embedding mechanism to map the non-uniform time label and the uniform reference time label sequence into two corresponding sets of time embedding vector sequences; with the help of a learnable parameter matrix, the similarity of the two sets of vectors is calculated to obtain a weight relationship matrix from each original time point to the reference time point; using the weight relationship matrix and the subsequent linear mapping layer, the non-uniform representation sequence is mapped into a high-level representation sequence with uniform time intervals.

[0027] Furthermore, in step 6, a complete encoder-decoder structure based on the attention mechanism is used to calculate the time domain relationship between the decoder state vector at each prediction time point and the encoder high-level representation at each historical time point.

[0028] Furthermore, in step 7, a sequence of molten iron silicon content of any prediction time length is outputted through a decoder structure.

[0029] The existing non-uniform data processing method of time-domain linear interpolation cannot consider the dynamic change characteristics of the time series before and after the interpolation point, so the reliability of the interpolated data is limited. The present invention uses the time label information of the data, and through a learnable mapping method, it mines the time domain relationship between the data and maps the data of the non-uniform time interval to uniform data. In the fifth step: design a uniform mapping layer to map the non-uniform encoder representation sequence to a uniform high-level representation sequence. The uniform mapping layer first uses a learnable time embedding mechanism to map the non-uniform time label and the uniform reference time label sequence to a corresponding time embedding vector sequence, and form two time embedding matrices. The two time embedding matrices are similarly calculated by learnable parameters to obtain the weight relationship matrix K from each original time point to the reference time point. Using K and the subsequent linear mapping layer, the non-uniform representation sequence is mapped to a high-level representation sequence of uniform time intervals. The present invention can model the original industrial data of non-uniform sampling, make full use of the effective information contained in the time label, and improve the prediction accuracy of the silicon content of molten iron.

[0030] The existing prediction model for silicon content in molten iron cannot directly integrate the original heterogeneous industrial data, and the data preprocessing operation based on downsampling is prone to information loss, which affects the prediction accuracy of silicon content in molten iron. The present invention uses the structural design of the neural network to enable high-frequency uniformly sampled industrial data and low-frequency non-uniformly sampled data to be simultaneously input into the prediction model, and perform data fusion to avoid the problem of effective information loss caused by downsampling. In the fourth step: on the basis of the original sequence-to-sequence architecture, a data fusion layer is added to the encoder. The data fusion layer first sends high-frequency input data, such as oxygen-enriched flow, top pressure, etc., to a single-layer LSTM network (long short-term memory network). At this time, the hidden state of each time step of the LSTM is the representation of the input data up to the current time step. According to the low-frequency historical molten iron silicon content time label, the LSTM hidden layer state of the corresponding time step is extracted, so that the high-frequency input data is converted into low-frequency representation data (representation data refers to the vector sequence output by the first layer of LSTM network in the data fusion layer). Furthermore, the extracted representation is spliced ​​with the corresponding historical molten iron silicon content, and the spliced ​​data is sent to the second-layer LSTM network to obtain the encoder representation H (encoder representation refers to the vector at a certain time step output by the data fusion layer), which contains 22 groups of data such as oxygen-enriched flow rate and top pressure, as well as information on the historical molten iron silicon content, thus realizing information fusion. The present invention can directly integrate and process the most original heterogeneous industrial data - sensor data and manual test data, prevent information loss caused by preprocessing operations, improve the prediction accuracy of the model, and the end-to-end model from original data to predicted data is easier to implement than the prediction method that requires multi-step processing.

[0031] Most of the existing technologies are single-step predictions, which cannot flexibly handle different prediction time lengths, and are inconsistent with the uneven prediction time length of molten iron silicon content. The present invention uses a complete sequence-to-sequence architecture based on the attention mechanism, and the prediction model can output molten iron silicon content prediction data of any length, which can meet the needs of any prediction time length. In the sixth step: use the attention mechanism to better characterize the influence relationship between the input and output variables of the system in the time domain. The dependency relationship between the output data at different time steps and the input data at different time steps is different, and the attention weight calculated by the attention mechanism can well characterize this relationship of different strengths. The encoder high-level representation H contains the information of the input variable up to a certain time step, and the decoder state vector s contains the information of the output variable up to a certain time step. Therefore, the attention mechanism takes the high-level representation sequence and s at different time steps as input, and calculates the attention weights between the output data at different time steps and the input data at different time steps. First, a series of H is convolved in the time domain to obtain a series of key values ​​k. Since the key value k is calculated from H of a certain time length (convolution kernel length), it contains richer time domain features relative to H. Perform a dot multiplication of k of all input time steps and s of the previous output time step, and send all the calculated results to the Softmax layer to obtain the attention weight w of the current time step, where w represents the attention of the current time step to the data of different input time steps. Use w to perform a weighted sum of H of all different input time steps to obtain the context feature vector z of the current output time step, and z is sent to the encoder for calculation of the current output time step. Step 7: Output a predicted output sequence of any length through the decoder structure. The decoder is a GRU network (Gated Recurrent Unit Network), the input of the network is the context feature vector z of the current time step and the silicon content of the molten iron in the previous step, the hidden state is represented by s, and the output is the decoder state vector "s" at the current time step and the predicted value of the silicon content in the molten iron. With the help of the complete encoder-decoder structure based on the attention mechanism, the model can continue to calculate the silicon content in the molten iron at the next time step. First, the decoder state vector s returns to the attention mechanism layer, and the time domain dependency relationship is calculated again with the encoder high-level representation H that has been calculated to obtain z for the next time step. Then, according to the latest s and z, the predicted value of the silicon content in the molten iron at the next time step can be calculated. Through cyclic calculation, the prediction model can output the predicted data of the silicon content in the molten iron of any time length, which overcomes the problem of uneven prediction time length and overcomes the bottleneck that the classic molten iron silicon content model can only perform single-point prediction. The present invention enables the prediction model to output the predicted data of the silicon content in the molten iron at any future time point, improves the flexibility and availability of the prediction, and is more valuable for on-site production workers.

[0032] No complete encoder-decoder structure has been found for the prediction of silicon content in molten iron. Variable duration prediction is not only related to the decoder, but also to the use of a complete encoder-decoder structure. The complete structure includes an encoder and a decoder. The key point is that the decoder will return a series of decoder state vectors s to the attention mechanism layer for calculating a series of context feature vectors z; the existing technology only predicts one time point, so it only calculates one z and one s, so there is no need to return s to the attention mechanism layer to calculate the next z and further obtain the next s.

[0033] Compared with the prior art, the present invention has the following obvious substantial features and significant advantages:

[0034] 1. The present invention has a high application value in predicting the silicon content of molten iron in blast furnaces of the steel industry.

[0035] 2. The present invention can effectively integrate heterogeneous industrial data and mine time tag information, so that the model has a high prediction hit rate for molten iron silicon content, can meet the prediction needs of actual industrial sites, and has a high reference value.

[0036] 3. The present invention can directly process original industrial sensor data and manual test data in an end-to-end manner and predict the silicon content of molten iron, thus avoiding cumbersome data processing and having strong practicality.

[0037] 4. The prediction time length of the present invention is relatively flexible, can meet the variable prediction needs of the production site, and has a high reference value.

[0038] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is an overall flow chart of a method for predicting silicon content in molten iron according to a preferred embodiment of the present invention;

[0040] Figure 2 It is an encoder data fusion layer of a preferred embodiment of the present invention;

[0041] Figure 3 It is an encoder uniform mapping layer of a preferred embodiment of the present invention;

[0042] Figure 4 is an attention mechanism of a preferred embodiment of the present invention;

[0043] Figure 5 It is a decoder structure of a preferred embodiment of the present invention;

[0044] Figure 6This is the model prediction effect of a preferred embodiment of the present invention;

[0045] Figure 7 This is a comparison of the test set accuracy during the training process of different data sets in a preferred embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0047] In the drawings, components with the same structure are indicated by the same numerical reference numerals, and components with similar structures or functions are indicated by similar numerical reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present invention does not limit the size and thickness of each component. In order to make the illustration clearer, the thickness of the components is appropriately exaggerated in some places in the drawings.

[0048] In this example, a domestic steel plant with a capacity of 2650 m 3 Take the actual blast furnace production data collected from January 1, 2021 to December 31, 2021 of a blast furnace as an example.

[0049] A variable time prediction method for silicon content in blast furnace hot metal based on heterogeneous data fusion and time information embedding. Figure 1 As shown, the following steps are included:

[0050] Step 1: Preprocess the raw data collected by the steel plant. First, remove and complete the abnormal data (outliers and missing values). Due to sensor failure and blast furnace shutdown and maintenance, the raw data contains outliers and missing values. The outliers are identified using the box plot method. Short-term abnormal data is determined to be sensor failure, and the average value of the normal data at both ends of the abnormal time is used to fill it. Long-term abnormal data is determined to be blast furnace shutdown and maintenance, so all data during this period are discarded.

[0051] The padded data is digitally filtered to smooth out abnormal peaks.

[0052] Except for the sampling interval of molten iron silicon content, which is approximately 20 minutes to 40 minutes, the sampling interval of all other data (oxygen enrichment flow, top pressure, etc.) is 1 minute. Therefore, the unit step of the recurrent neural network processing data corresponds to 1 minute of actual time.

[0053] The data is normalized to zero-mean to avoid the influence of different data magnitudes on the model accuracy.

[0054] Step 2: Use the Pearson correlation coefficient, Spearman correlation coefficient and maximum mutual information coefficient to analyze the correlation between the silicon content of molten iron and other variables such as oxygen enrichment flow rate and top pressure, and screen the input variables of the model based on the analysis results and expert experience. Since the sampling frequency of the silicon content of molten iron is lower than that of other variables, the time labels of the silicon content data [t1, t2, t3, …, t n ] as the benchmark, the high-frequency data is sampled according to a certain lag time lag, and the sampling time points are [t1-lag, t2-lag, t3-lag,…, t n -lag]. Finally, 22 variables were selected as the input of the model (excluding the historical silicon content of molten iron), including: CO2, actual coal injection amount in the previous hour, cold air flow, actual wind speed, oxygen enrichment flow, oxygen enrichment rate, actual coal injection amount in this hour, bosh gas index, bosh gas volume, hot air temperature, set coal injection amount, permeability index, resistance coefficient, No. 1 top pressure, No. 2 top pressure, top temperature downcomer, top temperature northeast, top temperature southeast, top temperature northwest, top temperature southwest, blast kinetic energy, and blast humidity.

[0055] Step 3: Construct the data set. The earliest time label in all data corresponds to step 0, 1 minute later is step 1, and nmin later corresponds to step n. After the first two steps, the minimum unit of all data event labels is 1 minute. Except for the time period corresponding to the discarded data, each time step corresponds to a set of 22 input data such as oxygen enrichment flow rate and top pressure, but does not necessarily correspond to the molten iron silicon content data.

[0056] Select a time series of molten iron silicon content with a step length [y n(1) ,y n(2) ,y n(3) ,…,y n(step) ], according to the corresponding time steps [n(1),n(2),n(3),…,n(step)], select the time series [x of 22 variables such as oxygen enrichment flow rate and top pressure at the time steps [n(1)-40,…,n(1)-1,n(1),n(1)+1,…,n(step-1)] n(1)-40, …,x n(1)-1 ,x n(1) ,x n(1)+1 ,…,x n(step-1) ], thus forming a "piece of data" in the data set. Since the sampling time intervals of the silicon content in molten iron are different, the step lengths corresponding to the silicon content time series of the molten iron with step lengths are different, so the length of each piece of data is different. The time series corresponding to the 22 variables such as oxygen enrichment flow rate and top pressure in each piece of data [x n(1)-40 ,…,x n(1)-1 ,x n(1) ,x n(1)+1 ,…,xn(step-1) ] and the historical series of molten iron silicon content [y n(1) ,y n(2) ,y n(3) ,…,y n(step-1) ] will be used as the input of the model, y n(step) as the prediction output of the model.

[0057] The entire data set contains 12,508 data items, including 10,000 data items in the training set, 1,308 data items in the validation set, and 1,200 data items in the test set. The training set is used for model training and parameter optimization, the validation set is used for hyperparameter adjustment, and the test set is used to test the actual performance of the model. The following steps 4 to 6 describe the processing of "one piece of data" in the data set.

[0058] Step 4: Based on the original sequence-to-sequence architecture, a data fusion layer is added to the encoder. The data fusion layer can directly use 22 sets of data with higher frequency, such as oxygen enrichment flow and top pressure, as well as historical molten iron silicon content data with lower frequency and unevenness, thereby avoiding information loss caused by excessive processing of the original data. The structure of the data fusion layer is as follows: Figure 2 As shown, first, the 22 groups of input data [x n(1)-40 ,…,x n(1)-1 ,x n(1) ,x n(1)+1 ,…,x n(step-1) ], and sent it to a single-layer LSTM network to extract the hidden state [h n(1) ,h n(2) ,h n(3) ,…,h n(step-1) ]. Since the hidden state of LSTM depends on all input data up to the current step, it can be considered that [h n(1) ,h n(2) ,h n(3) ,…,h n(step-1) ] is the representation of the original input data at time steps [n(1), n(2), n(3), …, n(step-1)], which contains all the information of the original data. This operation converts the high-frequency input data into low-frequency representation data and realizes the alignment with the historical molten iron silicon content at uneven time steps. Next, the aligned and spliced ​​data [ <h n(1) ,y n(1) >, <h n(2) ,y n(2) >, <h n(3) ,y n(3) >,…, <h n(step-1) ,y n(step-1)>] is sent to the second layer LSTM network to obtain the encoder representation sequence [H n(1) ,H n(2) ,H n(3) ,…,H n(step-1) ]. This sequence contains 22 sets of data such as oxygen enrichment flow, top pressure, and historical information on molten iron silicon content, achieving information fusion.

[0059] Step 5: Use the time tag embedding mechanism to map the representation sequence of non-uniform time intervals to uniform time points. The structure of the uniform mapping layer is as follows: Figure 3 As shown, firstly, given a uniformly distributed reference time step [n ref (1),n ref (2),…,n ref (step-1)], the original time step and the reference time step are transformed into two sets of time embedding vector sequences through the time embedding layer and the linear mapping layer, forming the time embedding matrices φ and φ ref The function of the temporal embedding layer The calculation corresponding to the linear mapping layer is as follows, where ω and β are learnable parameters and d determines the embedding vector Dimension.

[0060]

[0061] Furthermore, with the help of learnable parameter matrices W and V, the weight relationship matrix K between each original time step and the reference time step can be obtained according to the following similarity calculation formula.

[0062]

[0063] The matrix K can transform the original non-uniform representation sequence [H n(1) ,H n(2) ,H n(3) ,…,H n(step-1) ] is mapped into a uniform high-level representation sequence at the reference time step according to the following formula [H nref(1) ,H nref(2) ,H nref(3) ,…,H nref(step-1) ].

[0064]

[0065] The uniform mapping layer makes full use of the original time labels and embeds the non-uniform time information into the representation sequence to achieve the modeling of the non-uniform time domain relationship of the original data. While obtaining a high-level representation sequence with uniform time distribution, it also rationalizes the downstream standard neural network modeling method based on equal time steps. ref The difference is therefore abbreviated as n.

[0066] Step 6: Use the attention mechanism to better characterize the influence relationship between the system input and output variables in the time domain. Figure 4 As shown in Figure 1, the overall structure of the encoder including the data fusion layer and the attention mechanism. n(1) ,H n(2) ,H n(3) ,…,H n(step-1) ] performs a 1D convolution in the time domain to obtain the key value [k n(1) ,k n(2) ,k n(3) ,…,k n(step-1) ], the query value is the decoder state vector s. The initial hidden layer state s n(step-1) It is obtained by performing a 1D convolution in the time domain on the encoder high-level representation H of the last few time steps. The decoder state vector s at time step n (step-1) and any subsequent time step is combined with [k n(1) ,k n(2) ,k n(3) ,…,k n(step-1) ] to perform point multiplication, and then pass through the softmax layer to obtain the attention weight [w n(1) ,w n(2) ,w n(3) ,…,w n(step-1) ], and use this weight to perform weighted summation on the high-level representation of the encoder to obtain the context feature vector z corresponding to time step n (step-1) and the time steps thereafter. Since the attention weights calculated by the decoder state vector s at different time steps are different, the corresponding context feature vector z is also different. Corresponding to [s n(step-1) ,s n(step-1)+1 ,s n(step-1)+2 ,…,s n(step)-1 ], we can calculate [z n(step-1) ,z n(step-1)+1 ,z n(step-1)+2 ,…,z n(step)-1 ].

[0067] The weights calculated by the attention mechanism represent the dependencies between input and output variables in the time domain. n(1) is the representation of all information from the input variable x to step n(1), s n(step-1) It contains the information of the output variable of step n(step-1). If w is calculated by the attention mechanism n(1) If the value of is large, it means that the input data of step n(1) has a strong relationship with the output data of step n(step-1); if w n(1)The value of is small, and vice versa. Each time step from n(step-1) to n(step) is different, and the outputs of different time steps naturally have different temporal dependencies with the input data. Therefore, the attention mechanism is calculated at each time step, and the obtained attention weights well describe the dependency between the output variables and input variables at different time steps.

[0068] Step 7: Output a predicted output sequence of arbitrary length through the decoder structure. Figure 5 The structure of the decoder is n(step-1) Initialize the decoder hidden state and convert the context feature vector z calculated at the current time step into n(step-1)+1 The predicted output y of the previous time step n(step-1) Together as input into the GRU network to obtain the hidden state s n(step-1)+1 , and then pass through the linear output layer to get the predicted output y n(step-1)+1 . n(step-1)+1 Re-send it to the attention mechanism described in step 5 to calculate the z required for the next time n(step-1)+2 , together with y n(step-1)+1 Together as the input of the next time step. After the above cycle calculation, the predicted output sequence [y n(step-1)+1 ,y n(step-1)+2 ,y n(step-1)+3 ,…,y n(step) ]. It should be noted that the original silicon content of molten iron only has real data at time steps n(step-1) and n(step), and there is no real data at time steps [n(step-1)+1, n(step-1)+2,…, n(step)-1]. Therefore, when calculating the prediction accuracy, only the predicted value y is considered. n(step) and the true value y n(step) In addition, in the data set, the length of each data item [n(step-1), n(step)] is different, which is caused by the uneven sampling of the silicon content in the molten iron. However, the structure of the recurrent neural network in the decoder can obtain the predicted output sequence of any length after n(step-1) through cyclic calculation, which naturally overcomes the problem of uneven prediction time length and realizes the prediction of molten iron silicon content with flexible time length.

[0069] Step 8: Calculate the loss function based on the predicted value and the true value of the silicon content in molten iron, and then perform model back propagation and parameter optimization. The training set is used to optimize the model parameters. The optimization algorithm uses Adams, and the loss function uses the mean absolute value loss; the validation set is used to adjust the model hyperparameters. The main model hyperparameters are step, time domain convolution length, hidden layer variable dimension, key value dimension, and data batch size; the test set is used to test the actual performance of the model. The experimental results are as follows Figure 6 As shown in the figure, the comparison between the predicted value and the actual value of the silicon content in molten iron and the error range diagram. The square solid line in the figure represents the predicted value, the triangle solid line represents the actual value, the solid line represents the error between the two, and the error between the two dotted lines is less than ±0.1.

[0070] There are three indicators for evaluating the prediction model of molten iron silicon content, one of which is the hit rate (HR):

[0071]

[0072] Among them, N is the number of prediction samples, and the error e is k Satisfies the following, where y true is the true value of silicon content:

[0073]

[0074] The second is the root mean square error (RMSE):

[0075]

[0076] The third is the Mean Absolute Percent Error (MAPE):

[0077]

[0078] The model prediction hit rate is 95.15%, and the root mean square error is 0.0443. The model has achieved very ideal prediction results. In addition, according to the traditional data preprocessing method, the high-frequency sensor data is averaged in time, so that the sensor data corresponds to the test value data of the silicon content in molten iron one by one, and then linear interpolation is performed to obtain uniform data to construct the data set. Figure 7 The comparison of the accuracy of the model trained by the dataset with the same frequency of each type of data after downsampling and linear interpolation and the dataset containing heterogeneous data of different frequencies shows that the model index trained by the original dataset without excessive preprocessing (the dataset containing heterogeneous industrial data of different frequencies) is better. It can be concluded that the model designed by the present invention can well preserve the information of the original data, prevent the damage of the original data by data preprocessing operations such as time domain downsampling and linear interpolation, and ensure the prediction accuracy of the model.

[0079] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A flexible prediction method for silicon content in molten iron of blast furnace based on data fusion and time embedding, characterized in that: The following steps are involved: Step 1: Preprocess the original data; Step 2: perform correlation analysis to obtain sensor data most relevant to the silicon content of molten iron; Step 3: Construct a data set, where each piece of data consists of a time series of silicon content in molten iron and sensor data in a corresponding time range; Step 4: The high-frequency sensor data is fused with the low-frequency and non-uniform historical molten iron silicon content data through the data fusion layer of the encoder, and a representation sequence corresponding to the input time step of the historical molten iron silicon content is extracted; Step 5: Map the non-uniform representation sequence to a uniform time point by using the non-uniform time tag information through the encoder uniform mapping layer to obtain a high-level representation sequence with uniform time intervals; Step 6: Send the high-level representation sequence output by the uniform mapping layer and the decoder state vector of the previous prediction time step to the attention mechanism layer, calculate the attention weight of each high-level representation of the current prediction time step, and further calculate the context feature vector of the current time step; Step 7: Send the context feature vector of the current time step and the predicted value of silicon content in molten iron at the previous time step to the decoder, and calculate the decoder state vector and the predicted value of silicon content in molten iron at the current time step; Step 8: Calculate the loss function based on the predicted value and true value of the molten iron silicon content, and then perform model back propagation and parameter optimization.

2. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: The step 1, data preprocessing includes abnormal data processing, missing value filling, filtering, and normalization.

3. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: The step 2 samples the high-frequency data based on the time tags of the low-frequency data, so as to achieve the correspondence of heterogeneous data under different lag time settings.

4. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: In the step 2, the Pearson correlation coefficient, the Spearman correlation coefficient and the maximum mutual information coefficient are comprehensively used to perform correlation analysis.

5. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: The step 4 directly processes the original high-frequency uniform sensor data and the low-frequency non-uniform historical molten iron silicon content data to obtain a representation sequence.

6. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: The step 5 uses a time tag embedding mechanism to map the representation sequence of non-uniform time intervals to uniform time points.

7. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: The step 5 utilizes the time tag information of the data and uses a learnable mapping method to mine the time domain relationship between the data and map the sequence of non-uniform time intervals into a uniform sequence.

8. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: In step 5, the uniform mapping layer uses a learnable time embedding mechanism to map the non-uniform time tags and the uniform reference time tag sequence into two corresponding sets of time embedding vector sequences; With the help of a learnable parameter matrix, the similarity of the two groups of vectors is calculated to obtain the weight relationship matrix from each original time point to the reference time point; using the weight relationship matrix and the subsequent linear mapping layer, the non-uniform representation sequence is mapped into a high-level representation sequence with uniform time intervals.

9. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: The step 6 adopts a complete encoder-decoder structure based on the attention mechanism to calculate the time domain relationship between the decoder state vector of each predicted time point and the encoder high-level representation of each historical time point.

10. The method for flexible prediction of silicon content in molten iron of blast furnace based on data fusion and time embedding according to claim 1, characterized in that: The step 7 is to output a sequence of molten iron silicon content of any prediction time length through a decoder structure.

Images