Intelligent operation optimization method for municipal solid waste incineration process

Abstract

The present invention belongs to the field of municipal solid waste management, and also belongs to the field of intelligent optimization. Provided is an intelligent operation optimization method for a municipal solid waste incineration process. The method comprises: collecting historical data in a municipal solid waste incineration process to construct a sample data set; using nitrogen oxide emissions and combustion efficiency as optimization objectives to establish an index model based on a self-organizing radial basis function neural network, so as to characterize a mapping relationship between operation variables and the optimization objectives in the municipal solid waste incineration process; and using the established index model as an evaluation function of a multi-objective optimization algorithm, and using a multi-objective particle swarm optimization algorithm to obtain optimal set values of the operation variables, so as to improve the combustion efficiency while reducing nitrogen oxide emission concentration. In the present invention, an intelligent operation optimization method for a municipal solid waste incineration process is established, such that the combustion efficiency and denitrification efficiency can be synergistically improved, and technical support can be provided for efficient and environment-friendly operation of the municipal solid waste incineration process.

Description

Intelligent operation optimization method for municipal solid waste incineration process

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on July 31, 2023, with application number 202310951828.4 and invention name “Intelligent Operation Optimization Method for Urban Solid Waste Incineration Process”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present invention relates to the fields of urban solid waste management and intelligent optimization, and in particular to an intelligent operation optimization method for an urban solid waste incineration process. Background Art

[0003] The accumulation of large amounts of solid waste seriously pollutes the soil, air, and water environment and is also harmful to human health. For municipal solid waste that cannot be directly recycled, the main treatment method is incineration technology. However, with the rapid development of the municipal solid waste incineration industry, more and more pollutants are emitted, such as nitrogen oxides. NO generated during the incineration process x It is difficult to remove effectively and poses a dual threat to the ecological environment and human health.

[0004] At present, the operation of municipal solid waste incineration plants mostly relies on manual experience to distribute air and materials and add reducing agents such as urea to avoid NO x Exceeding emission standards can lead to unstable combustion in the incinerator and secondary pollution. Therefore, designing intelligent operation optimization methods for the municipal solid waste incineration process and accurately and effectively setting and adjusting operating variables is an important means to ensure the reduction of NOx emissions and the improvement of combustion efficiency during the municipal solid waste incineration process. This has always been a technology and research focus that urgently needs to be implemented.

[0005] Summary of the Invention

[0006] The purpose of the present invention is to provide an intelligent operation optimization method for the municipal solid waste incineration process, with the optimization goals of improving combustion efficiency and reducing NOx emission concentration, and seeking the optimal set values ​​of operating variables (including grate speed, primary air inlet flow, primary air outlet pressure, and secondary air inlet flow) during the operation of municipal solid waste incineration while meeting actual constraints, thereby providing guidance for actual industrial processes.

[0007] To achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for intelligent operation optimization of a municipal solid waste incineration process, comprising:

[0009] Step 1: Collect historical data of the municipal solid waste incineration process to construct a sample dataset and divide it into a training set and a test set;

[0010] Step 2: Establishing a nitrogen oxide index model, a carbon monoxide index model, and a carbon dioxide index model based on a self-organizing radial basis function neural network based on the training set to characterize the mapping relationship between the operating variables and the optimization objectives in the solid waste incineration process, wherein the operating variables include grate speed, primary air inlet flow rate, primary air outlet pressure, and secondary air inlet flow rate; and the optimization objectives include reducing nitrogen oxide emissions and improving combustion efficiency, wherein the combustion efficiency is calculated based on the output values ​​of the carbon monoxide index model and the carbon dioxide index model;

[0011] Step 3: Using the nitrogen oxide index model, the carbon monoxide index model, and the carbon dioxide index model as evaluation functions of the multi-objective particle swarm optimization algorithm to evaluate the quality of solutions in the population;

[0012] Step 4: Using the multi-objective particle swarm optimization algorithm, obtain the Pareto optimal solution set of the operating variables;

[0013] Step 5: Determine the optimal setting values ​​of the operating variables from the Pareto optimal solution set based on the utility function.

[0014] Optionally, step 1 specifically includes: collecting and storing historical data of the urban solid waste incineration process through detection equipment, the historical data including grate speed, primary air inlet flow, primary air outlet pressure, secondary air inlet flow, NOx concentration, CO concentration and CO2 concentration, and constructing the sample data set based on the collected historical data.

[0015] Optionally, the method for constructing the indicator model in step 2 includes:

[0016] Step 21: training a self-organizing radial basis function neural network according to the training set, wherein the number of neurons in the hidden layer of the self-organizing radial basis function neural network at the initial moment is 0;

[0017] Step 22: adding a first hidden layer neuron to the self-organizing radial basis function neural network based on the training sample corresponding to the expected output value with the largest absolute value in the training set data;

[0018] Step 23: adjusting the parameters of the self-organizing radial basis function neural network with the first hidden layer neurons added based on a second-order learning algorithm;

[0019] Step 24: Based on the training set data and the output of the self-organizing radial basis function neural network, calculate the error vector of the training set, wherein the error vector includes the error value of each training sample; determine the error peak point, which is the error value with the largest absolute value; and based on the training sample corresponding to the error peak point, add the next hidden layer neuron to the self-organizing radial basis function neural network;

[0020] Step 25: adjusting the parameters of the self-organizing radial basis function neural network with the next hidden layer neurons added based on a second-order learning algorithm;

[0021] Step 26: Repeat the steps of adding a next hidden layer neuron based on the error peak point and adjusting the parameters of the self-organizing radial basis function neural network with the next hidden layer neuron added, until the learning accuracy of the self-organizing radial basis function neural network reaches a preset accuracy or the number of hidden layer neurons in the self-organizing radial basis function neural network reaches a preset number, and the nitrogen oxide index model, the carbon monoxide index model, or the carbon dioxide index model is constructed;

[0022] Wherein, when constructing the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model, the input and output of the self-organizing radial basis function neural network are different, and the input is determined from the training set.

[0023] Optionally, step 4 specifically includes:

[0024] Step 41: Initializing the initial position and initial velocity of each particle in the particle swarm: the initial velocity of each particle is zero, and the initial position of each particle is determined based on a constraint condition, which is set based on the boundary value of the training set data;

[0025] Step 42: Calculating the fitness value of each particle based on the nitrogen oxide index model, the carbon monoxide index model, and the carbon dioxide index model for evaluating the particle;

[0026] Step 43: establishing an external archive based on an adaptive grid mechanism, and storing non-dominated particles in the external archive based on the fitness value of each particle;

[0027] Step 44: updating the particle position and velocity of each particle within the constraints, wherein the method for selecting the global optimal particle used in the updating process is based on the external archive established in step 3, and using an adaptive grid algorithm to select the global optimal particle in the external archive to enhance the diversity of the algorithm;

[0028] Step 45: For each particle after the position and velocity are updated, repeat the above steps of evaluating the fitness value, updating the external archive, and updating the particle position and velocity until a preset number of iterations is reached, and output the Pareto optimal solution set corresponding to each non-dominated particle in the external archive at the last iteration.

[0029] Optionally, the weight coefficient of the utility function in step 5 is set according to actual conditions and decision preferences.

[0030] Compared with the prior art, the present invention has the following beneficial effects:

[0031] The intelligent operation optimization method for the municipal solid waste incineration process provided by the present invention, on the one hand, establishes an accurate and effective municipal solid waste incineration process performance index model based on the good nonlinear mapping ability of the self-organizing radial basis function neural network, which can realize the optimization of NO x The precise characterization of the dynamic characteristics of emissions and combustion efficiency lays a solid foundation for the subsequent optimization of operational variables. Furthermore, a fitness evaluation function is established using a self-organizing radial basis function neural network indicator model. An iterative solution based on a multi-objective particle swarm optimization algorithm is used to obtain optimized operational variables, achieving a synergistic improvement in combustion and denitrification efficiency. This enhances the intelligence level of the municipal solid waste incineration process and promotes its efficient and green operation.

[0032] Figures in the specification

[0033] The present invention will be further described below in conjunction with the accompanying drawings:

[0034] FIG1 is a flow chart of a method for optimizing the intelligent operation of a municipal solid waste incineration process according to the present invention;

[0035] FIG2 is a schematic diagram of the structure of a self-organizing radial basis function neural network provided by the present invention;

[0036] FIG3 is a schematic diagram of the test results of the nitrogen oxide index model provided by the present invention;

[0037] FIG4 is a schematic diagram of the test results of the carbon monoxide index model provided by the present invention;

[0038] FIG5 is a schematic diagram of the test results of the carbon dioxide index model provided by the present invention;

[0039] FIG6 is a second flow chart of the intelligent operation optimization method for the municipal solid waste incineration process provided by the present invention;

[0040] FIG7 is a schematic diagram of the optimization results of NOx concentration provided by the present invention;

[0041] FIG8 is a schematic diagram of the optimization results of the combustion efficiency provided by the present invention. DETAILED DESCRIPTION

[0042] The following is a detailed description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work shall fall within the scope of protection of the present invention.

[0043] The purpose of the present invention is to provide an intelligent operation optimization method for the municipal solid waste incineration process, with the optimization goals of improving combustion efficiency and reducing NOx emission concentration, and seeking the optimal set values ​​of operating variables (including grate speed, primary air inlet flow, primary air outlet pressure, and secondary air inlet flow) during the operation of municipal solid waste incineration while meeting actual constraints, thereby providing guidance for actual industrial processes.

[0044] 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.

[0045] The following describes the intelligent operation optimization method for the municipal solid waste incineration process of the present invention in conjunction with Figures 1 to 8. As shown in Figure 1, the method includes:

[0046] Step 1: Collect historical data of the municipal solid waste incineration process to construct a sample dataset and divide it into a training set and a test set;

[0047] Step 2: Establishing a nitrogen oxide index model, a carbon monoxide index model, and a carbon dioxide index model based on a self-organizing radial basis function neural network based on the training set to characterize the mapping relationship between the operating variables and the optimization objectives in the solid waste incineration process, wherein the operating variables include grate speed, primary air inlet flow rate, primary air outlet pressure, and secondary air inlet flow rate; and the optimization objectives include reducing nitrogen oxide emissions and improving combustion efficiency, wherein the combustion efficiency is calculated based on the output values ​​of the carbon monoxide index model and the carbon dioxide index model;

[0048] Step 3: Using the nitrogen oxide index model, the carbon monoxide index model, and the carbon dioxide index model as evaluation functions of the multi-objective particle swarm optimization algorithm to evaluate the quality of solutions in the population;

[0049] Step 4: Using the multi-objective particle swarm optimization algorithm, obtain the Pareto optimal solution set of the operating variables;

[0050] Step 5: Determine the optimal setting values ​​of the operating variables from the Pareto optimal solution set based on the utility function.

[0051] Specifically, the solid waste incineration historical data is collected and stored through the collection equipment. The historical data can be collected once every preset time interval. The preset time can be, for example, 30 seconds, 25 seconds, 40 seconds, etc., which is not specifically limited here. Each time the data is collected, the grate speed, primary air inlet flow, primary air outlet pressure, secondary air inlet flow, NO x Concentration value, CO concentration value, CO2 concentration value and other related variable data.

[0052] After collecting enough historical data, for example, after collecting 1,000 sets of historical data, a sample data set is constructed based on the 1,000 sets of historical data, and a training set and a test set are divided from the sample data set. For example, the first 700 sets of historical data or the last 700 sets of historical data from the 1,000 sets of data are obtained as the training set, and the remaining data is used as the test set.

[0053] After determining the training set, a nitrogen oxide index model, a carbon monoxide index model, and a carbon dioxide index model based on the self-organizing radial basis function (SORBF) neural network were established to characterize the mapping relationship between the operating variables and the optimization objectives in the solid waste incineration process. The operating variables include grate speed, primary air inlet flow, primary air outlet pressure, and secondary air inlet flow. The optimization objectives include reducing nitrogen oxide concentration and improving combustion efficiency.

[0054] According to the optimization target and the calculation formula of combustion efficiency, the output variables of the nitrogen oxide index model, carbon monoxide index model and carbon dioxide index model are determined to be the NOx concentration, CO concentration value and CO2 concentration value at time t respectively.

[0055] The calculation formula for combustion efficiency is:

[0056] The methods for constructing the NOx model, CO model, and CO2 index model are the same, but the inputs and outputs of each model are different. Input variables of the NOx index model Represents the nitrogen oxides (NO x ) concentration value, x2 represents the grate speed at the tth moment, x3, x4, x5 represent the primary air inlet flow rate, primary air outlet pressure, and secondary air inlet flow rate at the tth moment, respectively, and T represents the transposition symbol; the input variables of the carbon monoxide index model are represents the CO concentration value at time t-1, x2 represents the grate speed at time t, x3, x4, x5 represent the primary air inlet flow, primary air outlet pressure, and secondary air inlet flow at time t respectively; the input variables of the carbon dioxide index model are represents the CO2 concentration value at the moment t-1, x2 represents the grate speed at the moment t, x3, x4, and x5 represent the primary air inlet flow rate, primary air outlet pressure, and secondary air inlet flow rate at the moment t, respectively.

[0057] The nitrogen oxide index model, carbon monoxide index model and carbon dioxide index model are used as evaluation functions of the multi-objective particle swarm optimization algorithm to evaluate the quality of solutions in the population, and then solve to obtain the Pareto optimal solution set of the operating variables. A utility function set based on demand is used to finally determine the optimal setting value of the operating variable from the Pareto optimal solution set.

[0058] The intelligent operation optimization method for the municipal solid waste incineration process provided by the present invention, on the one hand, establishes an accurate and effective municipal solid waste incineration process performance index model based on the good nonlinear mapping ability of the self-organizing radial basis function neural network, which can realize the optimization of NO x Accurate characterization of the dynamic characteristics of emissions and combustion efficiency lays a solid foundation for the subsequent optimization of operational variables. Furthermore, a fitness evaluation function is established using a self-organizing radial basis function neural network indicator model. An iterative solution based on a multi-objective particle swarm optimization algorithm is used to obtain optimized operational variables, achieving a synergistic improvement in combustion and denitrification efficiency. This enhances the intelligence level of the municipal solid waste incineration process and promotes its efficient and green operation.

[0059] In one embodiment, step 1 specifically includes: collecting and storing historical data of the municipal solid waste incineration process through detection equipment, the historical data including grate speed, primary air inlet flow, primary air outlet pressure, secondary air inlet flow, NOx concentration, CO concentration and CO2 concentration, and constructing the sample data set based on the collected historical data.

[0060] In one embodiment, the method for constructing the indicator model in step 2 includes:

[0061] Step 21: training a self-organizing radial basis function neural network according to the training set, wherein the number of neurons in the hidden layer of the self-organizing radial basis function neural network at the initial moment is 0;

[0062] Step 22: adding a first hidden layer neuron to the self-organizing radial basis function neural network based on the training sample corresponding to the expected output value with the largest absolute value in the training set data;

[0063] Step 23: adjusting the parameters of the self-organizing radial basis function neural network with the first hidden layer neurons added based on a second-order learning algorithm;

[0064] Step 24: Based on the training set data and the output of the radial basis function neural network, calculate the error vector of the training set, wherein the error vector includes the error value of each training sample; determine the error peak point, wherein the error peak point is the error value with the largest absolute value; and based on the training sample corresponding to the error peak point, add the next hidden layer neuron to the self-organizing radial basis function neural network;

[0065] Step 25: adjusting the parameters of the self-organizing radial basis function neural network with the next hidden layer neurons added based on a second-order learning algorithm;

[0066] Step 26: Repeat the steps of adding a next hidden layer neuron based on the error peak point and adjusting the parameters of the self-organizing radial basis function neural network with the next hidden layer neuron added, until the learning accuracy of the self-organizing radial basis function neural network reaches a preset accuracy or the number of hidden layer neurons in the self-organizing radial basis function neural network reaches a preset number, and the nitrogen oxide index model, the carbon monoxide index model, or the carbon dioxide index model is constructed;

[0067] Wherein, when constructing the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model, the input and output of the self-organizing radial basis function neural network are different, and the input is determined from the training set.

[0068] Specifically, the above steps are a general method for constructing a nitrogen oxide index model, a carbon monoxide index model, and a carbon dioxide index model. When constructing different index models, the input and output of the self-organizing radial basis function (SORBF) neural network model are different, where the input is determined from the training set. Figure 2 shows the structure of the SORBF neural network model.

[0069] The NOx index model / carbon monoxide index model / carbon dioxide index model outputs NO x The calculation formula for the CO / CO2 concentration value is:

[0070] Among them, y is the output of the SORBF neural network model, x is the input variable of the nitrogen oxide index model / carbon monoxide index model / carbon dioxide index model (after normalization), w j is the output weight from the jth hidden layer neuron to the output neuron, Φ j is the output of the jth hidden layer neuron, and J is the total number of hidden layer neurons.

[0071] The method of constructing the indicator model based on SORBF neural network is as follows:

[0072] (1) When constructing the index model, the number of neurons in the hidden layer of the SORBF neural network at the initial moment is 0.

[0073] (2) At the initial moment, the training sample corresponding to the maximum absolute residual of the current SORBF neural network is the training sample k1 corresponding to the expected output value with the largest absolute value: k1=argmax[||y h1 ||,||y h2 ||,…,||y hp ||,…,||y hP ||] (3)

[0074] Among them, P represents the number of training samples in the training set, y hp Represents the expected output value of the pth training sample. The expected output value is the actual NO in the historical data. x / CO / CO2 concentration value.

[0075] The first hidden layer neurons added to the SORBF neural network are set as follows: σ1=1 (6)

[0076] Among them, c1 is the center vector of the first hidden layer neuron, w1 is the output weight of the first hidden layer neuron, σ1 is the width of the first hidden layer neuron, is the input vector of training sample k1, is the expected output value of training sample k1.

[0077] (3) Based on the second-order learning algorithm, the parameters of the SORBF neural network with the first hidden layer neurons added are adjusted as follows: Ψ(η+1)=Ψ(η)-(H(η)+λI(η)) -1 Ω(η)) (7)

[0078] Where η is the number of iteration steps for parameter adjustment (for example, the total number of iteration steps can be set to 50), H is the Hessian matrix, λ is the learning rate (for example, it can be 0.01), I is the identity matrix, Ω is the gradient vector, and Ψ is the parameter that needs to be adjusted for the SORBF neural network: Ψ(η)=[c1(η),σ1(η),w1(η)] (8)

[0079] In order to reduce the computational complexity, the Hessian matrix H is converted into the sum of P Hessian sub-matrices h, and the gradient vector Ω is converted into the sum of P gradient sub-vectors g, as shown below:

[0080] When the network parameters are adjusted for the ηth time, the p-th Hessian matrix h p (η) and the gradient subvector g p (η) is calculated as follows:

[0081] Among them, e p (η) is the network output value y of the p-th training sample output by the SORBF neural network at the η-th adjustment p (η) and the expected output value y hp The difference, j p (η) is the Jacobian vector, and T is the transpose sign.

[0082] e p The calculation method of (η) is as follows: p (η) = y p (η)-y hp (13)

[0083] The Jacobian vector is calculated as follows:

[0084] Among them, [c 11 (η),c 12 (η),c 13 (η),c 14 (η),c 15 (η)] is the center vector at the ηth iteration, c 11 (η), c 12 (η), c 13 (η), c 14 (η) and c 15 (η) represents the element in the center vector at the ηth iteration, w1(η) and σ1(η) are the weight and width at the ηth iteration, respectively. p (η) is the network output at the ηth iteration.

[0085] (4) Calculate the error vector e of the training set, which includes the error value of each training sample: e=[e1,e2,…,e p ,…,e P ] (15)

[0086] The error value e of the pth training sample p The calculation is as follows: e p =y p -y hp (16)

[0087] Among them, y p is the network output value of the pth training sample output by the SORBF neural network, y hp is the expected output value of the pth training sample.

[0088] Based on the error vector e of the training set, find the training sample k corresponding to the current error peak point. The error peak point is the error value with the largest absolute value: k = argmax||e|| (17)

[0089] Based on the training sample corresponding to the error peak point, the next (i.e., the lth) hidden layer neuron is added to the SORBF neural network, and the center vector c of the hidden layer neuron is l and output weight w l As follows: c l =x k (18) w l =y hk (19)

[0090] Among them, x k is the input vector of the kth training sample, y hk is the expected output value of the kth training sample.

[0091] In order to avoid structural redundancy of the neural network model, it is hoped that the existing hidden layer neurons in the SORBF neural network have little influence on the newly added hidden layer neurons. Therefore, the width σ of the newly added hidden layer neurons is l The settings are as follows: c min =argmin(dist(c l ,c j≠l )) (20) σ l =a||c l -c min || (21)

[0092] Among them, c min is the center vector of the hidden layer neuron closest to the lth hidden layer neuron in the SORBF neural network, c j is the center vector of the other hidden layer neurons excluding the lth hidden layer neuron in the SORBF neural network. a is adjusted according to different index models. For example, the values ​​of a for the nitrogen oxide index model, the carbon monoxide index model, and the carbon dioxide index model can be: 0.8, 0.65, and 0.8, respectively.

[0093] (5) After adding the first hidden layer neuron, the next hidden layer neuron is added to the SORBF neural network. When adding the next neuron, it is added based on the training sample corresponding to the error peak point in all training samples in the training set.

[0094] After the next neuron is added, the parameters of the SORBF neural network with the next hidden layer neuron added are adjusted based on the second-order learning algorithm. The specific adjustment method is the same as the method for adjusting the parameters of the SORBF neural network with the first hidden layer neuron added.

[0095] Repeat the above steps of adding a next hidden layer neuron based on the error peak point and adjusting the parameters of the SORBF neural network model with the next hidden layer neuron added until the learning accuracy of the SORBF neural network model reaches a preset accuracy or until the number of hidden layer neurons in the RBF neural network model reaches a preset number, and the nitrogen oxide index model, carbon monoxide index model or carbon dioxide index model is constructed, that is:

[0096] If the number of neurons in the hidden layer of the SORBF neural network reaches J max , or when the learning accuracy MSE is less than E0, the construction of the nitrogen oxide index model, the carbon monoxide index model or the carbon dioxide index model is completed, wherein the J corresponding to the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model is max They can be 13, 10, and 10 respectively.

[0097] The mean square error (MSE) can be used to measure the learning accuracy of the network. The specific calculation is as follows:

[0098] Among them, y hp is the expected output value of the pth training sample, y p is the network output value of the pth training sample output by the SORBF neural network model, and P is the total number of training samples.

[0099] The nitrogen oxide index model, the carbon monoxide index model or the carbon dioxide index model is constructed based on a SORBF neural network whose learning accuracy reaches a preset accuracy or the number of hidden layer neurons reaches a preset number.

[0100] For example, after establishing the NOx, CO, and CO2 indicator models, these indicator models can be verified to determine whether they meet the requirements. Specifically, the verification can be performed by dividing the sample dataset into a test set. For example, the sample dataset can be divided into a training set and a test set according to a preset ratio, such as 7 / 3 or 8 / 2, with the training set having a higher ratio.

[0101] The test samples in the test set are used as the input of the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model. The outputs of the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model are NOx The fitting values ​​of concentration, CO concentration and CO2 concentration are used to quantitatively evaluate the fitting accuracy of the nitrogen oxide index model, carbon monoxide index model and carbon dioxide index model respectively using the root mean square error and average percentage error. The root mean square error RMSE and average percentage error MAPE are calculated as follows:

[0102] Among them, M is the total number of test samples in the test set, y m is the network output value of the nitrogen oxide index model / carbon monoxide index model / carbon dioxide index model for the mth test sample, y hm is the expected output value of the mth test sample. The lower the RMSE and MAPE, the higher the fitting accuracy.

[0103] Figures 3, 4, and 5 show the test results of the nitrogen oxide index model, the carbon monoxide index model, and the carbon dioxide index model, respectively. In the figure, the x-axis represents the number of test samples (unit: pieces / sample), and the y-axis represents the NO x , CO, CO2 concentration values, in mg / m 3 The network output in the figure is the fitting value of each indicator model. The root mean square error (RMSE) and average percentage error (MAPE) are used to quantitatively evaluate the fitting accuracy of each indicator model. For the nitrogen oxide indicator model, the root mean square error (RMSE) obtained using the test set is 9.3939, and the average percentage error (MAPE) is 5.5603%. For the carbon monoxide indicator model, the root mean square error (RMSE) obtained using the test set is 0.5368, and the average percentage error (MAPE) is 15.4444%. For the carbon dioxide indicator model, the root mean square error (RMSE) obtained using the test set is 0.1866, and the average percentage error (MAPE) is 2.2523%.

[0104] In one embodiment, step 4 specifically includes:

[0105] Step 41: Initializing the initial position and initial velocity of each particle in the particle swarm: the initial velocity of each particle is zero, and the initial position of each particle is determined based on a constraint condition, which is set based on the boundary value of the training set data;

[0106] Step 42: Calculating the fitness value of each particle based on the nitrogen oxide index model, the carbon monoxide index model, and the carbon dioxide index model for evaluating the particle;

[0107] Step 43: establishing an external archive based on an adaptive grid mechanism, and storing non-dominated particles in the external archive based on the fitness value of each particle;

[0108] Step 44: updating the particle position and velocity of each particle within the constraints, wherein the method for selecting the global optimal particle used in the updating process is based on the external archive established in step 3, and using an adaptive grid algorithm to select the global optimal particle in the external archive to enhance the diversity of the algorithm;

[0109] Step 45: For each particle after the position and velocity are updated, repeat the above steps of evaluating the fitness value, updating the external archive, and updating the particle position and velocity until a preset number of iterations is reached, and output the Pareto optimal solution set corresponding to each non-dominated particle in the external archive at the last iteration.

[0110] Specifically, as shown in FIG6 , the method for obtaining the Pareto optimal solution set of the operating variables using the multi-objective particle swarm optimization algorithm includes:

[0111] (1) Initialize the initial position and initial velocity of each particle in the particle swarm. The initial velocity of each particle is zero, and the initial position of each particle is determined based on the constraint conditions.

[0112] There are four operating variables, so the dimension of the operating variables is 4. The highest value of the operating variable in the training data is used as the upper bound, and the lowest value of the operating variable is used as the lower bound. The constraint condition is that the position of the particle must be within the upper and lower bounds of the operating variables. For example, the initial velocity v of the particle is i,d and the initial position z i,d They are: z i,d =LB d +r*(UB d -LB d ) (25)

[0113] Among them, LB d Represents the lower bound of the d-th dimension operation variable, UB d represents the upper bound of the d-th dimension operation variable, r is a random number between 0 and 1, i = 1, 2, ..., Q (Q is the number of particles in the particle swarm, which can be 200), d = 1, 2, 3, 4, z i,d represents the position of the i-th particle under the d-th dimension operation variable, v i,d represents the velocity of the i-th particle under the d-th dimension operation variable.

[0114] (2) After initialization, the fitness value of each particle is calculated based on the nitrogen oxide index model, the carbon monoxide index model, and the carbon dioxide index model. The calculation method is:

[0115] Each particle in the particle group is input as a variable into the nitrogen oxide index model, carbon monoxide index model and carbon dioxide index model obtained in the above embodiment to calculate NO x Concentration value, CO concentration value and CO2 concentration value. Define fitness functions f1(z) and f2(z) (i.e., objective functions) to calculate the fitness value of each particle, where z represents the position vector of each particle, and the fitness function is:

[0116] Among them, g nox To convert the NO output by the nitrogen oxide index model x The concentration value is obtained by denormalization, g co The CO concentration value output by the carbon monoxide index model is denormalized and calculated, g co2 The CO2 concentration value output by the carbon dioxide index model is denormalized and calculated. The value of each particle calculated based on the above fitness function is the fitness value of each particle. Based on this fitness value, the non-dominated particles in the particle group can be determined.

[0117] (3) Establish an external archive based on the adaptive grid mechanism and store the current non-dominated particles.

[0118] For the external archive, an adaptive grid mechanism is used to generate a grid in the current objective function space of each particle. The grid is used as an index to define each particle. The grid index of each particle can be determined according to the value of the objective function. The grid definition of the external archive is as follows:

[0119] Among them, grid.LB μ The lower boundary of the grid, grid.UB μ is the upper boundary of the grid, β is a coefficient (β can be 0.1), μ=1,2…K, K represents the number of objective functions, which is 2 in this embodiment.

[0120] According to grid.LB μ and grid.UB μ The grid is evenly divided into multiple subgrids in each objective function dimension, and the coordinates of each subgrid in different objective function dimensions (P1, P2, ..., P K ) is used to define each particle, where P1 represents the ranking position of the particle in the first objective function dimension, and P2 represents the ranking position of the particle in the second objective function dimension.

[0121] (4) Determine the individual optimal position of each particle and the global optimal particle position at the current iteration.

[0122] For the selection of the global optimal particle, based on the external archive, the density of subgrids containing more than one particle is compared to reduce the probability of selecting a subgrid containing more non-dominated particles. The roulette wheel method is applied to obtain the subgrid for selecting the global optimal particle. Once the subgrid is selected, a non-dominated particle is randomly selected from the subgrid as the global optimal particle.

[0123] (5) Update the speed and position of the particles. If the current iteration is the it-th iteration, the updated speed and position of each particle for the it+1-th iteration are: z i,d (i t +1)=z i,d (it)+v i,d (it+1)

[0124] Among them, v i,d (it) is the velocity of the i-th particle under the d-th dimension operation variable at the it-th iteration, v i,d (it+1) is the velocity of the i-th particle under the d-th dimension operation variable at the it+1-th iteration, z i,d (it) is the position of the ith particle under the d-th dimension operation variable at the it-th iteration, z i,d (it+1) is the position of the ith particle under the d-th dimension operation variable at the it+1th iteration, p i,d (it) represents the individual optimal position of the i-th particle at the it-th iteration (the individual optimal position is the optimal position searched by the i-th particle up to the it-th iteration), g i,d (it) represents the position of the global optimal particle at the it-th iteration, r1 and r2 are random numbers between 0 and 1, ω, The flight parameters of the particle swarm algorithm, ω is the inertia weight, ω decays significantly with the number of iterations, and the decay rate is 0.99. is the learning factor, the learning factor The values ​​can be 1 and 2 respectively.

[0125] In order to ensure that the updated particles meet the actual constraints, it is necessary to check the updated particle positions and compare the updated particle positions z i,d (it+1) is the boundary limit of the constraint condition. If it is greater than the upper bound, the position of this dimension takes the upper bound UB d If it is less than the lower bound, the position takes the lower bound LB d , as shown in the following formula:

[0126] (6) After updating the position and velocity of each particle, the above steps of calculating the fitness value to updating the position and velocity are repeated until a preset number of iterations is reached. The Pareto solution set corresponding to each non-dominated particle in the external archive at the last iteration is output. The preset number of iterations can be, for example, 100, 120, 150, etc., which is not specifically limited here. After obtaining the Pareto solution set, a utility function is set, and the optimal operating variable is determined from the Pareto solution set based on the utility function.

[0127] In one embodiment, the weight coefficient of the utility function in step 5 is set according to actual conditions and decision preferences.

[0128] For example, a utility function is set to determine the optimal setting solution for the operating variables including grate speed, primary air inlet flow rate, primary air outlet pressure, and secondary air inlet flow rate. The specific method is as follows:

[0129] In order to determine the optimal setting value of the operating variable from the Pareto optimal solution set obtained by the multi-objective particle swarm optimization algorithm, a method for determining the optimal setting value based on the Pareto optimal solution set is designed. For example, the preferences of each optimization target can be determined according to the current actual situation and the decision maker's preference information, and the utility function of the solution in the Pareto optimal solution set can be defined to determine the optimal solution. Among them, the utility function U of the rth solution in the Pareto optimal solution set is r Defined as:

[0130] Among them, R represents the number of solutions in the Pareto solution set, δ μ is the weight coefficient of the μth optimization objective, which can be determined based on the decision-making process and the actual situation of the incineration plant. In this embodiment, there are two optimization objectives. To achieve balanced optimization of the two optimization objectives, for example, the ratio of the weight value of the first optimization objective to the weight value of the second optimization objective can be set to 0.3:0.7. In actual incineration projects, it is necessary to optimize the operating variables once every certain period of time (for example, every 4 minutes). In the Pareto solution set obtained from each optimization, the solution with the minimum utility function value is determined as the optimal setting solution for the operating variables.

[0131] As shown in Figures 7 and 8, after optimization based on the above method, NO x The optimization results of concentration and combustion efficiency are shown in the figure. The x-axis represents the optimization time in minutes, and the y-axis represents NO x Concentration and combustion efficiency, in mg / m 3 and %, respectively. The dotted line and * represent the actual values ​​before optimization, and the solid line and o represent the NO values ​​after optimization. xConcentration and combustion efficiency output values. As can be seen from the figure, the optimized NO x The concentration was compared with the actual NO before optimization. x The concentration decreased significantly. After optimization, NO x The concentration can be reduced by an average of 10.56%; the combustion efficiency after optimization is significantly improved compared with the actual combustion efficiency before optimization, and the combustion efficiency after optimization can be increased by an average of 8.24%.

[0132] In the above embodiment, by setting the utility function, the final optimized set value of the operating variable can be obtained according to the actual situation and decision preference, so as to achieve a synergistic improvement in combustion efficiency and denitrification efficiency.

[0133] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the purpose of the present invention.

Claims

1. A method for optimizing intelligent operation of a municipal solid waste incineration process, characterized in that: include: Step 1: Collect historical data of the municipal solid waste incineration process to construct a sample data set, and divide it into a training set and a test set; Step 2: According to the training set, a nitrogen oxide index model, a carbon monoxide index model and a carbon dioxide index model based on a self-organizing radial basis function neural network are established to characterize the mapping relationship between the operating variables and the optimization objectives in the solid waste incineration process, wherein the operating variables include grate speed, primary air inlet flow, primary air outlet pressure, and secondary air inlet flow, and the optimization objectives include reducing nitrogen oxide emissions and improving combustion efficiency, wherein the combustion efficiency is calculated through the output values ​​of the carbon monoxide index model and the carbon dioxide index model; Step 3: using the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model as evaluation functions of the multi-objective particle swarm optimization algorithm to evaluate the quality of solutions in the population; Step 4: Using the multi-objective particle swarm optimization algorithm to obtain the Pareto optimal solution set of the operating variables; Step 5: Determine the optimal setting value of the operating variable from the Pareto optimal solution set based on the utility function.

2. The intelligent operation optimization method for urban solid waste incineration process according to claim 1 is characterized in that: The step 1 specifically includes: collecting and storing historical data of the urban solid waste incineration process through detection equipment, the historical data including grate speed, primary air inlet flow, primary air outlet pressure, secondary air inlet flow, NOx concentration, CO concentration and CO2 concentration, and constructing the sample data set based on the collected historical data.

3. The intelligent operation optimization method for urban solid waste incineration process according to claim 1 is characterized in that: The method for constructing the indicator model in step 2 includes: Step 21: training a self-organizing radial basis function neural network according to the training set, wherein the number of neurons in the hidden layer of the self-organizing radial basis function neural network at the initial moment is 0; Step 22: adding a first hidden layer neuron to the self-organizing radial basis function neural network based on the training sample corresponding to the expected output value with the largest absolute value in the training set data; Step 23: adjusting the parameters of the self-organizing radial basis function neural network with the first hidden layer neurons added based on a second-order learning algorithm; Step 24: Based on the training set data and the output of the self-organizing radial basis function neural network, the error vector of the training set is calculated, and the error vector includes the error vector of each training sample. value; determining the error peak point, the error peak point being the error value with the largest absolute value, and adding a next hidden layer neuron to the self-organizing radial basis function neural network based on the training sample corresponding to the error peak point; Step 25: adjusting the parameters of the self-organizing radial basis function neural network with the next hidden layer neurons added based on a second-order learning algorithm; Step 26: Repeat the above steps of adding a next hidden layer neuron based on the error peak point and adjusting the parameters of the self-organizing radial basis function neural network with the next hidden layer neuron added, until the learning accuracy of the self-organizing radial basis function neural network reaches a preset accuracy or until the number of hidden layer neurons of the self-organizing radial basis function neural network reaches a preset number, the nitrogen oxide index model, the carbon monoxide index model or the carbon dioxide index model is constructed; Wherein, when constructing the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model, the input and output of the self-organizing radial basis function neural network are different, and the input is determined from the training set.

4. The intelligent operation optimization method for the municipal solid waste incineration process according to claim 1 is characterized in that: The step 4 specifically includes: Step 41: Initializing the initial position and initial velocity of each particle in the particle swarm: the initial velocity of each particle is zero, and the initial position of each particle is determined based on a constraint condition, and the constraint condition is set based on the boundary value of the training set data; Step 42: Calculating the fitness value of each particle based on the nitrogen oxide index model, the carbon monoxide index model and the carbon dioxide index model for evaluating the particle; Step 43: establishing an external archive based on an adaptive grid mechanism, and storing non-dominated particles in the external archive based on the fitness value of each particle; Step 44: updating the particle position and velocity of each particle within the constraint conditions, wherein the method for selecting the global optimal particle used in the updating process is based on the external archive established in step 3, and using an adaptive grid algorithm to select the global optimal particle in the external archive to enhance the diversity of the algorithm; Step 45: For each particle after updating its position and velocity, repeat the above steps of evaluating the fitness value, updating the external archive, and updating the particle position and velocity until the preset number of iterations is reached, and output the Pareto minimum value corresponding to each non-dominated particle in the external archive at the last iteration. Optimal solution set.

5. The intelligent operation optimization method for municipal solid waste incineration process according to claim 1 is characterized in that: The weight coefficient of the utility function in step 5 is set according to actual conditions and decision preferences.

Images