A waste heat valve control optimization method based on fusion driving

Abstract

The application discloses a kind of waste heat valve control optimization methods based on fusion drive, comprising the following steps: S1: the grate cooler of certain cement plant, AQC boiler, generator arrangement data monitoring point acquisition data, collection time interval is 5ms, total 397766 historical data points;S2: waste heat recovery text data is conceptually abstracted, and waste heat recovery mechanism knowledge is obtained;S3: fusion waste heat recovery mechanism knowledge and historical data, construct knowledge graph model based on fuzzy set;S4: construct the LSTM valve opening optimization model based on time protection mechanism, propose time protection mechanism algorithm, determine the optimal adjustment frequency of valve, and predict the change trend of waste heat boiler parameter, reduce the delay of adjusting valve to temperature change;S5: when new data is transmitted, the valve adjustment frequency is judged by time protection mechanism algorithm, then the parameter change trend is predicted, knowledge reasoning is carried out by knowledge graph model based on fuzzy set, and valve opening is recommended.

Description

Technical Field

[0001] This invention relates to the field of waste heat recovery technology, and specifically to a waste heat valve control optimization method based on fusion-driven operation. Background Technology

[0002] Traditional waste heat recovery control technologies are divided into mechanism modeling methods and data-driven methods. Mechanism modeling is based on the physical and chemical reaction principles of industrial processes, and establishes relevant models according to thermodynamic laws, material balance, and other theories. For example, the article [Yin Q et al. Optimization design of heat recovery systems on rotary kilnsusing genetic algorithms[J]. Applied Energy, 2017, 202:153-168] derives a multi-objective optimization model for the heat recovery system to obtain optimal design parameters and reduce heat loss; the article [Chen Q et al. Analternative Energy flow model for analysis and optimization of heat transfer systems[J]. International Journal of Heat and Mass Transfer, 2017, 108:712-720] proposes an energy flow model for a single heat exchanger and heat exchanger network, describing the system-level heat transfer characteristics and optimizing the thermal management system; the article [Ahmad R et al. Mass and energy balance in grate cooler of cementplant[J]. International Journal of Scientific Engineering and Technology, 2013, [2(7):631-637] A first-principles model was established to simulate the changes in gas, solid temperature, and wall temperature loss to understand the impact of various design parameters on cement waste heat recovery. The effectiveness of the model was verified through simulation experiments. The above method achieved good results, but with the increase of system complexity, it is difficult for this type of method to explore the deep relationships between data. With the rapid development of technologies such as the Industrial Internet, equipment generates massive amounts of operating data, and data-driven system control optimization methods have become a research hotspot. Data-driven system control optimization methods mainly adjust and optimize equipment parameters by deeply mining the intrinsic relationships between historical data. Commonly used data-driven methods include neural networks and genetic algorithms.For example, the article [Liu Qiang et al. Performance optimization of medium and low temperature waste heat power generation system based on BP neural network strategy [J]. Chemical Industry Management, 2019(01):109-110.] applies BP neural network to the modeling of low temperature waste heat system, which can improve the utilization efficiency of low temperature waste heat; the article [Ali A et al. Power prediction of waste heat recovery system for a cementplant using back propagation neural network and its thermodynamic modeling [J]. International Journal of Energy Research, 2021, 45(6): 9162-9178.] uses BPNN neural network to develop a regression-based prediction model with an accuracy of up to 99.9%; the article [Liu Jing et al. Data fusion driven valve regulation method for waste heat boiler [J]. Journal of Yanshan University, 2021, 45(01):76-86+94.] proposes a data fusion driven valve regulation method for waste heat boiler, which is based on AQC waste heat boiler valve regulation historical data modeling to maximize waste heat reuse. The above methods have achieved good results, but in practical applications, data-driven methods require comprehensive data coverage of all working conditions, and the time required for effective data accumulation is long, which is difficult to meet. Summary of the Invention

[0003] To address the above issues, a fusion-driven waste heat valve control optimization method is proposed. This method integrates mechanistic knowledge and data knowledge to construct a knowledge graph model based on fuzzy sets, materializes valve opening knowledge, establishes an LSTM valve opening optimization model based on a time protection mechanism, and proposes a time protection mechanism algorithm to determine the optimal valve adjustment frequency, thereby improving the waste heat recovery rate while protecting the equipment.

[0004] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0005] A waste heat valve control optimization method based on fusion-driven operation includes the following steps:

[0006] S1: Data monitoring points were set up for the grate cooler, AQC boiler and generator of a cement plant to collect data on bypass valve opening, mixing flue gas regulating valve (cold air valve), flue gas temperature before the combined superheater, flue gas pressure in the outlet flue, flue gas pressure before the combined superheater, main steam temperature and high-pressure steam flow. The data collection time interval was 5ms, and a total of 397,766 historical data points were collected.

[0007] S2: Abstract the concepts from the waste heat recovery text data, define valve openings for some equipment states, and obtain knowledge of the waste heat recovery mechanism.

[0008] S3: Integrate knowledge of waste heat recovery mechanisms with historical data to construct a knowledge graph model based on fuzzy sets;

[0009] S4: Construct an LSTM valve opening optimization model based on a time protection mechanism to determine the optimal valve regulation frequency and predict the parameter change trend;

[0010] S5: When new data is received, the valve adjustment frequency is determined, the parameter change trend is predicted, and knowledge reasoning is performed using a knowledge graph model based on fuzzy sets to recommend the valve opening.

[0011] Furthermore, in step S3, knowledge of waste heat recovery mechanisms and historical data are integrated to construct a knowledge graph model based on fuzzy sets, including the following steps:

[0012] 1-1) Based on the principle of waste heat recovery and correlation analysis, the characteristics affecting valve opening are obtained, and the attributes of these characteristics are extracted sequentially from historical data. An initial entity is formed based on the characteristics and characteristic attributes.

[0013] 1-2) Based on the vocabulary found in expert experience, determine the state set V = {"high", "relatively high", "relatively low", "low"}. For different features, determine the optimal value set corresponding to the state set, and determine the membership function.

[0014] For a variable with a state value of "high", its membership function takes the following form:

[0015]

[0016] in, Represents the membership function. The current entity feature value, The membership degree is the current entity's state being "high". This is the optimal value corresponding to the state value "high". This represents the lowest acceptable feature value for a state of "high". The highest acceptable feature value for a state of "high" is where... ,

[0017] For variables with state values ​​of "higher" or "lower", their membership function takes the following form:

[0018]

[0019] in: Represents the membership function. The current entity feature value, The membership degree is the degree to which the current entity's state is "higher" ("lower"). This is the optimal value corresponding to the state value "higher" ("lower"); The lowest acceptable feature value for a state of "higher" ("lower"); The highest acceptable feature value for a state of "higher" ("lower") is where ,

[0020] For a variable with a state value of "low", its membership function takes the following form:

[0021]

[0022] in, Represents the membership function. The current entity feature value, The membership degree is "low" for the current entity state. This is the optimal value corresponding to the state value "low". This represents the lowest acceptable feature value for a state of "low". The highest acceptable feature value for a state of "low" is where ;

[0023] 1-3) After calculating the membership function of an entity, determine the state set of the entity. Assign different weights to features in different state sets based on the degree of influence of features on the entity's valve opening, and calculate the similarity between entities. Let...

[0024] in, , For entity feature vectors, , For entity features, The number of features of an entity. , , , for Entity feature values , , , for Entity feature values for , Feature weight vector, , , , for , The weight values ​​of the corresponding features, for , similarity, for The 1 eigenvalue, for The 1 eigenvalue, for , No. The weights corresponding to each feature;

[0025] 1-4) After calculating the similarity between entities, multiple entities with similarity less than a set threshold are merged into a set entity, and the feature vector of the set entity is the average of the vectors of multiple old entities;

[0026] 1-5) Calculate the enthalpy and ν in the waste heat gas, as shown in the following formula:

[0027]

[0028] in, The enthalpy of the working fluid, The internal energy of matter. For pressure, For volume, For working fluid 㶲, For the working fluid temperature, For ambient temperature, Specific heat capacity of the working fluid;

[0029] 1-6) The quality of the data is determined based on the enthalpy and γ of the waste heat gas. The calculation formula is shown below:

[0030] in, For quality, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲 For enthalpy of data under similar operating conditions, This refers to data under similar operating conditions.

[0031] In order to obtain , Based on game theory, an objective function is established, using a combination of weighted indicators. and , The objective is to minimize the sum of deviations, and to find the optimal linear combination coefficients. , The weighted combination of indicators at this point is the optimal weighted combination. The objective function and constraints are as follows:

[0032] in, Represents the computation of vectors The L-2 norm, To improve the quality of data under similar operating conditions, For enthalpy of data under similar operating conditions, For data under similar working conditions, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲

[0033] According to the principle of differentiation, we can obtain the result from the first derivative that yields the minimum value of the above equation. , The value will , Normalization can be performed to obtain , The final mass calculation formula is as follows:

[0034]

[0035] in, To ensure the final quality of data under similar operating conditions, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲 For enthalpy of data under similar operating conditions, For data under similar working conditions;

[0036] 1-7) Determine the number of categories for continuous variables based on variance goodness of fit and silhouette coefficient, perform k-meas clustering, and then select high-quality data under similar working conditions to be placed into a decision tree for classification to obtain the relationships between entities;

[0037] 1-8) Establish a knowledge graph model based on fuzzy sets according to entities and relationships.

[0038] 4. Further, in step S4, an LSTM valve opening optimization model based on a time protection mechanism is constructed to determine the optimal valve adjustment frequency and predict the parameter change trend. The steps are as follows:

[0039] 2-1) Construct an LSTM prediction model that integrates convolutional neural networks. The model includes convolutional layers, pooling layers, LSTM layers, Dropout layers, and BN layers. Train the model.

[0040] 2-2) Determine the protection window time based on the on-site working conditions. If the temperature and rate of change do not exceed the specified threshold within the time protection window, continue monitoring; otherwise, exit the time protection period.

[0041] 2-3) Input the boiler parameters within a preset time period before the time to be predicted into the trained model to predict the boiler parameters at the time to be predicted;

[0042] 2-4) Based on the predicted parameters of the boiler, the valve opening is recommended by knowledge reasoning through a knowledge graph model based on fuzzy sets.

[0043] The beneficial effects of adopting the above technical solution are as follows:

[0044] This invention proposes a fusion-driven waste heat valve control optimization method (OWF). First, addressing the difficulty in integrating waste heat recovery mechanism knowledge with equipment operation data, a fuzzy set-based knowledge graph model is constructed. This model extracts rules from mechanism knowledge and attributes from historical data, with rules and attributes jointly forming initial entities. Simultaneously, to accelerate graph retrieval, an expert experience algorithm (EAF) based on fuzzy sets is proposed, aggregating some entities into set entities to reduce the number of entities. Furthermore, an evaluation criterion (EST) based on thermal mechanism knowledge is proposed, using this criterion to select high-quality data from historical data and mine the relationships between entities, thereby establishing a fuzzy set-based knowledge graph model. Second, addressing the issue of frequent adjustments causing easy wear and tear on boiler valves, an LSTM valve opening optimization model based on a time protection mechanism is constructed. This model proposes a time protection mechanism algorithm to determine the optimal frequency of valve adjustment. Further, an LSTM prediction algorithm fused with a convolutional neural network (LSTM-CNN) is proposed to predict the changing trends of waste heat boiler parameters, reducing the delay between valve adjustment and temperature changes, enabling timely valve adjustment, and reducing equipment risk.

[0045] Compared with traditional waste heat recovery methods, this invention proposes: (1) a knowledge graph model based on fuzzy set, which extracts rules from mechanism knowledge and attributes from equipment operation data. The two are combined to form an initial entity. For a large number of entities, an expert experience algorithm based on fuzzy set is established to finally realize the construction of valve opening knowledge graph. This solves the problems of complex mechanism knowledge and difficulty in modeling, slow data knowledge accumulation and difficulty in merging the two in traditional waste heat recovery methods; (2) a valve opening optimization model based on LSTM is proposed. This model first proposes a time protection mechanism algorithm to determine the optimal frequency of valve adjustment. Furthermore, it proposes an LSTM prediction model that integrates convolutional neural networks to predict the changing trend of waste heat boiler parameters, adjust the valve in time, and reduce equipment risk. This model not only reduces the number of valve adjustments in traditional waste heat recovery, but also reduces the impact of delay on the equipment; (3) a judgment standard based on thermal mechanism knowledge is proposed, which can generate more heat and reduce "entropy pollution" to the environment while ensuring equipment safety.

[0046] The waste heat valve control optimization method based on fusion-driven proposed in this invention was applied to the production dataset of a cement plant for experimentation. The results showed that the waste heat recovery rate and equipment safety were effectively improved after using the method, and intelligent decision-making on the opening degree of the waste heat recovery valve was realized. Attached Figure Description

[0047] Figure 1 This is a framework diagram of a waste heat valve control optimization method based on fusion-driven approach;

[0048] Figure 2 It is a knowledge graph model constructed based on fuzzy sets;

[0049] Figure 3 It is a decision tree relationship extraction graph that integrates the EST standard;

[0050] Figure 4 This is a diagram showing the construction of an LSTM valve opening optimization model based on a time protection mechanism.

[0051] Figure 5 This is a diagram illustrating the time protection mechanism;

[0052] Figure 6 This is partial data from a cement plant;

[0053] Figure 7 It is the Pearson correlation coefficient;

[0054] Figure 8 It is the variance goodness of fit;

[0055] Figure 9 It is the profile coefficient.

[0056] Figure 10 This is a line graph comparing OV-LSTM and LSTM predictions;

[0057] Figure 11 This is a recommended flow chart for waste heat valve opening.

[0058] Figure 12 This is a comparison chart of enthalpy values ​​before and after applying the waste heat valve control optimization method based on fusion-driven approach;

[0059] Figure 13 This is a comparison chart of bypass valve openings when the temperature is too low;

[0060] Figure 14 This is a comparison chart of the opening degree of the air conditioning valve when the temperature is too high. Detailed Implementation

[0061] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0062] This invention uses a knowledge graph based on fuzzy sets as its carrier and an LSTM valve opening optimization model based on a time protection mechanism as its main algorithm framework. The model is as follows: Figure 1 As shown, it includes the following steps:

[0063] S1: Data monitoring points were set up for the grate cooler, AQC boiler and generator of a cement plant to collect data on bypass valve opening, mixing flue gas regulating valve (cold air valve), flue gas temperature before the combined superheater, flue gas pressure in the outlet flue, flue gas pressure before the combined superheater, main steam temperature and high-pressure steam flow. The data collection time interval was 5ms, and a total of 397,766 historical data points were collected.

[0064] S2: Abstract the concepts from the waste heat recovery text data, define valve openings for some equipment states, and obtain knowledge of the waste heat recovery mechanism.

[0065] S3: Integrate waste heat recovery mechanism knowledge with historical data to construct a knowledge graph model based on fuzzy sets. The process is as follows: Figure 2 As shown, the specific steps are as follows;

[0066] 1-1) Based on the principle of waste heat recovery and correlation analysis, the characteristics affecting valve opening are obtained, and the attributes of these characteristics are extracted sequentially from historical data. An initial entity is formed based on the characteristics and characteristic attributes.

[0067] 1-2) Based on the vocabulary found in expert experience, determine the state set V = {"high", "relatively high", "relatively low", "low"}. For different features, determine the optimal value set corresponding to the state set, and determine the membership function.

[0068] For a variable with a state value of "high", its membership function takes the following form:

[0069] in, Represents the membership function. The current entity feature value, The membership degree is the current entity's state being "high". This is the optimal value corresponding to the state value "high". This represents the lowest acceptable feature value for a state of "high". The highest acceptable feature value for a state of "high" is where... ,

[0070] For variables with state values ​​of "higher" or "lower", their membership function takes the following form:

[0071] in, Represents the membership function. The current entity feature value, The membership degree is the degree to which the current entity's state is "higher" ("lower"). This is the optimal value corresponding to the state value "higher" ("lower"); The lowest acceptable feature value for a state of "higher" ("lower"); The highest acceptable feature value for a state of "higher" ("lower") is where ,

[0072] For a variable with a state value of "low", its membership function takes the following form:

[0073] in, Represents the membership function. The current entity feature value, The membership degree is "low" for the current entity state. This is the optimal value corresponding to the state value "low". This represents the lowest acceptable feature value for a state of "low". The highest acceptable feature value for a state of "low" is where ;

[0074] 1-3) After calculating the membership function of an entity, determine the state set of the entity. Assign different weights to features in different state sets based on the degree of influence of features on the entity's valve opening, and calculate the similarity between entities. Let...

[0075] in, , For entity feature vectors, , For entity features, The number of features of an entity. , , , For entity feature values, , , , for Entity feature values for , Feature weight vector, , , , for The weight values ​​of the corresponding features, for , The similarity is The 1 eigenvalue, for The 1 eigenvalue, for , No. The weights corresponding to each feature;

[0076] 1-4) After calculating the similarity between entities, multiple entities with similarity less than a set threshold are merged into a set entity, and the feature vector of the set entity is the average of the vectors of multiple old entities;

[0077] 1-5) After obtaining the entities in the graph, extract the relationships between entities based on historical data. The process is as follows: Figure 3 As shown, the enthalpy and ν in the waste heat gas are first calculated, as shown in the following formula:

[0078]

[0079] in, The enthalpy of the working fluid, The internal energy of matter. For pressure, For volume, For working fluid 㶲, For the working fluid temperature, For ambient temperature, Specific heat capacity of the working fluid;

[0080] 1-6) The quality of the data is determined based on the enthalpy and γ of the waste heat gas. The calculation formula is shown below:

[0081] in, For quality, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲 For enthalpy of data under similar operating conditions, For data under similar working conditions,

[0082] In order to obtain , Based on game theory, an objective function is established, using a combination of weighted indicators. and , The objective is to minimize the sum of deviations, and to find the optimal linear combination coefficients. , The weighted combination of indicators at this point is the optimal weighted combination. The objective function and constraints are as follows:

[0083]

[0084] in, Represents the computation of vectors The L-2 norm, To improve the quality of data under similar operating conditions, For enthalpy of data under similar operating conditions, For data under similar working conditions, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲

[0085] According to the principle of differentiation, we can obtain the result from the first derivative that yields the minimum value of the above equation. , The value will , Normalization can be performed to obtain , The final mass calculation formula is as follows:

[0086] in, To ensure the final quality of data under similar operating conditions, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲 For enthalpy of data under similar operating conditions, For data under similar working conditions;

[0087] 1-7) Determine the number of categories for continuous variables based on variance goodness of fit and silhouette coefficients, perform k-meas clustering, and then select high-quality data under similar working conditions to be placed into a decision tree for classification to obtain the relationships between entities.

[0088] 1-8) Establish a knowledge graph model based on fuzzy sets according to entities and relationships;

[0089] S4: Construct an LSTM valve opening optimization model based on a time protection mechanism, determine the optimal valve regulation frequency, and predict the parameter change trend. The flowchart is as follows: Figure 4 As shown, the steps are as follows:

[0090] 2-1) Construct an LSTM prediction model that integrates convolutional neural networks. The model includes convolutional layers, pooling layers, LSTM layers, Dropout layers, and BN layers. Train the model.

[0091] 2-2) Determine the protection window time based on the on-site working conditions. If the temperature and rate of change do not exceed the specified threshold within the time protection window, monitoring continues; otherwise, the time protection period ends. The flowchart is as follows: Figure 5 As shown;

[0092] 2-3) Input the boiler parameters within a preset time period before the time to be predicted into the trained model to predict the boiler parameters at the time to be predicted;

[0093] 2-4) Based on the predicted parameters of the boiler, the valve opening is recommended by knowledge reasoning through a knowledge graph model based on fuzzy sets;

[0094] S5: When new data is received, the valve adjustment frequency is determined, the parameter change trend is predicted, and knowledge reasoning is performed using a knowledge graph model based on fuzzy sets to recommend the valve opening.

[0095] 1. Data Description

[0096] The experiment used real data collected from a waste heat recovery system in a cement plant. The data collection period was from August 3rd to August 27th, 2020, with a collection interval of 5ms, totaling 397,766 samples. The data was divided into three parts: grate cooler data, AQC boiler data, and generator data. Grate cooler data included bypass valve opening; AQC boiler data mainly included bypass valve opening, AQC boiler mixing flue gas regulating valve (cold air valve), AQC boiler combined superheater inlet flue gas temperature, AQC boiler outlet flue gas pressure, and AQC combined superheater inlet flue gas pressure; generator data included generator power, main steam temperature, and AQC high-pressure steam flow rate. Selected data are shown below. Figure 6 As shown.

[0097] 2. Experimental Procedure

[0098] Based on the principle of waste heat recovery, the relevant characteristics are selected as shown in Table 1:

[0099] Table 1 Feature Parameters

[0100] serial number name unit 1 Main steam temperature ℃ 2 AQC High-Pressure Steam Flow kPa 3 AQC bypass valve % 4 AQC Boiler Mixing Air Flue Regulating Valve (Cold Air Valve) % 5 Flue gas temperature before the combined superheater of AQC furnace ℃ 6 AQC furnace outlet flue gas pressure kPa 7 Flue gas pressure in front of the combined superheater of AQC furnace kPa 8 Delay of regulating valve to saturated steam change s

[0101] The time delay between the regulating valve and the saturated steam change cannot be directly obtained from the database. Therefore, the delay between the two is calculated using Pearson correlation technology, and the correlation coefficient curve is shown below. Figure 7 As shown, the correlation coefficient is highest and tends to level off when the delay is between 120s and 150s; therefore, the delay is considered to be 150s. Initial entities are established based on features and feature attributes. There are a total of 397,766 initial entities.

[0102] To reduce the number of entities, fuzzy judgment is performed on entity features based on the EAF algorithm. The expert experience related to the flue gas temperature before the AQC combined superheater is summarized as follows:

[0103] Experience 1: When the flue gas temperature before the AQC combined superheater is high and the change tends to be stable or the temperature drops, the valve remains unchanged;

[0104] Experience 2: When the flue gas temperature in front of the AQC combined superheater is high, the cold air valve needs to be opened;

[0105] Experience 3: When the flue gas temperature before the AQC combined superheater is low and trending upward, the valve remains unchanged;

[0106] Experience 4: When the flue gas temperature before the AQC combined superheater is low and tends to be stable, open the bypass valve.

[0107] Based on expert experience and membership functions, entities related to the above experience were given the feature "temperature state value", with values ​​of {"high", "relatively high", "relatively low", "low"}. Further integration was performed based on the similarity between entities, resulting in a reduction of 20,420 entities, a decrease of 5.1%, thus improving retrieval efficiency.

[0108] The value of k is determined based on the variance goodness of fit and the silhouette coefficient. Figure 8 It can be seen that the flue gas temperature and main steam temperature curves before the AQC combined superheater tend to flatten out at points 12, 13, 14, and 15, while the AQC pressure differential curve tends to flatten out at points 14, 15, 16, and 17. Next, the profile coefficients for these points are calculated, and the results are as follows: Figure 9 As shown in the figure. Therefore, it can be determined that the flue gas temperature before the combined superheater is divided into 13 categories, the main steam temperature into 15 categories, and the AQC pressure difference into 16 categories.

[0109] The samples used for relation extraction were factory data filtered according to the EST standard. The sample size was 164,185, containing four conditional attributes and two decision attributes. The decision attributes were discrete, while the conditional attributes, AQC (Area-QC) combined with the superheater inlet flue gas temperature and main steam temperature, were continuous. The improved algorithm aggregated the continuous values ​​of the conditional attributes into discrete values ​​using the k-means algorithm.

[0110] To verify the effectiveness of the method, a decision tree incorporating the EST standard was compared with a traditional ID3 decision tree, and the results are shown in Table 2. Experiments demonstrate that the accuracy of the proposed algorithm on both the training and test sets is higher than that of the traditional algorithm.

[0111] Table 2 Comparison of Decision Tree Accuracy

[0112] serial number method Training set accuracy Test set accuracy 1 Traditional ID3 decision tree 87.2% 87.1% 2 Decision tree integrating EST standards 88.9% 88.6%

[0113] Evaluation metrics are commonly used in regression tasks to calculate prediction error. Several evaluation criteria were used to determine how the predicted data deviated from the target (actual) data during the fitting period, to measure the performance of the model in this study. The prediction error measurements used in this paper are mean absolute error (MAE) and root mean square error (RMSE).

[0114] To verify the effectiveness of OV-LSTM time series prediction, an experiment was designed to compare the prediction results of this model with those of other models. Figure 10 As shown, it can be seen that the prediction value of OV-LSTM is significantly closer to the true value than the prediction value of LSTM.

[0115] To further determine the accuracy of OV-LSTM predictions, the MAE and RMSE values ​​of OV-LSTM and LSTM were calculated, and the results are shown in Table 3. The table shows that the MAE and RMSE values ​​of OV-LSTM are both lower than those of LSTM, further confirming the effectiveness of OV-LSTM time series predictions.

[0116] Table 3 Comparison of OV-LSTM and LSTM predictions

[0117] Model MAE RMSE OV-LSTM 0.121 0.126 LSTM 0.123 0.146

[0118] Next, to verify the effectiveness of the time protection mechanism in practical applications, a simulation experiment was conducted using data from August. The experimental results are shown in Table 4. The number of valve adjustments decreased significantly after adding the time protection mechanism. After adding the time protection mechanism, the input data was analyzed and predicted to determine whether valve adjustment was necessary, thus reducing unnecessary valve adjustments.

[0119] Table 4. Number of valve adjustments before and after adding the time protection mechanism

[0120] Model Bypass valve adjustment times Number of times the air cooler valve is adjusted After adding a time protection mechanism 445 120 No time protection mechanism added 495 13

[0121] The OWF waste heat valve control process is as follows: Figure 11 As shown, the comparative experiment selected a half-month dataset from a cement plant as input, with the sampling period from August 3rd to August 16th. After data preprocessing of this sample segment, it was input into the OWF model to obtain the recommended valve opening.

[0122] To verify the effectiveness of OWF in improving waste heat recovery rate (represented by enthalpy) in practical applications, the waste heat recovery rate after applying this method was compared with the actual adjustment situation. The horizontal axis represents the date, and the vertical axis represents the average enthalpy value of that day. The comparison results are as follows: Figure 12 As shown. Based on the comparison results, except for August 5th, 10th, 11th, and 13th when the enthalpy values ​​were close, the enthalpy values ​​after applying OWF were significantly higher than the original enthalpy values ​​on other days. That is, in 12 out of the 16 days compared, the enthalpy values ​​after applying this method were significantly higher than the original enthalpy values, representing a probability of 75%. This demonstrates the effectiveness of the method proposed in improving waste heat recovery rate. To verify the effectiveness of OWF protection equipment, a comparative experiment was conducted to compare the valve opening recommended by this method with the valve opening adjusted by the employee. According to... Figure 13It can be seen that when the temperature is below the safety threshold, except for August 4th, the probability of opening the valve and raising the temperature using this method is higher than that of employees. That is, when the temperature is below the safety threshold, the OWF method is 94% effective in improving equipment safety, and the average probability of improving equipment safety is increased by 52.12%. According to... Figure 14 It can be seen that, except for August 5th, when the temperature is above the safety threshold, the probability of opening the valve to cool down using this method is greater than or equal to that of the employee. That is, when the temperature is above the safety threshold, the OWF method is 94% effective in improving equipment safety, and the average probability of improving equipment safety is increased by 43.88%. Therefore, this experiment verifies the effectiveness of OWF in protecting equipment.

[0123] 3. Conclusion

[0124] To address the challenge of integrating mechanistic and data knowledge in traditional waste heat valve control technologies, this paper proposes a fusion-driven optimization method for waste heat valve control, consisting of two parts. The first part extracts classification rules based on industrial mechanistic knowledge and node attributes based on historical data. These rules and attributes together form initial entities, and a fuzzy set-based expert experience algorithm is established to reduce the number of entities, thus constructing a valve opening knowledge graph. The second part proposes a time-protection mechanism algorithm to determine the optimal frequency of valve adjustment, reducing the number of adjustments. Furthermore, it uses an LSTM prediction model fused with a convolutional neural network to predict parameter trends, enabling timely valve adjustment and reducing equipment risk. Experimental results demonstrate that this method effectively integrates mechanistic and data knowledge, achieving higher efficiency and greener manufacturing processes.

[0125] The foregoing detailed examples of the present invention are merely preferred embodiments and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the present invention should still fall within the patent coverage of the present invention.

Claims

1. A waste heat valve control optimization method based on fusion-driven operation, characterized in that, Includes the following steps: S1: Set up data monitoring points for the grate cooler, AQC boiler, and generator in the cement plant to collect data on bypass valve opening, mixing flue gas regulating valve, flue gas temperature before the combined superheater, flue gas pressure in the outlet flue, flue gas pressure before the combined superheater, main steam temperature, and high-pressure steam flow. S2: Abstract the waste heat recovery text data conceptually, define valve opening for equipment status, and obtain knowledge of waste heat recovery mechanism; S3: Integrate knowledge of waste heat recovery mechanisms with historical data to construct a knowledge graph model based on fuzzy sets; S4: Construct an LSTM valve opening optimization model based on a time protection mechanism to determine the optimal valve regulation frequency and predict the parameter change trend; S5: When new data is received, the valve adjustment frequency is determined, and then the parameter change trend is predicted. Knowledge reasoning based on a fuzzy set-based knowledge graph model is used to recommend the valve opening. In step S3, the waste heat recovery mechanism knowledge and historical data are integrated to construct a fuzzy set-based knowledge graph model, including the following steps: 1-1) Based on the principle of waste heat recovery and correlation analysis, the characteristics affecting valve opening are obtained, and the attributes of these characteristics are extracted sequentially from historical data. An initial entity is formed based on the characteristics and characteristic attributes. 1-2) Based on the vocabulary found in expert experience, determine the state set V = {"high", "relatively high", "relatively low", "low"}. For different features, determine the optimal value set corresponding to the state set, and determine the membership function. For a variable with a state value of "high", its membership function takes the following form: in, Represents the membership function. The current entity feature value, This represents the membership degree of the entity whose current state is "high". This is the optimal value corresponding to the state value "high". This represents the lowest acceptable feature value for a state of "high". The highest acceptable feature value for a state of "high" is where... , For variables with state values ​​of "higher" or "lower", their membership function takes the following form: in: Represents the membership function. The current entity feature value, This represents the membership degree of the entity whose current state is "higher" (or "lower"). The optimal value corresponding to the state value "higher" ("lower"); The lowest acceptable feature value for a state of "higher" ("lower"); The highest acceptable feature value for a state of "higher" ("lower") is where... , For variables with a state value of "low", their membership function takes the following form: in, Represents the membership function. The current entity feature value, The membership degree is "low" for the current entity state. This is the optimal value corresponding to the state value "low". This represents the lowest acceptable feature value for a state of "low". The highest acceptable feature value for a state of "low" is where ; 1-3) After calculating the membership function of an entity, determine the state set of the entity; assign different weights to features in different state sets according to the degree of influence of features on the valve opening of the entity, and calculate the similarity between entities; let... in, , For entity feature vectors, , For entity features, The number of features of an entity. , , , for Entity feature values , , , for Entity feature values for , Feature weight vector, for , The weight values ​​of the corresponding features, for , similarity, for The 1 eigenvalue, for The 1 eigenvalue, for , The first The weights corresponding to the features; 1-4) After calculating the similarity between entities, multiple entities with similarity less than a set threshold are merged into a set entity. The feature vector of the set entity is the average of the vectors of the multiple old entities. 1-5) Calculate the enthalpy and ν in the waste heat gas, as shown in the following formula: in, The enthalpy of the working fluid, The internal energy of matter. For pressure, For volume, For working fluid 㶲, For the working fluid temperature, For ambient temperature, Specific heat capacity of the working fluid; 1-6) The quality of the data is determined based on the enthalpy and γ of the waste heat gas. The calculation formula is shown below: in, For quality, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲 For enthalpy of data under similar operating conditions, For data under similar working conditions, In order to obtain , Based on game theory, an objective function is established, using a combination of weighted indicators. and , The objective is to minimize the sum of deviations, and to find the optimal linear combination coefficients. , The weighted combination of indicators at this point is the optimal weighted combination. The objective function and constraints are as follows: in, Represents the computation of vectors The L-2 norm, To improve the quality of data under similar operating conditions, For enthalpy of data under similar operating conditions, For data under similar working conditions, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲 According to the principle of differentiation, we can obtain the result from the first derivative that yields the minimum value of the above equation. , The value will , Normalization can be performed to obtain , The final mass calculation formula is as follows: in, To ensure the final quality of data under similar operating conditions, As weight, The weight corresponding to enthalpy, The weight corresponding to 㶲 For enthalpy of data under similar operating conditions, For data under similar working conditions; 1-7) Determine the number of categories for continuous variables based on variance goodness of fit and silhouette coefficient, perform k-meas clustering on them, and then select high-quality data under similar working conditions to put into a decision tree for classification to obtain the relationship between entities. 1-8) Establish a knowledge graph model based on fuzzy sets according to entities and relationships.

2. The waste heat valve control optimization method based on fusion drive according to claim 1, characterized in that: In step S4, an LSTM valve opening optimization model based on a time protection mechanism is constructed to determine the optimal valve adjustment frequency and predict the parameter change trend. The steps are as follows: 2-1) Construct an LSTM prediction model that integrates convolutional neural networks. The model includes convolutional layers, pooling layers, LSTM layers, Dropout layers, and BN layers. Train the model. 2-2) Determine the protection window time based on the on-site working conditions. If the temperature and rate of change do not exceed the specified threshold within the time protection window, continue monitoring; otherwise, exit the time protection period. 2-3) Input the boiler parameters within a preset time period before the time to be predicted into the trained model to predict the boiler parameters at the time to be predicted; 2-4) Based on the predicted parameters of the boiler, the valve opening is recommended by knowledge reasoning through a knowledge graph model based on fuzzy sets.

Images