Temperature prediction method for municipal solid waste incinerator

Abstract

The invention discloses a method for predicting the furnace temperature of a municipal solid waste incinerator, which relates to the technical field of predicting key parameters of a municipal solid waste incineration process. The method predicts the furnace temperature of a solid waste incinerator by inputting variables such as the grate speed of each section, the primary air flow rate, the secondary air flow rate, and the fan pressure, and guides on-site operators to timely and accurately judge the changing trend of the furnace temperature. The method mainly comprises the following steps: (1) designing a heterogeneous feature set for establishing a furnace temperature prediction model for a municipal solid waste incinerator; (2) data preprocessing; (3) dividing the training set of the grate temperature base model based on the heterogeneous feature set of step 1; (4) determining the training set of the furnace temperature secondary learning model based on step 1; (5) parameter initialization; (6) constructing the grate temperature base model and the furnace temperature secondary training model by using a regularized random configuration network algorithm; and (7) performing denormalization processing on the output of the furnace temperature secondary training model.

Description

Technical Field

[0001] The present invention relates to the technical field of prediction of key parameters of a municipal solid waste incineration process, and more specifically, to a method for predicting furnace temperature of a municipal solid waste incinerator. Background Art

[0002] Waste incineration power generation technology can achieve the harmlessness, reduction and resource utilization of urban domestic waste. Therefore, this technology has become one of the main ways for us to deal with urban domestic waste. It is worth noting that, on the one hand, due to the composition of urban domestic waste in my country and the living habits of the people, the change of temperature during the incineration process is difficult to accurately predict. On the other hand, since the incineration process has a strong hysteresis, it is difficult for on-site operators to judge the changing trend of the furnace temperature in time, which may cause the furnace temperature to be out of control. Accurately predicting the temperature of the incinerator can provide guidance for the stable control of the incineration process of urban domestic waste. Therefore, the research results of the present invention have important practical significance.

[0003] The incinerator temperature usually refers to the average temperature of the flue gas in a combustion chamber, which is one of the key indicators to determine whether the incineration process is reasonable. When the incinerator temperature is lower than 850℃, the incineration process will produce dioxin gas with high carcinogenicity. Excessive temperature will cause coking, equipment damage and other faults in the incinerator. It can be seen that it is very necessary to control the incinerator temperature within a reasonable range. Accurate prediction of the incinerator temperature can guide on-site operators to adjust operating variables such as feed amount, grate speed, and primary air flow, so as to ensure that the incinerator temperature is within the normal operating range.

[0004] At present, the prediction methods of the furnace temperature of municipal solid waste incinerators mainly include two types: mechanism prediction model and data prediction model. The mechanism prediction model establishes a mathematical model of the furnace temperature based on the physical and chemical reactions such as mass conservation and energy conservation in the incinerator. However, due to the strong nonlinearity and strong coupling of the incineration process, the prediction accuracy of the mechanism model is difficult to guarantee. With the vigorous development of sensor technology, data-based prediction models have been widely used, such as BP neural network and support vector regression. However, due to the problems of slow convergence speed of BP neural network, easy to fall into local optimality, and poor processing ability of support vector regression for large samples, the furnace temperature prediction model established by the above method still has certain limitations.

[0005] As an emerging data-driven modeling method, the random configuration network has a universal approximation property for nonlinear mapping relationships, and the network structure is constructed using an incremental method, which greatly avoids the impact of manual settings on model performance. In addition, since the input weights and biases of the hidden layer of the random configuration network are randomly generated through a supervision mechanism, the model has an extremely fast training speed. Therefore. This method has been effectively applied in the field of parameter prediction. However, due to the large number of characteristic variables that affect the furnace temperature variables during the incineration process, the use of a single random configuration network model to predict the furnace temperature may have problems such as complex network structure and poor model generalization ability. Therefore, when faced with high-dimensional characteristic variables of the incineration process, it is very necessary to study a method for predicting the furnace temperature of an incinerator with strong generalization ability. Summary of the invention

[0006] In response to the above problems, the present invention provides a method for predicting the furnace temperature of an urban domestic waste incineration furnace based on a quadratic training integrated random configuration network, which can predict the furnace temperature of the incinerator according to input variables such as grate speed, primary air volume, and secondary air volume, thereby providing guidance for the stable control of the incineration process.

[0007] In order to achieve the above object, the present invention adopts the following technical scheme:

[0008] A method for predicting furnace temperature of a municipal solid waste incinerator, characterized in that it comprises the following steps:

[0009] (1) Design a heterogeneous feature set for establishing a temperature prediction model for municipal solid waste incinerators:

[0010] For the Tianxiong SN mechanical grate furnace, the input variables that affect the furnace temperature mainly include the following four types of input variables: grate speed x 1 -x 18 , primary air flow x 19 -x 32 , secondary air volume x 33 , fan pressure x 34 -x 35 The specific variables are shown in Table 1. In addition, the grate temperature x 36 -x 47 It is an intermediate variable that affects the furnace temperature, totaling 12. According to the operating mechanism of the incineration process, x 1 -x 35 The input variables are divided into 12 groups of heterogeneous features, and the specific division form is shown in the following formula:

[0011]

[0012] Among them, X 1 -X 12They represent 12 groups of input heterogeneous characteristic variables, which are used as input variables of the subsequent grate temperature base model. The corresponding 12 output variables are x 36 -x 47 .

[0013] In addition, according to the operating mechanism of the incineration process, the input and output correspondence between the grate temperature and the furnace temperature is shown in the following formula:

[0014] x 48 (t+1)=f(x 36 (t),x 37 (t),x 38 (t),x 39 (t),x 40 (t),x 41 (t),x 42 (t),x 43 (t),x 44 (t),x 45 (t),x 46 (t),x 47 (t),x 48 (t)) (2)

[0015] The variables are listed in Table 1. f(·) represents the mapping relationship between grate temperature and furnace temperature, which is obtained by secondary training using a regularized random configuration network. t represents the current moment, and t+1 represents the next moment.

[0016] (2) Data preprocessing:

[0017] All historical data of the incineration process are subjected to Z-SCORE standardization, as shown in the following formula:

[0018]

[0019] Among them, x n,q The qth characteristic variable of the nth sample in the historical data, n = 1, 2, ..., N, N represents the number of samples; q = 1, 2, ..., Q, Q represents the number of characteristic variables in the training set, here Q is 48. represents the qth characteristic variable of the nth sample after standardization. mean(x q ) represents the sample mean of the qth feature variable, std(x q ) represents the sample standard deviation of the qth characteristic variable. After standardization, the training set consisting of the historical data of the incineration process is represented by D.

[0020] (3) Divide the training set of the grate temperature base model based on the heterogeneous feature set in step 1:

[0021] According to the heterogeneous feature partitioning of formula (1) in step 1, the historical data of the municipal waste incineration process is divided into 12 training subsets, as shown in the following formula:

[0022]

[0023] Where D represents the training set consisting of N historical data of the incineration process; D p represents the training set of the p-th base model; X p represents the input variable of the p-th base model training set; y p represents the output variable of the pth base model training set, and N represents the sample size of each training set.

[0024] (4) Determine the training set of the furnace temperature secondary learning model based on step 1:

[0025] According to formula (2) in step 1, the training set of the furnace temperature secondary training model is determined as shown in the following formula:

[0026] D 2nd ={(X 2nd (n),y 2nd (n))},n=1,...,N (5)

[0027] Among them, D 2nd represents the training set of the furnace temperature secondary training model, X 2nd Represents the input variables of the training set, y 2nd Represents the output variable of the training set, where the output is the furnace temperature.

[0028] (5) Parameter initialization: Let the maximum number of hidden layer neurons of the random configuration network algorithm be L max =10, the maximum number of configurations is T max =100, the expected training error is ε=0.001, the parameter configuration range of the input weight and bias of the hidden layer neurons is [-λ, +λ], where λ=1; the regularization penalty factor is α=0.1.

[0029] (6) The regularized random configuration network algorithm is used to construct the grate temperature base model and the secondary training model of the furnace temperature: The regularized random configuration network learning process is mainly divided into two stages: random configuration of hidden layer neuron parameters and evaluation of output weights. For the objective function f:R d(p) →R, d(p) represents the number of input variables of the p-th base model.

[0030] The output expression of the model can be expressed as follows:

[0031]

[0032] Among them, X prepresents the input of the p-th grate temperature base model training set; H represents the model hidden layer output matrix; T represents the matrix transposition operation; β = {β 1 ,β 2 ,…,β L} represents the output weight of the hidden layer neurons; w j and b j denote the input weight and bias of the jth hidden layer neuron, represents the activation function of the jth hidden layer neuron, which is the Sigmoid function here, j = 1, ..., L, L is the number of hidden layer neurons. The input weights and biases in formula (6) can be randomly generated in [-λ, +λ] according to the supervision mechanism of formula (7), and the output weights β of the model can be re-determined using formula (8) * .

[0033]

[0034]

[0035] Among them, φ L represents the output of the Lth hidden layer neuron; represents the network residual when the number of neurons in the hidden layer is L-1, {μ L} is a non-negative real number sequence, μ L =(1-r) / L and the value of the constant r is r={0.9,0.99,0.999}; b g Represents the upper bound of the hidden layer neuron activation function, where b g =1; H L It represents the hidden layer output matrix of the model when the number of hidden layer neurons is L; Y is the output of the training set; I represents the identity matrix; α represents the regularization penalty factor, which is 0.1 here.

[0036] Based on the above regularized random configuration network training process, 12 grate temperature base models and 1 furnace temperature secondary training model are established using the 12 training subsets in step 3 and the training set in step 4, respectively.

[0037] (7) De-standardize the output of the furnace temperature secondary training model:

[0038] Since the output of the secondary training model of furnace temperature is a standardized value, the model output is de-standardized using the following formula to obtain the final furnace temperature prediction value.

[0039]

[0040] in, Represents the output value of the furnace temperature secondary training model, std(y 2nd) represents the standard deviation of the furnace temperature in the training set, mean(y 2nd ) represents the mean value of furnace temperature in the training set, Represents the furnace temperature prediction value output by the final model.

[0041] Table 1 Variable details

[0042]

[0043]

[0044] Compared with the prior art, the present invention has the following advantages: 1. The present invention adopts a regularized random configuration network learning algorithm to construct a base model, which ensures the accuracy of the base model while reducing the complexity of the base model. 2. By designing heterogeneous feature variables, the model complexity and calculation complexity of high-dimensional feature variables are effectively reduced, thereby better ensuring the generalization ability of the model. 3. The mapping relationship between grate temperature and furnace temperature is directly established, thereby more reasonably ensuring the important role of each section of grate temperature in furnace temperature forecasting. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a strategy diagram of the furnace temperature prediction method of the municipal domestic waste incinerator of the present invention.

[0046] Figure 2 Comparison of the prediction results of the furnace temperature prediction method of the present invention on the test samples DETAILED DESCRIPTION

[0047] The sample data comes from 900 sets of data generated during the incineration process of a municipal solid waste incineration power plant, including 700 training samples and 200 test samples, with a sample time interval of 5 minutes. Figure 1 The specific implementation modes of the present invention are further described.

[0048] A method for predicting furnace temperature of a municipal solid waste incinerator, characterized in that it comprises the following steps:

[0049] (1) Design a heterogeneous feature set for establishing a temperature prediction model for municipal solid waste incinerators:

[0050] For the Tianxiong SN mechanical grate furnace, the input variables that affect the furnace temperature mainly include the following four types of input variables: grate speed x 1 -x 18 , primary air flow x 19 -x 32 , secondary air volume x 33 , fan pressure x 34 -x 35 The specific variables are shown in Table 1. In addition, the grate temperature x 36 -x47 It is an intermediate variable that affects the furnace temperature, totaling 12. According to the operating mechanism of the incineration process, x 1 -x 35 The input variables are divided into 12 groups of heterogeneous features, and the specific division form is shown in the following formula:

[0051]

[0052] Among them, X 1 -X 12 They represent 12 groups of input heterogeneous characteristic variables, which are used as input variables of the subsequent grate temperature base model. The corresponding 12 output variables are x 36 -x 47 .

[0053] In addition, according to the operating mechanism of the incineration process, the input and output correspondence between the grate temperature and the furnace temperature is shown in the following formula:

[0054] x 48 (t+1)=f(x 36 (t),x 37 (t),x 38 (t),x 39 (t),x 40 (t),x 41 (t),x 42 (t),x 43 (t),x 44 (t),x 45 (t),x 46 (t),x 47 (t),x 48 (t)) (2)

[0055] The variables are listed in Table 1. f(·) represents the mapping relationship between grate temperature and furnace temperature, which is obtained by secondary training using a regularized random configuration network. t represents the current moment, and t+1 represents the next moment.

[0056] (2) Data preprocessing:

[0057] All historical data of the incineration process are subjected to Z-SCORE standardization, as shown in the following formula:

[0058]

[0059] Among them, x n,q The qth feature variable of the nth sample in the historical data, n = 1, 2, ..., 700; q = 1, 2, ..., Q, Q represents the number of feature variables in the training set, here Q is 48. represents the qth characteristic variable of the nth sample after standardization. mean(xq ) represents the sample mean of the qth feature variable, std(x q ) represents the sample standard deviation of the qth characteristic variable. After standardization, the training set consisting of the historical data of the incineration process is represented by D.

[0060] (3) Divide the training set of the grate temperature base model based on the heterogeneous feature set in step 1:

[0061] According to the heterogeneous feature partitioning of formula (1) in step 1, the historical data of the municipal waste incineration process is divided into 12 training subsets, as shown in the following formula:

[0062]

[0063] Where D represents the training set consisting of 700 historical data of incineration processes; D p represents the training set of the p-th base model; X p represents the input variable of the p-th base model training set; y p Represents the output variable of the p-th base model training set.

[0064] (4) Determine the training set of the furnace temperature secondary learning model based on step 1:

[0065] According to formula (2) in step 1, the training set of the furnace temperature secondary training model is determined as shown in the following formula:

[0066] D 2nd ={(X 2nd (n),y 2nd (n))},n=1,...,700 (5)

[0067] Among them, D 2nd represents the training set of the furnace temperature secondary training model, X 2nd Represents the input variables of the training set, y 2nd Represents the output variable of the training set, where the output is the furnace temperature.

[0068] (5) Parameter initialization: Let the maximum number of hidden layer neurons of the random configuration network algorithm be L max =10, the maximum number of configurations is T max =100, the expected training error is ε=0.001, the parameter configuration range of the input weight and bias of the hidden layer neurons is [-λ, +λ], where λ=1; the regularization penalty factor is α=0.1.

[0069] (6) The regularized random configuration network algorithm is used to construct the grate temperature base model and the secondary training model of the furnace temperature: The regularized random configuration network learning process is mainly divided into two stages: random configuration of hidden layer neuron parameters and evaluation of output weights. For the objective function f:Rd(p) →R, d(p) represents the number of input variables of the p-th base model.

[0070] The output expression of the model can be expressed as follows:

[0071]

[0072] Among them, X p represents the input of the p-th grate temperature base model training set; H represents the model hidden layer output matrix; T represents the matrix transposition operation; β = {β 1 ,β 2 ,…,β L} represents the output weight of the hidden layer neurons; w j and b j denote the input weight and bias of the jth hidden layer neuron, represents the activation function of the jth hidden layer neuron, which is the Sigmoid function here, j = 1, ..., L, L is the number of hidden layer neurons. The input weights and biases in formula (6) can be randomly generated in [-λ, +λ] according to the supervision mechanism of formula (7), and the output weights β of the model can be re-determined using formula (8) * .

[0073]

[0074]

[0075] Among them, φ L represents the output of the Lth hidden layer neuron; represents the network residual when the number of neurons in the hidden layer is L-1, {μ L} is a non-negative real number sequence, μ L =(1-r) / L and the value of the constant r is r={0.9,0.99,0.999}; b g Represents the upper bound of the hidden layer neuron activation function, where b g =1; H L It represents the hidden layer output matrix of the model when the number of hidden layer neurons is L; Y is the output of the training set; I represents the identity matrix; α represents the regularization penalty factor, which is 0.1 here.

[0076] Based on the above regularized random configuration network training process, 12 grate temperature base models and 1 furnace temperature secondary training model are established using the 12 training subsets in step 3 and the training set in step 4, respectively.

[0077] (7) De-standardize the output of the furnace temperature secondary training model:

[0078] Since the output of the secondary training model of furnace temperature is a standardized value, the model output is de-standardized using the following formula to obtain the final furnace temperature prediction value.

[0079]

[0080] in, Represents the output value of the furnace temperature secondary training model, std(y 2nd ) represents the standard deviation of the furnace temperature in the training set, mean(y 2nd ) represents the mean value of furnace temperature in the training set, Represents the furnace temperature prediction value output by the final model.

[0081] Due to the complexity of the incineration process, the accuracy of the furnace temperature mechanism prediction model is difficult to guarantee. As the number of characteristic variables of the incineration process increases, the complexity of the furnace temperature prediction method model based on data increases rapidly, resulting in poor generalization ability of the model. The present invention proposes a furnace temperature prediction method for urban domestic waste incinerators based on heterogeneous feature secondary training integrated random configuration network. The complexity of the grate temperature base model is reduced by designing heterogeneous features, and the generalization ability of the model is guaranteed by using a regularized random configuration network method. In addition, through the secondary training of the direct mapping relationship between the grate temperature and the incinerator temperature, the importance of the grate temperature in furnace temperature prediction is emphasized, thereby having a furnace temperature prediction model with strong generalization ability. The model can predict the furnace temperature value at the next moment in real time, so as to timely guide the on-site operators to adjust the incinerator operating variables. We use 200 sets of historical data of the incineration process as test samples to test the furnace temperature prediction method proposed in the present invention. The test results show that the test root mean square error of the prediction method proposed in the present invention is 21.3℃, and the prediction result comparison chart is as follows. Figure 2 As shown, Figure 2 The horizontal axis is the number of samples, and the vertical axis is the furnace temperature forecast value, in degrees Celsius.

Claims

1. A method for predicting furnace temperature of a municipal solid waste incinerator, characterized in that: The specific steps include: (1) Design a heterogeneous feature set for establishing a temperature prediction model for municipal solid waste incinerators: For the Tianxiong SN mechanical grate furnace, the input variables that affect the furnace temperature mainly include the following four types of input variables: grate speed x1-x 18 , primary air flow x 19 -x 32 , secondary air volume x 33 , fan pressure x 34 -x 35 , the specific variables are shown in Table 1; in addition, the grate temperature x 36 -x 47 There are 12 intermediate variables that affect the furnace temperature. According to the operating mechanism of the incineration process, x1-x 35 The input variables are divided into 12 groups of heterogeneous features, and the specific division form is shown in the following formula: Among them, X1-X 12 They represent 12 groups of input heterogeneous characteristic variables, which are used as input variables of the subsequent grate temperature base model. The corresponding 12 output variables are x 36 -x 47 ; In addition, according to the operating mechanism of the incineration process, the input and output correspondence between the grate temperature and the furnace temperature is shown in the following formula: x 48 (t+1)=f(x 36 (t),x 37 (t),x 38 (t),x 39 (t),x 40 (t),x 41 (t),x 42 (t),x 43 (t),x 44 (t),x 45 (t),x 46 (t),x 47 (t),x 48 (t)) (2) The variables are listed in Table 1. f(·) represents the mapping relationship between grate temperature and furnace temperature, which is obtained by secondary training using a regularized random configuration network. t represents the current moment, and t+1 represents the next moment. (2) Data preprocessing: All historical data of the incineration process are subjected to Z-SCORE standardization, as shown in the following formula: Among them, x n,q The qth characteristic variable of the nth sample in the historical data, n = 1, 2, ..., N, N represents the number of samples; q = 1, 2, ..., Q, Q represents the number of characteristic variables in the training set, here Q is 48; represents the qth characteristic variable of the nth sample after standardization; mean(x q ) represents the sample mean of the qth feature variable, std(x q ) represents the sample standard deviation of the qth characteristic variable; after standardization, the training set consisting of the historical data of the incineration process is represented by D; (3) Divide the training set of the grate temperature base model based on the heterogeneous feature set in step 1: According to the heterogeneous feature partitioning of formula (1) in step 1, the historical data of the municipal waste incineration process is divided into 12 training subsets, as shown in the following formula: Where D represents the training set consisting of N historical data of the incineration process; D p represents the training set of the p-th base model; X p represents the input variable of the p-th base model training set; y p represents the output variable of the pth base model training set, and N represents the sample size of each training set; (4) Determine the training set of the furnace temperature secondary learning model based on step 1: According to formula (2) in step 1, the training set of the furnace temperature secondary training model is determined as shown in the following formula: D 2nd ={(X 2nd (n),y 2nd (n))},n=1,...,N (5) Among them, D 2nd represents the training set of the furnace temperature secondary training model, X 2nd Represents the input variables of the training set, y 2nd represents the output variable of the training set, where the output is the furnace temperature; (5) Parameter initialization: Let the maximum number of hidden layer neurons of the random configuration network algorithm be L max =10, the maximum number of configurations is T max =100, the expected training error is ε=0.001, the parameter configuration range of the input weight and bias of the hidden layer neurons is [-λ, +λ], where λ=1; the regularization penalty factor is α=0.1; (6) The regularized random configuration network algorithm is used to construct the grate temperature base model and the secondary training model of the furnace temperature: The regularized random configuration network learning process is mainly divided into two stages: random configuration of hidden layer neuron parameters and evaluation of output weights; for the objective function f:R d(p) →R, d(p) represents the number of input variables of the p-th base model; The output expression of the model can be expressed as follows: Among them, X p represents the input of the p-th grate temperature base model training set; H represents the model hidden layer output matrix; T represents the matrix transposition operation; β={β1,β2,…,β L } represents the output weight of the hidden layer neurons; w j and b j denote the input weight and bias of the jth hidden layer neuron, represents the activation function of the jth hidden layer neuron, which is the Sigmoid function here, j = 1, ..., L, L is the number of hidden layer neurons; the input weights and biases in formula (6) can be randomly generated in [-λ, +λ] according to the supervision mechanism of formula (7), and the output weight β of the model can be re-determined using formula (8) * ; Among them, φ L represents the output of the Lth hidden layer neuron; represents the network residual when the number of neurons in the hidden layer is L-1, {μ L } is a non-negative real number sequence, μ L =(1-r) / L and the value of the constant r is r={0.9,0.99,0.999}; b g Represents the upper bound of the hidden layer neuron activation function, where b g =1; H L represents the hidden layer output matrix of the model when the number of hidden layer neurons is L; Y is the output of the training set; I represents the identity matrix; α represents the regularization penalty factor, which is 0.1 here; Based on the above regularized random configuration network training process, 12 grate temperature base models and 1 furnace temperature secondary training model are established using the 12 training subsets in step 3 and the training set in step 4 respectively; (7) De-standardize the output of the furnace temperature secondary training model: Since the output of the secondary training model of furnace temperature is a standardized value, the model output is de-standardized using the following formula to obtain the final furnace temperature prediction value; in, Represents the output value of the furnace temperature secondary training model, std(y 2nd ) represents the standard deviation of the furnace temperature in the training set, mean(y 2nd ) represents the mean value of furnace temperature in the training set, It represents the furnace temperature prediction value output by the final model; Table 1 Variable details 。

Images