Flue gas oxygen content prediction method in MSWI process

Abstract

The invention provides a method for predicting the oxygen content of flue gas in the MSWI process, and relates to the technical field of urban solid waste incineration. Comprising the following steps: constructing a final flue gas oxygen content prediction model; inputting manipulated variables into the final flue gas oxygen content prediction model to obtain the flue gas oxygen content; a construction method of the final flue gas oxygen content prediction model comprises the following steps: constructing a flue gas oxygen content main model according to EnTSFRT; on the basis of LRDT, a compensation model is constructed according to the error value of the flue gas oxygen content main model; constructing an initial flue gas oxygen content prediction model according to the compensation model and the flue gas oxygen content main model; and optimizing hyper-parameters of the initial flue gas oxygen content prediction model by using a BO algorithm to obtain a final flue gas oxygen content prediction model. The method solves the problems that in the prior art, prediction of the flue gas oxygen content in the MSWI process is poor in interpretability, low in model precision, complex in manual parameter adjustment and the like.

Description

Technical Field

[0001] The invention relates to the technical field of municipal solid waste incineration, and in particular to a method for predicting the oxygen content of flue gas in a MSWI process. Background Art

[0002] Flue gas oxygen content is one of the important parameters in the municipal solid waste incineration (MSWI) process. Its stable control is closely related to the incineration efficiency and pollutant emissions of the MSWI process. However, existing methods have problems such as poor interpretability, low model accuracy and complex manual parameter adjustment. Summary of the invention

[0003] In order to overcome the shortcomings of the prior art, the purpose of the present invention is to provide a method for predicting the oxygen content of flue gas in the MSWI process. The present invention solves the problems in the prior art of poor interpretability of the prediction of oxygen content in flue gas in the MSWI process, low model accuracy and complex manual parameter adjustment.

[0004] To achieve the above object, the present invention provides the following solutions:

[0005] A method for predicting oxygen content in flue gas during a MSWI process, comprising:

[0006] Construct a prediction model for the final flue gas oxygen content;

[0007] Inputting manipulated variables into the final flue gas oxygen content prediction model to obtain the flue gas oxygen content;

[0008] The method for constructing the final flue gas oxygen content prediction model is:

[0009] Construct the main model of flue gas oxygen content based on EnTSFRT;

[0010] Based on the LRDT, a compensation model is constructed according to the error value of the main model of flue gas oxygen content;

[0011] Constructing an initial flue gas oxygen content prediction model according to the compensation model and the flue gas oxygen content main model;

[0012] The BO algorithm is used to optimize the parameters of the initial flue gas oxygen content prediction model to obtain a final flue gas oxygen content prediction model.

[0013] Preferably, the constructing of the main model of flue gas oxygen content according to EnTSFRT comprises:

[0014] Construct multiple TSFRT sub-models;

[0015] Integrating the multiple TSFRT sub-models in parallel to obtain EnTSFRT;

[0016] The main model of flue gas oxygen content is constructed based on EnTSFRT.

[0017] Preferably, the expression of the main model of flue gas oxygen content is:

[0018]

[0019] in, is the output value of the main model of flue gas oxygen content, F is the EnTSFRT model, is the weight matrix of TSFRT in the EnTSFRT model, U, K, θ leaf is the minimum number of samples, l is the width matrix of the membership function, δ is the center matrix of the membership function, ω is the consequent weight parameter matrix,

[0020] Preferably, the optimization of the parameters of the initial flue gas oxygen content prediction model using the BO algorithm includes:

[0021] Construct the final optimization model;

[0022] Inputting a hyperparameter set into the final optimization model to obtain an optimized hyperparameter set;

[0023] The construction method of the optimization model is:

[0024] Set the boundary space and construct the initial data set in the set space;

[0025] Constructing an initial optimization model according to the initial data set;

[0026] The parameters of the initial optimization model are adjusted using maximum likelihood estimation to obtain a final optimization model.

[0027] Preferably, the final flue gas oxygen content prediction model is expressed as:

[0028]

[0029] in, is the output value of the final flue gas oxygen content prediction model, and is the error output value of the compensation model.

[0030] Preferably, the hyperparameters include:

[0031] Output parameters, minimum number of samples and number of decision trees of the main model of flue gas oxygen content and minimum number of samples and regular term coefficients of the compensation model.

[0032] The present invention discloses the following technical effects:

[0033] The present invention provides a method for predicting flue gas oxygen content in a MSWI process, including: constructing a final flue gas oxygen content prediction model; inputting manipulated variables into the final flue gas oxygen content prediction model to obtain flue gas oxygen content; the method for constructing the final flue gas oxygen content prediction model is: constructing a main flue gas oxygen content model according to EnTSFRT; constructing a compensation model according to the error value of the main flue gas oxygen content model based on LRDT; constructing an initial flue gas oxygen content prediction model according to the compensation model and the main flue gas oxygen content model; optimizing the parameters of the initial flue gas oxygen content prediction model using the BO algorithm to obtain the final flue gas oxygen content prediction model. The present invention improves the interpretability of the model based on the main-complementary integrated modeling of the tree algorithm; proposes a hyperparameter selection method for the main-complementary tree integrated model based on the BO algorithm to reduce the complexity and time-consuming of manual parameter adjustment. The model performance is enhanced; the BO optimized main-complementary tree integrated model for flue gas oxygen content is constructed for the first time, providing support for related intelligent control research. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0035] Figure 1 A flow chart of a method for predicting oxygen content in flue gas during a MSWI process provided by an embodiment of the present invention;

[0036] Figure 2 A schematic diagram of an overall modeling strategy provided by an embodiment of the present invention;

[0037] Figure 3 A schematic diagram of a master model modeling strategy based on EnTSFRT provided in an embodiment of the present invention;

[0038] Figure 4 A schematic diagram of a compensation model modeling strategy based on LRDT provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0040] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0041] like Figure 1 As shown, the present invention provides a method for predicting oxygen content in flue gas during a MSWI process, comprising:

[0042] Step 100: constructing a final flue gas oxygen content prediction model;

[0043] Step 200: inputting manipulated variables into the final flue gas oxygen content prediction model to obtain the flue gas oxygen content;

[0044] The method for constructing the final flue gas oxygen content prediction model is:

[0045] Step 101: constructing a main model of flue gas oxygen content according to EnTSFRT;

[0046] Step 102: constructing a compensation model based on LRDT and according to the error value of the main model of flue gas oxygen content;

[0047] Step 103: constructing an initial flue gas oxygen content prediction model according to the compensation model and the flue gas oxygen content main model;

[0048] Step 104: Optimize the parameters of the initial flue gas oxygen content prediction model using the BO algorithm to obtain a final flue gas oxygen content prediction model.

[0049] Further, such as Figure 2 As shown, the expression of the main model of flue gas oxygen content is:

[0050]

[0051] in, is the output value of the main model of flue gas oxygen content, F is the EnTSFRT model, is the weight matrix of TSFRT in the EnTSFRT model, U, K, θ leaf is the minimum number of samples, l is the width matrix of the membership function, δ is the center matrix of the membership function, ω is the consequent weight parameter matrix, U is the input variable matrix, and K represents the total number of fuzzy rules.

[0052] Specifically, Figure 2 In the equation, u represents MV, the manipulated variable, and y represents the true value of flue gas oxygen content. Indicates the flue gas oxygen content output value of the EnTSFRT main model, Represents the error value of the EnTSFRT main model, represents the error output value of the LRDT compensation model, represents the error value of the LRDT compensation model, Indicates the flue gas oxygen content output value of BO-EnTSFRT-LRDT. The relationship between the variables is as follows:

[0053]

[0054]

[0055] Furthermore, the main model of flue gas oxygen content is constructed according to EnTSFRT, including:

[0056] Construct multiple TSFRT sub-models;

[0057] Integrating the multiple TSFRT sub-models in parallel to obtain EnTSFRT;

[0058] The main model of flue gas oxygen content is constructed based on EnTSFRT.

[0059] The EnTSFRT model requires parallel integration of multiple TSFRT sub-models. The TSFRT sub-model mainly includes a screening layer and a fuzzy reasoning layer. The screening layer is used for feature screening, and the fuzzy reasoning layer uses TS fuzzy reasoning. Figure 3 As shown, Figure 3 middle is the input of EnTSFRT, u n represents MV (manipulated variable), n represents the nth input sample, y represents the true value of flue gas oxygen content, N represents the total number of data set samples, M represents the input feature dimension, U j represents the input variable matrix of the jth subset sample, D j represents the jth training subset obtained by random feature sampling of D, and represents the membership degree on the t+1th and t+2th non-leaf nodes, f TSFRT represents the TSFRT model, Indicates the output value of EnTSFRT.

[0060] Furthermore, the features of D are randomly sampled to obtain a training subset D of M features. j , which can be expressed as:

[0061]

[0062] Where D j represents the jth training subset, is the input variable matrix, Represents the true value vector of flue gas oxygen content.

[0063] Correspondingly, from the perspective of input features, u containing M input featuresj It can be expressed as:

[0064]

[0065] The subset {D 1 ,…,D j ,…D Tn}.

[0066] Furthermore, the construction of the j-th TSFRT sub-model is described as an example.

[0067] The dataset D j As the input of the screening layer, and determine the best split position.

[0068] Get D j Each recursive coordinate point (j, m) in , and calculate its mean square error (MSE) value to minimize MSE to determine the segmentation node. The estimation process of the loss function value is as follows:

[0069]

[0070] In the formula, Φ represents the data set D j The loss value, Represents the left subset (or right subset ), Represents the left subset The real vector of Represents the right subset The real vector of Indicates input u j The clear membership function at the tth non-leaf node, Represents the left subset The average of the target values, Represents the right subset The average value of the target value. and Also expressed as The calculation is as follows:

[0071]

[0072] In the formula, express (or )

[0073] In formula (5) is calculated as:

[0074]

[0075] In the formula, μ cs (u j ) indicates input u j The clear membership function of Represents sample u j The split operation on the tth non-leaf node, φ t The segmentation threshold can be expressed as φ t =u j,m Assume that the TSFRT model of the present invention consists of T nodes, so the number of non-leaf nodes is The membership set sum is

[0076] D j Divide into left and right subsets as follows:

[0077]

[0078] In the formula, Indicates that the clear set is obtained using the minimum MSE The first membership degree of is defined by equation (7), and the first element of the clear set is described as:

[0079]

[0080] Repeat the above steps until the minimum number of samples of the leaf node is reached Reaching the set threshold, A clear set can be expressed as:

[0081]

[0082] It can be further expressed as:

[0083]

[0084] according to Determine the training data for the node And construct TS fuzzy reasoning, It is expressed as follows:

[0085]

[0086] In the formula, is the training data set for fuzzy reasoning, Indicates leaf nodes, Indicates The number of samples in a leaf node.

[0087] In the fuzzy reasoning layer, K rules are defined to represent the local linear relationship between the input variables and the target, and the kth rule is defined as follows:

[0088]

[0089] In the formula, y k represents the output of the fuzzy rule consequence, that is, the weight of the kth fuzzy rule. (13) is further simplified as follows:

[0090]

[0091] in, express The variables in express The membership function describes right The degree of membership.

[0092] The present invention adopts Gaussian function as the membership function, as follows:

[0093]

[0094] In the formula, δ m,k represents the center of the membership function, Represents the width of the membership function.

[0095] Calculate the activation strength of the kth fuzzy rule for m input variables as follows:

[0096]

[0097] In the formula, α k represents the product output of the kth fuzzy rule, Represents the membership degree of the kth fuzzy rule.

[0098] Further normalization operation:

[0099]

[0100] According to (13) and (17), y k and Perform linear combination and accumulation to obtain the predicted output value as follows:

[0101]

[0102] Among them, g(·) represents the consequent output function, and ω represents the consequent weight parameter.

[0103] Repeat the above steps until you get leaf nodes, by obtaining The fuzzy inference constructs a multi-input single-output model, and the simplified TSFRT model with multiple inputs and single output is as follows:

[0104]

[0105] In the formula, represents the jth TSFRT model, U={U 1 ,…,U j ,…,U Tn}, θ leaf represents the minimum number of samples, ω={ω 1 ,ω 2 ,…,ω m} represents the consequent weight parameter matrix, δ represents the center matrix of the membership function, l represents the width matrix of the membership function, and j represents the jth TSFRT model.

[0106] In addition, the present invention adopts the least square method to update the subsequent weight.

[0107] Expand (18) and merge them as follows:

[0108]

[0109] In the formula,

[0110] Given the input matrix and the output matrix, the TS consequent weights can be estimated as follows:

[0111] ω=((U * ) T U * ) -1 (U * ) T y (21)

[0112] Where U * by u * constitute.

[0113] The least squares method is used for parameter learning, and the recursive calculation is as follows:

[0114]

[0115]

[0116] In the formula, ω 0 is the weight of the subsequent element at the initial moment, is a random vector, i represents the i-th recursive process, H i is the intermediate variable in the recursive process, H 0 Defined as H 0 =γI, λ represents a large positive value, and I represents the unit matrix.

[0117] Furthermore, first, according to the TSFRT modeling process, Tn TSFRT models are constructed. Defined as

[0118] Then, in order to estimate the weight with the minimum training error, the present invention adopts the following optimal problem to calculate the pseudo-inverse:

[0119]

[0120] Where ν represents the regularization term coefficient.

[0121] The weight matrix is ​​calculated using the ridge regression theory, that is, the Moore-Penrose inverse matrix. The calculation results are as follows:

[0122]

[0123] Finally, the output of EnTSFRT is as follows:

[0124]

[0125] Further, the compensation model based on LRDT (LRDT-based compensation model modeling module):

[0126] This module consists of two parts: feature selection and linear regression. The structure is as follows Figure 4 As shown:

[0127] Figure 4 The feature selection module performs feature screening to determine the input features of the leaf nodes, where the input feature of the sth leaf node is The linear regression module uses the linear regression method to calculate the predicted output of the leaf node, as shown in formula (30), where samples represents the number of node samples. represents the indicator function, Represents the node judgement of the first non-leaf node, represents the input feature of the sth leaf node, express The number of samples, express The dimension of represents the weight vector, and J(·) represents the cost function.

[0128] Furthermore, first, given the data set Represents the true error value of the EnTSFRT master model.

[0129] D E Divide into left and right subsets, as follows:

[0130]

[0131] The difference from (8) is that in (27) represents the indicator function, which is expressed as follows:

[0132]

[0133] According to (5), the node judger can be determined, and then the coordinates of the optimal segmentation position can be determined.

[0134] There are T paths from the first non-leaf node to the leaf node. The input features of the leaf node are determined according to the nodes on the path, as follows:

[0135]

[0136] Further, linear regression module: This module uses linear regression method to calculate the predicted output of leaf nodes, as follows:

[0137]

[0138] In order to ensure that the weight vector does not diverge during the flue gas oxygen content modeling process, the present invention adopts the Tikhonov regularized least squares cost function The method to solve the weight vector is as follows:

[0139]

[0140] Relative to The gradient of can be expressed as follows:

[0141]

[0142] make Available as follows:

[0143]

[0144] Finally, (30) can be used to calculate the output value of the leaf node.

[0145] Furthermore, the optimization of the parameters of the initial flue gas oxygen content prediction model using the BO algorithm includes:

[0146] Construct the final optimization model;

[0147] Inputting a hyperparameter set into the final optimization model to obtain an optimized hyperparameter set;

[0148] The construction method of the optimization model is:

[0149] Set the boundary space and construct the initial data set in the set space;

[0150] Constructing an initial optimization model according to the initial data set;

[0151] The parameters of the initial optimization model are adjusted using maximum likelihood estimation to obtain a final optimization model.

[0152] Specifically, BO is a type of optimization method based on machine learning, which aims to optimize the objective function, that is, to minimize the objective function. The present invention uses the BO algorithm to optimize the RMSE value of the five hyperparameters in the main-supplement tree ensemble model, among which LSM α , MinSamples1 and Tn are the main model hyperparameters, and MinSamples2 and β are the compensation model hyperparameters.

[0153] The optimization process of the BO algorithm is as follows:

[0154] Step 1: Set the boundary space;

[0155] Set the input sample of BO The value range of includes the upper boundary (Upperboundary, UB) and lower boundary (Lowerboundary, LB) of the sample space, and D represents the number of sample features, which is set as follows:

[0156]

[0157] Among them, {x 1 ;L;x 5} respectively represent as follows:

[0158]

[0159] Step 2: Construct the initial data set;

[0160] Randomly initialize P init Input samples And observe the objective function value corresponding to the sample Construct the initial data set.

[0161] Step 3: Build / train the GPR model;

[0162] The Gaussian process regression (GPR) model is constructed using the data set. The GPR model takes the known data as a prior probability distribution that satisfies the multivariate normal distribution, which is described as follows:

[0163] Γ~N(ζ 0 ,Σ 0 )(36)

[0164] Among them,0 =[ζ 0 (x 1 ),L,ζ 0 (x p )] is the mean function, p is the number of samples in the current data set, Σ 0 is the covariance matrix, expressed as follows:

[0165]

[0166] Among them, σ 2 is the variance of the measurement noise, Σ 0 (x a ,x b ) is the kernel function between the a-th and b-th sample points, calculated as follows:

[0167]

[0168] Among them, x ao represents the o-th dimension feature of the a-th sample, σ Γ is the kernel amplitude of the kernel function, η o Represents the length ratio of the o-th dimension of the sample.

[0169] For a new sample point in the sample space, the posterior probability distribution of the sample point can be calculated through the prior probability distribution. According to the definition of GP, the joint Gaussian distribution Γ of the known data and the predicted value It also obeys the Gaussian distribution and can be expressed as:

[0170]

[0171] in, is the covariance of the training and testing sets, is the test set covariance.

[0172] Since the training set is known, the given Γ can be calculated based on (39): The conditional distribution of is as follows:

[0173]

[0174] Among them, p and as follows:

[0175]

[0176] Among them, p is the posterior mean, is the posterior variance.

[0177] Step 4: Estimate the parameters of the GPR model (optimization model);

[0178] Estimate the hyperparameters in the mean function and covariance matrix. Use Maximum Likelihood Estimation (MLE) to estimate the above parameters to improve the model's fit and prediction performance.

[0179] Step 5: Sampling;

[0180] To find the optimal point x op So that the objective function value Γ(x op ) is minimized and the expected improvement (EI) sampling function is used to sample the objective function.

[0181] Step 6: Update the dataset;

[0182] Calculate the new sample point x * The corresponding objective function value Γ(x * ), and update the data set X = {x 1 ;L;x p ;x *} and Γ=[Γ(x 1 ),L,Γ(x p ),Γ(x * )] T .

[0183] Repeat steps 3-6 until the preset maximum number of iterations is reached, and find the optimal sample point x that meets the requirements from the data set. op .

[0184] Furthermore, Loop represents the number of optimization iterations, Ω represents the EI sampling function, ratio represents the exploration ratio, and q represents the qth optimization iteration number. Indicates that when the input of BO is x p The output value corresponding to the mixed model.

[0185] Furthermore, the pseudo code of the algorithm used in the present invention is shown in Table 1. Table 1 is as follows:

[0186] Table 1 Pseudo code of BO-EnTSFRT-LRDT algorithm

[0187]

[0188]

[0189]

[0190] Furthermore, this experiment uses the operating data of an actual MSWI plant in Beijing on a certain day. The data set contains a total of 857 samples, which, after empirical processing, include four MVs (primary air volume, secondary air volume, feeder average speed, drying grate average speed) and one controlled variable (flue gas oxygen content).

[0191] The evaluation indicators used in the present invention are RMSE and mean absolute error (MAE), as follows:

[0192]

[0193] Set the BO parameter to Loop = 20, Ω represents the sampling function, the present invention uses the EI function, and the ratio is selected as 0.5, and the optimal hyperparameter is LSM α =1.7504, MinSamples1=42, Tn=200, MinSamples2=30 and β=0.30434.

[0194] In order to further illustrate the effectiveness of the method proposed in the present invention, the method proposed in the present invention is compared with random forest (RF), LRDT, EnTSFRT-LRDT, and BO-EnTSFRT. The parameters of each model are set as follows: in the modeling process of Random Forest, the minimum number of samples is set to 50 and the number of features is set to 4; in the modeling process of LRDT, the minimum number of samples is set to 9 and the regularization coefficient is set to 0.4; in the modeling process of EnTSFRT-LRDT, the LSM is set to α =2.4, MinSamples1=2, Tn=132, MinSamples2=9, β=0.5; During the BO-EnTSFRT modeling process, set LSM α =1.2266, MinSamples1=66, Tn=137.

[0195] Based on the above model building, the test was conducted on the test set and repeated 20 times. The calculation results of the performance evaluation indicators of different methods are shown in Table 2. Table 2 is as follows:

[0196] Table 2 Statistical results of performance evaluation indicators of different methods

[0197]

[0198] As shown in Table 2, compared with the BO-EnTSFRT method, the RMSE and MAE values ​​of the method proposed in the present invention on the training set, validation set and test set are smaller, which shows the effectiveness of the error modeling module based on LRDT; compared with the EnTSFRT-LRDT method, the RMSE and MAE values ​​of the method proposed in the present invention on the validation set and test set are smaller, which shows the feasibility of selecting the optimal parameters using the BO algorithm. In addition, by comparing the method proposed in the present invention with LRDT and RF respectively, it can be seen that the RMSE and MAE of the method proposed in the present invention on the validation set and test set are smaller than those of the LRDT method, and the RMSE and MAE on the training set, validation set and test set are smaller than those of the RF method.

[0199] The above comparative experiments illustrate the effectiveness and feasibility of the LRDT module and the BO method of the present invention, and also prove the accuracy and effectiveness of the method proposed in the present invention.

[0200] The present invention uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only used to help understand the method and core ideas of the present invention. At the same time, for those skilled in the art, according to the ideas of the present invention, there will be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting oxygen content in flue gas during MSWI process, characterized in that: include: Construct a prediction model for the final flue gas oxygen content; Inputting manipulated variables into the final flue gas oxygen content prediction model to obtain the flue gas oxygen content; The method for constructing the final flue gas oxygen content prediction model is: Construct the main model of flue gas oxygen content based on EnTSFRT; Based on the LRDT, a compensation model is constructed according to the error value of the main model of flue gas oxygen content; Constructing an initial flue gas oxygen content prediction model according to the compensation model and the flue gas oxygen content main model; The BO algorithm is used to optimize the hyperparameters of the initial flue gas oxygen content prediction model to obtain a final flue gas oxygen content prediction model.

2. A method for predicting oxygen content in flue gas during MSWI process according to claim 1, characterized in that: The main model of flue gas oxygen content is constructed according to EnTSFRT, including: Construct multiple TSFRT sub-models; Integrating the multiple TSFRT sub-models in parallel to obtain EnTSFRT; The main model of flue gas oxygen content is constructed based on EnTSFRT.

3. The method for predicting oxygen content in flue gas during MSWI process according to claim 1, characterized in that: The expression of the main model of flue gas oxygen content is: in, is the output value of the main model of flue gas oxygen content, F is the EnTSFRT model, is the weight matrix of TSFRT in the EnTSFRT model, U, K, θ leaf is the minimum number of samples, l is the width matrix of the membership function, δ is the center matrix of the membership function, ω is the consequent weight parameter matrix, U is the input variable matrix, and K represents the total number of fuzzy rules.

4. The method for predicting oxygen content in flue gas during MSWI process according to claim 1, characterized in that: The optimization of the parameters of the initial flue gas oxygen content prediction model using the BO algorithm includes: Construct the final optimization model; Inputting a hyperparameter set into the final optimization model to obtain an optimized hyperparameter set; The construction method of the optimization model is: Set the boundary space and construct the initial data set in the set space; Constructing an initial optimization model according to the initial data set; The parameters of the initial optimization model are adjusted using maximum likelihood estimation to obtain a final optimization model.

5. The method for predicting oxygen content in flue gas during MSWI process according to claim 3, characterized in that: The final flue gas oxygen content prediction model expression is: in, is the output value of the final flue gas oxygen content prediction model, is the error output value of the compensation model.

6. The method for predicting oxygen content in flue gas during MSWI process according to claim 3, characterized in that: The hyperparameters include: Output parameters, minimum number of samples and number of decision trees of the main model of flue gas oxygen content and minimum number of samples and regular term coefficients of the compensation model.

Images