Intelligent regulation and control method and system for dust removal system of iron and steel enterprise

Abstract

The invention discloses an intelligent regulation and control method and system for a dust removal system of an iron and steel enterprise, and the method comprises the following steps: collecting real-time data according to a set frequency, including the operation parameters of each dust production point, the opening degree of a valve, and the operation parameters and exhaust condition parameters of a dust remover, dust conveying equipment and a fan, and dividing the data into operation variables and regulation variables; a hybrid neural network model is constructed, the hybrid neural network model comprises an LSTM-GPT-Neo backbone network, a DenseNet branch network and a meta learning feature fusion correction layer, and the LSTM-GPT-Neo backbone network fuses an LSTM architecture and a GPT-Neo architecture; inputting time sequence data to carry out model training and optimization; and inputting the data into the trained hybrid neural network model for calculation and analysis, and outputting the opening degree of the valve and the optimal operation parameters of the fan. Under the condition that the dust removal efficiency and stability are guaranteed, the electricity utilization cost of the steel plant can be remarkably reduced, and the problem of equipment loss caused by excessive dust emission or excessive air volume due to insufficient air volume is avoided.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent control, and in particular to an intelligent regulation method and system for a dust removal system in an iron and steel enterprise. Background Art

[0002] The iron and steel industry is a key area for energy consumption and pollutant emissions. A large amount of dust is generated in each process of the iron and steel production process. Improving the operating energy efficiency of the dust removal system is crucial for energy conservation, consumption reduction and environmental protection compliance of iron and steel enterprises. In the process of continuously deepening the work of the dual-carbon goal, improving the energy efficiency of the dust removal system is the key work for energy conservation and consumption reduction in iron and steel plants.

[0003] Traditional dust removal system fans often adopt a fixed air volume operation mode, and there are often problems such as excessive dust emissions due to insufficient air volume, or unnecessary energy consumption and equipment losses due to excessive air volume. At present, the regulation of the air volume of the fan depends more on manual adjustment. Limited by the experience of personnel, the regulation is often inaccurate and cannot be adjusted in real time, resulting in unstable operation, unqualified emissions or waste of energy. In addition to the fan, the operation of the valves in the pipeline and the ash cleaning and conveying equipment all face the problems of automation and precise control. In the era of intelligence, production big data is a valuable asset, and the data in the actual production operation can provide a strong basis for production control and potential tapping, especially in terms of energy conservation and consumption reduction, the collection and utilization of real-time data are more important. By adjusting the opening degree of valves, the operating parameters of fans and ash cleaning and conveying equipment, etc. in real time according to parameters such as real-time air volume, air pressure, dust generation amount, dust generation points, dust collector resistance, and ash cleaning and conveying operation status during the iron and steel plant production process, the operating efficiency of the dust removal system can be improved, and the electricity cost of the iron and steel plant can be effectively reduced, and the carbon emissions of iron and steel enterprises can be reduced. Summary of the Invention

[0004] The object of the present invention is to overcome the above-mentioned disadvantages of the prior art, and provide an intelligent regulation method and system for a dust removal system in an iron and steel enterprise, aiming to solve problems such as energy waste in the traditional dust removal system in the iron and steel industry, so as to realize the efficient and energy-saving operation of the dust removal system.

[0005] The present invention is realized through the following technical solutions: An intelligent regulation method for a dust removal system in an iron and steel enterprise includes the following steps: S1. Collect real-time data at a set frequency, and divide the collected data into operating variables and adjustment variables; the collected data includes the operating parameters at each dust generation point, the valve opening degree K1 of the valve between each dust generation point and the dust collector, the operating parameters of the dust collector, the operating parameters of the ash conveying equipment, the operating parameters of the fan, and the exhaust condition parameters; S2. Construct a hybrid neural network model, which includes an LSTM-GPT-Neo backbone network, a DenseNet branch network, and a meta-learning feature fusion correction layer. The LSTM-GPT-Neo backbone network is located at the front end of the model and is used to calculate and output the valve opening and the optimal operating parameter distribution of the fan according to the input operating variable data. It integrates the LSTM and GPT-Neo architectures. LSTM is used to extract high-level temporal features from the input operating variable data, and the temporal features contain key information about the device operating state, so that GPT-Neo can perform more complex reasoning. The GPT-Neo architecture enhances the ability to capture long-distance dependencies and complex patterns by introducing the idea of a large-scale pre-trained language model to perform semantic-level understanding and analysis of the operating variable data. The DenseNet branch network is used to calculate and analyze the input adjustment variable data and output dust removal performance parameters, which are used to characterize the mapping relationship between the adjustment variable and the dust removal efficiency. The meta-learning feature fusion correction layer adopts a meta-learning strategy and is used to dynamically adjust the fusion weights and fusion methods according to the outputs of the LSTM-GPT-Neo backbone network and the DenseNet branch network, as well as different working conditions and historical data, so as to adapt to various complex and changeable situations. S3. Collect the time series data of the operating variables and adjustment variables to form a data set, divide the data set into a training set and a validation set in chronological order, train the hybrid neural network model through the training set, and then optimize the hybrid neural network model through the validation set. S4. Input the operating variable data and adjustment variable data into the trained hybrid neural network model for calculation and analysis, output the valve opening K1 and the optimal operating parameters of the fan, and perform real-time control on the operation of the valve and the fan according to these parameters. The optimal operating parameters of the fan include the rotational speed r of the fan, the opening K2 of the inlet valve of the fan, and the opening K3 of the outlet valve of the fan.

[0006] Further, the operating parameters of the dust generation point include the air volume q1, the wind pressure p, the temperature t1, the moisture content wet, and the dust concentration m1. The operating parameters of the dust collector include the dust collector resistance zu, the ash amount c1 in the ash hopper, the cleaning cycle T1, the cleaning time T2, and the cleaning intensity F. The operating parameters of the ash conveying equipment include the ash conveying speed v and the ash conveying amount c2. The operating parameters of the fan include the rotational speed r of the fan, the opening K2 of the inlet valve of the fan, the opening K3 of the outlet valve of the fan, the temperature t2, and the current I. The exhaust condition parameters include the exhaust air volume q2 discharged from the dust collector and the dust concentration m2 in the gas discharged from the dust collector. The operating variables include the air volume q1, air pressure p, temperature t1, moisture content wet, dust concentration m1, dust generation point s, resistance zu of the dust collector, ash quantity c1 in the ash hopper, temperature t2 of the fan, current I of the fan, air volume q2 and dust concentration m2 of the dust collector; the adjustment variables include the rotational speed r of the fan, valve opening K1, opening K2 of the inlet valve of the fan, opening K3 of the outlet valve of the fan, cleaning cycle T1, cleaning time T2, cleaning intensity F, ash conveying speed v and ash conveying quantity c2.

[0007] Further, assume that there are n dust generation points, s dust collectors, and u fans in the dust removal system, and there are f acquisition time points. The processing and representation methods for the operating variable data are as follows: If processed through the function y = Y(q1, p, t1, wet, m1, zu, c1, t2, I, q2, m2), the air volume q1, air pressure p, temperature t1, moisture content wet, and dust concentration m1 at the dust generation point can be represented as three-dimensional arrays, namely q1[i][j][k], p[i][j][k], t1[i][j][k], wet[i][j][k], m1[i][j][k] respectively, where i represents the dust generation point number, 0 ≤ i ≤ n, j represents the acquisition time point number, 0 ≤ j ≤ f, and k represents the data dimension; The resistance zu of the dust collector, ash quantity c1 in the ash hopper, discharged air volume q2 and dust concentration m2 of the dust collector are represented by three-dimensional arrays, namely zu[h][j][k], c1[h][j][k], q2[h][j][k], m2[h][j][k] respectively, where h represents the dust collector number, 0 ≤ h ≤ s, j represents the acquisition time point number, 0 ≤ j ≤ f, and k represents the data dimension; The temperature t2 and current I of the fan are represented by three-dimensional arrays, namely t2[g][j][k], I[g][j][k] respectively, where g represents the fan number, 0 ≤ g ≤ u, j represents the acquisition time point number, 0 ≤ j ≤ f, and k represents the data dimension.

[0008] The processing and representation methods for the adjustment variable data are as follows: The adjustment variables are processed by the function t=T(r, K1, K2, K3, T1, T2, F, v, c2). The speed r of each fan, valve opening K1, fan inlet valve opening K2, fan outlet valve opening K3, cleaning cycle T1, cleaning time T2, cleaning intensity F, ash conveying speed v and ash conveying amount c2 can be expressed as three-dimensional arrays r[g][j][k], K1[g][j][k], K2[g][j][k], K3[g][j][k], T1[g][j][k], T2[g][j][k], F[g][j][k], v[g][j][k], c2[g][j][k], where g is the fan number, 0≤g≤u. For the ash conveying equipment, g is the fan number corresponding to the ash conveying equipment, j is the collection time point, 0≤j≤f, and k is the data dimension.

[0009] The operating variables and adjustment variable data need to be preprocessed. The preprocessing method is: Clean the data, identify and remove obviously abnormal data, and interpolate missing values. If the missing data accounts for a small proportion, it can be directly filled, and if the missing data accounts for a large proportion, it can be directly discarded; Then the data of different dimensions and ranges are normalized and converted into dimensionless numbers; Then the dimensionless data is represented as a vector X=[x1,x2,…,x n ], where x i is the i-th input parameter, and n is the total number of input parameters.

[0010] Furthermore, it also includes: S5, by monitoring the air volume q1, wind pressure p, dust concentration m1 at the dust generating point after regulation, the energy consumption of the dust collector, ash conveying equipment and fan, and exhaust status parameters, the energy-saving effect of the regulation, the stability of the system operation and the dust removal effect are evaluated, and the evaluation results are input into the hybrid neural network model for iterative optimization to continuously improve the regulation strategy of the hybrid neural network model.

[0011] Furthermore, the LSTM-GPT-Neo backbone network is located at the front end of the model, and its output data represents the possible optimal values ​​of the valve opening K1, the speed r of the fan, the opening K2 of the fan's air inlet valve, and the opening K3 of the air outlet valve in the form of probability distribution, as well as the confidence interval and uncertainty estimation; The LSTM unit adopts a stacked structure of four LSTM layers, each LSTM layer includes 512 neurons, and is combined with an improved variant of the gated recurrent unit GRU to capture complex time series features in the running variable data; the calculation and analysis method of the LSTM unit is: Each LSTM layer is used to capture the time series features of the parameters, and a residual connection is introduced between each LSTM layer. A skip connection is added between two LSTM layers, and the input of the previous LSTM layer is directly added element-wise to the output of the next LSTM layer. The output is: , where, is the output hidden state after the residual connection and skip connection operations, is the hidden state after being processed by the current LSTM layer at time step t, is the hidden state at the previous time step t−1; The neurons in the LSTM layer are connected by weights. Each neuron performs weighted summation and activation function transformation on the input data. The output h of the j-th neuron j is expressed as: , where w ij is the weight from the i-th input parameter to the j-th neuron, b j is the bias of the j-th neuron, and f is the activation function ReLU, expressed as: f(x)=max(0,x); The DenseNet branch network includes four DenseBlock layers and Transition layers located between every two DenseBlock layers. Each DenseBlock layer includes three convolutional layers and a skip connection. Each Transition layer consists of a convolutional layer and a pooling layer; The meta-learning feature fusion and correction layer includes a meta-learner MAML and a fusion module. MAML obtains the adapted weights through gradient update according to the working condition features. The fusion module performs weighted averaging on the dynamic weights and finally outputs a corrected prediction value as the output data of the entire hybrid neural network model.

[0012] Furthermore, in step S3, the training method of the hybrid neural network model is: S3-1. Collect historical operation sequence data, including the operation parameters at each dust generation point, the valve opening K1 of the valve between each dust generation point and the dust collector, the operation parameters of the dust collector, the operation parameters of the ash conveying equipment, the operation parameters of the fan, and the exhaust condition parameters; S3-2. After removing the abnormal operation data (such as unqualified dust removal effect, excessive fan temperature or current, abnormal operation of the dust collector, etc.), evaluate the historical operation data, and screen out the historical operation data with excellent energy-saving effect, system operation stability, and dust removal effect; S3-3. The screened data are used as a data set, and randomly divided into a training set and a validation set in a certain ratio according to time training; the data in the data set are preprocessed and input into the hybrid neural network model for training. The supervised learning method is used, and the optimal operating parameters of the wind turbine are used as the target variable. The weights are updated through the Adam optimization algorithm. The parameter update formula is: , in, θ is the parameter to be optimized, η is the learning rate, and are the bias-corrected first-order and second-order moment estimates, ϵ A small value (usually 10) is added to prevent division by zero. -8 ).

[0013] Furthermore, in step S3-2, the method of data evaluation and screening is: calculate the average energy consumption of the system in each time period, the rate of change of air volume and pressure at the dust removal point, and the dust removal effect (dust removal efficiency), first sort according to the dust removal effect, and after eliminating the bottom 40% of the data, sort the remaining data according to the average energy consumption, and after eliminating the bottom 30% of the data, sort the remaining data according to the rate of change of air volume and pressure, and after eliminating the bottom 20% of the data, obtain the screened data.

[0014] Furthermore, in step S3-3, the parameters of model training are set as follows: the learning rate is set to 0.001, the maximum number of pre-training iterations is 500, and the maximum number of fine-tuning iterations is 200; the sparsity parameter is set to 0.4, which helps to reduce redundant information; the sparsity penalty parameter is 2, which is used to control the penalty intensity of sparsity; the activation function uses the ReLU function; and the fine-tuning loss function uses the mean square error loss function MSE.

[0015] Furthermore, in step S5, the method of inputting the evaluation result into the hybrid neural network model for iterative optimization includes: S5-1. Define positive and negative samples based on actual business needs and evaluation objectives; S5-2. After determining the positive and negative samples, a confusion matrix is ​​constructed based on the evaluation data obtained from monitoring; S5-3. Based on the constructed confusion matrix, the evaluation index of the hybrid neural network model is calculated, and then the evaluation index is input into the hybrid neural network model. The current performance of the model is measured through the loss function, and the control strategy is continuously improved.

[0016] Furthermore, the evaluation result includes the following parameters: Air volume tracking error rate : Calculated according to the following formula: , where q pred is the target air volume at the predicted dust generation point, and q post is the measured air volume at the dust generation point after intelligent regulation; The air pressure stability index s p is calculated according to the following formula: , where σ(p post ) is the standard deviation of the air pressure at the dust generation point after intelligent regulation, and μ(p post ) is the mean value of the air pressure at the dust generation point after intelligent regulation; The dust concentration compliance rate R m is calculated according to the following formula: where T is the total number of sampling times during the statistical period, II() is the indicator function, and m t is the dust concentration in the gas discharged from the dust collector at the t-th sampling.

[0018] The energy saving rate e: is calculated according to the following formula: , where E1 is the total energy consumption of the dust removal system before intelligent regulation, and E2 is the total energy consumption of the dust removal system before and after intelligent regulation; The air volume change rate Q: is calculated according to the following formula: , where q1 is the air volume at the dust removal point before intelligent regulation, and q2 is the air volume at the dust removal point after intelligent regulation; The air pressure change rate P: is calculated according to the following formula: , where p1 is the air pressure at the dust removal point before intelligent regulation, and p2 is the air pressure at the dust removal point after intelligent regulation; The dust removal efficiency η: is calculated according to the following formula: , where m in is the dust concentration entering the dust collector, and m out is the dust concentration discharged from the dust collector; The change in dust removal efficiency: is calculated according to the following formula: , where η1 is the dust removal efficiency before intelligent regulation, and η2 is the dust removal efficiency after intelligent regulation.

[0019] An intelligent control system for the dust removal system of an iron and steel enterprise, which is used to implement the intelligent control method for the dust removal system of the above iron and steel enterprise, includes a dust removal equipment, a data acquisition module, an intelligent control module, an operation control module and a monitoring and optimization module; The dust removal equipment includes an air collecting hood, a conveying pipeline, a valve, a dust collector, an ash conveying equipment, a fan and a chimney. A plurality of the air collecting hoods are arranged at the dust generating points. The air collecting hoods are connected to the dust collector through the conveying pipeline. The valve is arranged on the conveying pipeline. The dust collector is arranged at the rear end of the air collecting hood. The dust collector is provided with a dust cleaning equipment, which is used to clean the dust accumulated in the dust collector. The ash conveying equipment is arranged at the outlet of the ash hopper of the dust collector and is used to convey the accumulated dust. The air inlet of the fan is connected to the dust collector, and the air outlet of the fan is connected to the chimney; The data acquisition module includes a plurality of sensors and a data collector. The sensors are respectively arranged at the positions of the conveying pipeline, the valve and the dust collector connected to the air collecting hood at each dust generating point. The data collector is connected to each of the sensors, the valve, the dust collector, the ash conveying equipment and the fan, and is used to collect the operation parameters at each dust generating point, the valve opening of the valve, the operation parameters of the dust collector, the operation parameters of the ash conveying equipment and the operation parameters of the fan. The data acquisition module is connected to the intelligent control module and the monitoring and optimization module to transmit the collected data; A hybrid neural network model is embedded in the intelligent control module, which is used to calculate and analyze the collected data to obtain the optimal operation parameters of the fan; The operation control module is connected to the intelligent control module and is used to receive the optimal operation parameters of the fan obtained by the intelligent control module. The operation control module is also connected to the fan and is used to adjust the rotation speed of the fan, the opening of the air inlet valve and the opening of the air outlet valve according to the optimal operation parameters of the fan obtained by the intelligent control module to control the size of the air volume of the fan; The monitoring and optimization module includes a data analysis unit, a closed-loop feedback unit and a data storage unit. The data analysis unit is used to calculate and analyze the collected data to evaluate the energy-saving effect of the control, the stability of the system operation and the dust removal effect. The closed-loop feedback unit is used to transmit the evaluation result to the intelligent control module to perform iterative optimization on the hybrid neural network model to form a dynamic optimization closed loop. The data storage unit is used to store the collected data and the data obtained by calculation and analysis.

[0020] Further, the sensor includes an air volume measuring instrument, a wind pressure gauge, a dust concentration sensor, a temperature and humidity meter, a pressure sensor, and a material level meter. The air volume measuring instrument is respectively arranged on the conveying pipelines at each dust-producing point and at the dust collector, and is used to measure the air volume q1 at the dust-producing point and the discharged air volume q2 discharged from the dust collector. The wind pressure gauge is arranged on the conveying pipeline at the dust-producing point and is used to measure the wind pressure p at the dust-producing point. The dust concentration sensors are respectively arranged on the conveying pipeline at the dust-producing point and at the dust collector, and are used to measure the dust concentration m1 at the dust-producing point and the dust concentration m2 discharged from the dust collector. The temperature and humidity meter is arranged on the conveying pipeline at the dust-producing point and is used to measure the temperature t1 and the moisture content wet at the dust-producing point. The pressure sensor and the material level meter are both arranged in the dust collector and are used to measure the resistance zu of the dust collector and the ash amount c1 in the ash hopper.

[0021] The dust-producing points are the coke oven coal charging and coke discharging sites, the secondary dedusting of the converter, and the blast furnace tapping yard. The gas hoods are arranged at the positions above the coke oven chamber furnace head, the coke guide machine, the coke pusher, above the converter furnace mouth, and above the blast furnace tapping hole.

[0022] The intelligent regulation module further includes a data cleaning unit, a dust collector operation optimization unit, and a dust conveying equipment operation optimization unit. The data cleaning unit is used to eliminate abnormal data by using the 3σ principle and perform normalization processing on the data. The dust collector operation optimization unit is used to analyze the equipment operation status of the dust collector based on historical data and real-time data through a fault diagnosis algorithm, so as to dynamically adjust the dust cleaning cycle T1, the dust cleaning time T2, and the dust cleaning intensity F of the dust collector. The dust conveying equipment operation optimization unit is used to analyze the equipment operation status of the dust conveying equipment through a fault diagnosis algorithm, so as to dynamically adjust the dust conveying speed v and the dust conveying amount c2 of the dust conveying equipment. The operation regulation module is simultaneously connected to the dust collector and the dust conveying equipment, so as to dynamically adjust the dust cleaning cycle T1, the dust cleaning time T2, and the dust cleaning intensity F of the dust collector, as well as the dust conveying speed v and the dust conveying amount c2 of the dust conveying equipment according to the data of the intelligent regulation module.

[0023] The present invention adjusts parameters such as the opening degree of valves and the control of fans in real time by collecting parameters such as the real-time air volume, air pressure, dust production amount, dust production points, dust collector resistance, and the operation status of ash cleaning and ash transportation at each key node in the steel plant. It uses an intelligent regulation algorithm model to accurately calculate the optimal operation parameters of the fan, and automatically adjusts the rotation speed and air volume output (inlet and outlet valves) of the fan, so that the fan always operates under the best working conditions. While ensuring the dust removal efficiency and stability, it can significantly reduce the electricity cost of the steel plant, ensure that the dust removal system can operate stably and efficiently under different working conditions, and avoid problems such as excessive dust emissions caused by insufficient air volume or equipment wear caused by excessive air volume. At the same time, it can also intelligently regulate the operation parameters of the dust collector and ash transportation equipment, making the operation conditions of these equipment more reasonable, greatly reducing the wear of the dust removal filter material, dust removal air ducts, and ash transportation equipment, and extending the service life of equipment such as dust collectors and fans. The training of the intelligent regulation algorithm model of the present invention can use historical operation data and data obtained by correcting historical operation data. The data can be automatically collected, providing rich data for the model's data set and improving the ease of operation of model training. The intelligent regulation system of the present invention can automatically adjust according to changes in working conditions, has strong adaptability to working conditions with unstable processing air volume, does not require frequent manual intervention, improves the flexibility and response speed of the system, realizes real-time monitoring, intelligent regulation, and optimized management of the dust removal system, and meets the requirements of the green, low-carbon, and high-quality development of the steel industry. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 It is a structural block diagram of the dust removal equipment in the regulation system of the embodiment of the present invention.

[0025] Figure 2 It is a structural block diagram of the regulation system of the embodiment of the present invention.

[0026] Figure 3 It is a schematic flowchart of the regulation method of the embodiment of the present invention.

[0027] Figure 4 It is a working principle block diagram of the regulation method of the embodiment of the present invention.

[0028] Figure 5 It is a framework structure diagram of the hybrid neural network model in the embodiment of the present invention.

[0029] Reference numerals: 1 - air collecting hood; 2 - conveying pipeline; 3 - valve; 4 - dust collector; 5 - ash cleaning equipment; 6 - ash transportation equipment; 7 - fan; 8 - chimney. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0030] An intelligent regulation system for the dust removal system of a steel enterprise, as Figure 2 shown, includes dust removal equipment, a data collection module, an intelligent regulation module, an operation regulation module, and a monitoring and optimization module.

[0031] As Figure 1 shown, the dust removal equipment includes a gas hood 1, a conveying pipeline 2, a valve 3, a dust collector 4, an ash conveying device 6, a fan 7 and a chimney 8. A plurality of the gas hoods 1 are arranged at the dust generating points. The gas hood 1 is connected to the dust collector 4 through the conveying pipeline 2. The valve 3 is arranged on the conveying pipeline 2. The dust collector 4 is provided with a dust cleaning device 5 for cleaning the dust accumulated in the dust collector 4. The ash conveying device 6 is arranged below the dust collector 4 for conveying the accumulated dust. The air inlet of the fan 7 is connected to the dust collector 4, and the air outlet of the fan 7 is connected to the chimney 8. The dust-removed gas is discharged from the chimney 8.

[0032] One or more gas hoods 1 can be arranged at one dust generating point, depending on the area and dust concentration of the dust generating point. For iron and steel enterprises, the dust generating points are mainly concentrated in processes such as coke oven coal charging, coke pushing, converter secondary dust removal, and blast furnace tapping yard. The gas hood 1 can be arranged at positions such as the furnace head of the carbonization chamber, the coke quenching machine, the coke pusher, above the converter furnace mouth, and above the blast furnace tapping hole. The dust collector 4 can be a bag type dust collector 4, an electrostatic dust collector 4, a centrifugal dust collector 4 or a composite dust collector 4, etc., and is arranged at the rear end of the gas hood 1. The corresponding dust cleaning device 5 can be a pulse jet system, a wet scrubbing system, a mechanical vibration system or a reverse air blowing dust cleaning system. Currently, the more commonly used is the pulse bag type dust collector 4. The ash conveying device 6 can be a conventional conveying device, such as a screw conveyor, a pneumatic conveying system, a star-shaped discharger, etc. In iron and steel enterprises, some dust has recycling value and can be conveyed to the corresponding section for further treatment, such as being conveyed to the ash bunker through an ash conveying pipeline. The ash conveying device 6 is arranged at the ash hopper outlet of the dust collector 4 and can be operated and controlled according to the real-time ash volume to avoid continuous operation even in the case of no ash or less ash volume, resulting in energy waste and equipment wear. The fan 7 is an axial flow fan or a centrifugal fan and is arranged between the outlet of the dust collector 4 and the chimney 8.

[0033] The data acquisition module includes a plurality of sensors and a data collector. The sensors are respectively arranged at the positions of the conveying pipeline 2, the valve 3 and the dust collector 4 connected to the gas hood 1 at each dust generating point. The data collector is connected to the sensors, the valve 3, the dust collector 4, the ash conveying device 6 and the fan 7 for collecting the operating parameters at each dust generating point, the valve opening of the valve 3, the operating parameters of the dust collector 4, the operating parameters of the ash conveying device 6 and the operating parameters of the fan 7. The data acquisition module is connected to the intelligent regulation module and the monitoring and optimization module to transmit the collected data.

[0034] The intelligent regulation module is embedded with a hybrid neural network model for calculating and analyzing the collected data to obtain the optimal operating parameters of the fan.

[0035] The operation control module is connected to the intelligent control module and is used to receive the optimal operating parameters of the fan obtained by the intelligent control module. The operation control module is also connected to the fan and is used to adjust the speed and air volume of the fan according to the optimal operating parameters of the fan obtained by the intelligent control module. The opening of the air inlet valve and the air outlet valve of the fan can be automatically adjusted by the actuator to control the size of the fan air volume so that it reaches the opening percentage required by the optimal operating parameters; at the same time, the frequency converter is used to automatically adjust the speed of the fan so that it reaches the speed required by the optimal operating parameters, thereby controlling the air volume output.

[0036] The monitoring and optimization module includes a data analysis unit, a closed-loop feedback unit and a data storage unit. The data analysis unit is used to perform calculations and analysis based on the collected data to evaluate the energy-saving effect of the regulation, the stability of the system operation and the dust removal effect, so as to ensure the effectiveness and reliability of the regulation method; the closed-loop feedback unit is used to transmit the evaluation results to the intelligent regulation module, perform iterative optimization on the hybrid neural network model, continuously improve the regulation strategy, and form a dynamic optimization closed loop; the data storage unit is used to store the collected data and the data obtained by calculation and analysis.

[0037] Of course, a communication module should also be provided to realize data transmission between these modules through wired or wireless communication.

[0038] Specifically, the sensor includes an air volume measuring instrument, an air pressure gauge, a dust concentration sensor, a thermometer and hygrometer, a pressure sensor and a level meter. The air volume measuring instrument is respectively arranged at the conveying pipe 2 and the dust collector 4 at each dust generating point, and is used to measure the air volume q1 at the dust generating point and the exhaust air volume q2 discharged from the dust collector 4. The air pressure gauge is arranged at the conveying pipe 2 at the dust generating point, and is used to measure the wind pressure p at the dust generating point. The dust concentration sensor is respectively arranged at the conveying pipe 2 and the dust collector 4 at the dust generating point, and is used to measure the dust concentration m1 at the dust generating point and the dust concentration m2 discharged from the dust collector 4. The thermometer and hygrometer is arranged at the conveying pipe 2 at the dust generating point, and is used to measure the temperature t1 and moisture content wet at the dust generating point. The pressure sensor and the level meter are both arranged in the dust collector 4, and are used to measure the resistance zu of the dust collector 4 and the ash amount c1 in the ash hopper. Other data that need to be collected are directly connected to the corresponding equipment. The equipment is equipped with corresponding detection units to detect the data in real time. For example, the speed, temperature and current value of the fan can be detected and controlled by the control unit of the fan, and the operating parameters of the dust collector and ash conveying equipment can also be detected and controlled by the control unit, such as the cleaning cycle T1 and cleaning time T2 set by the dust collector.

[0039] The intelligent control module also includes a data cleaning unit, a dust collector operation optimization unit and an ash conveying equipment operation optimization unit. The data cleaning unit is used to eliminate abnormal data using the 3σ principle. Then it is eliminated, and the data is normalized.

[0040] The dust collector operation optimization unit is used to analyze the equipment operation status of the dust collector based on historical data and real-time data through a fault diagnosis algorithm, so as to dynamically adjust the dust cleaning cycle T1, dust cleaning time T2 and dust cleaning intensity F of the dust collector, taking into account both dust removal and energy conservation and consumption reduction.

[0041] The ash conveying equipment operation optimization unit is used to analyze the equipment operation status of the ash conveying equipment through a fault diagnosis algorithm, timely detect abnormal conditions of the equipment, and optimize the operation strategy of the ash conveying equipment according to the operation status of the equipment and the overall requirements of the dust removal system. It mainly dynamically adjusts the ash conveying speed v and ash conveying volume c2 of the ash conveying equipment, improves the operation efficiency and reliability of the equipment, and reduces energy consumption.

[0042] The fault diagnosis algorithm compares the real-time data with the set normal range. If it deviates from the normal range, it means that the equipment is operating abnormally. Then it is necessary to adjust the operation status of the dust collector or ash conveying equipment according to the preset rules, mainly the operation status of the dust cleaning equipment and ash conveying equipment. When the resistance zu of the dust collector 4 reaches a certain value, the dust cleaning equipment can be started in real time, and the dust cleaning time T2 and dust cleaning intensity F can be dynamically adjusted according to the size of the resistance zu. When the ash volume c1 in the ash hopper accumulates to a certain height, the ash conveying equipment can be started in real time, and the ash conveying speed v can be dynamically adjusted according to the ash volume c1.

[0043] The operation control module is simultaneously connected to the dust collector and the ash conveying equipment to dynamically adjust the dust cleaning cycle T1, dust cleaning time T2 and dust cleaning intensity F of the dust collector, as well as the ash conveying speed v and ash conveying volume c2 of the ash conveying equipment according to the data of the intelligent control module.

[0044] An intelligent control method for a dust removal system in an iron and steel enterprise, such as Figure 3 , includes the following steps: S1. Collect real-time data at a set frequency, which can be set to collect once every 1 minute or 3 minutes, such as Figure 1 . Figure 4, the collected data includes the operating parameters at each dust generation point, the valve opening degree K1 of the valve between each dust generation point s and the dust collector, the operating parameters of the dust collector, the operating parameters of the ash conveying equipment, the operating parameters of the fan, and the exhaust condition parameters. The operating parameters of the dust generation point include the air volume q1, the air pressure p, the temperature t1, the moisture content wet, and the dust concentration m1. The operating parameters of the dust collector include the dust collector resistance zu, the ash volume c1 in the ash hopper, the cleaning cycle T1, the cleaning time T2, and the cleaning intensity F. The operating parameters of the ash conveying equipment include the ash conveying speed v and the ash conveying volume c2. The operating parameters of the fan include the fan speed r, the opening degree K2 of the fan inlet valve, the opening degree K3 of the fan outlet valve, the temperature t2, and the current I. The exhaust condition parameters include the exhaust air volume q2 discharged from the dust collector and the dust concentration m2 in the gas discharged from the dust collector.

[0045] The dust removal effect is affected by the temperature and humidity at the dust generation point. The dust collector resistance of the dust collector affects the dust removal effect on the one hand and determines the operation of the cleaning and ash conveying equipment on the other hand. The fan speed can adjust the air volume and power output. Monitoring the temperature t2 and current I of the fan can ensure that the fan operates within the normal range. The collection of these data can comprehensively grasp the operation status of the entire dust removal system to achieve the monitoring of the corresponding equipment and environment, such as understanding the operation status of the dust collector and the ash conveying equipment, and controlling accordingly. At the same time, these data can also become historical operation data, serving as the training data set of the network model and the basis for intelligent regulation.

[0046] The collected data is divided into operating variables and adjustment variables. The operating variables include the air volume q1, the air pressure p, the temperature t1, the moisture content wet, the dust concentration m1, the dust collector resistance zu, the ash volume c1 in the ash hopper, the temperature t2 of the fan, the current I of the fan, the exhaust air volume q2 of the dust collector, and the dust concentration m2 at each dust generation point s. The adjustment variables include the fan speed r, the valve opening degree K1, the opening degree K2 of the fan inlet valve, the opening degree K3 of the fan outlet valve, the cleaning cycle T1, the cleaning time T2, the cleaning intensity F, the ash conveying speed v, and the ash conveying volume c2. The operating variables can be processed through the function y = Y(q1, p, t1, wet, m1, zu, c1, t2, I, q2, m2). The adjustment variables can be processed through the function t = T(r, K1, K2, K3, T1, T2, F, v, c2). The energy-saving parameters are processed through the function e = E(Y, T).

[0047] In the case where there are multiple dust - generating points, multiple dust collectors, multiple fans, and corresponding different collection times, arrays and matrices can be used to represent various variables. Suppose there are n dust - generating points, s dust collectors, u fans, and f collection time points. For operating variables, the air volume q1, air pressure p, temperature t1, moisture content wet, and dust concentration m1 at each dust - generating point can be represented as three - dimensional arrays, namely q1[i][j][k], p[i][j][k], t1[i][j][k], wet[i][j][k], m1[i][j][k], where i (0 ≤ i ≤ n) represents the dust - generating point number, j (0 ≤ j ≤ f) represents the collection time point number, and k represents the data dimension. The resistance zu of the dust collector, the ash volume c1 in the ash hopper, the exhaust air volume q2 and dust concentration m2 of the dust collector are represented by three - dimensional arrays, which are zu[h][j][k], c1[h][j][k], q2[h][j][k], m2[h][j][k] respectively, where h (0 ≤ h ≤ s) represents the dust - collector number. The temperature t2 and current I of the fan are also represented by three - dimensional arrays, which are t2[g][j][k], I[g][j][k] respectively, where g (0 ≤ g ≤ u) represents the fan number. The operating - variable function can be expressed as: y[j]=Y(q1[j],p[j],t1[j],wet[j],m1[j],zu[j],c1[j],t2[j],I[j],q2[j],m2[j]), where j represents the collection time point.

[0048] For adjustment variables, the rotational speed r of the fan, the valve opening K1, the opening K2 of the inlet valve of the fan, the opening K3 of the outlet valve of the fan, the cleaning cycle T1, the cleaning time T2, the cleaning intensity F, the ash - conveying speed v, and the ash - conveying volume c2 can also be represented by arrays according to different devices such as fans and dust collectors and the collection time. If it is assumed that the adjustment variables of different devices can be independently controlled, for the rotational speed r of each fan, it can be represented as a three - dimensional array r[g][j][k], where g is the fan number, 0 ≤ g ≤ u, j is the collection time point, and k is the data dimension; the valve openings K, cleaning - related parameters, and ash - conveying - related parameters can be represented similarly, which are K1[g][j][k], K2[g][j][k], K3[g][j][k], T1[g][j][k], T2[g][j][k], F[g][j][k], v[g][j][k], c2[g][j][k] respectively. For the ash - conveying equipment, g is the fan number corresponding to the ash - conveying equipment. The adjustment - variable function can be expressed as t[j]= T(r[j],K[j],T1[j],T2[j],F[j],v[j],c2[j]), where j represents the collection time point.

[0049] The energy-saving parameter function can be expressed as e[j] = E(y[j], t[j]), that is, the corresponding energy-saving parameters are calculated according to the operating variables and adjustment variables at each acquisition time point. In this way, through the form of arrays and matrices, the acquisition data and related variables under different working conditions and different times can be clearly managed and processed.

[0050] S2. Construct a hybrid neural network model, such as Figure 5 , and the hybrid neural network model is composed of an LSTM-GPT-Neo backbone network, a DenseNet branch network, and a meta-learning feature fusion and correction layer.

[0051] The LSTM-GPT-Neo backbone network is located at the front end of the model and integrates the LSTM and GPT-Neo architectures. Its input is the operating variable data. The task of LSTM is to extract high-level temporal features from the operating variable data, and these features contain the key information of the equipment operating state, so that the subsequent GPT-Neo can perform more complex inferences. LSTM adopts a four-layer stacked structure, with 512 neurons in each layer, combined with an improved variant of the gated recurrent unit (GRU), which can better capture the complex time series features in the operating variable data. The GPT-Neo architecture introduces the idea of a large-scale pre-trained language model to perform semantic-level understanding and analysis of the operating variable data, enhancing the ability to capture long-distance dependencies and complex patterns. The input of this backbone network is the operating variable data after deep preprocessing (including data cleaning, normalization, feature engineering, outlier detection, and missing value imputation, etc.), covering the air volume q1, wind pressure p, temperature t1, moisture content wet, dust concentration m1, dust generation point s, resistance zu of the dust collector, ash volume c1 in the ash hopper, temperature t2 of the fan, current I of the fan, air volume q2 and dust concentration m2 of the dust collector, etc. After processing, mathematical transformations and feature combinations are used in the hidden layer to calculate the high-order statistical features, non-linear relationship features, and comparison features with historical data of each parameter, and these features are spliced to form a new feature array. The final output layer calculates and outputs the optimal operating parameter distribution of the valve and the fan according to these feature arrays, and gives the possible optimal values of multiple parameters such as valve opening, fan speed, air volume, current, and power in the form of a probability distribution, and at the same time outputs the confidence interval and uncertainty estimation of the parameters.

[0052] The DenseNet branch network includes multiple DenseBlock layers (dense blocks) and Transition layers (transition layers) located between every two DenseBlock layers. Each DenseBlock layer includes multiple convolutional layers and a skip connection, and each Transition layer is composed of a convolutional layer and a pooling layer. The DenseNet branch network inputs the regulated variable data, calculates and analyzes the input regulated variable data, and outputs the dust removal performance parameters, which are used to characterize the mapping relationship between the regulated variable and the dust removal efficiency.

[0053] The input of the meta-learning feature fusion correction layer comes from the outputs of the LSTM - GPT - Neo backbone network and the DenseNet branch network. The meta-learning feature fusion correction layer includes a meta-learner and a fusion module. Using the meta-learning strategy, according to the outputs of the LSTM-GPT-Neo backbone network and the DenseNet branch network, the fusion weights and fusion methods are dynamically adjusted according to different working conditions and historical data to adapt to various complex and changeable situations. The fused data is further processed by a deep neural network to output a predicted value, which is used as the output data of the entire hybrid neural network model, providing a more accurate, intelligent, and forward-looking decision-making basis for subsequent fan operation regulation and the optimization of the entire dust removal system, ensuring that the system can achieve an efficient and energy-saving dust removal effect under various complex working conditions, and at the same time providing strong support for the interpretability and generalization ability of the model.

[0054] Meta-learning can quickly adapt to new tasks by learning the commonalities of multiple tasks or data distributions. In the hybrid neural network model of the present invention, a meta-learner is added to achieve fast learning by improving the optimization process itself, and the fusion weights can be quickly adjusted according to the working conditions. Model-Agnostic Meta-Learning (MAML) can be used to optimize the initial parameters, enabling the model to adapt to new tasks with only a few gradient updates. For new working conditions, a small number of samples are used for fine-tuning (Few-Shot Adaptation), and the fusion strategy is dynamically adjusted to correct the output value. The MAML optimizes the meta-learner, enabling it to quickly adapt to new working conditions. The meta-learner inputs the working condition features and outputs the initial fusion weights, and then through gradient updates, the adapted weights are obtained. The fusion module performs weighted averaging on the dynamic weights and finally outputs a corrected predicted value as the output data of the entire hybrid neural network model.

[0055] The hybrid neural network model is responsible for capturing the time series features and long-term dependencies of the parameters. The model is trained using historical data with a supervised learning method, taking the optimal operating parameters of the fan (such as rotational speed and air volume output) as the target variables, and adjusting the model parameters through optimization algorithms (such as gradient descent method) so that the system can accurately regulate the operating parameters of the fan. The model starts from data input, processes through multiple LSTM layers, obtains the final prediction result through a fully connected layer, then calculates the loss and updates the model parameters, forming a complete training cycle to continuously improve the accuracy of the model in regulating the optimal operating parameters of the fan.

[0056] S3. Collect the time series data of the operating variables and adjustment variables to form a data set, divide the data set into a training set and a validation set in chronological order, train the hybrid neural network model through the training set, and then optimize the hybrid neural network model through the validation set.

[0057] S4. Input the operating variables and adjustment variables data into the trained hybrid neural network model for calculation and analysis, output the valve opening and the optimal operating parameters of the fan, and perform real-time regulation on the operation of the valve and the fan according to these parameters. The operating parameters of the fan mainly include the rotational speed of the fan, the opening of the inlet valve, and the opening of the outlet valve.

[0058] Automatically adjusting the operating parameters of the fan to match the current operating conditions of the dust removal system can achieve energy-saving effects. The automatic regulation system can accurately adjust the rotational speed and air volume output of the fan through the driving device connected to the fan, ensure that the fan can operate efficiently under different working conditions, avoid excessive energy consumption, and at the same time ensure the stability and reliability of the dust removal effect.

[0059] S5. Evaluate the energy-saving effect of the regulation, the stability of the system operation, and the dust removal effect by monitoring the air volume q1, wind pressure p, dust concentration m1 at the dust generation point after regulation, the energy consumption of the dust collector, ash conveying equipment, and the fan, and the exhaust condition parameters, and input the evaluation results into the hybrid neural network model for iterative optimization to continuously improve the regulation strategy of the hybrid neural network model.

[0060] In this embodiment, the LSTM unit adopts a four-layer LSTM stacked structure, and each layer of LSTM includes 512 neurons; the DenseNet branch network includes four DenseBlock layers, and each DenseBlock layer includes three convolutional layers; the output dust removal performance vector can have five dimensions.

[0061] The method for the LSTM unit to perform calculation and analysis is: The LSTM backbone network includes an input layer, a hidden layer, and an output layer. The hidden layer includes multiple LSTM layers. Each LSTM layer is used to capture the time series features of parameters, and a residual connection is introduced between LSTM layers. A skip connection is added between two LSTM layers, and the input of the previous LSTM layer is directly added element-wise to the output of the next LSTM layer. The output is: , where, is the output hidden state after the residual connection and skip connection operations, is the hidden state processed by the current LSTM layer at time step t, is the hidden state at the previous time step (time step t−1).

[0062] The neurons in the LSTM layer are connected by weights. Each neuron performs weighted summation and activation function transformation on the input data. The output h j of the j-th neuron is expressed as: , where w ij is the weight from the i-th input parameter to the j-th neuron, and b j is the bias of the j-th neuron. f is the activation function ReLU, expressed as: f(x)=max(0,x).

[0063] In step S3, the training method of the hybrid neural network model is as follows: (1) Collect historical operation sequence data, including operation parameters at each dust generation point, valve opening K1 of the valve between each dust generation point and the dust collector, operation parameters of the dust collector, operation parameters of the ash conveying equipment, operation parameters of the fan, and exhaust condition parameters. (2) After removing abnormal operation data (such as unqualified dust removal effect, excessive fan temperature or current, abnormal operation of the dust collector, etc.), evaluate the historical operation data, and select the historical operation data with excellent energy-saving effect, system operation stability, and dust removal effect. The methods for data evaluation and screening can be: calculate the average energy consumption of the system, the change rate of air volume and air pressure at the dust removal point, and the dust removal effect (dust removal efficiency) in each time period. First, sort according to the dust removal effect, remove a certain proportion (such as 40%) of the data with the lowest ranking, then sort the remaining data according to the average energy consumption, remove a certain proportion (such as 30%) of the data with the lowest ranking, and then sort the remaining data according to the change rate of air volume and air pressure, and remove a certain proportion (such as 20%) of the data with the lowest ranking to obtain the selected data.

[0064] (3) Use the selected data as a dataset and randomly divide it into a training set and a validation set according to a certain ratio; preprocess the data in the dataset and then input it into the hybrid neural network model for training. Adopt the supervised learning method, set the learning rate to 0.001. A smaller learning rate can make the model converge more stably; the maximum number of pre-training iterations is 500; the maximum number of fine-tuning iterations is 200; set the sparse parameter to 0.4, which helps to reduce redundant information; the sparse penalty term parameter is 2, which is used to control the penalty strength of sparsity; use the ReLU (Rectified Linear Unit) function as the activation function; select the mean squared error loss function (MSE) as the fine-tuning loss function. Take the optimal operating parameters of the fan (such as rotational speed and air volume output) as the target variable, and update the weights through the Adam optimization algorithm. The parameter update formula is: , where, θ is the parameter to be optimized, η is the learning rate, and are the first-order and second-order moment estimates after bias correction respectively, ϵ is a very small value added to prevent division by zero (usually 10 -8 ).

[0065] The data input into the model needs to be preprocessed. The data preprocessing method is: Clean the data, identify and eliminate obviously abnormal data, such as abnormally high or low data caused by sensor failures, and impute missing values. For cases where the proportion of missing data is small, it can be directly filled. For cases where the proportion of missing data is large, it is directly discarded; Then normalize the data with different dimensions and ranges to convert it into dimensionless numbers; Then represent the dimensionless data as a vector X = [x1, x2, …, x n , where x i is the i-th input parameter and n is the total number of input parameters.

[0066] In step S5, the method of inputting the evaluation results into the hybrid neural network model for iterative optimization includes: When performing iterative optimization, it is first necessary to clarify the definitions of positive and negative samples based on actual business requirements and evaluation goals; after determining the positive and negative samples, construct a confusion matrix based on the evaluation data obtained through monitoring; based on the constructed confusion matrix, calculate metrics such as recall rate and precision rate. Taking the energy-saving effect as an example, the higher the recall rate, the stronger the model's ability to predict energy consumption reduction, and the more actual energy consumption reduction situations it can capture; input these calculated recall rate, precision rate and other metrics into the hybrid neural network model, measure the current performance of the model through the loss function, and then continuously improve the control strategy to make the intelligent control of the entire dust removal system more accurate and efficient.

[0067] Adjust the hyperparameters of the model. The hyperparameters include learning rate, batch size, number of LSTM layers, number of neurons, etc.; if the data shows complex long-term dependencies and the model cannot capture features well, then consider increasing the number of LSTM layers. For example, when the model shows overfitting, that is, the performance on the training set is good, but the performance on new data or the test set is poor, and after analysis, it is found that it is due to the model being too complex, then the number of LSTM layers should be reduced.

[0068] Find the parameter combination that optimizes the model's prediction performance through methods such as grid search, random search, or Bayesian optimization. Change the connection methods between LSTM layers. The connection methods are skip connection, gated connection, etc.

[0069] In step S5, through the effect monitoring module, real-time monitor parameters such as the air volume, air pressure, and energy consumption after the fan is adjusted. The system compares and analyzes the monitored adjusted parameters with the parameters before adjustment, such as the change in air volume before and after adjustment, the difference in air pressure before and after adjustment, etc., calculates the energy consumption difference before and after the fan is adjusted, and evaluates the energy-saving effect by comparing the power consumption data before and after adjustment. At the same time, analyze the changes in air volume and air pressure, and combine the dust generation data to evaluate the stability and dust removal effect of the dust removal system, such as the improvement of the dust removal efficiency. The evaluation results can be used for model optimization.

[0070] The evaluation results include the following parameters: Air volume tracking error rate : Calculate according to the following formula: , where q pred is the predicted target air volume, and q post is the measured air volume after adjustment. Here, the air volume refers to the air volume at the dust generation point.

[0071] Air pressure stability index s p : Calculate according to the following formula: , where, σ(p post ) is the standard deviation of the adjusted air pressure, and μ(p post ) is the mean value of the adjusted air pressure. Here, the air pressure refers to the air pressure at the dust generation point.

[0072] The compliance rate R of the dust concentration m : is calculated according to the following formula: , where, T is the total number of sampling times within the statistical period, II() is the indicator function, and m t is the dust concentration in the gas discharged from the dust collector during the t-th sampling.

[0073] The energy saving rate e: is calculated according to the following formula: , where, E1 is the total energy consumption of the dust removal system before intelligent regulation, and E2 is the total energy consumption of the dust removal system before and after intelligent regulation.

[0074] The air volume change rate Q: is calculated according to the following formula: , where, q1 is the air volume at the dust removal point before adjustment, and q2 is the air volume at the dust removal point after adjustment.

[0075] The air pressure change rate P: is calculated according to the following formula: , where, p1 is the air pressure at the dust removal point before adjustment, and p2 is the air pressure at the dust removal point after adjustment.

[0076] The dust removal efficiency η: is calculated according to the following formula: , where, m in is the dust concentration entering the dust collector, and m out is the dust concentration discharged from the dust collector; The change in dust removal efficiency: is calculated according to the following formula: , where, η1 is the dust removal efficiency before intelligent regulation, and η2 is the dust removal efficiency after intelligent regulation.

[0077] Applying this hybrid neural network model to intelligently control the dust removal system, after running for one month, significant effects have been achieved compared with before the control. From the perspective of energy consumption, the overall energy consumption of the system has been saved by 25%. Before the control, the air volume fluctuation range was ±10%, and after the control, it stabilized at ±3%. The air pressure fluctuation decreased from ±15% before the control to ±5%, greatly improving the stability of the system operation. From the perspective of model evaluation parameters, the accuracy reached 88%, the loss value was 0.12, the recall rate was 85%, and the precision rate was 86%. The F1 score calculated based on the recall rate and precision rate was 85.5%. These actual operation effect data and model evaluation parameters fully prove the effectiveness and superiority of this hybrid neural network model in the intelligent control of the dust removal system.

[0078] The above detailed description is a specific description of the feasible embodiments of the present invention. This embodiment is not intended to limit the patent scope of the present invention. Any equivalent implementation or change without departing from the present invention shall be included in the patent scope of this case.

Claims

1. An intelligent control method for dust removal system in a steel enterprise, characterized in that: The steps include: S1. Collect real-time data at a set frequency, and divide the collected data into operating variables and regulating variables; the collected data include operating parameters at each dust generating point, valve opening K1 of the valve between each dust generating point and the dust collector, operating parameters of the dust collector, operating parameters of the ash conveying equipment, operating parameters of the fan and exhaust status parameters; S2. Construct a hybrid neural network model, which includes an LSTM-GPT-Neo backbone network, a DenseNet branch network, and a meta-learning feature fusion correction layer; the LSTM-GPT-Neo backbone network is located at the front end of the model, and is used to perform calculations and analysis based on the input operating variable data, and output the valve opening K1 and the optimal operating parameter distribution of the fan. It integrates the LSTM and GPT-Neo architectures. LSTM is used to extract high-level time series features from the input operating variable data. The time series features contain equipment operating status information. The GPT-Neo architecture introduces large-scale pre-training The idea of ​​language model is to understand and analyze the operation variable data at the semantic level, and enhance the ability to capture long-distance dependencies and complex patterns; the DenseNet branch network is used to calculate and analyze the input adjustment variable data and output dust removal performance parameters, which are used to characterize the mapping relationship between the adjustment variable and the dust removal efficiency; the meta-learning feature fusion correction layer adopts a meta-learning strategy to dynamically adjust the fusion weight and fusion mode according to the output of the LSTM-GPT-Neo backbone network and the DenseNet branch network, and according to different working conditions and historical data, so as to adapt to various complex and changeable situations; S3, collecting time series data of operating variables and adjustment variables to form a data set, dividing the data set into a training set and a validation set in chronological order, training the hybrid neural network model with the training set, and then optimizing the hybrid neural network model with the validation set; S4. Input the operating variable data and the regulating variable data into the trained hybrid neural network model for calculation and analysis, output the valve opening K1 and the optimal operating parameters of the fan, and perform real-time regulation on the operation of the valve and the fan based on the parameters; the optimal operating parameters of the fan include the fan speed r, the opening K2 of the air inlet valve and the opening K3 of the air outlet valve.

2. The intelligent control method for dust removal system of a steel enterprise according to claim 1 is characterized in that: The operating parameters of the dust generating point include air volume q1, wind pressure p, temperature t1, moisture content wet and dust concentration m1; the operating parameters of the dust collector include dust collector resistance zu, ash content c1 in the ash hopper, cleaning cycle T1, cleaning time T2 and cleaning intensity F; the operating parameters of the ash conveying equipment include ash conveying speed v and ash conveying volume c2; the operating parameters of the fan include fan speed r, fan inlet valve opening K2, fan outlet valve opening K3, temperature t2 and current I; the exhaust status parameters include exhaust air volume q2 discharged from the dust collector and dust concentration m2 in the gas discharged from the dust collector; The operating variables include the air volume q1, wind pressure p, temperature t1, moisture content wet, dust concentration m1, dust generating point s, resistance zu of the dust collector, ash amount c1 in the ash hopper, temperature t2 of the fan, current I of the fan, air volume q2 and dust concentration m2 of the dust collector at each dust generating point; the regulating variables include the fan speed r, valve opening K1, fan inlet valve opening K2, fan outlet valve opening K3, cleaning cycle T1, cleaning time T2, cleaning intensity F, ash conveying speed v and ash conveying amount c2.

3. The intelligent control method for dust removal system of a steel enterprise according to claim 2 is characterized in that: Assume that there are n dust generating points, s dust collectors, u fans, and f collection time points in the dust removal system. The processing and representation method of the operating variable data is as follows: Through the function y=Y(q1,p,t1,wet,m1,zu,c1,t2,I,q2,m2), the air volume q1, wind pressure p, temperature t1, moisture content wet, and dust concentration m1 at the dust generating point can be expressed as a three-dimensional array, namely q1[i][j][k], p[i][j][k], t1[i][j][k], wet[i][j][k], m1[i][j][k], where i represents the dust generating point number, 0≤i≤n, j represents the collection time point number, 0≤j≤f, and k represents the data dimension; The resistance zu of the dust collector, the amount of ash in the hopper c1, the exhaust air volume q2 of the dust collector, and the dust concentration m2 are represented by three-dimensional arrays, namely zu[h][j][k], c1[h][j][k], q2[h][j][k], and m2[h][j][k], where h represents the dust collector number, 0≤h≤s, j represents the collection time point number, 0≤j≤f, and k represents the data dimension; The temperature t2 and current I of the fan are represented by three-dimensional arrays, t2[g][j][k] and I[g][j][k], respectively, where g represents the fan number, 0≤g≤u, j represents the acquisition time point number, 0≤j≤f, and k represents the data dimension; The processing and presentation method of the adjustment variable data is: The adjustment variables are processed by the function t=T(r,K1, K2, K3,T1,T2,F,v,c2). The speed r, valve opening K1, fan inlet valve opening K2, fan outlet valve opening K3, cleaning cycle T1, cleaning time T2, cleaning intensity F, ash conveying speed v and ash conveying amount c2 of each fan can be expressed as three-dimensional arrays r[g][j][k], K1[g][j][k], K2[g][j][k], K3[g][j][k], T1[g][j][k], T2[g][j][k], F[g][j][k], v[g][j][k], c2[g][j][k], where g is the fan number, 0≤g≤u. For the ash conveying equipment, g is the fan number corresponding to the ash conveying equipment, j is the collection time point, 0≤j≤f, and k is the data dimension.

4. The intelligent control method for dust removal system of a steel enterprise according to claim 1 is characterized in that: The LSTM-GPT-Neo backbone network is located at the front end of the model, and its output data represents the possible optimal values ​​of the valve opening K1, the speed r of the fan, the opening K2 of the fan's air inlet valve, and the opening K3 of the air outlet valve in the form of probability distribution, as well as the confidence interval and uncertainty estimation; The LSTM unit adopts a stacked structure of four LSTM layers, each LSTM layer includes 512 neurons, and is combined with an improved variant of the gated recurrent unit GRU to capture complex time series features in the running variable data; The calculation and analysis method of the LSTM unit is: Each LSTM layer is used to capture the time series characteristics of the parameters, and a residual connection is introduced between each LSTM layer. A skip connection is added between two LSTM layers. The input of the previous LSTM layer is directly added to the output of the next LSTM layer element by element. The output is: ， in, is the output hidden state after residual connection and skip connection operations, is the hidden state after being processed by the current LSTM layer at time step t, is the hidden state of the previous time step t−1; The neurons in the LSTM layer are connected by weights. Each neuron performs weighted summation and activation function conversion on the input data. The output h of the jth neuron is j It is expressed as: , where w ij is the weight of the i-th input parameter to the j-th neuron, b j is the bias of the jth neuron, f is the activation function ReLU, expressed as: f(x)=max(0,x); The DenseNet branch network includes four DenseBlock layers and a Transition layer located between every two DenseBlock layers, each DenseBlock layer includes three convolutional layers and one skip connection, and each Transition layer consists of a convolutional layer and a pooling layer; The meta-learning feature fusion correction layer includes a meta-learner MAML and a fusion module. MAML obtains adapted weights through gradient updating according to working condition characteristics, and the fusion module performs weighted averaging on dynamic weights.

5. The intelligent control method for dust removal system of a steel enterprise according to claim 1 is characterized in that: In step S3, the training method of the hybrid neural network model is: S3-1. Collect historical operation sequence data, including the operation parameters at each dust generating point, the valve opening K1 of the valve between each dust generating point and the dust collector, the operation parameters of the dust collector, the operation parameters of the ash conveying equipment, the operation parameters of the fan and the exhaust status parameters; S3-2. After eliminating abnormal operation data, evaluate the historical operation data and select historical operation data with excellent energy saving effect, system operation stability and dust removal effect; S3-3. The screened data are used as a data set, and randomly divided into a training set and a validation set in a certain ratio according to time training; the data in the data set are preprocessed and input into the hybrid neural network model for training. The supervised learning method is used, and the optimal operating parameters of the wind turbine are used as the target variable. The weights are updated through the Adam optimization algorithm. The parameter update formula is: ， in, θ is the parameter to be optimized, η is the learning rate, and are the bias-corrected first-order and second-order moment estimates, ϵ A small value added to prevent division by zero.

6. The intelligent control method for dust removal system in a steel enterprise according to claim 5 is characterized in that: In the step S3-2, the method of data evaluation and screening is as follows: calculate the average energy consumption of the system in each time period, the rate of change of air volume and air pressure at the dust removal point, and the dust removal effect, first sort according to the dust removal effect, remove the bottom 40% of the data, and then sort the remaining data according to the average energy consumption, remove the bottom 30% of the data, and then sort the remaining data according to the rate of change of air volume and air pressure, and remove the bottom 20% of the data to obtain the screened data; In the step S3-3, the parameters of the model training are set as follows: the learning rate is set to 0.001, the maximum number of pre-training iterations is 500, and the maximum number of fine-tuning iterations is 200; The sparsity parameter is set to 0.4, which helps to reduce redundant information; the sparsity penalty parameter is 2, which is used to control the penalty intensity of sparsity; the activation function uses the ReLU function; and the fine-tuning loss function uses the mean square error loss function MSE.

7. The intelligent control method for dust removal system in a steel enterprise according to claim 1 is characterized in that: Also includes: S5. By monitoring the air volume q1, wind pressure p, dust concentration m1, energy consumption of dust collector, ash conveying equipment and fan, and exhaust status parameters at the dust generating point after regulation, the energy-saving effect of regulation, the stability of system operation and the dust removal effect are evaluated, and the evaluation results are input into the hybrid neural network model for iterative optimization to continuously improve the regulation strategy of the hybrid neural network model; Among them, the method of inputting the evaluation results into the hybrid neural network model for iterative optimization includes: S5-1. Define positive and negative samples based on actual business needs and evaluation objectives; S5-2. After determining the positive and negative samples, a confusion matrix is ​​constructed based on the evaluation data obtained from monitoring; S5-3. Based on the constructed confusion matrix, the evaluation index of the hybrid neural network model is calculated, and then the evaluation index is input into the hybrid neural network model. The current performance of the model is measured through the loss function, and the control strategy is continuously improved.

8. The intelligent control method for dust removal system in a steel enterprise according to claim 7 is characterized in that: The evaluation results include the following parameters: Air volume tracking error rate : Calculated according to the following formula: ， Among them, q pred is the target air volume at the predicted dust generating point, q post The measured air volume at the dust-generating point after intelligent control; Wind pressure stability index p : Calculated according to the following formula: ， Among them, σ(p post ) is the standard deviation of wind pressure at the dust generating point after intelligent control, μ(p post ) is the average wind pressure at the dust generating point after intelligent control; Dust concentration compliance rate R m : Calculated according to the following formula: ， Where T is the total number of sampling times in the statistical period, II() is the indicative function, and m t is the dust concentration in the gas discharged from the dust collector at the tth sampling time; Energy saving rate e: calculated according to the following formula: ， Among them, E1 is the total energy consumption of the dust removal system before intelligent regulation, and E2 is the total energy consumption of the dust removal system before and after intelligent regulation; Air volume change rate Q: calculated according to the following formula: ， Among them, q1 is the air volume at the dust removal point before intelligent control, and q2 is the air volume at the dust removal point after intelligent control; Wind pressure change rate P: calculated according to the following formula: ， Among them, p1 is the wind pressure at the dust removal point before intelligent regulation, and p2 is the wind pressure at the dust removal point after intelligent regulation; Dust removal efficiency η: calculated according to the following formula: ， Among them, m in is the dust concentration entering the dust collector, m out The dust concentration discharged from the dust collector; Change in dust removal efficiency: calculated according to the following formula: ， Among them, η1 is the dust removal efficiency before intelligent regulation, and η2 is the dust removal efficiency after intelligent regulation.

9. An intelligent control system for dust removal system of a steel enterprise, used to implement the intelligent control method for dust removal system of a steel enterprise as claimed in any one of claims 1 to 8, characterized in that: Including dust removal equipment, data acquisition module, intelligent control module, operation control module and monitoring optimization module; The dust removal equipment includes an air collecting hood, a conveying pipeline, a valve, a dust collector, an ash conveying device, a fan and a chimney. A plurality of the air collecting hoods are arranged at dust generating points. The air collecting hoods are connected to the dust collectors through the conveying pipelines. The valves are arranged on the conveying pipelines. The dust collectors are arranged at the rear end of the air collecting hoods. The dust collectors are provided with ash cleaning devices, which are used to clean the dust accumulated in the dust collectors. The ash conveying device is arranged at the outlet of the dust collector hopper and is used to convey the accumulated dust. The air inlet of the fan is connected to the dust collector, and the air outlet of the fan is connected to the chimney. The data acquisition module includes a plurality of sensors and a data collector, wherein the sensors are respectively arranged at the conveying pipes, valves and dust collectors connected to the air collecting hood at each dust generating point, and the data collector and each of the sensors, valves, dust collectors, ash conveying equipment and fans are used to collect the operating parameters at each dust generating point, the valve opening of the valve, the operating parameters of the dust collector, the operating parameters of the ash conveying equipment and the operating parameters of the fan; The data acquisition module is connected to the intelligent control module and the monitoring optimization module to transmit the collected data; The intelligent control module is embedded with a hybrid neural network model, which is used to perform calculations and analysis based on the collected data to obtain the optimal operating parameters of the fan; The operation control module is connected to the intelligent control module and is used to receive the optimal operating parameters of the fan obtained by the intelligent control module. The operation control module is also connected to the fan and is used to adjust the speed of the fan, the opening of the air inlet valve and the opening of the air outlet valve according to the optimal operating parameters of the fan obtained by the intelligent control module, so as to control the air volume of the fan; The monitoring and optimization module includes a data analysis unit, a closed-loop feedback unit and a data storage unit. The data analysis unit is used to perform calculations and analysis based on the collected data to evaluate the energy-saving effect of the regulation, the stability of the system operation and the dust removal effect. The closed-loop feedback unit is used to transmit the evaluation results to the intelligent regulation module, perform iterative optimization on the hybrid neural network model, and form a dynamic optimization closed loop. The data storage unit is used to store the collected data and the data obtained by calculation and analysis.

10. The intelligent control system for dust removal system of a steel enterprise according to claim 9, characterized in that: The sensors include an air volume measuring instrument, an air pressure gauge, a dust concentration sensor, a thermometer and hygrometer, a pressure sensor and a material level meter. The air volume measuring instrument is respectively arranged at the conveying pipeline and the dust collector at each dust generating point, and is used to measure the air volume q1 at the dust generating point and the exhaust air volume q2 discharged from the dust collector. The air pressure gauge is arranged at the conveying pipeline at the dust generating point, and is used to measure the wind pressure p at the dust generating point. The dust concentration sensor is respectively arranged at the conveying pipeline and the dust collector at the dust generating point, and is used to measure the dust concentration m1 at the dust generating point and the dust concentration m2 discharged from the dust collector. The thermometer and hygrometer is arranged at the conveying pipeline at the dust generating point, and is used to measure the temperature t1 and the moisture content wet at the dust generating point. The pressure sensor and the material level meter are both arranged in the dust collector, and are used to measure the resistance zu of the dust collector and the ash amount c1 in the ash hopper. The dust generating points are the places for coal loading, coke guiding and coke discharging of coke ovens, the secondary dust removal of converters and the iron-making site of blast furnaces. The gas collecting hoods are arranged at the furnace head of the carbonization chamber, the coke interceptor, the coke pusher, above the converter furnace mouth and above the iron-making site of the blast furnace.

Images