A Soft Measurement Method For Dioxin Emission Of Grate Furnace MSWI Process Based On Simplified Deep Forest Regression Of Residual Fitting Mechanism

Abstract

The invention provides a soft measurement method for dioxin emission of grate furnace MSWI process based on simplified deep forest regression of residual fitting mechanism. The highly toxic pollutant dioxin (DXN) generated in the solid waste incineration process is a key environmental index which must be subjected to control. The rapid and accurate soft measurement of the DXN emission concentration is an urgent affair for reducing the emission control of the pollutants. The method comprises the following steps: firstly, carrying out feature selection on a high-dimensional process variable by adopting mutual information and significance test; then, constructing a simplified deep forest regression (SDFR) algorithm to learn a nonlinear relationship between the selected process variable and the DXN emission concentration; and finally, designing a gradient enhancement strategy based on a residual error fitting (REF) mechanism to improve the generalization performance of a layer-by-layer learning process. The method is superior to other methods in the aspects of prediction precision and time consumption.

Description

TECHNICAL FIELD

The invention belongs to the field of solid waste incineration.

BACKGROUND

Municipal solid waste (MSW) treatment aims to achieve harmlessness, reduction and resource utilization, of which MSW incineration (MSWI) is currently the main method. However, MSWI process is also one of the main industrial processes currently emitting dioxins (DXN), a highly toxic organic pollutant, accounting for approximately 9% of total emissions. MSWI mainly uses technologies such as grate furnaces, fluidized beds and rotary kilns, among which grate furnace technology accounts for the largest proportion. The optimized operation of the MSWI process based on the grate furnace has an important contribution to the reduction of DXN emissions. Therefore, it is necessary to conduct high-precision real-time detection of DXN emission concentration.

Data-driven soft measurement technology can effectively solve the above problems, that is, using machine learning or deep learning methods to characterize the correlation between easily measurable process variables and DXN emission concentrations. This usually requires determining a mapping function to predict DXN emission concentrations. For example, genetic programming is combined with neural network (NN) to model DXN emissions, but it is not suitable for different types of incineration plants; the design is based on back-propagation NN (BPNN).), but its portability is not good, and BPNN has serious over-fitting problems when facing small sample problems; it adopts selective integration and evaluation variable projection importance strategies, and uses support vector machines and The nuclear latent structure mapping algorithm selects valuable process variables to construct the DXN soft sensor model, but it cannot represent depth features.

Based on 12 years of DXN data of an 800-ton grate furnace, a simplified deep forest regression (SDFR) method (SDFR-ref) with high accuracy and short time-consuming residual fitting mechanism was proposed. The main innovations of this article include: using decision trees to replace complex forest algorithms, thereby reducing the size of the deep forest ensemble model; using a residual fitting strategy with learning factors between cascade layers to give the model higher predictive performance; Mutual information (MI) and significance test (ST) are used for feature selection to simplify the input of the soft sensor model. In China incineration is the main MSW treatment method, and its typical process is shown in FIG. 1.

As shown in FIG. 1, the MSWI process flow based on the grate furnace includes six stages: solid waste storage and transportation, solid waste incineration, waste heat boiler, steam power generation, flue gas purification and flue gas emission. At present, MSWI factories are mainly concentrated in coastal areas, and more than 90% of them use grate furnaces. The grate-type MSWI process has the advantages of large daily processing capacity, stable operation, and low DXN emission concentration. The detection of DXN emission concentration in this article is aimed at the “smoke G3” position in the flue gas emission stage.

The MSWI plant studied in this article was ignited and put into operation in 2009.

From 2009 to 2013, the emission level of DXN was not higher than China's environmental emission standard (GB18485 2001), which is Ing I-TEQ/Nm³ (oxygen content is 11%). Correspondingly, the number of DXN detections increased year by year, and finally stabilized at 4 times/year. Since 2014, my country has revised the emission limit of DXN (updated from Ing I-TEQ/Nm³ to 0.1 ng I-TEQ/Nm³). Obviously, increasingly stringent emission restrictions have led to a gradual increase in the number of DXN tests by enterprises and governments, and the operating costs of enterprises have also increased accordingly.

SUMMARY

The invention aims to explore how to use MSWI process data and limited DXN detection data to establish a DXN soft measurement model to provide key indicator data for MSWI companies' for their DXN emission reduction optimization control and cost reduction.

The invention proposes a modeling strategy based on feature selection and SDFR-ref. The structure is shown in FIG. 2.

As can be seen from FIG. 2, the proposed modeling strategy includes a feature selection module based on MI and ST and a SDFR module based on the residual fitting mechanism. Feature selection module selects the corresponding features by calculating the MI value and ST value of each feature; for SDFR module, Layer-k represents the k-th layer model, ŷ₁^Regvoc represents the output vector of the first layer model, v₁^Augfea represents the augmented regression vector of the second layer input, {circumflex over (y)}_k^Regvoc represents the average value of y_k^Regvoc, α is the remaining learning rate between each layer; x and X^Xsel respectively represents the process data before and after feature selection; y, ŷ and e are the true value, predicted value and prediction error respectively.

In addition, {δ_MI, δ_SL, θ, T, α, K} represents the learning parameter set of the proposed SDFR-ref, where: δ_MI represents the threshold of MI, δ_SL represents the threshold of significance level, and θ represents the minimum sample in the leaf node number, T represents the number of decision trees in each layer of the model, a is the learning rate in the gradient boosting process, and K represents the number of layers. The globally optimized selection of these learning parameters can improve the synergy between different modules, thereby improving the overall performance of the model. Therefore, the proposed modeling strategy can be formulated as solving the following optimization problem:

$\begin{array}{cc}\min \mathrm{RMSE}\left(F^{\mathrm{SDFR}-\mathrm{ref}}(·)\right)=& \left(1\right)\end{array}$ ${\sqrt{\frac{1}{N}\sum _{n=1}^N((\frac{1}{N}\sum _{n=1}^N{y}_n+\alpha \frac{1}{T}\sum _{k=1}^K\sum _{t=1}^T\left[c_{1,l}^{\mathrm{CART}},\dots ,c_{T,l}^{\mathrm{CART}}\right]}{I}_{R^{M\times N}}\left(X^{\mathrm{Xse}l}\right))-y_n)}^2$ $s.t.\{\begin{array}{c}{X}^{Xse1}=f_{\mathrm{FeaSe}l}\left(D,{\delta }_{MI},{\delta }_{\mathrm{SL}}\right)\\ 0<\alpha \le 2\\ 1\le T\le 500\\ 1\le \theta \le N\\ 1\le K\le 20\\ 0\le {\delta }_{\mathrm{MI}}\le 1\\ 0\le {\delta }_{\mathrm{SL}}\le 1\end{array}$

Among them, F^SDFR-ref(⋅) represents the SDFR-ref model; f_FeaSel(⋅) represents the nonlinear feature selection algorithm proposed in this article; N represents the number of modeling samples; y_n represents the n-th true value; c_1,l^CART represents predicted value of l-th leaf node of first CART, C_T,l^CART represents the predicted value of the l-th leaf node of T-th CART; D={X, y|X∈R^N×M, y∈R^N×1} represents the original modeling data, which is also represents the input of the feature selection algorithm, M is the number of original features; I_RM×N(X^Xsel) is the indicator function, when X^Xsel ∈R^M×N, I_RM×N(X^sel)=1, when X^Xsel ∈R^M×N, I_RM×N(X^Xsel)=0.

4.1 Feature Selection Based on MI and ST

MI and ST are used to calculate the information correlation between the original features (process variables) DXN values, and achieve the best selection of features through preset thresholds.

For the input data set, the nonlinear feature selection algorithm f_FeaSel(⋅) proposed in the invention is defined as follows:

$\begin{array}{cc}{D}^{\mathrm{Sel}}=f_{\mathrm{FeaSel}}\left(D,{\delta }_{MI},{\delta }_{SL}\right)& \left(2\right)\end{array}$

Among them, D^Sel={X^Sel, y|X∈R^N×MSel, y∈R^N×1} respectively represent the output of the proposed feature selection algorithm, and M^Sel is the number of selected features.

In fact, MI does not need to assume the potential joint distribution of the data. MI provides an information quantification measure of the degree of statistical dependence between random variables, and estimates the degree of interdependence between two random variables to express shared information. The calculation process is as follows:

$\begin{array}{cc}{I}_i^{\mathrm{MI}}\left(x_i,y\right)=\sum _{x_{n,i}}\sum _{y_n}p\left(x_{n,i},y_n\right){\log }_2\frac{p\left(x_{n,i},y_n\right)}{p\left(x_{n,i}\right),p\left(y_n\right)}& \left(3\right)\end{array}$

Among them, x_i is the i-th eigenvector of x, x_n,i is the n-th value of the i-th eigenvector, y represents the joint probability density; p(x_n,i) and p(y_n) represent the marginal probability density of x_n,i and y_n.

If the MI value of a feature is greater than the threshold δ_MI, it is regarded as an important feature constituting the preliminary feature set X^MI. Furthermore, ST is used to analyze the correlation between the selected features based on MI and remove collinear features.

The Pearson coefficient value PCoe between the selected features x_i^MI and x_j^MI is calculated as follows:

$\begin{array}{cc}\mathrm{PCoe}=\frac{\sum _{n=1}^N\left(x_{n,i}^{\mathrm{MI}}-{\overline{x}}_i^{\mathrm{MI}}\right)\left(x_{n,j}^{\mathrm{MI}}-{\overline{x}}_j^{\mathrm{MI}}\right)}{{\left\{\sum _{n=1}^N{\left(x_{n,j}^{\mathrm{MI}}-{\overline{x}}_i^{\mathrm{MI}}\right)}^2\sum _{n=1}^N{\left(x_{n,j}^{\mathrm{MI}}-{\overline{x}}_j^{\mathrm{MI}}\right)}^2\right\}}^{1/2}}& \left(4\right)\end{array}$

Among them, x_i^MI and x_j^MI represent the average value of x_i^MI and x_i^MI respectively, x_n,i^MI and x_n,j^MI represent the n-th value of x_i^MI and x_j^MI. Z-test is used to calculate the Z_test value between features x_i^MI and x_j^MI:

$\begin{array}{cc}{z}_{\mathrm{test}}=\frac{{\stackrel{\_}{x}}_i^{\mathrm{MI}}-{\stackrel{\_}{x}}_j^{\mathrm{MI}}}{\sqrt{S_i^2/N_i+S_j^2/N_j}}& \left(5\right)\end{array}$

Among them, S_i and S_j represent the standard deviation of x_i^MI and x_j^MI; N_i and N_j represent the number of samples of x_i^MI and x_j^MI.

Furthermore, the p-value is obtained by looking up the Z_test value in the table. At this point, we assume in H₀ that there is no linear relationship between the i-th and j-th features, and the Pearson coefficient PCoe is regarded as the alternative hypothesis H₁. Based on the comparison of p-value and significance level δ_SL, the final selected X^Xsel including the preferred features is determined. The criteria are expressed as follows:

$\begin{array}{cc}\{\begin{array}{cc}\begin{array}{c}\mathrm{Accept}{H}_1(\mathrm{linearly}\mathrm{dependent}),\\ \mathrm{reject}{H}_0\left(\mathrm{linearly}\mathrm{independent}\right)\end{array}& p-\mathrm{value}<{\delta }_{\mathrm{SL}}\\ \begin{array}{c}\mathrm{Accept}{H}_0(\mathrm{linearly}\mathrm{independent}),\\ \mathrm{reject}{H}_1\left(\mathrm{linearly}\mathrm{dependent}\right)\end{array}& p-\mathrm{value}>{\delta }_{\mathrm{SL}}\end{array}& \left(6\right)\end{array}$

Based on the above assumptions, the collinear features selected by MI are removed, thereby reducing the impact of data noise on the training model.

4.2 SDFR (SDFR-Ref) Based on Residual Fitting Mechanism

4.2.1 First Layer Implementation

The training set after feature selection is recorded as D^Sel. The SDFR algorithm replaces the forest algorithm in the original DFR with a decision tree, that is, CART. Each layer contains multiple decision trees, and the tree nodes are divided using the squared error minimization criterion. The minimum loss function of this process is expressed as follows:

$\begin{array}{cc}{\mathrm{Split}}^{\mathrm{CART}}=\min \left[\sum _{x_i^{\mathrm{Xsel}}\in R_{\mathrm{Left}}}{\left(y_{\mathrm{Left}}-c_{\mathrm{Left}}^{\mathrm{CART}}\right)}^2+\sum _{x_i^{X\mathrm{se}l}\in R_{\mathrm{Right}}}{\left(y_{\mathrm{Right}}-c_{\mathrm{Right}}^{\mathrm{CART}}\right)}^2\right]& \left(7\right)\end{array}$

Among them, c_Left^CART and c_Right^CART are the outputs of R_Left and R_Right nodes respectively; y_Left and y_Right represent the true values in R_Left and R_Right nodes respectively.

Specifically, the nodes are determined in the following way:

$\begin{array}{cc}\{\begin{array}{c}{R}_{\mathrm{Left}}\left(j,s\right)=\left\{x^{\mathrm{Sel}}|x_j^{\mathrm{Sel}}\le s\right\}\\ R_{\mathrm{Right}}\left(j,s\right)=\left\{x^{\mathrm{Se}1}|x_j^{\mathrm{Se}1}>s\right\}\end{array}& \left(8\right)\end{array}$

Among them, j and s represent segmentation features and segmentation values respectively; x_j^Sel is the j-th eigenvalue of the selected feature x^Sel. Therefore, CART can be expressed as:

$\begin{array}{cc}{h}_1^{\mathrm{CART}}\left(x^{\mathrm{Sel}}\right)=\sum _{l=1}^L{c}_l^{\mathrm{CART}}{I}_{R_l^{\mathrm{CART}}}\left(x^{\mathrm{Sel}}\right)& \left(9\right)\end{array}$

Among them, L represents the number of CART leaf nodes, c_l^CART represents the output of the l-th leaf node of CART, and I_RlCART(x^Sel) is the indicator function, when x^Sel ∈R_l^CART, I_Rl^CART (x^Sel)=1, when x^Sel∉R_l^CART, I_Rl^CART(x^Sel)=0.

The first-level model containing multiple CARTs is represented as follows:

$\begin{array}{cc}{f}_1^{\mathrm{SDFR}}\left(x^{\mathrm{Se}1}\right)=\frac{1}{T}\sum _{t=1}^T{h}_{1,t}^{\mathrm{CART}}(·)& \left(10\right)\end{array}$

Among them, f₁^SDFR(⋅) represents the first layer model in SDFR, T represents the number of CARTs in each layer model, h_1,t^CART(⋅) represents the t-th CART model in layer 1.

Furthermore, the first-layer regression vector ŷ₁^Regvec from the first-layer model f₁^SDFR(⋅) is expressed as follows:

$\begin{array}{cc}{\widehat{y}}_1^{\mathrm{Regvec}}=\left[h_{1,1}^{\mathrm{CART}}(·),\dots ,h_{1,T}^{\mathrm{CART}}(·)\right]=\left[c_{1,l}^{\mathrm{CART}},\dots ,c_{T,l}^{\mathrm{CART}}\right]& \left(11\right)\end{array}$

Among them, c_1,l^CART represents the predicted value of the l-th leaf node of the first CART, represents the predicted value of the l-th leaf node of the T-th CART.

The augmented regression vector V₁^Augfea is obtained by merging the layer regression vectors ŷ₁^Regvec and is expressed as follows:

$\begin{array}{cc}{v}_1^{\mathrm{Augfea}}=f_{\mathrm{FeaCom}}^1\left({\hat{y}}_1^{\mathrm{Regvec}},x^{\mathrm{Sel}}\right)& \left(12\right)\end{array}$

Among them, f_FeaCom¹(⋅) represents the eigenvector combination function.

v₁^Augfea is then used as the feature input for the next layer. In the invention, the DXN true value is no longer used in subsequent cascade modules, but the new true value is recalculated through the gradient boosting strategy. Therefore, the invention uses the following formula to calculate the loss function of the squared error:

$\begin{array}{cc}{L}_1^{\mathrm{SDFR}}\left(y_n^{\left(1\right)},{f_1^{\mathrm{SDFR}}(·)}_n\right)=\frac{1}{2}\sum _{i=1}^N{\left(y_n^{\left(1\right)}-{f_1^{\mathrm{SDFR}}(·)}_n\right)}^2& \left(13\right)\end{array}$

Among them, L₁^SDFR(⋅) represents the squared error loss function in SDFR-ref; y_n⁽¹⁾ represents the n-th true value of the first layer training set.

The loss function L₁^SDFR is further used to calculate the gradient direction as shown below.

$\begin{array}{cc}{\sigma }_{1,n}^{\mathrm{SDFR}}=-{\left[\frac{\partial L\left(y_n^{\left(1\right)},f_1^{\mathrm{SDFR}}(·)\right)}{\partial f_1^{\mathrm{SDFR}}(·)}\right]}_{f_1^{\mathrm{SDFR}}(·)=f_0^{\mathrm{SDFR}}(·)}& \left(14\right)\end{array}$

Among them, σ_1,n^SDFR is the gradient of the nth true value of layer 1; f₀^SDFR(⋅) represents the arithmetic mean of the initial true value, that is

$f_0^{\mathrm{SDFR}}(·)=\frac{1}{N}\sum _{n=1}^N{y}_n,$

y_n represents the n-th true value.

Then, the objective function is:

$\begin{array}{cc}{f}_1^{\mathrm{SDFR}}\left(x^{\mathrm{Sel}}\right)=f_0^{\mathrm{SDFR}}(·)+\alpha \sum _{t=1}^T\left[c_{1,l}^{\mathrm{CART}},\dots ,c_{T,l}^{\mathrm{CART}}\right]I_R\left(x^{\mathrm{Sel}}\right)& \left(15\right)\end{array}$

Among them, f₁^SDFR(⋅) is the first layer model; α represents the learning rate; I_R(x^Sel) represents when x^Sel ∈R, I_R(x^Sel)=1, when x^Sel∈R, I_R(x^Sel)=0.

Therefore, the true value of the second level is:

$\begin{array}{c}\left(16\right)\end{array}$ $y_2=y-f_0^{\mathrm{SDFR}}(·)-\alpha f_1^{\mathrm{SDFR}}(·)=y_1-\alpha f_1^{\mathrm{SDFR}}(·)=y_1-\alpha {\stackrel{\_}{\hat{y}}}_1^{\mathrm{Regvec}}$

Among them, y₁ is the true value of the first layer model, that is, y₁=y, y is the true value vector of DXN; {circumflex over (y)}₁^Regvec represents the mean value of the first layer regression vector.

4.2.2 k-th Layer Implementation

The training set of the k-th layer based on the augmented regression vector of the (k−1)-th layer is expressed as D_k^Sel={{v_(k−1),n^Augfea}_n=1^N, y_k}, v_(k−1)^Augfea is the augmented regression vector of the (k−1)-th layer, and y_k is the k-th true value.

First, establish the k-th level decision tree h_k^CART(⋅) according to formulas (7) and (8). The k-th level model is expressed as follows:

$\begin{array}{cc}{f}_k^{\mathrm{SDFR}}\left(v_{\left(k-1\right),i}^{\mathrm{Augfea}}\right)=\frac{1}{T}\sum _{t=1}^T{h}_{k,t}^{\mathrm{CART}}(·)& \left(17\right)\end{array}$

Among them, f_k^SDFR(⋅) represents the k-th layer model, and h_k,t^CART(⋅) represents the k-th layer of the t-th CART model.

Then, the augmented regression vector v_k^Augfea of the k-th layer is expressed as follows:

$\begin{array}{cc}{v}_k^{\mathrm{Augfea}}=f_{\mathrm{FeaCom}}^k\left({\hat{y}}_1^{\mathrm{Regvec}},x^{\mathrm{Sel}}\right)& \left(18\right)\end{array}$

Among them, ŷ_k^Regvec represents the regression vector of the k-th layer, that is, ŷ_k^Regvec=[h_k,1^CART(⋅), . . . , h_k,T^CART(⋅)].

Then, calculate the gradient σ_k^SDFR according to formulas (12) and (13). The true value of (k+1)-th layer is expressed as follows:

$\begin{array}{cc}{y}_{k+1}=y_1-\alpha \left({\overline{\widehat{y}}}_1^{\mathrm{Regvec}}+\dots +{\overline{\widehat{y}}}_k^{\mathrm{Regvec}}\right)& \left(19\right)\end{array}$

4.2.3 K-th Layer Implementation

The K-th layer is the last layer of the SDFR-ref training process, that is, the preset maximum number of layers, and its training set is D_K^Sel={{v_(K−1),n^Augfea}_n=1^N,y_K}.

First, build a decision tree model h_K^CART(⋅) through the training set D_K^Sel and further obtain the K-th layer model f_K^SDFR(⋅). Then, calculate the K-th layer regression vector ŷ_K^Regvec according to the input augmented regression vector v_(K-1)^Augfea, which is expressed as follows:

$\begin{array}{cc}{\hat{y}}_K^{\mathrm{Regvec}}=\left[h_{K,1}^{\mathrm{CART}}(·),\dots ,h_{K,T}^{\mathrm{CART}}(·)\right]& \left(20\right)\end{array}$

Among them, h_k,1^CART(⋅) represents the first CART model of the K-th layer, h_K,T^CART(⋅) represents the T-th CART model of the K-th layer.

Finally, the output value after gradient boosting with learning rate α is:

$\begin{array}{cc}{y}_k=y_1-\alpha \sum _{k=1}^{\left(K-1\right)}{\overline{\widehat{y}}}_k^{\mathrm{Regvec}}& \left(21\right)\end{array}$

Among them, {circumflex over (y)}_k^Regvec represents the mean value of the k-th layer regression vector.

4.2.4 Prediction Output Implementation

After multiple layers are superimposed, each layer is used to reduce the residual of the previous layer. Finally, the SDFR-ref model can be expressed as:

$\begin{array}{cc}{F}^{\mathrm{SDFR}-\mathrm{ref}}\left(x^{\mathrm{Sel}}\right)=\sum _{k=1}^K{f}_k^{\mathrm{SDFR}}(·)=\alpha \sum _{k=1}^K\sum _{t=1}^T\left[c_{1,l}^{\mathrm{CART}},\dots ,c_{T,l}^{\mathrm{CART}}\right]I_R\left(x^{\mathrm{Sel}}\right)& \left(22\right)\end{array}$

Among them, I_R(x^Sel) means I_R(x^Sel)=1 when x^Sel∈R, and I_R(x^Sel)=0 when x^Sel∉R.

Since F^SDFR-ref(⋅) is calculated based on addition, the final predicted value cannot be simply averaged. Therefore, it is necessary to first calculate the mean value of the regression vector of each layer. Taking layer 1 as an example, it is as follows:

$\begin{array}{cc}\begin{array}{c}{\hat{y}}_1^{\mathrm{add}}=\frac{1}{T}\sum _{t=1}^T{\hat{y}}_1^{\mathrm{Regvec}}\\ =\frac{1}{T}\sum _{t=1}^T\left[h_1^{\mathrm{CART}}(·),\dots ,h_T^{\mathrm{CART}}(·)\right]\\ =\frac{1}{T}\sum _{t=1}^T\left[c_{1,l}^{\mathrm{CART}},\dots ,c_{T,l}^{\mathrm{CART}}\right]\end{array}& \left(23\right)\end{array}$

Add K predicted values to get the final predicted value, as shown below:

$\begin{array}{cc}\widehat{y}=\frac{1}{N}\sum _{n=1}^N{y}_n+\alpha \frac{1}{T}\sum _{k=1}^K\sum _{t=1}^T\left[c_{1,l}^{\mathrm{CART}},\dots ,c_{T,l}^{\mathrm{CART}}\right]I_{R^{M\times N}}\left(X^{\mathrm{Sel}}\right)& \left(24\right)\end{array}$

Among them, ŷ is the predicted value of SDFR-ref model; means I_R(x^Sel)=1 when x^Sel ∈R, and I_R(x^Sel)=0 when. x^Sel∉R

DESCRIPTION OF DRAWINGS

FIG. 1 is the typical process flow of MSWI based on grate furnace;

FIG. 2 is the modeling strategy proposed in the invention.

EMBODIMENTS

This embodiment uses a real DXN data set to verify the effectiveness of the proposed method. The DXN data comes from the actual MSWI process of an incineration plant in Beijing in the past 12 years, including 141 samples and 116 process variables. The process variables cover the four stages of MSWI, namely solid waste incineration, waste heat boiler, flue gas purification and flue gas emission, and Table 1 shows the detailed information.

TABLE 1
type of the procedure variable
	stage
		waste
	solid waste	heat	flue gas	flue gas
procedure variable	incineration	boiler	treatment	emission
Temperature	42	5	6	/
Velocity	18	/	/	/
Flux	15	5	6	/
Pressure	2	7	/	/
Liquid level	/	1	/	/
Concentration	/	/	1	8
Total	77	18	13	8
	116

The sample sizes of the training, validation and test sets are respectively ½, ¼ and ¼ of the original sample data.

TABLE 2
Abbreviations of procedure variables
Stage	procedure variables	uint	Abbreviations
solid waste	combustion temperature 1	° C.	T1
incineration	combustion temperature 2	° C.	T2
	combustion temperature 3	° C.	T3
	maximum temperature at which a grate burns	° C.	T4
	temperature if the dry grate left inlet	° C.	T5
	temperature if the dry grate right inlet	° C.	T6
	temperature in the left side of the drying and	° C.	T7
	burning sections of inner grate wall
	temperature in the left side of the drying and	° C.	T8
	burning sections of outer grate wall
	temperature in the right side of the drying and	° C.	T9
	burning sections of inner grate wall
	temperature in the right side of the drying and	° C.	T10
	burning sections of outer grate wall
	left inner temperature of combustion grate 1-1	° C.	T11
	left outer temperature of combustion grate 1-1	° C.	T12
	right inner temperature of combustion grate 1-1	° C.	T13
	right outer temperature of combustion grate 1-1	° C.	T14
	left inner temperature of combustion grate 1-2	° C.	T15
	left outer temperature of combustion grate 1-2	° C.	T16
	left inner temperature of combustion grate 2-1	° C.	T17
	left outer temperature of combustion grate 2-1	° C.	T18
	Outlet air temperature of primary air preheater	° C.	T19
	air temperature of the combustion grate inlet	° C.	T20
	temperature of cooling air outlet	° C.	T21
flue gas	temperature of fluidization fan outlet	° C.	T22
purification
solid waste	air flux of the left combustion grate	km3N/h	LAF1
incineration
waste heat	cooling water flux of the secondary superheater	t/h	CWF1
boiler
flue gas	supply flux of urea solvent	L/h	FUS1
treatment	Bag pressure difference	kPa	BP1
flue gas	O2 concentration of CEMS system	%	OC1
purification	Dust concentration of CEMS system	mg/m3N	DC1
	HCL concentration of CEMS system	mg/m3N	HC1
	CO2 concentration of CEMS system	%	CC1

First calculate the MI value between the 116 process variables and the DXN emission concentration. The invention sets the threshold value δ_Xsel of MI=0.75 to ensure that the amount of information between the selected process variable and the DXN emission is as large as possible, the initial number of features selected is 30; Further, the significance level is set δ_SL=0.1 and the final selected process variable are T2, T4, T5, T6, T7, T9, T10, T16, T20, T21, LAF1, FUS1, DC1 and CC1, 14 in total. The linear correlation between the selected process variables is weak, which demonstrates the effectiveness of the method used.

In this embodiment, the hyperparameters of SDFR-ref are empirically set as follows: the minimum number of samples is 3, the number of random feature selections is 11, the number of CARTs is 500, the number of layers is 500, and the learning rate is 0.1. RF, BP neural network (BPNN), XGBoost, DFR, DFR-clfc and ImDFR modeling methods are used for experimental comparison. The parameter settings are as follows: 1) RF: the minimum number of samples is 3, the number of CART is 500, and the random feature selection is 11; 2) BPNN: The number of hidden layer neurons is 30, the convergence error is 0.01, the algebra is 1500, and the learning rate is 0.1; 3) XGBoost: The minimum number of samples is 3, the number of XGBoost is 10, the regularization coefficient is 1.2, and the learning rate is 0.8; 4) DFR and DFR-clfc: the minimum number of samples is 3, the number of CART is 500, the number of random feature selection is 11, and the number of RF and CRF is 2 respectively.

The performance of the modeling method is evaluated using RMSE and R², which are defined as follows:

$\begin{array}{cc}\mathrm{RMSE}=\sqrt{\sum _{n=1}^N{\left(y_n-{\widehat{y}}_n\right)}^2/\left(N-1\right)}& \left(25\right)\end{array}$

$\begin{array}{cc}{R}^2=1-\sum _{n=1}^N{\left(y_n-{\widehat{y}}_n\right)}^2/\sum _{n=1}^N{\left(y_n-\overline{y}\right)}^2& \left(26\right)\end{array}$

Among them, y_n represents the n-th true value, ŷ_n represents the n-th predicted value, y represents the average output value, and N represents the number of samples.

On this basis, 30 repeated experiments were conducted on seven methods, and Table 3 shows the statistical results. Table 4 gives the statistical results of training time.

TABLE 3
statistical results
	RMSE	R2
		mean		optimum	mean		optimum
Method	set	value	variance	value	value	variance	value
RF	training	1.0993E−02	2.5498E−08	1.0704E−02	8.5783E−01	1.7106E−05	8.6522E−01
	Validation	1.9794E−02	3.9919E−08	1.9471E−02	5.1479E−01	9.6301E−05	5.3056E−01
	Test	1.6775E−02	6.1264E−08	1.6349E−02	5.9723E−01	1.4143E−04	6.1750E−01
BPNN	training	3.0495E−03	6.5539E−07	2.8748E−03	9.8832E−01	9.5015E−05	9.9028E−01
	Validation	3.2603E−02	2.4818E−04	2.1896E−02	−6.1325E−01	4.0544E+00	4.0635E−01
	Test	3.1648E−02	2.1475E−04	1.8531E−02	−7.3037E−01	3.3001E+00	5.0856E−01
XGBoost	training	1.0125E−02	0.0000E+00	1.0125E−02	8.7942E−01	3.1877E−31	8.7942E−01
	Validation	2.5207E−02	1.2452E−35	2.5207E−02	2.1325E−01	1.9923E−32	2.1325E−01
	Test	1.9748E−02	1.2452E−35	1.9748E−02	4.4189E−01	5.1004E−32	4.4189E−01
DFR	training	1.1508E−02	7.8541E−09	1.1347E−02	8.4422E−01	5.7639E−06	8.4855E−01
	Validation	2.0654E−02	1.0405E−08	2.0463E−02	4.7175E−01	2.7248E−05	4.8151E−01
	Test	1.7762E−02	1.6786E−08	1.7558E−02	5.4852E−01	4.3515E−05	5.5883E−01
DFR-	training	7.9183E−03	1.7761E−06	5.5822E−03	9.2423E−01	6.7227E−04	9.6335E−01
clfc	Validation	2.0084E−02	1.4533E−07	1.9410E−02	5.0034E−01	3.6156E−04	5.3348E−01
	Test	1.6968E−02	9.9144E−08	1.6430E−02	5.8785E−01	2.3681E−04	6.1370E−01
ImDFR	training	7.7000E−03	/	/	9.2420E−01	/	/
	Validation	2.3700E−02	/	/	1.3120E−01	/	/
	Test	1.7900E−02	/	/	6.6360E−01	/	/
SDFR-	training	6.6200E−04	4.7281E−09	5.2456E−04	9.9950E−01	1.2323E−08	9.9970E−01
ref	Validation	2.1700E−02	6.9600E−07	2.0200E−02	4.1450E−01	2.1000E−03	4.9700E−01
	Test	1.4500E−02	6.5875E−07	1.3100E−02	6.9780E−01	1.2000E−03	7.5300E−01

TABLE 4
the statistical results of training time
		Time
	Method	mean value	variance	optimum value
	RF	5.4138E+01	6.2333E−01	5.3153E+01
	XGBoost	9.7248E+01	3.5522E−01	9.6595E+01
	DFR	4.8513E+02	2.2753E+04	2.3745E+02
	DFR-clfc	8.2871E+02	1.0154E+05	3.4013E+02
	SDFR-ref	3.7039E+01	1.5538E+00	3.4474E+01

It can be seen from Table 3: 1) In the training set, the proposed method SDFR-ref has the average (6.6200E−04 and 9.9950E−01) and the best values (5.2456E−04 and 9.9970E−01) of RMSE and R² has optimal results; since no randomness is introduced, the variance statistics of XGBoost is almost 0; 2) In the validation set, SDFR-ref has no obvious advantage, and its performance is only better than BPNN, XGBoost and ImDFR; RF, DFR and The generalization performance of DFR-clfc is almost the same; 3) In the test set, SDFR-ref has the best measurement accuracy (1.4500E−02) and fitting performance (6.9780E−01).

To sum up, SDFR-ref has more powerful learning capabilities compared with classic learning methods (RF, BPNN and XGBoost). In addition, SDFR-ref contrasts deep learning methods (DFR, DFR-clfc, ImDFR) to further enhance the implementation of the model based on the simplified forest algorithm. The performance of SDFR-ref in the test set also shows that its generalization ability is stronger than other methods. Therefore, the proposed method is effective for DXN prediction of MSWI processes.

Table 4 shows that the method proposed in the invention has a greater advantage in the average training time compared with the method that is also a decision tree.

The invention proposes a method based on SDFR-ref to predict the DXN emission concentration in the MSWI process based on the grate furnace. The main contributions are as follows: 1) The feature selection module based on mutual information and significance test effectively reduces the computational complexity and improves the prediction performance; 2) The decision tree is used instead of the forest algorithm in the deep integration structure, which has excellent training speed and learning ability. For DFR and DFR clfc; 3) Due to the introduction of residual fitting, the prediction accuracy of SDFR-ref is further improved. Experimental results show that compared with traditional ensemble learning and deep ensemble learning, SDFR-ref has better modeling accuracy and generalization performance, and the training cost is lower than the state-of-the-art ensemble models. Therefore, SDFR-ref is easier for practical applications.

Claims

1. A soft measurement method for dioxin emission of grate furnace MSWI process based on simplified deep forest regression of residual fitting mechanism, comprising: a feature selection module based on Mutual information (MI) and significance test (ST) and a simplified deep forest regression (SDFR) module based on the residual fitting mechanism; wherein the feature selection module selects corresponding features by calculating MI value and ST value of each feature; for the SDFR module, Layer-k represents the k-th layer model, ŷ1Regvoc represents the output vector of the first layer model, v1Augfea represents augmented regression vector of the second layer input, {circumflex over (y)}kRegvoc represents the average value of ŷkRegvoc, α is the remaining learning rate between each layer; X and XXsel respectively represents the process data before and after feature selection; y, ŷ and e are the true value, predicted value and prediction error respectively;in addition, {δMI, δSL, θ, T, α, K} represents the learning parameter set of proposed SDFR-ref, where: δMI represents the threshold of MI, δSL represents the threshold of significance level, and θ represents the minimum sample in the leaf node number, T represents the number of decision trees in each layer of the model, a is the learning rate in the gradient boosting process, and K represents the number of layers; the globally optimized selection of these learning parameters being capable of improving the synergy between different modules, thereby improving the overall performance of the model; wherein the proposed modeling strategy is formulated as solving the following optimization problem: