Intelligent optimization control method of furnace temperature based on multi-loop improved SNA-PID

Through the multi-loop ISNA-PID control method and particle swarm optimization algorithm, the stability of furnace temperature control and pollutant emission optimization problems during the MSWI process are solved, and the stable control of furnace temperature and the reduction of pollutant emission concentration are achieved, which improves MSWI Safety and environmental protection of the process.

CN115857585BActive Publication Date: 2024-11-19BEIJING UNIV OF TECH
View PDF 6 Cites 0 Cited by

Patent Information

Application Number
CN202211628882.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-12-18
Publication Date
2024-11-19
Estimated Expiration
2042-12-18

AI Technical Summary

Technical Problem

During the existing MSWI process, it is difficult to achieve stability in the furnace temperature control, resulting in insufficient combustion, frequent emissions of toxic and harmful substances and equipment failures, and difficult to optimize the concentration of pollutant emissions.

Method used

An intelligent optimization control method for furnace temperature based on multi-loop ISNA-PID is proposed, including a controlled object model for furnace temperature, a multi-loop intelligent controller and pollutant emission optimization algorithm, using wavelet denoising, least squares decision tree and particle group Optimization algorithm builds a control framework for industrial practice.

Benefits of technology

It realizes stable control of furnace temperature, reduces NOx and CO2 emission concentrations, reduces equipment failures, improves combustion efficiency and environmental protection indicators, and meets the intelligent optimization control needs of the MSWI process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN115857585B_ABST
    Figure CN115857585B_ABST
Patent Text Reader

Abstract

The present invention provides an intelligent optimization control method for furnace temperature based on multi-loop improved SNA-PID. Most actual industrial sites rely on domain experts to operate according to pollutant emission concentrations based on experience, which leads to large fluctuations in furnace temperature and difficulty in optimizing pollutant emission concentrations. First, facing the problem of process object modeling with furnace temperature control as the goal, a multi-input single-output controlled object model based on the least squares decision tree (LSDT) is established; then, facing the problem of furnace temperature being related to multiple operating variables, a multi-loop controller based on ISNA-PID is proposed; finally, facing the optimization setting problem with the goal of reducing the pollutant emission concentration related to furnace temperature, a single-input single-output indicator model based on the CART tree algorithm is established, and the particle swarm optimization (PSO) algorithm is used to solve the NOx and CO 2 The emission optimization model obtains the furnace temperature setting value. The present invention can realize the autonomous optimization of the furnace temperature setting value and reduce the pollutant emission concentration.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to a furnace temperature intelligent optimization control method Background Art

[0002] The symbols in this paper are shown in Table 1.

[0003] Table 1 Meaning of symbols in the text

[0004]

[0005] Compared with other municipal solid waste (MSW) treatment technologies, MSW incineration (MSWI) technology has become the mainstream treatment technology worldwide due to its advantages in harmlessness, reduction and resource utilization. Its main advantage is to reduce the use of land resources and achieve energy recycling. As of September 2022, China has put into operation 811 MSWI power plants, excluding Taiwan Province of China, and the MSW treatment capacity accounts for more than 50% of the total. As a typical process industry, the MSWI process involves complex physical and chemical reactions. It is a multi-input and multi-output system with the characteristics of multi-loop, strong coupling, and nonlinearity. Whether it can operate stably depends on the coordinated control between the controlled variables and many operating variables, and is related to interference factors such as the differences in MSW components, the operation and maintenance management of the equipment, and the experience of the operating engineer. Generally, the control of the MSWI process follows the "3T+E" principle, that is, the furnace temperature is greater than 850℃, the flue gas residence time is greater than 2 seconds, a certain flue gas turbulence intensity, and sufficient air volume. Theoretically, the "3T+E" principle can effectively ensure that harmful substances produced during the combustion process are fully decomposed, and control the generation of toxic pollutants such as acid gases and dioxins from the source.

[0006] In the above-mentioned "3T+E" principle, furnace temperature is particularly important. It is also closely related to other controlled variables and process parameters of the MSWI process, such as being proportional to boiler steam flow, inversely proportional to flue gas oxygen content, and directly related to dioxin decomposition and regeneration and combustion efficiency. It can be seen that furnace temperature is an important indicator to ensure the safe operation of the MSWI process, and is also a key factor in minimizing the emission concentration of various pollutants in the exhaust gas. Therefore, it is very necessary to study the stable control of furnace temperature in the MSWI process.

[0007] Since the automatic combustion control (ACC) system needs to be "localized" according to factors such as the characteristics of MSW and the quality of personnel in different countries, most of China's MSWI power plants currently adopt a manual operation mode that relies on experts in the field. In view of the many problems faced by MSWI power plants, such as complex and fluctuating MSW components and difficult to measure calorific value, unstable combustion process, strong volatility of pollution emissions, frequent failures caused by ultra-high furnace temperature operation and high operation and maintenance costs, the development of intelligent optimization control technology is the key to achieving intelligent, low-carbon and harmless incineration. The author believes that in the case of difficulty in conducting research on intelligent optimization control algorithms for furnace temperature and conducting industrial field verification, there are currently three problems to be solved: 1) A controlled object model with operating variables as input for furnace temperature control; 2) A furnace temperature loop controller that conforms to the actual status of industrial process operation; 3) A setting optimization algorithm that can minimize the emission concentrations of multiple pollutants related to furnace temperature.

[0008] In summary, this paper proposes an intelligent optimization control method for furnace temperature based on multi-loop ISNA-PID from the perspective of modeling, control and optimization. The main innovations are: 1) An intelligent optimization control framework for furnace temperature control in the MSWI process is proposed, including a controlled object model for furnace temperature control, a multi-loop ISNA-PID controller based on industrial practice, and an optimization setting algorithm based on PSO to reduce NOx and CO2 emission concentrations; 2) A multi-input single-output controlled object model that is aimed at furnace temperature control and conforms to industrial practice is established to provide support for the subsequent research on intelligent optimization control algorithms; 3) A multi-loop furnace temperature controller based on industrial practice is established, and the input and weight update algorithm of the traditional SNA-PID controller is improved, so that the error increment has a larger weight share, thereby ensuring a smaller output oscillation to maintain the stability of operation and production; 4) Based on the above model and controller, a single-input single-output NOx and CO2 indicator model is established, and the PSO algorithm is used to optimize the furnace temperature setting value with the minimum comprehensive pollutant emission concentration as the optimization target.

[0009] The process flow of a MSWI power plant in Beijing is as follows: Figure 1 shown.

[0010] Depend on Figure 1It can be seen that MSW is transported by vehicles, weighed on a scale and discharged into a solid waste pool. After 3 to 7 days of biological fermentation and dehydration, the MSW in the solid waste pool is put into the hopper by a grab bucket; then, the feeder pushes the MSW to the grate, and goes through three stages of drying, combustion and incineration. In order to ensure that harmful substances in the high-temperature flue gas are fully decomposed and burned, the flue gas temperature should be controlled above 850°C, the flue gas stays for more than 2 seconds and has sufficient flue gas turbulence; then, the high-temperature flue gas enters the waste heat boiler, and the high-temperature steam generated by heat exchange The steam turbine generator set is driven to generate electricity; then, the flue gas mixed with slaked lime and activated carbon enters the deacidification reactor, where a neutralization reaction occurs and dioxins and heavy metals in the flue gas are adsorbed; then, the flue gas particles, neutralization reactants and activated carbon adsorbents are removed in the bag filter; finally, the waste gas containing dust, CO, NOx, SO2, HCL, HF, Hg, Cd and other substances such as dioxins is discharged into the atmosphere through the chimney, and the ash produced by incineration is transported to the slag pit by a slag scoop machine and then transported to a designated location for landfill by vehicles.

[0011] Ensuring that the furnace temperature meets the standard is the primary prerequisite for achieving stable operation of the MSWI process. The reasons are as follows: MSW has a complex composition. If it is not burned fully, a large amount of toxic and harmful substances will be produced, which need to be completely decomposed at high temperatures; secondly, the furnace temperature is closely related to the steam flow rate, which is used to drive the turbine to generate electricity; thirdly, the high operation of the furnace temperature for a long period of time will cause burn-through, coking, ash accumulation and other phenomena, which will cause frequent failures of the incineration equipment. In addition, maintaining a suitable furnace temperature is extremely beneficial to reducing the emission concentration of environmental indicators such as NOx and CO2 in the pollutants directly related to it. Therefore, stable control of the furnace temperature is extremely critical for the MSWI process. Summary of the invention

[0012] The furnace temperature intelligent optimization control method based on multi-loop INSA-PID proposed in this paper consists of three parts: the controlled object model for furnace temperature, the multi-loop intelligent controller for furnace temperature and the setting optimization for furnace temperature. Figure 2 shown.

[0013] Among them, y FT represents the optimal setting value of furnace temperature; U represents the output matrix of multi-loop controller; Represents the output value of the furnace temperature model; and Represents the output values ​​of NOx and CO2 models. The remaining symbols are described below.

[0014] Due to problems such as sensor aging and electromagnetic interference during the acquisition process, the different degrees of noise in the acquired data will seriously affect the model and control accuracy. This paper proposes a controlled object model for furnace temperature, including wavelet-based data denoising and a furnace temperature model based on the least squares decision tree (LSDT). Its structure is as follows: Figure 3 shown.

[0015] in, and They represent the original data of primary air volume, secondary air volume, feeder grate average speed, drying grate average speed and ammonia injection volume respectively; u PriAir 、u SecAir 、u Feeder 、u Dry and u NH3H2O They represent the data of primary air volume, secondary air volume, average speed of feeder grate, average speed of drying grate and ammonia injection volume after wavelet denoising respectively; Represents the furnace temperature model output.

[0016] 1) Data denoising based on wavelet

[0017] Compared with Fourier transform, wavelet analysis solves the problem that Fourier transform is difficult to handle non-stationary signals, and has the characteristics of simple algorithm and small amount of calculation. The hard threshold denoising expression of wavelet threshold method is as follows:

[0018]

[0019] The soft threshold denoising expression is as follows:

[0020]

[0021]

[0022] Among them, w θ and Indicates the wavelet coefficients before and after denoising; threshold M is the signal length.

[0023] From equations (1)-(3), we can see that the hard threshold retains the elements whose absolute values ​​are greater than the threshold and sets the elements whose absolute values ​​are less than the threshold to 0; the soft threshold sets the elements whose absolute values ​​are less than the threshold to 0 and shrinks the other non-zero elements to 0. The advantage of the soft threshold is good smoothness, which is suitable for the process data denoising used in this paper.

[0024] 2) Furnace temperature model based on LSDT

[0025] Combined with industrial practice, in order to simplify the description, all input characteristics of the furnace temperature model are recorded as

[0026] First, take the mth input feature As the segmentation variable, take the nth sel The value corresponding to the sample Cut as the split point and divide the input feature space into two regions and as follows:

[0027]

[0028] Based on the following criteria, the optimal segmentation variables and segmentation points are found by traversal.

[0029]

[0030] in, and Representing regions and True value of the sample; and Representing regions and The average true value of the samples in .

[0031] Then, repeat the above process until the number of leaf node samples is less than the empirically set threshold θ FT (Determined by the number of training samples, generally one tenth of the number of samples is taken as the threshold).

[0032] Finally, the input feature space is divided into K FT area and marked as

[0033] Furthermore, the LSDT-based FT model is as follows:

[0034]

[0035] in, is the weight of the kth leaf node; I(·) is the indicator function. The function value is 1 if it exists, otherwise it is 0. represents the number of intermediate nodes, and

[0036] The loss function for building the above model is defined as follows:

[0037]

[0038] Among them, λ is the regularization coefficient.

[0039] Convert it into vector form and transform the loss function into Taking the derivative and setting it to 0, we get:

[0040]

[0041] Among them, U model Represents the model input feature vector.

[0042] In view of the dynamic characteristics and strong nonlinearity of the MSWI process, field experts in the actual field adjust multiple operating variables to achieve control of the furnace temperature. Therefore, this paper first calculates the correlation coefficient between the operating variables such as primary air volume, secondary air volume, and average speed of the feeder and the furnace temperature, and takes the primary air with a lower correlation coefficient as the disturbance variable, and proposes the following Figure 4 The 4-loop improved SNA-PID (ISNA-PID) controller is shown.

[0043] Take the jth operating variable As an example (j=1,···,4), the structure of ISNA-PID controller is as follows Figure 5 shown.

[0044] The original SNA-PID controller needs to convert the error signal into a state, and its input is:

[0045]

[0046]

[0047] Where, e(k) represents the error value; and FT They represent the furnace temperature setting value and the model output value respectively; k represents the number of iterations.

[0048] In order to increase the weight of the error increment, this paper improves the input z3 of the SNA-PID controller to reduce the output oscillation, as follows:

[0049] z3(k)=[e(k)-e(k-1)] 2 (11)

[0050] Therefore, the output of the ISNA-PID controller for the jth manipulated variable can be expressed as:

[0051]

[0052] in, represents the increment of the jth operation variable; represents the neuron gain coefficient of the jth manipulated variable; represents the i-th neuron weight coefficient of the j-th operating variable at time k, expressed as:

[0053]

[0054] From equations (18) and (19), it can be seen that the SNA-PID controller tracks the set value under the supervised Hebb learning algorithm by continuously adjusting the neuron weight coefficient online, as follows:

[0055]

[0056] Based on formula (20), this paper repeatedly uses the error increment to update the weighted coefficient. The modified weighted coefficient update formula is as follows:

[0057]

[0058] in, and Respectively represent the weight coefficients of the proportional, integral and differential neurons in the jth manipulated variable, and They represent the learning coefficients of the proportional, integral and differential neurons in the jth manipulated variable respectively.

[0059] Optimizing furnace temperature settings

[0060] 1) NOx and CO2 index model based on CART tree

[0061] Here, only the effect of furnace temperature on NOx and CO2 is considered. In order to increase the interpretability of the model, a single-input and single-output indicator model is constructed using the CART tree.

[0062] The NOx modeling process is as follows:

[0063] First, take the nth sel The value corresponding to the sample Cut as the split point and divide the input feature space into two regions and as follows:

[0064]

[0065] Based on the following criteria, the optimal segmentation point is found by traversal.

[0066]

[0067] in, and Respectively represent and The true value of the samples in the region; and Respectively represent and The average of the true values ​​of the samples in the region.

[0068] Then, repeat the above process to grow until the number of leaf node samples is less than the empirically set threshold θ NOx (Determined by the number of training samples, generally one tenth of the number of samples is taken as the threshold).

[0069] Finally, the input feature space is divided into K NOx area and marked as

[0070] Furthermore, the NOx model constructed based on the CART tree is as follows:

[0071]

[0072]

[0073] in, Indicates area The number of samples included in ; Indicates area Middle The true value of the sample; I(·) is the indicator function, when The function value is 1 if it exists, otherwise it is 0.

[0074] The CO2 modeling process is as follows:

[0075] First, take the nth sel The value corresponding to the sample Cut as the split point and divide the input feature space into two regions and as follows:

[0076]

[0077] Based on the following criteria, the optimal segmentation point is found by traversal.

[0078]

[0079] in, and Respectively represent and The true value of the samples in the region; and Respectively represent and The average of the true values ​​of the samples in the region.

[0080] Then, repeat the above process to grow until the number of leaf node samples is less than the empirically set threshold (Determined by the number of training samples, generally one tenth of the number of samples is taken as the threshold).

[0081] Finally, the input feature space is divided into area and marked as

[0082] Furthermore, the CO2 model constructed based on the CART tree is as follows:

[0083]

[0084]

[0085] in, Indicates area The number of samples included in ; Indicates area Middle The true value of samples; represents the empirically set threshold; I(·) is the indicator function. The function value is 1 if it exists, otherwise it is 0.

[0086] 2) Multi-objective optimization model

[0087] The multi-objective model for optimizing furnace temperature setting with NOx and CO2 emission concentrations as optimization targets is as follows:

[0088]

[0089] in, represents the comprehensive emission concentration of pollutants; w Pollutant represents the weight vector; γ Pollutant Represents the pollutant emission concentration output vector; ω NOx , Respectively represent the weights of NOx and CO2 emission concentrations.

[0090] To minimize the objective function γ mix ,In this section, based on the single-input and single-output NOx and CO2 indicator model established based on the CART tree, the particle swarm optimization (PSO) algorithm is used to search the optimal furnace temperature setting value.

[0091] As one of the swarm intelligence algorithms, the PSO algorithm simulates the predation behavior of bird flocks, regards the problem to be optimized as a predatory bird flock, and the solution space as the flight space of the bird flock. The position of each bird in the space is a particle of the PSO algorithm in the solution space. It uses the mutual cooperation and information sharing between individuals in the population to find the optimal solution. It has the advantages of simple principle, few adjustment parameters, and fast operation speed, and is widely used in engineering optimization problems. The particle speed and position update formula is:

[0092]

[0093] Where t represents the number of iterations; ω represents the inertia factor; r1 and r2 are random numbers between (0,1); c1 and c2 represent learning factors, which are usually 2; x i (t), pbest i (t) and gbest(t) represent the current position, individual optimum and global optimum of the ith particle at the tth iteration, respectively.

[0094] Since the furnace temperature setting is usually an integer, this article rounds the speed update formula as follows:

[0095]

[0096] Where q represents the particle number. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 Process flow chart of a MSWI power plant in Beijing

[0098] Figure 2 Strategy diagram of furnace temperature intelligent optimization control method based on multi-loop ISNA-PID

[0099] Figure 3 Structure diagram of the controlled object model for furnace temperature

[0100] Figure 4 Structure diagram of multi-loop intelligent controller for furnace temperature

[0101] Figure 5 ISNAPID controller structure diagram

[0102] Figure 6 Data preprocessing results

[0103] Figure 7 Furnace temperature model fitting curve

[0104] Figure 8 Tracking result curve

[0105] Fig. 9 Error curve

[0106] Fig.10 Controller output curve

[0107] Fig.11 Controller parameter change curve

[0108] Fig.12 NOx model fitting curve

[0109] Fig.13 CO2 model fitting curve

[0110] Fig.14Optimal furnace temperature setting and pollutant emission concentration curve under different NOx weights

[0111] Fig.15 Controller tracking curve DETAILED DESCRIPTION

[0112] This paper uses process data from a Beijing MSWI power plant that runs continuously for a total of 16 hours on a certain day to verify the accuracy of the proposed intelligent optimization control framework.

[0113] After taking the 60s average of the process data, the data was preprocessed and 957 sets of data were finally obtained. To ensure the smoothness of the data, wavelet soft threshold denoising was adopted, and the relevant parameters were set as follows: the number of decomposition layers was 3, the heuristic threshold Heursure rule and the signal length were 957. The denoising results are shown in Figure 2. Figure 6 shown.

[0114] Depend on Figure 6 It can be seen that since the average speed of the feeder and the average speed of the drying grate are constant changes and there is no smooth fluctuation process, no denoising is performed on them.

[0115] 2) Furnace temperature model

[0116] The root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (R-square) are used to evaluate the performance of the established model. The specific formula is:

[0117]

[0118]

[0119]

[0120] in, is the model prediction value, is the model truth value, is the model mean.

[0121] The relevant parameters are set as follows: Minimum number of samples is 10 and the regularization coefficient λ FT To verify the effectiveness of the model, CART tree, RF and BPNN models are used for comparison, among which: CART tree model sets the minimum number of samples 10; RF model sets the minimum number of samples is 10 and the number of decision trees The BPNN model sets the number of neurons in the hidden layer to 11, the number of iterations to 1000, the convergence error to 0.1, and the learning rate to 0.5. The comparison results of the fitting curves of the training set, validation set, and test set are shown in the figure. Figure 7The comparison results of model performance indicators are shown in Table 2.

[0122] Table 2 Comparison results of performance indicators of controlled object models

[0123]

[0124] From the above modeling results, we can know that:

[0125] (1) The data after wavelet soft threshold denoising is smoother than the original data, which is conducive to the subsequent research of intelligent optimization control algorithm;

[0126] (2) For the furnace temperature model, the LSDT model in this paper has the best performance compared with other models, with good fitting ability and modeling accuracy. The reason is that weights are introduced into the tree structure model to improve the smoothing performance of the model and correct the shortcomings of the leaf node mean output of the tree structure model.

[0127] The integrated square error (ISE), integrated absolute error (IAE), and maximum deviation (DEV max ) to evaluate the controller performance, and the calculation formula is as follows:

[0128]

[0129]

[0130] DEV max =max{|e(t)|} (32)

[0131] Among them, t0 and t f Indicates the control start and end time.

[0132] Based on the furnace temperature model mentioned above, this paper takes secondary air volume, feeder grate speed, drying grate speed and ammonia injection as operating variables, primary air volume as disturbance, furnace temperature as controlled variable, and uses 4 ISNA-PID controllers for control experiments. It is compared with the traditional SNA-PID controller and PID controller, among which: the neuron gain coefficient of SNA-PID / ISNA-PID controller is 0.05, and the proportional, differential and integral neuron learning rates are 0.3, 0.1 and 0.2 respectively; for PID controller, the proportional, differential and integral coefficients of secondary air volume, feeder speed, drying grate speed and ammonia injection controller are 0.05, 0.005 and 0.001, 0.01, 0.005 and 0.001, 0.01, 0.005 and 0.001 and 0.05, 0.005 and 0.001 respectively.

[0133] Tracking results and error curves are shown in Figures 8-9 shown.

[0134] Depend on Figures 8-9 It can be seen that under the condition of variable set value, all three controllers can effectively track the set value, but the ISNA-PID controller has smaller output oscillation and error. The output and parameter adjustment curves of SNA-PID and ISNA-PID controllers are shown in Figures 10-11 shown.

[0135] Depend on Figures 10-11 It can be seen that compared with the SNA-PID controller, the ISNA-PID controller has a smaller parameter variation range, which in turn reduces the controller output variation range accordingly, but the two have the same control effect, indicating that the ISNA-PID controller has the advantage of reducing frequent fluctuations of the controller.

[0136] This paper reuses the error increment in the controller input and weighted coefficient update formula. Although it increases the adjustment time, it reduces the output oscillation amplitude of the controlled variable and is more suitable for MSWI industrial processes.

[0137] To evaluate the controller performance, ISNA-PID, SNA-PID and PID controllers are compared, and the statistical results are shown in Table 3.

[0138] Table 3 Controller performance evaluation results

[0139]

[0140] From the above experimental results, we can know that:

[0141] (1) In the variable set value comparison test, all three controllers can track the set value in the presence of disturbance, but the ISNA-PID controller has smaller output oscillation and error. The reason is that the weight of the error increment is increased in the ISNA-PID controller, so when approaching the set value, the controller's output increment is slowed down and the controlled variable fluctuates within a small range to achieve tracking.

[0142] (2) From the controller performance evaluation results in Table 3, it can be seen that the ISNA-PID controller has the best control performance in terms of ISE and IAE indicators, but its DEV max The reason is that the ISNA-PID controller reduces the oscillation of the controlled variable by slowing down the output amplitude of the controller.

[0143] Setting optimization results

[0144] 1) Indicator Model

[0145] In this paper, the NOx and CO2 indicator models are single-input and single-output. Considering the interpretability of the model and the extraction of knowledge rules, this paper adopts the CART tree algorithm to model and sets the minimum number of samples. To verify the effectiveness of the model, it is compared with LSDT, RF and BPNN models, where: LSDT model sets the minimum number of samples is 10 and the regularization coefficient λ NOx 0.5; RF model sets the minimum number of samples is 10 and the number of decision trees The BPNN model sets the number of hidden layer neurons to 3, the number of iterations to 1000, the convergence error to 0.1 and the learning rate to 0.5. The comparison results of the test set fitting curve are as follows: Fig.12 shown.

[0146] Setting the minimum number of samples for building a CO2 model using CART tree 10; LSDT model sets the minimum number of samples is 10 and the regularization coefficient 0.5; RF model sets the minimum number of samples is 10 and the number of decision trees The BPNN model sets the number of hidden layer neurons to 3, the number of iterations to 1000, the convergence error to 0.1 and the learning rate to 0.5. The comparison results of the test set fitting curve are as follows: Fig.13 shown.

[0147] For the indicator model, although the CART tree algorithm is inferior to other algorithms in generalization performance indicators, it performs well on the fitting curve. At the same time, compared with other modeling algorithms, it has the advantages of fast training speed, simple structure, and strong interpretability, which is convenient for mapping causal relationships in industrial data.

[0148] 2) Multi-objective optimization model

[0149] According to the established multi-objective optimization model, the PSO algorithm is used to search for the optimal furnace temperature setting value to minimize the comprehensive emission concentration of pollutants. The PSO parameters are set as follows: the population size ps is 30, the number of iterations iter is 20, the inertia factor ω is 0.8, the learning factor c1=c2=2, and the maximum particle velocity V max The particle position interval is [900,950] and only integers are taken.

[0150] The optimal furnace temperature setting value and pollutant emission concentration curve under different weights are as follows: Fig.14 shown.

[0151] Depend on Fig.14 It can be seen that when ω NOx =0.0788 and When the concentration of both pollutants can reach the lowest level, the following results are obtained: the optimal setting value is 935℃ and the NOx emission concentration is 103.59mg / m 3 The N and CO2 emission concentrations are 7.9255%; accordingly, the controller tracking curve is as follows Fig.15 shown.

[0152] In summary, for the index model established based on the CART tree algorithm, the PSO algorithm realizes the optimal setting of the optimal furnace temperature to minimize the comprehensive emission concentration of pollutants and can provide guidance for operation and production.

[0153] In view of the intelligent optimization control method of furnace temperature based on multi-loop ISNA-PID proposed in this paper, this section uses actual field data to verify each module and achieves good results. Further analysis and summary are as follows: 1) The LSDT algorithm is used to establish a controlled object model with furnace temperature control as the goal. The experimental results show that the controlled object model can reflect the furnace temperature characteristics of the actual MSWI process; 2) Four ISNA-PID controllers are used to control the constructed furnace temperature model. The experimental results show that the ISNA-PID controller has the advantages of simple structure, small output oscillation and strong anti-interference ability, which is more in line with the actual industrial site; 3) The CART tree algorithm is used to establish the NOx and CO2 index model. Although the CART tree algorithm performs poorly in generalization performance indicators, the curve fitting is better, and the structure is simple and easy to explain. This part can be further studied in the future; and the PSO is used to solve the constructed MSWI process multi-objective optimization model, which can obtain the optimal furnace temperature setting value under different pollutant weighting coefficients, so as to achieve the purpose of reducing the emission concentration of pollutants directly related to the furnace temperature.

[0154] In summary, for the MSWI process, this paper first established a multi-input and single-output furnace temperature controlled object model based on LSDT and a single-input and single-output NOx and CO2 indicator model based on the CART tree algorithm, and verified it using the proposed intelligent optimization control framework. The strategies of each part of the framework can be improved according to actual needs, which provides a basis for the research on intelligent optimization control of the MSWI process.

[0155] As a typical process industry, the MSWI process involves complex physical and chemical reactions. It is a multi-input and multi-output system with the characteristics of multiple loops, strong coupling and nonlinearity. Industrial practice shows that the stable operation of the process depends on the coordinated control of many operating variables and controlled variables. This paper proposes an intelligent optimization control method for furnace temperature based on multi-loop ISNA-PID. Based on actual industrial data, the method is verified to achieve the purpose of reducing pollutant emission concentrations by optimizing the set value and stabilizing the furnace temperature. Main contributions: Based on the MSWI process, the experience of domain experts and existing research, an intelligent optimization control framework for MSWI process furnace temperature control is proposed from the perspective of modeling, control and optimization, including a process object model for furnace temperature control, a multi-loop ISNA-PID controller based on industrial practice, and a PSO-based NOx and CO2 emission reduction optimization algorithm. The experimental results show that the furnace temperature model constructed by the LSDT algorithm has good fitting ability; the multi-loop ISNA-PID controller can track the furnace temperature set value, and the controller can reduce the output oscillation by slowing down the output increment when approaching the set value; the PSO algorithm can take the minimum NOx and CO2 emission concentration as the optimization target, obtain the optimal furnace temperature set value of the ISNA-PID controller, and reduce the comprehensive emission concentration of pollutants.

[0156] In the future, we will focus on studying process object models and multi-input multi-output controllers that can represent the MSWI process more comprehensively and closer to industrial reality, and conduct in-depth research on intelligent optimization control algorithms for the MSWI process.

Claims

1. The intelligent optimization control method of furnace temperature based on multi-loop SNA-PID is characterized by: It consists of three parts: a controlled object model for furnace temperature, a multi-loop intelligent controller for furnace temperature, and setting optimization for furnace temperature; y FT represents the optimal setting value of furnace temperature; U represents the output matrix of multi-loop controller; Represents the output value of the furnace temperature model; and Indicates the output values ​​of NOx and CO2 models, and the remaining symbols are described below; The controlled object model for furnace temperature includes wavelet-based data denoising and a furnace temperature model based on the least squares decision tree. and They represent the original data of primary air volume, secondary air volume, feeder grate average speed, drying grate average speed and ammonia injection volume respectively; u PriAir 、u SecAir 、u Feeder 、u Dry and u NH3gH2O They represent the data of primary air volume, secondary air volume, average speed of feeder grate, average speed of drying grate and ammonia injection volume after wavelet denoising respectively; Represents the furnace temperature model output; 1) Data denoising based on wavelet The soft threshold denoising expression is as follows: Among them, w θ and Indicates the wavelet coefficients before and after denoising; threshold M is the signal length; The soft threshold is to set the elements whose absolute value is less than the threshold to 0, and the other non-zero elements shrink to 0; 2) Furnace temperature model based on least squares decision tree All input features of the furnace temperature model are recorded as First, take the mth input feature As the segmentation variable, take the nth sel The value corresponding to the sample Cut as the split point and divide the input feature space into two regions and as follows: Based on the following criteria, the optimal segmentation variables and segmentation points are found by traversal; in, and Representing regions and True value of the sample; and Representing regions and The average value of the true value of the samples in ; Then, repeat the above process until the number of leaf node samples is less than the empirically set threshold θ FT , take one tenth of the sample size as the threshold; Finally, the input feature space is divided into K FT area and marked as Furthermore, the FT model based on the least squares decision tree is as follows: in, is the weight of the kth leaf node; I(·) is the indicator function. The function value is 1 if it exists, otherwise it is 0. represents the number of intermediate nodes, and The loss function for building the above model is defined as follows: Among them, λ is the regularization term coefficient; Convert it into vector form and transform the loss function into Taking the derivative and setting it to 0, we get: Among them, U model Represents the model input feature vector; First, calculate the correlation coefficients among primary air volume, secondary air volume, average speed of feeder and furnace temperature, and take the primary air with lower correlation coefficient as disturbance quantity; The jth manipulated variable Where (j=1,…,4); The original SNA-PID controller needs to convert the error signal into a state, and its input is: Where, e(k) represents the error value; and FT They represent the furnace temperature setting value and the model output value respectively; k represents the number of iterations; The input x3 of the SNA-PID controller is improved to reduce output oscillation as follows: z3(k)=[e(k)-e(k-1)] 2 (10) Therefore, the ISNA-PID controller output for the jth manipulated variable is expressed as: in, represents the increment of the jth operation variable; represents the neuron gain coefficient of the jth manipulated variable; represents the i-th neuron weight coefficient of the j-th operating variable at time k, expressed as: From equations (18) and (19), it can be seen that the SNA-PID controller tracks the set value under the supervised Hebb learning algorithm by continuously adjusting the neuron weight coefficient online, as follows: Based on formula (20), the error increment is repeatedly used to update the weighted coefficient. The modified weighted coefficient update formula is as follows: in, and Respectively represent the weight coefficients of the proportional, integral and differential neurons in the jth manipulated variable, and Respectively represent the learning coefficients of the proportional, integral and differential neurons in the jth manipulated variable; Optimizing furnace temperature settings 1) NOx and CO2 index model based on CART tree Here, only the effect of furnace temperature on NOx and CO2 is considered. To increase the interpretability of the model, a single-input and single-output index model is constructed using CART tree; The NOx modeling process is as follows: First, take the nth sel The value corresponding to the sample Cut as the split point and divide the input feature space into two regions and as follows: Based on the following criteria, the optimal segmentation point is found by traversal; in, and Respectively represent and The true value of the samples in the region; and Respectively represent and The average value of the true value of the samples in the region; Then, repeat the above process to grow until the number of leaf node samples is less than the empirically set threshold θ NOx , take one tenth of the sample size as the threshold; Finally, the input feature space is divided into K NOx area and marked as Furthermore, the NOx model constructed based on the CART tree is as follows: in, Indicates area The number of samples included in ; Indicates area Middle The true value of the sample; I(·) is the indicator function, when The function value is 1 if it exists, otherwise it is 0; The CO2 modeling process is as follows: First, take the nth sel The value corresponding to the sample Cut as the split point and divide the input feature space into two regions and as follows: Based on the following criteria, the optimal segmentation point is found by traversal; in, and Respectively represent and The true value of the samples in the region; and Respectively represent and The average value of the true value of the samples in the region; Then, repeat the above process to grow until the number of leaf node samples is less than the empirically set threshold One tenth of the sample size is taken as the threshold; Finally, the input feature space is divided into area and marked as Furthermore, the CO2 model constructed based on the CART tree is as follows: in, Indicates area The number of samples included in ; Indicates area Middle The true value of samples; represents the empirically set threshold; I(·) is the indicator function. The function value is 1 if it exists, otherwise it is 0; 2) Multi-objective optimization model The multi-objective model for optimizing furnace temperature setting with NOx and CO2 emission concentrations as optimization targets is as follows: in, represents the comprehensive emission concentration of pollutants; w Pollutant represents the weight vector; γ Pollutant Represents the pollutant emission concentration output vector; ω NOx , Respectively represent the weights of NOx and CO2 emission concentrations; To minimize the objective function γ mix , based on the single-input and single-output NOx and CO2 indicator model established based on the CART tree, the particle swarm optimization algorithm is used to search for the optimal furnace temperature setting value; The particle speed and position update formula is: Where t represents the number of iterations; ω represents the inertia factor; r1 and r2 are random numbers between (0,1); c1 and c2 represent learning factors, with a value of 2; x i (t), pbest i (t) and gbest(t) represent the current position, individual optimality and global optimality of the i-th particle at the t-th iteration, respectively; Since the furnace temperature setting is usually an integer, the speed update formula is rounded as follows: Where q represents the particle number.

Citation Information

Patent Citations

  • Boiler heating surface wall temperature protection method and device

    CN109695892A

  • Boiler exhaust emission prediction model establishing method, prediction method and device

    CN111144609A

  • Hearth temperature control method based on event triggering in RBF-PID urban solid waste incineration process

    CN114296489A

  • Grate furnace MSWI process dioxin emission soft measurement method based on residual error fitting mechanism simplified deep forest regression

    CN114564894A

  • MSWI process combustion state identification method based on multi-feature fusion and improved cascade forest

    CN114882391A