A method for optimizing parameters of a grate cooler based on a clustering algorithm

Abstract

The present application relates to the technical field of grate cooler parameter control, and discloses a grate cooler equipment parameter optimization method based on a clustering algorithm, which comprises a modeling stage, two sample data sets are obtained according to the characteristics of the grate cooler, and two model groups are trained respectively based on the sample data sets, and a clustering cluster is obtained by using the model groups; an optimization stage, a sample set is obtained according to the established clustering cluster, candidate samples are obtained after the sample set is filtered, the candidate samples are evaluated by value scores, and an optimization value is obtained; and an updating stage, periodic updating data is obtained, the original sample set is replaced, the model is trained, and the original model is updated. The present application can automatically and real-timely monitor the changes of the working conditions of the grate cooler and give the value of the control parameter, reduce the workload of the operator on the grate cooler, can explain the reasons for the parameter adjustment of the grate cooler and the data sources, is more persuasive in results, improves the reliability, and effectively reduces the energy consumption.

Description

Technical Field

[0001] The invention relates to the technical field of grate cooler control, and in particular to a grate cooler equipment parameter optimization method based on a clustering algorithm. Background Art

[0002] Grate coolers are commonly used in cement production to cool hot gases and solidify clinker. Their unique grate structure and efficient cooling system enable rapid clinker cooling and heat recovery, improving production efficiency and energy utilization. The key control parameters of a grate cooler include the currents of the multiple cooling fans and the hydraulic motor that drives the grates. Typically, this hydraulic motor current is not directly controlled by an operator or algorithm, but rather by a dedicated PID or PLC unit that directly controls the hydraulic press based on a given grate speed. Therefore, the control parameters of a grate cooler can be simplified to the cooling fan current and grate speed. Currently, most grate cooler control in cement production requires an experienced operator team. However, due to the numerous parameters involved in controlling grate cooler operation and the strong coupling between these factors, control is complex, labor-intensive, and inefficient. Relying solely on experienced operators can lead to problems such as parameter adjustment lag and poor stability. In addition, the operation of the grate cooler consumes a lot of electricity because the operation of the cooling fan and grate bed requires a lot of electricity, which causes the energy consumption of the grate cooler to be high. When there are problems with the control parameters, the energy consumption will increase further.

[0003] Currently, a small number of cement production operations utilize automated methods to control grate coolers, utilizing rule engines, fuzzy reasoning, and predictive control to optimize control parameters such as fan current and grate speed. This has improved efficiency and reduced labor intensity to a certain extent. However, these typical optimization methods also have drawbacks. For example, the effectiveness of rule engines relies heavily on the completeness of experience and the significance of operating conditions. When experience is incomplete or operating conditions are unclear, their effectiveness can rapidly deteriorate. While fuzzy reasoning can effectively handle uncertainty and ambiguity in industrial control, it also relies heavily on predefined fuzzy rules. Traditional fuzzy reasoning systems generally lack self-learning capabilities, making it difficult to automatically optimize rules based on new data. Predictive control methods achieve preemptive adjustments to grate coolers by predicting relevant key controlled variables, theoretically offering good control results. However, their effectiveness relies heavily on the accuracy of the prediction model. Because unmeasurable variables, such as the temperature of the clinker leaving the rotary kiln, significantly influence the prediction of grate cooler controlled variables, establishing an accurate and reliable prediction model for grate coolers is difficult. In addition, intelligent methods such as fuzzy reasoning and predictive control often lack explainability, and it is difficult to gain the trust of the operator group by optimizing the control parameters. In severe cases, the operator group may even disable the use of the automated algorithm, making it unable to play an auxiliary role. There are always problems such as high labor intensity, heavy workload, high energy consumption of the grate cooler, and poor operating stability. Summary of the Invention

[0004] The present invention aims to provide a grate cooler equipment parameter optimization method based on clustering algorithm to solve the problem that the existing grate cooler control prediction results cannot be reasonably interpreted and analyzed to assist control, resulting in large workload, poor stability and high energy consumption of grate cooler control operations.

[0005] In order to achieve the above-mentioned purpose, the present invention adopts the following technical scheme: a method for optimizing the parameters of grate cooler equipment based on a clustering algorithm. This scheme sets up two groups of models in a targeted manner according to the structure of the grate cooler equipment, and conducts model training separately to ensure the accuracy and fit of the data. It can obtain data in real time and monitor the changes of the grate cooler, and give control parameter values ​​to reduce workload, thereby improving control efficiency and timeliness, ensuring the stable operation of the grate cooler, and further reducing energy consumption. At the same time, in the optimization stage, it can explain the reasons for the adjustment of the grate cooler parameters and the source of the data, so that the data output by the model is interpretable, which can be more in line with the experience judgment of the operating team and improve the credibility of the model. The optimization method of this scheme specifically includes the following steps:

[0006] Modeling phase:

[0007] S11, obtaining time series samples, dividing the time series samples into one-segment samples and two-segment samples according to the characteristics of the grate cooler, and extracting multidimensional features of the samples accordingly, which are used to train the working condition clustering model of the grate cooler to obtain a one-segment model group and a two-segment model group;

[0008] S12, obtaining clusters of historical samples according to each model group, calculating the cluster centers as the cluster features, dividing the sample data into corresponding clusters, and making the model group persistent;

[0009] Optimization stage:

[0010] S21, determine the current control state of the grate cooler and enter the optimization process according to the control state;

[0011] S22, obtaining the real-time operating condition data corresponding to each model group, calculating the cluster to which the real-time operating condition data belongs, and obtaining a sample set based on the cluster characteristics; filtering the sample set according to the set filtering mode to obtain candidate samples;

[0012] S23, calculating the value scores of the candidate samples according to the evaluation method, selecting several sample data with the highest value scores, taking their average as the optimal value and sending it to the control system;

[0013] Update phase:

[0014] S31, acquiring update data according to a set period to form update samples, and replacing part of the sample set to form an update set;

[0015] S32, setting initial cluster centers, using the update set to train the initial cluster centers until the model converges, and obtaining an updated model group;

[0016] S33, the updated model group is used in the optimization process.

[0017] The principles and advantages of this solution are:

[0018] In actual grate cooler control, since grate coolers have many control parameters and are mutually coupled, when constructing a prediction model for a grate cooler, the interrelationships between the parameters are taken into account and a comprehensive prediction model is constructed to ensure data accuracy and mutual influence, thereby ensuring the feasibility of the prediction model. At the same time, building a comprehensive model can also more effectively utilize limited resources and reduce the investment in manpower and material resources. Therefore, when constructing a prediction model for a grate cooler, the construction of multiple models is not considered. In addition, the coordination and management of multiple models can also be a more complex issue. Furthermore, the grate cooler itself is an equipment control system, so it is not considered that multiple prediction models need to be constructed for control, which increases the workload, and thus the complex construction of multiple models is generally not considered.

[0019] But in fact, in the actual operation of the grate cooler, the grate cooler itself contains 11 fans and 2 sections of grate plates. Although the main function of the fans is to increase the cooling effect, discharge heat from the system, and maintain the operating temperature of the equipment, the grate plates are mainly used to regulate the flow of the cooling medium. Therefore, a comprehensive approach is adopted during control. However, there are many types of fans. When facing different cooling positions, their corresponding parameters need to be adjusted, resulting in many control parameters and strong correlations, which makes it difficult to split and control them one by one. After careful analysis, this solution found that among these 11 fans and 2 sections of grate plates, some fans are subject to the same operating requirements and control parameters. Therefore, based on the cement burning process and the function of the grate cooler, we selected points that have a causal relationship with the control of the grate speed and cooling fan current of the grate cooler. In combination with the structural characteristics of the grate cooler, we established two control models for the grate cooler: the first stage control and the second stage control. We then grouped the measurement points and used them to train the models, making them more suitable for the actual operating status of the grate cooler and achieving precise control for different operating requirements, thereby further reducing the energy consumption of the grate cooler. At the same time, these two groups of models can automatically monitor the changes in the operating conditions of the grate cooler in real time and give the values ​​of the control parameters, reducing the workload of the operator on the grate cooler, making the reasons for the parameter adjustment of the grate cooler and the data source clear, thus being more convincing and improving the credibility of the control system. In this way, we can achieve the effect of effective prediction and control, realize reliable auxiliary control, truly and effectively reduce the workload and labor intensity of the operator group, and improve control efficiency.

[0020] At the same time, this solution increases the average secondary and tertiary air temperatures by 17.2°C compared to the operator group control, reduces the standard deviation by 6.1°C, and makes heat recovery more sufficient and more stable. From the perspective of energy consumption, the method proposed in this solution can save 3.4% of the system power consumption of the grate cooler compared to the operator group control.

[0021] In summary, the implementation of this solution has the following advantages:

[0022] (1) It can reduce the intervention of the operator team and greatly reduce the workload and labor intensity of the operator team;

[0023] (2) Ensure the stable operation of the grate cooler and reduce the power consumption of the grate cooler;

[0024] (3) Make the algorithm output values ​​interpretable, provide auxiliary analysis basis, improve the credibility of the results, and ensure usability.

[0025] Furthermore, in the optimization stage, the filtering modes include three types, namely upper and lower limit constraint filtering, change constraint filtering and similarity filtering; among them, upper and lower limit constraint filtering is to select samples within the upper and lower limits according to the set upper and lower limits of the parameter items; the change constraint filtering is to delete samples whose parameter changes exceed the threshold; the similarity filtering is to calculate the fusion similarity between historical samples and current samples, and retain the 5 samples with the highest similarity scores.

[0026] Furthermore, in the optimization stage, the evaluation method is to calculate the value score of the candidate sample according to the following formula:

[0027] Value score = (secondary air temperature + tertiary air temperature) / average fan current.

[0028] Furthermore, in the modeling stage, the time series samples are variables related to the control parameters of the grate cooler; among them, the first-stage samples are the data of the first five cooling fans corresponding to the first grate plate of the grate cooler; the second-stage samples are the data of the last six cooling fans corresponding to the second grate plate of the grate cooler, and the corresponding number of sample measurement points selected for the first stage is 22 and the number of sample measurement points for the second stage is 23.

[0029] Furthermore, in the modeling stage, it also includes grouping within the first stage sample and the second stage sample respectively; performing a histogram analysis on the given value of the raw material feed amount in the sample, grouping according to the analysis results, and then performing modeling processing within the group respectively to form multiple unit models. The multiple unit models constitute a first stage model group or a two stage model group.

[0030] Furthermore, the modeling phase also includes preprocessing of the time series sample data, which includes the following sub-steps:

[0031] S11.1, obtain the clustering characteristics of the time series sample data of each measuring point, including the mean μ, standard deviation σ, maximum value and minimum value of each measuring point;

[0032] S11.2, use the z-score method to normalize the values ​​of each measuring point, and the calculation expression is m ′ =(m-μ) / σ, where m is the original measurement point value, m ′ is the normalized measurement point value;

[0033] S11.3, use a sliding window with a window size of 10 and a step size of 5 to obtain the mean of the measurement point values.

[0034] Furthermore, in the modeling phase, the clustering method includes the following sub-steps:

[0035] S12.1, according to the set cluster radius, minimum number of samples and search step, traverse each sample in the data set, calculate the number of samples in each sample's neighborhood, and if the number of samples is not less than the minimum number of samples, classify them as core samples;

[0036] S12.2, taking any core sample as the center, search for other core samples within the cluster radius, and mark the samples within this range as a cluster based on the fusion similarity index;

[0037] S12.3, switch the central core sample and repeat S12.2 until all core samples are visited and unlabeled samples are removed; and the clustering effect is evaluated using the set clustering effect index.

[0038] Furthermore, the fusion similarity index includes sample similarity and measurement point similarity; the sample similarity is the sum of similarities of each measurement point constituting the sample; the measurement point similarity includes cosine similarity of measurement point statistics and sequence similarity of measurement point shapes.

[0039] Furthermore, the clustering effect index = intra-cluster divergence / inter-cluster divergence; the intra-cluster divergence is the average value of the distance between samples within a cluster; the inter-cluster divergence is the average value of the distance between samples between clusters.

[0040] Furthermore, the cosine similarity is expressed as In the formula, δ is the two measuring points and The cosine similarity of The measurement point sequence The statistical eigenvector of The measurement point sequence The statistical eigenvector of , ||·|| represents the modulus of the vector. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 The diagram is a structural diagram of a method for optimizing parameters of grate cooler equipment based on a clustering algorithm according to the present invention.

[0042] Figure 2 The present invention provides a distribution histogram of given values ​​of raw meal feed quantity in a method for optimizing grate cooler equipment parameters based on a clustering algorithm. DETAILED DESCRIPTION

[0043] The following is further described in detail through specific implementation methods:

[0044] Example 1

[0045] As attached Figure 1As shown in the figure, a clustering algorithm-based grate cooler equipment parameter optimization method in this embodiment is used. This method, based on a dynamic clustering algorithm, not only makes the output results interpretable, provides reliable parameter adjustment recommendations, reduces the workload of the operator team on the grate cooler, but also ensures real-time and stable adjustment of the grate cooler fan current and grate speed, achieving the effect of energy conservation. In this embodiment, the optimization method mainly includes three steps: modeling stage, optimization stage, and update stage, which are described in detail below.

[0046] 1. Modeling stage.

[0047] Data Acquisition: First, obtain relevant point data for the grate cooler from the DCS. In this embodiment, the start and end time of historical data can be set to, for example, June 1, 2023 to June 1, 2024. The DCS samples the measurement points every 2 seconds, excluding downtime periods. This results in a total of 11,059,200 samples, a massive amount of data. Therefore, measurement point selection is necessary within the acquired historical data.

[0048] Measuring point selection: According to the cement burning process and the function of the grate cooler, select the points that are causally related to the control of the grate speed of the grate cooler and the current of the cooling fan, including the feedback value of the O2 concentration at the outlet of the preheater C1, the feedback value of the negative pressure at the outlet of the preheater C1, the feedback value of the temperature at the outlet of the decomposition furnace, the feedback value of the negative pressure at the outlet of the decomposition furnace, the feedback value of the temperature in the smoke chamber at the tail of the kiln, the feedback value of the negative pressure in the smoke chamber at the tail of the kiln, the feedback value of the flame temperature of the fire-watching TV at the head of the kiln, the feedback value of the clinker temperature of the fire-watching TV at the head of the kiln, the feedback value of the rotary Controlled variables include kiln power feedback value, rotary kiln speed feedback value, kiln head hood negative pressure feedback value, secondary air temperature feedback value, tertiary air temperature feedback value, waste heat power generation temperature feedback value, grate cooler first-stage grate pressure feedback value, grate cooler second-stage grate pressure feedback value, as well as raw material feed quantity set value, grate cooler first-stage grate speed set value, grate cooler second-stage grate speed set value, grate cooler 11 cooling fan current set value, etc., a total of 30 measuring points, which are grouped after the measuring points are selected.

[0049] Measuring point grouping: In this embodiment, the grate cooler involved is a two-stage grate cooler, that is, there are two independently moving clinker cooling sections. The first grate plate of the grate cooler is cooled by the first five cooling fans, and the second grate plate is cooled by the last six cooling fans. Therefore, based on the structure of the grate cooler, two control models are established for the first stage control and the second stage control of the grate cooler. Since there are multiple models in each control stage, they are recorded as a first stage model group and a second stage model group. In this embodiment, the model principles, modeling, and usage steps of the first and second stage model groups are exactly the same.

[0050] The 30 measurement points obtained were grouped according to the two-stage control model. One of the model groups included 22 measurement points, including the O2 concentration feedback value at the outlet of preheater C1, the negative pressure feedback value at the outlet of preheater C1, the temperature feedback value at the outlet of decomposition furnace, the negative pressure feedback value at the outlet of decomposition furnace, the temperature feedback value of the smoke chamber at the tail of kiln, the negative pressure feedback value of the smoke chamber at the tail of kiln, the flame temperature feedback value of the fire-watching TV at the kiln head, the clinker temperature feedback value of the fire-watching TV at the kiln head, the power feedback value of the rotary kiln, the speed feedback value of the rotary kiln, the negative pressure feedback value of the kiln head hood, the secondary air temperature feedback value, the tertiary air temperature feedback value, the waste heat power generation temperature feedback value, the pressure feedback value under the first grate of the grate cooler, the raw meal feed amount set value, the grate speed set value of the first grate of the grate cooler, and the current set value of the cooling fans No. 1 to 5 of the grate cooler. The second-stage model group includes 23 measurement points: preheater C1 outlet O2 concentration feedback, preheater C1 outlet negative pressure feedback, precalciner outlet temperature feedback, precalciner outlet negative pressure feedback, kiln tail flue gas chamber temperature feedback, kiln head flame temperature feedback, kiln head clinker temperature feedback, rotary kiln power feedback, rotary kiln speed feedback, kiln head hood negative pressure feedback, secondary air temperature feedback, tertiary air temperature feedback, waste heat power generation temperature feedback, grate cooler second-stage grate pressure feedback, raw meal feed rate setpoint, grate cooler second-stage grate speed setpoint, and grate cooler cooling fan current setpoints 6-11. The controlled variables involved in the first- and second-stage model groups are used to describe the operating conditions and serve as the data source for feature extraction in the subsequent clustering model construction. After the measurement points are grouped, the corresponding time series samples are obtained and divided into first- and second-stage samples for sample construction.

[0051] Sample construction: Since the process of feeding raw materials from the preheater to the grate cooler to cool the clinker is a process with a large time lag, in this embodiment, the constructed working condition samples are in the form of time series and have a long time span. According to process experience, it is set to 30 minutes, and sampling is performed every 2 seconds, that is, each sample consists of 900 sampling points. Among them, each sampling point of the first-stage model group is a vector composed of the values ​​of the above 22 measuring points, that is, one sample is a 900×22 matrix; each sampling point of the second-stage model group is a vector composed of the values ​​of the above 23 measuring points, that is, one sample is a 900×23 matrix. A non-overlapping sliding window is used to construct the sample, that is, the number of time steps of the sampling window is 900, and the total number of samples corresponding to each model in the historical data is 11,512.

[0052] After sampling is completed, the samples are grouped. In this embodiment, in the cement burning system, the raw material feed rate is the most fundamental factor affecting the working conditions. Other control variables are mainly set to adapt to the raw material feed rate, and the same is true for the grate cooler control parameters. Therefore, in this embodiment, the raw material feed rate set values ​​in the historical data samples are histogram analyzed to draw a distribution histogram of the sample raw material feed rate set values. As shown in the attached figure, Figure 2As shown, the raw material feed rate setpoint exhibits a clear approximate mixed Gaussian distribution, with multiple typical clustering intervals. Therefore, in this embodiment, anchor points for the raw material feed rate setpoints are set based on the center values ​​of the feed rate clustering intervals, such as 500, 510, 520, 525, 530, 540, 550, 560, 570, 580, 590, 600, 605, 610, 615, 620, 625, 630, 635, 640, 645, and 650. Samples are divided into sample groups represented by the nearest anchor points according to the raw material feed rate setpoints, and modeling is performed separately. This facilitates rapid grouping of the acquired sample data, ensuring accuracy and stability. In this embodiment, during inference, the anchor point closest to the real-time raw material feed rate is first found, and then the clustering model of this anchor point is used for inference. As a result, multiple models exist in the first-stage model group and the second-stage model group based on the grouping of the raw material feed rate setpoints.

[0053] In this embodiment, the obtained measurement point data is preprocessed, and the preprocessing steps include:

[0054] 1) Count the clustering features of the time series sample data at each measuring point, including the mean μ, standard deviation σ, maximum value and minimum value of the measuring point, and generate some feature parameters as part of the subsequent clustering features.

[0055] 2) Use the z-score method to normalize the values ​​of each measuring point, and the calculation expression is m ′ =(m-μ) / σ, where m is the original measurement point value, m ′ is the normalized measurement point value.

[0056] 3) A sliding window with a window size of 10 and a step size of 5 is used to obtain the mean of the measurement point values, which reduces the volatility of the measurement point data and highlights the overall shape and trend of the measurement point data changes.

[0057] After the processing of the measuring point data is completed, the multi-dimensional features of the samples are extracted, and the fusion similarity is used as the sample similarity measurement indicator to perform sample clustering, which is used to train the working condition clustering model of the grate cooler to obtain the first-stage model group and the second-stage model group.

[0058] In this embodiment, the fusion similarity index includes sample similarity and measurement point similarity. Sample similarity is the sum of the similarities of the individual measurement points that make up the sample. Measurement point similarity includes the cosine similarity of the measurement point statistics and the sequence similarity of the measurement point shapes.

[0059] Among them, the cosine similarity of the measurement points is calculated using the feature vector composed of the statistical mean, standard deviation, maximum value and minimum value of the measurement points. For example, for a certain measurement point sequence and Assume that their statistical eigenvectors are and Then the cosine similarity is expressed as Here, ||·|| represents the modulus of the vector.

[0060] Sequence similarity is calculated using the Dynamic Time Wrapping (DTW) algorithm. Dynamic programming is used to find an optimal matching path between two measurement point sequences, minimizing the sum of the distances between corresponding points on the time axis of the two measurement point sequences. This minimum sum of distances is the sequence similarity. After calculating the cosine similarity and sequence similarity of a measurement point, the two similarities are added together to form the fused similarity for that measurement point.

[0061] The calculated fusion similarity is used to cluster samples, and clusters of historical samples are obtained according to each model group. In this embodiment, the samples in each raw material feed amount given value group in the first stage model group and the second stage model group are clustered respectively. In this embodiment, a density-based spatial clustering algorithm (Density-Based Spatial Clustering of Application with Noise, DBSCAN) is used to cluster based on the distribution density of the samples. There is no need to pre-specify the number of clusters to be clustered, and the clusters existing in the sample distribution can be automatically mined. The cluster center is calculated as the cluster feature, and the sample data is divided into the corresponding cluster to make the model group persistent. In this embodiment, the method for obtaining the cluster includes the following steps:

[0062] 1) Specify two hyperparameters: cluster radius ε and minimum number of samples minPts. In this embodiment, the optimal values ​​of these two hyperparameters need to be determined by grid search. The basis for determining the optimal values ​​is the final clustering effect.

[0063] The clustering effect is evaluated using a set clustering effect index: intra-cluster divergence / inter-cluster divergence. Intra-cluster divergence is the average distance between samples within a cluster; inter-cluster divergence is the average distance between samples between clusters. In this embodiment, the search ranges for the cluster radius ε and the minimum sample number minPts are set to 2.0-10.0 and 20-100, respectively, and the search step sizes are set to 0.5 and 5, respectively.

[0064] According to the set clustering radius, minimum number of samples and search step, each sample in the data set is traversed, and the number of samples in the neighborhood of each sample is calculated. In this embodiment, the neighborhood is an area whose distance from the sample is not greater than the clustering radius. When the number of samples in the neighborhood of each sample is not less than the minimum number of samples minPts, the sample is classified as a core sample.

[0065] 2) Taking any core sample as the center, that is, starting from any unvisited core sample, search for other core samples within its clustering radius. Repeat this step for each core sample in the domain until no new core samples appear in the neighborhood. The core samples within this range and the samples in the neighborhood are marked as a cluster according to the fusion similarity index.

[0066] 3) Switch the central core sample and repeat step 2) until all core samples have been visited. Unlabeled samples, that is, samples that do not belong to any cluster, are removed. The clustering effect is evaluated using the set clustering effect index to optimize the clustering effect. In this embodiment, after a grid search, the optimal cluster radius ε and minimum sample number minPts hyperparameters are set to 4.0 and 35, respectively. Under this hyperparameter combination, the intra-cluster divergence / inter-cluster divergence value is minimized.

[0067] At the same time, this embodiment also includes persisting the model. That is, after the samples grouped according to the given raw meal feed rate are clustered as described above, a total of several cluster models are generated for the first-stage model group and a total of several cluster models are generated for the second-stage model group. The samples are divided according to the cluster model and the cluster center of each cluster is calculated. The cluster model, cluster label, sample set corresponding to the cluster, and cluster center are saved to disk.

[0068] 2. Optimization stage.

[0069] Determine whether to enter the optimization process: obtain the data of the relevant points at the current moment from the DCS system to form a sample to describe the current working conditions. The sample is composed of multi-dimensional measurement point data of 900 current and historical time steps. The mean, standard deviation, maximum value, and minimum value of each measurement point of the sample are extracted, and normalization and smoothing operations are performed. Check whether there are abnormal working conditions based on real-time data. In this embodiment, the current grate cooler valve group pressure, grate plate temperature, and firing quality indicators are checked to see if they are abnormal. If they are abnormal, the control of the grate cooler is taken over by the work group or the expert control model. At the same time, check whether the current grate cooler system work group operation signal is valid. If there is a work group operation, it will remain under the control of the work group. Otherwise, enter the automatic optimization process of the grate cooler control parameters.

[0070] Load the saved clustering model, cluster labels, sample sets corresponding to the clusters, cluster centers, and other data from disk; obtain the raw material feed rate setpoint data for the real-time operating condition samples corresponding to each model group, and locate the current sample to the corresponding cluster model in the first and second stage model groups. In the corresponding cluster model, input the sample features required by the first and second stage model groups, calculate the fusion similarity between the sample features and all cluster centers, and the cluster with the highest fusion similarity is the cluster to which the current operating condition sample belongs. Obtain the corresponding cluster label.

[0071] According to the label of the cluster to which the current working condition sample belongs, all historical sample sets belonging to the cluster are obtained, and the sample sets are filtered according to the set filtering mode to obtain candidate samples.

[0072] In this embodiment, there are three filtering modes, namely upper and lower limit constraint filtering, change constraint filtering and similarity filtering. Among them, upper and lower limit constraint filtering is to select samples within the upper and lower limits according to the set upper and lower limits of the parameter items. That is, the first-stage grate speed given value, the second-stage grate speed, and the 11 fan current given values ​​of the sample shall not exceed the numerical upper and lower limits given by the current working group, and only samples within the upper and lower limits are retained. If the number of remaining samples after filtering is 0, all samples are retained, and the grate speed given value and current given value of these samples are modified to the upper limit value or the lower limit value. If it exceeds the upper limit value, it is changed to the upper limit value, and if it is lower than the lower limit value, it is changed to the lower limit value.

[0073] The variation constraint filter is to delete samples whose parameter variation exceeds the threshold. In this embodiment, a large change in the control variable in a short period of time will cause a sharp change in the working conditions and disrupt the steady state of the working conditions. Therefore, the actual operation requires that the absolute value of the change in the first-stage grate speed set value, the second-stage grate speed set value, and the 11 fan current set value shall not exceed the upper limit, the change in the single grate speed set value shall not exceed 0.4, and the change in the single fan current set value shall not exceed 5.0. According to this variation constraint, the samples in the sample set whose variation exceeds the upper limit are deleted. If the number of remaining samples after filtering is 0, all samples are retained, and the grate speed set value and current set value of these samples are modified to the upper limit of the current set value variation.

[0074] Similarity filtering is to calculate the fusion similarity between historical samples and current working condition samples, sort the historical samples from large to small by similarity, and retain the 5 samples with the highest similarity scores. If the number of remaining samples is less than 5, all samples are retained.

[0075] The value scores of the obtained candidate samples are calculated according to the evaluation method, and several sample data with the highest value scores are selected, and their average is taken as the optimal value and sent to the control system.

[0076] In this embodiment, the evaluation method is to calculate the value score of the candidate sample according to the following formula: Value score = (secondary air temperature + tertiary air temperature) / average fan current. In this embodiment, the value score is evaluated for each candidate sample after multiple filtering, that is, the higher the secondary air temperature, the higher the tertiary air temperature, and the lower the average fan current, the higher the value score of the sample. The candidate samples are sorted from large to small according to the value score, and the first-stage grate speed set value, the second-stage grate speed set value, and the 11 fan current set values ​​corresponding to the candidate sample with the highest value score are selected as the set values ​​of the grate cooler control parameters in the next time step and sent to the control system.

[0077] 3. Update phase.

[0078] In this embodiment, as the grate cooler continues to operate, new sample data is continuously generated. This new data represents the current operating status of the grate cooler and is more valuable than historical data. Therefore, updated data is acquired at a set periodicity to form updated samples. This updated set is then replaced with a portion of the sample set to form an updated set. The core clustering model of this solution is regularly updated online using the newly accumulated updated set.

[0079] In this embodiment, the update cycle is set to trigger the update process at 0:00 on the 1st of each month, and the operating condition data accumulated since the last update is constructed as an update sample. Based on the timestamp order and the "first in, first out" principle, the same number of old samples as the update sample set are discarded, and the update samples and the remaining old samples are merged into a new update set.

[0080] Using the existing cluster centers as the initial cluster centers, the model is retrained on the new update set until convergence, obtaining an updated model set, which is then used in the optimization process. In this embodiment, the update phase can track long-term changes in the overall operating state of the grate cooler, making the adjustment of the grate cooler control parameters more adaptive.

[0081] In this embodiment, the updating phase is decoupled from the optimization phase, and two processes are used to implement optimization and updating respectively. After the clustering model is updated, the new model can be applied to the optimization process in a timely manner.

[0082] In this embodiment, based on actual application scenarios, the recommended parameters obtained through model calculation are used to control the grate cooler in a production environment.

[0083] In this embodiment, the grate cooler is divided into two parts based on its equipment structure, and two sets of models are specifically configured. From a real-time perspective, this solution, deployed on a server, can automatically monitor changes in the grate cooler's operating conditions in real time and provide optimal values. However, the operator team must not only monitor the grate cooler but also the rest of the equipment, making it difficult to determine optimal current and grate pressure settings based on real-time data. This reduces the operator team's workload on the grate cooler. From a heat recovery perspective, the proposed method increases the mean secondary and tertiary air temperatures by 17.2°C compared to manual operation, and reduces the mean standard deviation by 6.1°C, resulting in more efficient and stable heat recovery. From an energy consumption perspective, the proposed method can reduce the grate cooler's system power consumption by 3.4% compared to when the operator team is operating. This also demonstrates that the optimization results of this solution are more stable than those when the operator team is operating.

[0084] The above is only an embodiment of the present invention, and the common knowledge such as the specific technical solutions and / or characteristics in the solution are 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 variations and improvements can be made, which should also be regarded as the scope of protection 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 description can be used to interpret the content of the claims.

Claims

1. A method for optimizing grate cooler equipment parameters based on clustering algorithm, characterized in that: The following steps are included: Modeling phase: S11, obtaining time series samples, dividing the time series samples into one-segment samples and two-segment samples according to the characteristics of the grate cooler, and extracting multidimensional features of the samples accordingly, which are used to train the working condition clustering model of the grate cooler to obtain a one-segment model group and a two-segment model group; It also includes preprocessing of time series sample data, which includes the following sub-steps: S11.1, obtain the clustering characteristics of the time series sample data of each measuring point, including the mean μ, standard deviation σ, maximum value and minimum value of each measuring point; S11.2, use the z-score method to normalize the values ​​of each measurement point. The calculation expression is: , where is the original measurement point value, is the normalized measurement point value; S11.3, use a sliding window with a window size of 10 and a step size of 5 to obtain the mean of the measurement point values; S12, obtaining clusters of historical samples according to each model group, calculating the cluster centers as the cluster features, dividing the sample data into corresponding clusters, and making the model group persistent; Optimization stage: S21, determine the current control state of the grate cooler and enter the optimization process according to the control state; S22, obtaining the real-time operating condition data corresponding to each model group, calculating the cluster to which the real-time operating condition data belongs, and obtaining a sample set based on the cluster characteristics; filtering the sample set according to the set filtering mode to obtain candidate samples; S23, calculating the value scores of the candidate samples according to the evaluation method, selecting several sample data with the highest value scores, taking their average as the optimal value and sending it to the control system; The evaluation method is to calculate the value score of the candidate sample according to the following formula: Value score = (secondary air temperature + tertiary air temperature) / average fan current; Update phase: S31, acquiring update data according to a set period to form update samples, and replacing part of the sample set to form an update set; S32, setting initial cluster centers, using the update set to train the initial cluster centers until the model converges, and obtaining an updated model group; S33, the updated model group is used in the optimization process.

2. The method for optimizing grate cooler equipment parameters based on clustering algorithm according to claim 1, characterized in that: In the optimization phase, the filtering modes include three types: upper and lower limit constraint filtering, change constraint filtering and similarity filtering; The upper and lower limit constraint filtering is to select samples within the upper and lower limits according to the set upper and lower limits of the parameter items; the change constraint filtering is to delete samples whose parameter changes exceed the threshold; the similarity filtering is to calculate the fusion similarity between historical samples and current samples, and retain the 5 samples with the highest similarity scores.

3. The method for optimizing grate cooler equipment parameters based on clustering algorithm according to claim 1, characterized in that: In the modeling stage, the time series samples are variables related to the control parameters of the grate cooler; among them, the first-stage samples are the data of the first five cooling fans corresponding to the first grate plate of the grate cooler; the second-stage samples are the data of the last six cooling fans corresponding to the second grate plate of the grate cooler. Correspondingly, 22 sample measurement points are selected for the first stage and 23 sample measurement points for the second stage.

4. The method for optimizing grate cooler equipment parameters based on clustering algorithm according to claim 1, characterized in that: In the modeling stage, it also includes grouping within the first stage sample and the second stage sample respectively; performing histogram analysis on the given values ​​of the raw material feed amount in the sample, grouping according to the analysis results, and then performing modeling processing within the group to form multiple unit models. The multiple unit models constitute a first stage model group or a two stage model group.

5. The method for optimizing grate cooler equipment parameters based on clustering algorithm according to claim 3, characterized in that: In the modeling phase, the clustering process includes the following sub-steps: S12.1, according to the set cluster radius, minimum number of samples and search step, traverse each sample in the data set, calculate the number of samples in each sample's neighborhood, and if the number of samples is not less than the minimum number of samples, classify them as core samples; S12.2, taking any core sample as the center, search for other core samples within the cluster radius, and mark the samples within this range as a cluster based on the fusion similarity index; S12.3, switch the central core sample and repeat S12.2 until all core samples are visited and unlabeled samples are removed; and the clustering effect is evaluated using the set clustering effect index.

6. The method for optimizing grate cooler equipment parameters based on clustering algorithm according to claim 5, characterized in that: The fusion similarity index includes sample similarity and measurement point similarity; the sample similarity is the sum of similarities of each measurement point constituting the sample; the measurement point similarity includes cosine similarity of measurement point statistics and sequence similarity of measurement point shapes.

7. The method for optimizing grate cooler equipment parameters based on clustering algorithm according to claim 5, characterized in that: The clustering effect index = intra-cluster divergence / inter-cluster divergence; the intra-cluster divergence is the average value of the sample distances within a cluster; the inter-cluster divergence is the average value of the sample distances between clusters.

8. The method for optimizing grate cooler equipment parameters based on clustering algorithm according to claim 6, characterized in that: The cosine similarity is expressed as , where For two measuring points and The cosine similarity of The measurement point sequence The statistical eigenvector of The measurement point sequence The statistical eigenvector of Represents the magnitude of a vector.

Images