Preheater outlet oxygen content target value optimizing method and system based on multi-step prediction

Abstract

The invention relates to the technical field of cement firing system control, and discloses a preheater outlet oxygen content target value optimization method based on multi-step prediction, which comprises the following steps: step 1, acquiring CO and O2 gas data, and dividing the CO and O2 gas data into a training set and a test set after processing; 2, performing feature selection in combination with a mutual information analysis method and a process principle, and performing standardization processing on data; 3, building a CO and O2 multi-step prediction initial model; 4, importing the training set into the initial model for training, and verifying the accuracy of the model by using the test set to obtain a prediction model; and 5, acquiring monitoring data in real time, obtaining a CO and O2 multi-step prediction result by using the prediction model, and transmitting the prediction result into an O2 target value optimization module to obtain a real-time O2 target recommendation value. According to the method, the problem caused by time lag of a sintering system is solved, the influence caused by data drift of a gas sensor can be adapted, the accuracy and reliability of the O2 target recommendation value are improved, and the power consumption of a high-temperature fan and the heat consumption of clinker are reduced.

Description

Technical Field

[0001] The invention relates to the technical field of cement burning system control, and in particular to a method and system for optimizing a target value of oxygen content at a preheater outlet based on multi-step prediction. Background Art

[0002] The O2 content at the outlet of preheater C1 is one of the important indicators to characterize the combustion state of the cement burning system, and plays a key role in the production of clinker. If the oxygen content is too high, it means that the system uses too much air volume, which will lead to unnecessary waste of energy and may cause an increase in the amount of exhaust gas in the combustion system, an increase in resistance and an increase in the power consumption of the high-temperature fan, thereby affecting the power consumption and cost of the system. On the other hand, if the oxygen content is too low, it means that the air volume used by the system is insufficient, the combustion is not complete, and a reducing atmosphere will be formed. The reducing atmosphere will have a serious impact on the burning of the clinker, causing the clinker to have a yellow core and reduce its grindability, seriously affecting the color and quality of the cement. In addition, too low an oxygen content may also cause adhesion and blockage of the pre-decomposition system, ringing and balling after the kiln, and even endanger the safe operation of the system. Therefore, the accuracy of the target value of the O2 content at the outlet of preheater C1 is of great significance to the production of clinker.

[0003] At present, various methods have been implemented for the determination of O2 content. For example, the patent number CN111721122A "A cement burning system and its automatic optimization control system" discloses an automatic optimization control system for cement burning system. This method divides the cement burning automatic optimization control system into five modules: rotary kiln control, pre-decomposition control, grate cooler control, predictive control, and automatic optimization control. Among them, the high-temperature fan control in the pre-decomposition control module controls the high-temperature fan speed according to the O2 content at the preheater outlet, the high-temperature fan inverter current, the preheater outlet CO content and the kiln feed amount. For example, the patent number CN113339307A "A method for controlling the high-temperature fan speed of a cement rotary kiln based on excess air coefficient" discloses a method for controlling the high-temperature fan speed of a cement rotary kiln based on excess air coefficient. This method calculates the excess air coefficient of the current system by using the excess air coefficient formula, locates the optimal excess air coefficient target value by combining the filtered data, and finally builds a relationship model and calculates the optimal high-temperature fan speed through the MPC algorithm. However, the current methods all have the following problems:

[0004] 1. The impact of time lag in the cement burning system is not taken into account. The cement burning process is a complex physical and chemical process with obvious time lag characteristics. For example, the exhaust gas content at the preheater outlet at the current moment can only reflect the combustion status of the burning section over a certain period of time in the past. If the system is adjusted according to the current exhaust gas content, the adjustment will be delayed.

[0005] 2. The drift of gas sensor data is not considered. In actual production, gas analyzer equipment is easily damaged due to high temperature, high pressure, and high dust environment. If the analyzer data is not maintained and calibrated in time after long-term use, it may also drift. In addition, air leakage in the system will also interfere with the analysis and use of data.

[0006] 3. The data of real-time working conditions are not taken into consideration. For different firing conditions, the gas concentration value at the outlet of preheater C1 will produce different data distributions, and the existing solution cannot provide an adaptive adjustment method.

[0007] 4. The exhaust gas content at the preheater outlet is affected by many factors. We cannot only consider the air volume, which makes the adjustment inaccurate and inflexible.

[0008] In response to the above problems, time series prediction methods are currently used to predict some variables that develop and change over time. By observing historical data from a time perspective, defining patterns, and making short-term or long-term predictions on how changes occur, it is determined how the target variable will change in the future. Therefore, the patent number CN114609986A "Cement decomposition furnace denitrification control optimization system and method based on predictive control" discloses a cement decomposition furnace denitrification control optimization method and system based on predictive control, in which the predictive control module uses the CNN-LSTM model to obtain the NOx concentration prediction value. The patent number CN113268871A "Cement chimney NOx prediction method based on multivariate time series deep network model" proposes a cement chimney NOx prediction method based on a multivariate time series deep network model. The method establishes a cement chimney NOx prediction model MT-LSTMs based on a multivariate time series long and short-term neural network LSTM to predict cement chimney NOx. The above models all use the recurrent neural network LSTM. Although the LSTM model has achieved certain results in short-term prediction, it has problems such as long signal propagation path, large hardware costs, excessive consumption of GPU resources and server storage resources, long inference time, and low efficiency in long-term series prediction. This has also limited the effect of long-term series prediction in the cement production process, affecting the accurate determination of the final O2 content target value. Summary of the invention

[0009] The present invention aims to provide a method and system for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction, so as to solve the problem in the prior art that the target value of O2 content at the outlet of preheater C1 is adjusted with deviation, so that the current combustion condition cannot reach the ideal state, resulting in a series of abnormal kiln conditions, increased power consumption of high-temperature fans, heat consumption of clinker and other problems.

[0010] To achieve the above purpose, the present invention adopts the following technical scheme: a method for optimizing the target value of oxygen content at the outlet of the preheater based on multi-step prediction. This scheme takes into account the influence of the time lag of the firing system and the drift phenomenon of the gas sensor data. By predicting the CO and O2 concentrations in the future, combining the prediction results with the adjustment strategy, a real-time O2 target recommendation is given to achieve early prediction and early adjustment, effectively solve the time lag problem of the firing system, and improve the accuracy of the O2 content recommendation. This scheme includes the following steps:

[0011] Step 1: Filter out CO and O2 gas data from the acquired historical data, and divide them into training set and test set after processing according to the set method;

[0012] Step 2: Combine mutual information analysis with process principles to select features and standardize the data;

[0013] Step 3: Build an initial model for multi-step prediction of CO and O2 based on Informer;

[0014] Step 4: Import the training set into the Informer initial model for training, and use the test set to verify the accuracy of the model to obtain the prediction model;

[0015] Step 5, real-time acquisition of system monitoring data, use of the prediction model to obtain CO, O2 multi-step prediction results, the prediction results are passed into the set O2 target value optimization module, and the real-time O2 target recommended value is obtained.

[0016] At the same time, the present solution also provides a preheater outlet oxygen content target value optimization system based on multi-step prediction, which is applied to the above-mentioned preheater outlet oxygen content target value optimization method based on multi-step prediction, including a data acquisition unit, a data processing unit, a model building unit, a data optimization unit and a result output unit;

[0017] The data acquisition unit is used to acquire historical data and perform data preprocessing; the data processing unit is used to filter out CO and O2 gas data from the historical data, and filter the CO and O2 gas data to form training set and test set data;

[0018] The model building unit is used to build an initial model for multi-step prediction of CO and O2 based on Informer, and obtain a prediction model after training and verifying the initial model;

[0019] The data optimization unit is used to optimize the CO and O2 multi-step prediction results obtained by the prediction model and obtain the real-time O2 target recommended value; the data optimization unit includes an optimization module;

[0020] The result output unit is used to output the real-time O2 target recommended value and the adjustment strategy, and the result output unit includes an adjustment strategy module and an output module.

[0021] The principles and advantages of this solution are:

[0022] This solution predicts the CO and O2 concentrations for a period of time in the future, combines the prediction results with the adjustment strategy to give a real-time O2 target recommendation, and makes an advance prediction of the CO and O2 values ​​at the outlet of preheater C1 based on the multi-step prediction model of Informer. Then, the CO concentration trend is calculated based on the prediction results combined with historical data, and finally the O2 target value is adjusted based on the CO concentration trend and future O2 concentration values. This method achieves advance prediction and advance adjustment, effectively solves the problems caused by the time lag of the firing system, and can adapt to the impact of gas sensor drift. Even if the operating conditions change and the data distribution of the preheater exhaust gas changes, this method can still give a suitable recommended value normally. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 It is a flow chart of the method for optimizing the target value of oxygen content at the outlet of the preheater based on multi-step prediction of the present invention.

[0024] Figure 2 It is a structural schematic diagram of the preheater outlet oxygen content target value optimization system based on multi-step prediction of the present invention. DETAILED DESCRIPTION

[0025] The following is further described in detail through specific implementation methods:

[0026] Embodiment 1

[0027] As attached Figure 1 As shown, the method for optimizing the target value of oxygen content at the outlet of the preheater based on multi-step prediction in this embodiment combines the time series prediction method based on deep learning with the "trial and error method" of the cement process to give an O2 target value suitable for the current working conditions. By predicting the CO and O2 concentrations for a period of time in the future, combining the predicted results with the adjustment strategy to give a real-time O2 target recommendation value, this method takes into account the impact of the time lag of the firing system and the drift of the gas sensor data, and effectively solves a series of abnormal kiln conditions caused by the O2 content not matching the current combustion conditions during the combustion process of the cement kiln, and effectively reduces the power consumption of the high-temperature fan and the heat consumption of the clinker. In this embodiment, the optimization method includes the following steps:

[0028] S1, filter out CO and O2 gas data from the acquired historical data, process them according to the set method, and divide them into training set and test set according to a certain ratio.

[0029] In this embodiment, historical data is data obtained by various system monitoring, which is the original data of the historical operation of the cement plant kiln, covering all links in the cement production process, such as raw material processing, clinker calcination, etc. These data are collected in real time through system monitoring equipment and sensors, and record key parameters in the cement production process, including temperature, pressure, flow rate and other information.

[0030] After obtaining the historical data, it is also necessary to process the historical data for outliers. In this embodiment, the outlier processing uses the 3σ rule to judge the outliers: , where is the average value of a feature in the data, and σ is the standard deviation of the feature. Outlier processing can be used to eliminate missing values, outliers, or non-steady-state values ​​caused by sensor or other equipment failures in the historical operation data of cement kilns, ensuring the accuracy and reliability of the data and facilitating subsequent analysis and prediction. Outlier processing can eliminate its impact on data analysis results and ensure the credibility of decisions and optimization measures based on accurate data.

[0031] In the processed historical data, CO and O2 gas data are selected, and the selected CO and O2 gas data are filtered to remove the abnormal values ​​of CO and O2 concentrations caused by backwashing of the gas analysis equipment to ensure the validity and accuracy of the data. In this embodiment, the filtering method includes the following sub-steps:

[0032] S1.1, determine outliers. First, give an upper threshold of O2 concentration, find the points in the O2 concentration sequence that are greater than the threshold, and mark them as outliers.

[0033] S1.2, determine the abnormal interval. With each abnormal value as the center, expand the window to the left and right respectively, until the points in the window are no longer monotonically increasing or decreasing, then mark the area as an abnormal interval, and the abnormal point is contained in the abnormal interval.

[0034] S1.3, after obtaining the abnormal interval, the values ​​in the abnormal interval in the CO and O2 series are replaced by sliding average to remove the abnormal values ​​of CO and O2 concentrations and ensure the validity and accuracy of the data.

[0035] After the filtering process is completed, the data are divided into training sets and test sets according to a certain ratio for future use.

[0036] S2, combine mutual information analysis method with process principles to perform feature selection and standardize the data.

[0037] In this embodiment, feature selection is to calculate the mutual information scores of CO, O2 data and other data points, wherein the mutual information score is determined by the mutual information correlation analysis method to determine the key variable data affecting the CO and O2 content at the C1 outlet, according to the calculation formula of the mutual information:

[0038]

[0039] Where X and Y are continuous variables, I(X; Y) is the mutual information output; p(x, y) is the joint probability density function of X and Y, and p(x) and p(y) are the marginal probability density functions of X and Y, respectively.

[0040] It represents the amount of information about another random variable given the known conditions of a random variable, so as to determine the correlation between random variables.

[0041] After obtaining the mutual information score, the characteristic variables are arranged in descending order according to the mutual information score and the variables with repeated measurement points are eliminated. Combined with the process principle, the characteristic variables with high mutual information scores and in line with the process principle are screened out as the input of the initial model.

[0042] After feature selection, the selected data needs to be standardized. In this embodiment, zero-mean normalization is used to complete the data processing, and its expression is:

[0043]

[0044] In the formula, x * The data is processed by standardization. is the mean of the original data; σ is the standard deviation of the original data. The standardized data can remove the unit restriction of the data and convert it into a dimensionless pure value, which is convenient for the comparison and weighting of indicators of different units or magnitudes.

[0045] S3, build an initial model for multi-step time series prediction of CO and O2 based on Informer.

[0046] In this embodiment, the main structure of the initial model based on Informer includes an input representation embedding layer, an encoding layer (Encoder) and a decoding layer (Decoder). Among them, the input representation embedding layer is composed of a convolutional layer and an embedding layer, including three independent parts: scalar projection, local timestamp embedding and global timestamp embedding.

[0047] The encoder layer is used to extract long-range dependencies of long-time series input. In this embodiment, the encoder layer uses the ProbSparse self-attention mechanism to calculate the attention value (attention) through the obtained query value (query) to optimize the calculation efficiency, which is expressed as follows

[0048]

[0049] Where A is the output of the attention mechanism; Q, K, and V represent query, key, and value matrices, respectively; is a sparse matrix obtained by sparsity measurement, which only includes the first u decisive query values; K T is the transpose of K, and d is the scaling factor. By calculating the dot product of the query Q and the key K, and using the sparsity measure, a sparse matrix is ​​obtained. Then perform Softmax operation on this sparse matrix and the transpose of K, and finally multiply the result with the value V to get the final output.

[0050] The decoding layer (Decoder) is used to decode the intermediate representation to output the final prediction result. The decoding layer includes a self-attention layer and a fully connected layer. In this embodiment, a generative reasoning process is adopted to decode all elements at once, which is more efficient than dynamic decoding.

[0051] S4, import the training set into the Informer initial model for training, and use the test set to verify the accuracy of the model to obtain the prediction model.

[0052] In this embodiment, the verification process also includes using the MSE loss function to measure the difference between the model prediction value and the actual value, and using the Adam optimizer for optimization. In terms of model prediction accuracy, MAE and MSE are used for evaluation to improve the accuracy of the prediction model.

[0053] S5, real-time acquisition of system monitoring data, use of the prediction model to obtain CO, O2 multi-step prediction results, the prediction results are passed to the set O2 target value optimization module, and the real-time O2 target recommended value is obtained.

[0054] In this embodiment, by acquiring the monitoring data of the DCS system in real time, that is, reading the monitoring data of the current cement production system from the DCS database in real time, including the data feature categories obtained through S4, the prediction model is used to obtain the multi-step prediction results of CO and O2, and the results are optimized.

[0055] In this embodiment, the O2 target value optimization module is also constructed, and the construction method includes the following sub-steps:

[0056] S5.1, take the average value of the predicted O2 concentration in the future time period T1 as the initial target value for each adjustment.

[0057] S5.2, splice the predicted CO concentration with the past CO concentration, and perform linear fitting on the spliced CO time series data to obtain the slope s and intercept p of the linear fitting of the CO concentration in this time period.

[0058] S5.3, take T2 as the step size, collect the CO data in the training set and perform linear fitting to obtain the statistical results. Set the CO slope thresholds s1 and s2 according to the statistical results, where s2 > s1.

[0059] S5.4, set T1 as the CO concentration upper limit control period, and s1 and s2 as the slope upper limits, that is, set the amplitude of the CO change trend within the time period T1, and observe the waiting time Δt s 。

[0060] S5.5, determine the change state of the CO slope s within the set time period, and adjust the O2 concentration target value according to the change situation according to the adjustment strategy to obtain the O2 target recommended value.

[0061] In this embodiment, the adjustment strategy includes an upward adjustment strategy and a downward adjustment strategy.

[0062] Among them, the upward adjustment strategy includes the following sub-steps:

[0063] Step a. When the slope s of the linear fitting of the CO data within the time period T2 exceeds the set slope threshold s1, that is, s > s1, calculate the adjustment index according to the following formula and adjust the O2 target value;

[0064]

[0065] Step b. After passing through Δt s observe whether the CO concentration meets the upper limit control, otherwise repeat step a

[0066] The downward adjustment strategy includes the following sub-steps:

[0067] Step A. After the system runs stably for T3 time, generate a BOOL-type random variable p, and calculate the adjustment index according to the p value Adjust the O2 target value,

[0068]

[0069] Step B. After passing through Δt s observe whether the change of the CO concentration within the time period T2 conforms to the set trend. If not, repeat step A; if the CO concentration slope hits the upper limit value and s < s1, then adjust the index for And the O2 target value will no longer be adjusted downward during the T4 time, otherwise it will jump to the upward adjustment strategy.

[0070] In this embodiment, the optimization method includes the steps of obtaining historical data from the DCS system database, processing abnormal values ​​of historical data, feature selection, building a multi-step prediction model of CO and O2 based on Informer, training and verifying the model, and optimizing the target value using the prediction results. The multi-step prediction model based on Informer is used to make advance predictions of the CO and O2 values ​​at the outlet of the preheater C1, and then the CO concentration trend is calculated based on the prediction results combined with historical data. Finally, the O2 target value is adjusted based on the CO concentration trend and the future O2 concentration value. This method achieves advance prediction and advance adjustment, effectively solves the problems caused by the time lag of the firing system, and can adapt to the influence of gas sensor drift. Even if the working conditions change and the data distribution of the preheater exhaust gas changes, the method can still give appropriate recommended values ​​normally, and has strong robustness.

[0071] Embodiment 2

[0072] In this embodiment, a system for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction is provided, which is applied to the method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction in Embodiment 1. Figure 2 As shown, it includes a data acquisition unit, a data processing unit, a model building unit, a data optimization unit and a result output unit.

[0073] The data acquisition unit is used to obtain historical data from the DCS system database and perform abnormal value processing on the historical data. In this embodiment, the key parameters in the cement production process, such as temperature, pressure, flow rate and other information, are collected in real time through system monitoring equipment and sensors, and recorded. The data processing unit is used to filter out CO and O2 gas data from historical data, and after filtering the CO and O2 gas data, divide them into training set and test set data according to a certain ratio.

[0074] The model building unit is used to build an initial model for multi-step prediction of CO and O2 based on Informer, and obtain a prediction model after training and verifying the initial model. In this embodiment, the main structure of the initial model includes an input representation embedding layer, an encoder encoding layer, and a decoder decoding layer. The encoding layer is used to extract robust long-range dependencies of long-time series inputs. The decoding layer is used to decode the intermediate representation to output the final prediction result. The loss function of the prediction model uses MSE loss, and the optimizer uses Adam optimizer. In terms of the accuracy of model prediction, MAE and MSE are used for evaluation.

[0075] The data optimization unit is used to optimize the CO and O2 multi-step prediction results obtained by the prediction model and obtain the real-time O2 target recommendation value. The data optimization unit includes an optimization module.

[0076] The result output unit is used to output the real-time O2 target recommended value and the adjustment strategy, and the result output unit includes an adjustment strategy module and an output module. Among them, the adjustment strategy module includes an upward adjustment strategy and a downward adjustment strategy, which will not be repeated here.

[0077] This solution predicts the CO and O2 concentrations in the future, and combines the prediction results with the adjustment strategy to give a real-time O2 target recommendation. This method takes into account the impact of the time lag of the firing system and the drift of the gas sensor data, and effectively solves a series of abnormal kiln conditions caused by the O2 content not matching the current combustion conditions during the combustion process of the cement kiln, thereby reducing the power consumption of high-temperature fans and the heat consumption of clinker.

[0078] The above is only an embodiment of the present invention, and the common knowledge such as the known specific technical solutions and / or characteristics in the solution is not described in detail here. It should be pointed out that for those skilled in the art, without departing from the technical solution of the present invention, several modifications and improvements can be made, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction, characterized in that: The following steps are involved: Step 1: Filter out CO and O2 gas data from the acquired historical data, and divide them into training set and test set after processing according to the set method; Step 2: Combine mutual information analysis with process principles to select features and standardize the data; Step 3: Build an initial model for multi-step prediction of CO and O2 based on Informer; Step 4: Import the training set into the Informer initial model for training, and use the test set to verify the accuracy of the model to obtain the prediction model; Step 5, real-time acquisition of system monitoring data, use of the prediction model to obtain CO, O2 multi-step prediction results, the prediction results are passed into the set O2 target value optimization module, and the real-time O2 target recommended value is obtained.

2. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 1, characterized in that: In step 5, the optimization module is constructed, and the construction method includes the following sub-steps: Step 5.1, take the average value of O2 concentration predicted in the future T1 time period as the initial target value for each adjustment; Step 5.2, splice the predicted CO concentration with the past CO concentration, and perform a linear fit on the spliced ​​CO time series data to obtain the slope s and intercept p of the linear fit of the CO concentration in this time period; Step 5.3, with T2 as the step length, collect CO data in the training set and perform linear fitting to obtain statistical results, and set CO slope thresholds s1 and s2 according to the statistical results, where s2>s1; Step 5.4, set T1 as the CO concentration upper limit control period, s1 and s2 as the slope upper limits, and observe the waiting time Δt s ; Step 5.5, determine the change state of the CO slope s within the set time period, adjust the O2 concentration target value according to the change according to the adjustment strategy, and obtain the O2 target recommended value.

3. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 2, characterized in that: The adjustment strategy includes an upward adjustment strategy and a downward adjustment strategy; wherein the upward adjustment strategy includes the following sub-steps: Step a. When the slope s of the linear fitting of CO data in the T2 period exceeds the set slope threshold s1, that is, s>s1, the adjustment index is calculated according to the following formula: and adjust the O2 target value; Step b. After Δt s Finally, observe whether the CO concentration meets the upper limit control, otherwise repeat step a.

4. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 3, characterized in that: The downward adjustment strategy includes the following sub-steps: Step A. After the system runs smoothly for T3 time, generate a BOOL random variable p and calculate the adjustment index according to the p value Adjust the O2 target value, Step B. After a lapse of Δt s observe whether the change in CO concentration within the T1 time period conforms to the set trend. If not, repeat Step A; if the CO concentration slope hits the upper limit value and s < s1, adjust the index to and no downward adjustment operation is performed on the O2 target value within the T4 time period, otherwise jump to the upward adjustment strategy.

5. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 1, characterized in that: In step 1, it also includes processing the abnormal values ​​of the historical data; the setting method is to filter the screened CO and O2 gas data, and the filtering method includes the following sub-steps: Step 1.1, determine the outliers; given the upper threshold of O2 concentration, find the points in the O2 concentration sequence that are greater than the threshold and mark them as outliers; Step 1.2, determine the abnormal interval; take each abnormal value as the center, expand the window to the left and right respectively, until the points in the window are no longer monotonically increasing or decreasing, and mark the area as the abnormal interval; In step 1.3, the values ​​in the abnormal interval of the CO and O2 series are replaced by sliding average.

6. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 1, characterized in that: The feature selection is to calculate the mutual information scores between CO, O2 data and other data points, arrange the feature variables in descending order according to the mutual information scores and eliminate variables with repeated measurement points, and combine the process principles to screen out feature variables with high mutual information scores and in line with the process principles as inputs to the initial model.

7. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 1, characterized in that: The data is standardized by using zero-mean normalization to complete the data processing, and its expression is In the formula, x * The data is processed by standardization. is the mean of the original data; σ is the standard deviation of the original data.

8. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 1, characterized in that: The initial model includes an input representation embedding layer, an encoding layer and a decoding layer; wherein the input representation embedding layer includes three independent parts: scalar projection, local timestamp embedding and global timestamp embedding; the encoding layer is used to extract long-range dependencies of long-time series inputs; the decoding layer is used to decode the intermediate representation to output the final prediction result, and the decoding layer includes a self-attention layer and a fully connected layer.

9. The method for optimizing the target value of oxygen content at the outlet of a preheater based on multi-step prediction according to claim 8, characterized in that: The encoding layer adopts the ProbSparse self-attention mechanism to calculate the attention value through the obtained query value, which is expressed as Where A is the output of the attention mechanism; Q, K, and V represent query, key, and value matrices, respectively; is a sparse matrix, including only the first u decisive query values; is the transpose of K, and d is the scaling factor.

10. The preheater outlet oxygen content target value optimization system based on multi-step prediction is characterized by: A method for optimizing a target value of oxygen content at a preheater outlet based on multi-step prediction as described in any one of claims 1 to 9, comprising a data acquisition unit, a data processing unit, a model building unit, a data optimization unit and a result output unit; The data acquisition unit is used to acquire historical data and perform data preprocessing; the data processing unit is used to filter out CO and O2 gas data from the historical data, and filter the CO and O2 gas data to form training set and test set data; The model building unit is used to build an initial model for multi-step prediction of CO and O2 based on Informer, and obtain a prediction model after training and verifying the initial model; The data optimization unit is used to optimize the CO and O2 multi-step prediction results obtained by the prediction model and obtain the real-time O2 target recommended value; the data optimization unit includes an optimization module; The result output unit is used to output the real-time O2 target recommended value and the adjustment strategy, and the result output unit includes an adjustment strategy module and an output module.

Images