Data-driven urban solid waste incineration process multi-objective operation optimization method

Abstract

The invention discloses a data-driven urban solid waste incineration process multi-objective operation optimization method, and belongs to the field of intelligent optimization. The method comprises the steps that firstly, nonlinear characteristics of the urban solid waste incineration process are captured through a self-organizing radial basis function neural network, accurate evaluation of operation performance index combustion efficiency and nitrogen oxide emission concentration is achieved, and an accurate optimization target model is provided for follow-up set value optimization solution; then, on the basis of the established optimization target model, a multi-target particle swarm optimization algorithm based on a double-file mechanism is designed, and an optimal solution set of set values of the primary air inlet flow and the secondary air inlet flow is obtained; and according to the optimal solution set, designing a utility function to determine optimal set values of the primary air inlet flow and the secondary air inlet flow, and keeping a relatively high combustion efficiency level while realizing that nitrogen oxide emission reaches the standard. According to the method, the combustion efficiency and the denitration efficiency can be synergistically improved, and efficient and green operation of the urban solid waste incineration process is promoted.

Description

Technical Field

[0001] Aiming at the problem of optimizing the operation of the municipal solid waste incineration process, the present invention designs a dual-file multi-objective particle swarm optimization algorithm, and proposes an operation optimization method for the municipal solid waste incineration process by using a data-driven online modeling strategy, so as to achieve compliance with nitrogen oxide emission standards and meet the growing demand for solid waste reduction. The present invention relates to both the field of solid waste incineration technology and the field of intelligent optimization, and in particular to a data-driven multi-objective operation optimization method for the municipal solid waste incineration process. Background Art

[0002] As the amount of urban solid waste increases year by year, incineration technology has gradually become the main method of urban solid waste treatment due to its efficient volume reduction and reduction effects. However, urban solid waste has the characteristics of low calorific value, high water content, complex composition and strong seasonality, which makes the urban solid waste incineration process unstable, low operating efficiency and high content of harmful gases such as nitrogen oxides produced after incineration. In order to achieve the goals of reducing, recycling and harmless treatment of urban solid waste incineration, it is necessary to optimize the key operating parameters in the operation process. At present, the operating parameters of the urban solid waste incineration process are mainly set by operators according to the current operating conditions and combined with expert experience. This method has great human subjectivity and lag, and cannot meet the actual requirements of efficient and low-emission operation of the urban solid waste incineration process. Therefore, optimizing the key operating parameters of the urban solid waste incineration process is crucial to improving the operating performance.

[0003] The municipal solid waste incineration process has the characteristics of high degree of operation continuity, numerous equipment, severe coupling, and strong nonlinearity. The optimization of the set values ​​of key variables in the municipal solid waste incineration process is a typical complex nonlinear optimization problem. In recent years, with the continuous development of intelligent modeling technology and intelligent optimization technology and their successful application in process industry, the solution to the problem of optimizing the setting of key variables in the municipal solid waste incineration process is essentially based on the process performance index model, designing the set value optimization algorithm of the operating variables, and realizing the optimization of performance indicators. With the development of data acquisition technology and artificial intelligence, data-driven optimization methods provide solutions to the operation optimization problem of the municipal solid waste incineration process. In data-driven optimization, the operation performance evaluation model is established by using machine learning technologies such as support vector regression and artificial neural network, and then a meta-heuristic intelligent optimization algorithm is designed to solve the optimization model. At present, there are few studies on data-driven operation optimization technology of municipal solid waste incineration process, and there is no multi-objective optimization method that meets industrial practice and can significantly improve the operation performance. Therefore, it is very necessary to design a data-driven multi-objective operation optimization method for municipal solid waste incineration process. Summary of the invention

[0004] The present invention provides a data-driven multi-objective operation optimization method for a municipal solid waste incineration process, which optimizes the optimal setting values ​​of operating variables (primary air inlet flow rate and secondary air inlet flow rate) during the operation of the municipal solid waste incineration process, provides guidance for the actual industrial process, and can meet the nitrogen oxide (Nitrogen Oxides, NOx) emission requirements of the municipal solid waste incineration process while maintaining efficient operation with optimal combustion efficiency.

[0005] The present invention adopts the following technical solutions and implementation steps:

[0006] Data-driven multi-objective operation optimization method for municipal solid waste incineration process.

[0007] It is characterized by comprising the following steps:

[0008] (1) 1080 sets of operation data of the municipal solid waste incineration process operation variables, optimization indicators (nitrogen oxide emission concentration and combustion efficiency) and other auxiliary variables (the right flue gas temperature of the primary combustion chamber, the left flue gas temperature of the primary combustion chamber, the right temperature of the primary combustion chamber, the left temperature of the primary combustion chamber, the right 1 air flow of the drying grate, the left 1 air flow of the drying grate, the right 1-1 air flow of the combustion section grate, the left 1-1 air flow of the combustion section grate, the amount of urea solution, the amount of lime feed, the amount of activated carbon storage bin feed, the main steam flow of the boiler outlet, the oxygen content of the inlet flue gas and the average temperature of the furnace) were collected through detection equipment to construct a sample data set, and the data was divided into a training set and a test set, of which 810 sets of data were used as the training set and 270 sets of data were used as the test data set;

[0009] (2) Establishing an optimization index evaluation model based on a self-organizing radial basis function neural network according to the training set to characterize the mapping relationship between the operating variables and the optimized target combustion efficiency and nitrogen oxide emission concentration in the solid waste incineration process, wherein the model input variables include the operating variables primary air inlet flow rate and secondary air inlet flow rate and other auxiliary variables, and the model output is the optimized index nitrogen oxide emission concentration and combustion efficiency, which is established as follows:

[0010] ①The initial number of neurons in the hidden layer is 0;

[0011] ② Calculate the network output of the training set samples, calculate the error between the network output value and the expected output value, determine the sample with the largest absolute error in the error vector, add a new hidden layer neuron in the self-organizing radial basis function neural network, and initialize the parameters of the newly added neuron based on the training sample corresponding to the maximum error value in the training set data, including the center and width of the neuron, to compensate for the current maximum absolute error;

[0012] ③ Based on the second-order learning algorithm, the parameters of the self-organizing radial basis function neural network with a newly added hidden layer of neurons are adjusted to achieve the purpose of training the network;

[0013] ④ Repeat the above steps ② and ③, that is, add a new hidden layer neuron based on the maximum network output error of the training sample and adjust the parameters of the self-organizing radial basis function neural network with a newly added hidden layer neuron, until the learning accuracy of the self-organizing radial basis function neural network reaches the expected training error or until the number of hidden layer neurons of the self-organizing radial basis function neural network reaches a preset number, the nitrogen oxide emission concentration index model and the combustion efficiency index model are constructed;

[0014] (3) Design a multi-objective particle swarm optimization algorithm based on a dual-file mechanism for the multi-objective operation optimization method of the municipal solid waste incineration process, and obtain the Pareto optimal solution set of the operating variables. The above-mentioned nitrogen oxide emission concentration index model and combustion efficiency index model are used as the optimization objective functions of the multi-objective particle swarm optimization algorithm based on a dual-file mechanism to evaluate the quality of the solutions in the population, specifically:

[0015] ① Parameter setting of multi-objective particle swarm optimization algorithm based on dual-archive mechanism: the particle swarm algorithm population size is set to 50, the external archives are divided into convergence archives and diversity archives, the size of each is set to 50, the maximum number of population iterations is set to 30, the maximum and minimum values ​​of inertia weight are set to 0.6 and 0.4 respectively, the maximum value of cognitive learning factor is 2.8, and the minimum value of social learning factor is 1.2;

[0016] ② Initialize the population and external archives of the multi-objective particle swarm optimization algorithm based on the dual-archive mechanism: randomly generate an initial population (including the speed and position of the particles), the speed of each particle in the population is initialized to 0, and the initial position is randomly generated in the decision variable search space, wherein the decision variable search space is determined based on the training set, and the position of each particle represents an optimized solution of the operating variable, and the target value of the solution corresponding to the particle is calculated by the above-mentioned indicator model; establish two empty archives as the convergence archive and the diversity archive, add the solution in the population to the convergence archive, and add the non-dominated solution in the population to the diversity archive;

[0017] ③ Determine the global optimal solution for updating the population: The solutions in the convergence archive and the diversity archive are used as parent solutions, and the offspring solutions are generated by simulating the genetic mating operation of binary crossover and polynomial mutation. The generated offspring solutions are used as the global optimal solution for updating the population;

[0018] ④ Update the speed and position of particles in the population, where the inertia weight, cognitive learning factor and social learning factor are adaptively adjusted based on the population evolution information;

[0019] ⑤ Use the updated population particles to maintain the convergence archive and diversity archive: add the solution in the population to the convergence archive, and ε The convergence index calculates the fitness value of the solutions in the archive, and sorts the solutions according to the fitness value. The worst individuals in the convergence archive will be iteratively deleted until the size of the convergence archive is met, and finally a convergence archive with a certain number of solutions is obtained; the non-dominated solutions in the population are added to the diversity archive. If the number of solutions in the diversity archive exceeds the archive size, the selection mechanism will be used to re-determine the solutions entering the diversity archive, where the boundary solutions with the maximum or minimum target value enter the diversity archive first, and then, the similarity of the solutions is measured by density estimation based on transfer and Euclidean distance, and the solutions with the greatest difference from the selected solutions are iteratively determined to enter the diversity archive until the size of the diversity archive is met;

[0020] ⑥ Repeat the above steps ③-⑤, that is, determine the global optimal solution for updating the population, update the speed and position of the particles in the population, and use the updated population particles to maintain the convergence file and the diversity file until the preset number of iterations is reached, and then output the Pareto optimization solution set corresponding to the particles in the diversity file;

[0021] (3) Based on the utility function method, the optimal setting values ​​of the primary air inlet flow rate and the secondary air inlet flow rate are determined from the obtained Pareto optimization solution set.

[0022] The creativity of the present invention is mainly reflected in:

[0023] (1) The present invention aims to improve the combustion efficiency and reduce the nitrogen oxide emission concentration in the municipal solid waste incineration process. On the basis of establishing a data-driven optimization target model, a multi-objective particle swarm optimization algorithm based on a dual-file mechanism is used to optimize the set values ​​of the primary air inlet flow rate and the secondary air inlet flow rate, thereby solving the complex problem of optimizing the setting of operating variables in the municipal solid waste incineration process. The present invention has the characteristics of significant optimization effect, and can achieve the nitrogen oxide emission compliance while maintaining a high combustion efficiency level;

[0024] (2) The present invention proposes a multi-objective particle swarm optimization algorithm based on a dual-file mechanism. Using the multi-objective particle swarm optimization algorithm as a carrier, a ε The convergence index and the dual-file update strategy based on transfer density estimation can preserve the elite solutions in a targeted manner while improving the diversity of evolution. A global optimal particle selection scheme based on genetic operation operators is established to further improve the search ability of the algorithm. The flight parameters of the particles are adjusted through the evolutionary information of the population to ensure the efficiency and robustness of the multi-objective particle swarm optimization algorithm based on the dual-file mechanism in the application of operation optimization of urban solid waste incineration process.

[0025] It should be noted in particular that the present invention uses the combustion efficiency and nitrogen oxide emission concentration of the urban solid waste incineration process as the optimized performance indicators, the primary air inlet flow rate and the secondary air inlet flow rate as the optimized operating variables, and adopts a multi-objective particle swarm optimization algorithm based on a dual-file mechanism to obtain the optimized set values ​​of the operating variables to achieve the optimization of the operating indicators. As long as the optimized performance indicators and optimization algorithms of the present invention are used to optimize the operation of the urban solid waste incineration process, they should fall within the scope of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 This is the data-driven multi-objective operation optimization framework of the municipal solid waste incineration process of the present invention

[0027] Figure 2 This is the combustion efficiency index model test result diagram of the present invention

[0028] Figure 3 This is the test result diagram of the nitrogen oxide emission concentration index model of the present invention

[0029] Figure 4 This is the flow chart of the multi-objective particle swarm optimization algorithm based on the dual-file mechanism of the present invention

[0030] Figure 5 This is the result diagram of the optimized setting value of the primary air inlet flow rate of the present invention

[0031] Figure 6 This is the result diagram of the secondary air inlet flow optimization setting value of the present invention

[0032] Figure 7 This is the combustion efficiency index optimization result diagram of the present invention

[0033] Figure 8 This is the optimization result diagram of the nitrogen oxide emission concentration index of the present invention DETAILED DESCRIPTION

[0034] The present invention obtains a data-driven multi-objective operation optimization method for a municipal solid waste incineration process. The method firstly captures the nonlinear characteristics of the municipal solid waste incineration process through a self-organizing radial basis function neural network, realizes accurate evaluation of the operation performance indicators combustion efficiency and nitrogen oxide emission concentration, and provides an accurate optimization target model for subsequent set value optimization solution; then, based on the established optimization target model, a multi-objective particle swarm optimization algorithm based on a dual-file mechanism is designed to obtain an optimization solution set of the primary air inlet flow rate and the secondary air inlet flow rate set values; according to the optimization solution set, a utility function is designed to determine the optimization set values ​​of the primary air inlet flow rate and the secondary air inlet flow rate, so as to realize the coordinated improvement of the combustion efficiency and the denitrification efficiency, thereby improving the intelligence level of the municipal solid waste incineration process and promoting the efficient and green operation of the municipal solid waste incineration process.

[0035] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the present invention.

[0036] A data-driven multi-objective operation optimization method for municipal solid waste incineration process is proposed. The municipal solid waste combustion efficiency and nitrogen oxide emission concentration are taken as optimization targets, and the primary air inlet flow rate and the secondary air inlet flow rate are taken as optimization operation variables. The data-driven multi-objective operation optimization framework is as follows: Figure 1 ;

[0037] The following steps are involved:

[0038] (1) 1080 sets of operation data of the municipal solid waste incineration process operation variables, optimization indicators (nitrogen oxide emission concentration and combustion efficiency) and other auxiliary variables (flue gas temperature on the right side of the primary combustion chamber, flue gas temperature on the left side of the primary combustion chamber, temperature on the right side of the primary combustion chamber, temperature on the left side of the primary combustion chamber, right air flow rate of the drying grate, left air flow rate of the drying grate, right air flow rate of the combustion section grate 1-1, left air flow rate of the combustion section grate 1-1, urea solution amount, lime feed amount, activated carbon storage bin feed amount, boiler outlet main steam flow rate, inlet flue gas oxygen content and furnace average temperature) were collected through detection equipment to construct a sample data set, and the data was divided into a training set and a test set, of which 810 sets of data were used as the training set and 270 sets of data were used as the test data set;

[0039] (2) Establishing an optimization index evaluation model based on a self-organizing radial basis function neural network according to the training set to characterize the mapping relationship between the operating variables and the optimized target combustion efficiency and nitrogen oxide emission concentration during the solid waste incineration process, wherein the model input variables are Including the operating variables primary air inlet flow rate and secondary air inlet flow rate and other auxiliary variables, where and Represent the optimized operating variables primary air inlet flow rate and secondary air inlet flow rate respectively, and the model output y r is the optimization index of nitrogen oxide emission concentration and combustion efficiency, where r = 1, 2, y 1 represents the combustion efficiency, y 2 Represents the nitrogen oxide emission concentration. Both optimization index models are established based on the self-organizing radial basis function neural network in the same way as shown below:

[0040] ①The initial number of neurons in the hidden layer is 0;

[0041] ② Calculate the network output of the training set samples,

[0042]

[0043] Among them, y r(x l ) is the lth training sample x in the training set l The network output, w h is the connection weight between the hth hidden layer neuron and the output layer neuron, φ h (x l ) is the hth hidden layer output of the lth training sample in the training set, and the calculation formula is

[0044]

[0045] Among them, c h and σ h are the center and width of the hth hidden layer neuron (h=1,2,…,H), respectively. ||·|| is used to calculate the Euclidean distance.

[0046] Calculate the error between the network output value and the expected output value,

[0047]

[0048] in, is the expected network output of the rth index of the lth training sample, and the network output error of all training samples is

[0049] e=[e1,e2,...,e l ,...,e L ] (4)

[0050] Among them, L represents the number of training samples, and its value is 810.

[0051] Determine the sample with the largest absolute error in the error vector,

[0052] e max (p) = arg{max||e||} (5)

[0053] The sample with the largest absolute error is determined to be the pth sample. A new hidden layer neuron is added to the self-organizing radial basis function neural network. The parameters of the newly added neuron, including the center and width of the neuron, are initialized based on the training sample corresponding to the maximum error value in the training set data to compensate for the current maximum absolute error. The center of the newly added kth neuron is initialized to

[0054] c k =x p (6)

[0055] Among them, x p is the training sample corresponding to the maximum error value, and the weight is set to the expected output of the training sample corresponding to the maximum error value

[0056]

[0057] The width of the newly added hidden layer neurons is determined by calculating the Euclidean distance between all hidden layer neurons of the self-organizing radial basis function neural network and the newly added kth hidden layer neuron.

[0058]

[0059] Among them, χ is a random number in the interval [0.4,0.6], dist(c k ,c i≠k ) means c k and c i≠k The Euclidean distance between represents the minimum Euclidean distance;

[0060] ③ Based on the second-order learning algorithm, the parameters of the self-organizing radial basis function neural network with a new hidden layer of neurons are adjusted. The update rule of the algorithm is:

[0061] Δ n+1 =Δ n -(Q n +λI) -1 g n (10)

[0062] Among them, λ is the learning coefficient, its value is set to 0.01, I is the identity matrix, Q n is a Hessian-like matrix, g n is the gradient vector, n represents the number of network update iterations, and its value is set to 30, Δ n Represents the parameters of the network to be updated, including the center and width of the hidden layer neurons, the connection weights between the hidden layer neurons and the output nodes, and the Hessian matrix and gradient vector are calculated according to formulas (11) and (12) respectively.

[0063]

[0064] Among them, j l is the Jacobian matrix row vector

[0065]

[0066] ④ Repeat the above steps ② and ③, that is, add a new hidden layer neuron based on the maximum network output error of the training sample and adjust the parameters of the self-organizing radial basis function neural network with a newly added hidden layer neuron, until the learning accuracy of the self-organizing radial basis function neural network reaches the expected training root mean square error of 0.001 or until the number of hidden layer neurons of the self-organizing radial basis function neural network reaches a preset number of 20, the training process of the self-organizing radial basis function neural network is completed, and the nitrogen oxide emission concentration index model and the combustion efficiency index model are constructed, and the expressions are as follows;

[0067]

[0068] Among them, y 1 (x) represents the output of the optimization index evaluation model for combustion efficiency, y 2 (x) represents the output of the optimization index evaluation model for nitrogen oxide emission concentration, x represents the input of the model, and the value of H is determined by the self-organization process of the above steps ② and ③;

[0069] The trained combustion efficiency and nitrogen oxide emission concentration models are tested using 270 groups of divided test data. The test samples of the test data set are used as the input of the combustion efficiency and nitrogen oxide emission concentration models, and the outputs are the predicted values ​​of combustion efficiency and nitrogen oxide emission concentration, respectively. The results are shown in the figure. Figure 2 and Figure 3 , Figure 2 and Figure 3 X-axis: test samples, in units of pieces, Figure 2 Y axis: combustion efficiency, in percentage, Figure 3 Y axis: NOx emission concentration, in mg / m 3 , Figure 2 and Figure 3 The dotted lines in the hollow circles are the actual outputs, and the solid lines in the solid circles are the network test outputs;

[0070] (3) Design a multi-objective particle swarm optimization algorithm based on a dual-file mechanism for the multi-objective operation optimization method of the municipal solid waste incineration process to obtain the Pareto optimal solution set of the operating variables. The flowchart of the algorithm is shown in the figure. Figure 4 As shown, the above-mentioned nitrogen oxide index model and combustion efficiency index model are used as the optimization objective function of the multi-objective particle swarm optimization algorithm based on the dual-file mechanism to evaluate the quality of the solutions in the population, specifically:

[0071] ① Parameter setting of multi-objective particle swarm optimization algorithm based on dual-archive mechanism: the particle swarm algorithm population size is set to 50, the external archives are divided into convergence archives and diversity archives, the size of each is set to 50, the maximum number of population iterations is set to 30, the maximum and minimum values ​​of inertia weight are set to 0.6 and 0.4 respectively, the maximum value of cognitive learning factor is 2.8, and the minimum value of social learning factor is 1.2;

[0072] ② Initialize the population and external archives of the multi-objective particle swarm optimization algorithm based on the dual-archive mechanism: randomly generate the initial population (including the speed and position of the particles),

[0073] v i (t) = [v i,1 (t),v i,2 (t)], z i (t) = [z i,1 (t),z i,2 (t)], i=1,2,...,50 (15)

[0074] Where t=1,2,...,30 represents the number of iterations, v i (t) and z i (t) represent the speed and position of the i-th particle in the t-th iteration population, respectively. The speed of each particle in the population is initialized to 0, and the initial position is randomly generated in the decision variable search space, where the decision variable search space is determined based on the training set. The position of each particle represents an optimized solution of the operating variable, and the target value of the solution corresponding to the particle is calculated by the above-mentioned indicator model, that is,

[0075] f i (t) = [f i 1 (t),f i 2 (t)],f1(t)=-y 1 (z i (t)),f2(t)=y 2 (z i (t)) (16)

[0076] where f i 1 (t) represents the combustion efficiency target value of the i-th particle, f i 2 (t) represents the target value of nitrogen oxide emission concentration of the i-th particle, y 1 (z i (t)) represents the output of the optimization index evaluation model for combustion efficiency, y 2 (z i(t)) represents the output of the optimization index evaluation model for nitrogen oxide emission concentration, and the model input is z i (t); Establish two empty archives as the convergence archive and the diversity archive, add the solutions in the population to the convergence archive, and add the non-dominated solutions in the population to the diversity archive;

[0077] ③ Determine the global optimal solution for updating the population: The solutions in the convergence archive and the diversity archive are used as parent solutions, and the offspring solutions are generated by simulating the genetic mating operation of binary crossover and polynomial mutation, with the parent solutions in the convergence archive as the parent solutions. and parent solutions in the diversity archive As an example, the detailed expressions of the simulated binary crossover operator are shown in formulas (17) and (18):

[0078]

[0079] in, and Represents two offspring individuals generated by the crossover operation. The parameter β in the formula is calculated by the following formula

[0080]

[0081] Where μ is a random number uniformly distributed in the interval [0,1], and η c is the crossover distribution index, whose value is 20. After the crossover operation, the newly generated offspring individuals have a mutation probability p m Perform mutation operation, p m The value is 0.5, based on the offspring individual Take as an example, the expression of the mutation operator is shown in formula (20):

[0082]

[0083] Where j = 1, 2, u j and l j Represent the maximum and minimum boundaries of the j-th dimension search space, respectively, based on the training set, the parameters Calculated by the following formula:

[0084]

[0085] where γ is a random number uniformly generated in the interval [0,1], and η m is the variation distribution index, whose value is 20. The final generated offspring solution is used as the global optimal solution of the updated population, that is,

[0086] ④ Update the speed and position of particles in the population,

[0087]

[0088] Among them, v i (t+1) and z i (t+1) is the velocity and position of the ith particle at the t+1th iteration, p i (t) and g i (t) represents the individual optimal and global optimal solutions of the ith particle, r1 and r2 are random numbers between 0 and 1; when t = 1, the inertia weight ω i (t) Randomly generate cognitive learning factor c in the interval [0.4, 0.6] 1,i (t) and social learning factor c 2,i (t) is randomly generated in the interval [1.2, 2.8]; when t>1, the inertia weight and learning factor are adaptively adjusted based on the population evolution information. First, the evolution information of the particles in the population is calculated, and the calculation formula is as follows:

[0089]

[0090]

[0091] Where M = 2 represents the number of targets. and Φ i (t) represent the evolutionary ability of the population and the evolutionary ability of the i-th particle in the population at the t-th iteration, respectively, and E i (t+1) represents the evolution efficiency of the i-th particle in the t+1th iteration of the population, and They represent the mth target value of the global optimal solution of the ith particle at the tth and t-1th iterations, respectively, m = 1, ..., M, and the particles in the population are sorted from small to large according to the target values ​​of each dimension. and They represent the mth target value of the second and second to last particle in the population at the tth iteration, respectively, and f i m (t) represents the mth target value of the ith particle in the population at the tth iteration. All the above target values ​​are normalized. Based on the population evolution information, i.e., population evolution ability, particle evolution ability and particle evolution efficiency, the self-organization parameter of the ith particle can be calculated by formula (26):

[0092]

[0093] Among them, Θ i (t+1) represents the self-organization parameter of the ith particle at the t+1th iteration, ε is a constant with a value of 0.01, and the inertia weight ω i (t), learning factor c 1,i(t) and c 2,i The adjustment strategy of (t) is shown in formulas (27), (28) and (29):

[0094]

[0095] Among them, ω min ,ω max Represent the maximum and minimum values ​​of the inertia weight, which are 0.6 and 0.4 respectively, c 1,max It represents the maximum value of cognitive learning factor, which is 2.8, c 2,min represents the minimum value of the social learning factor, which is 1.2, and T represents the maximum number of iterations of the population, which is 30;

[0096] ⑤ Use the updated population particles to maintain the convergence archive and diversity archive: add the solution in the population to the convergence archive, and ε Convergence index calculates the fitness value of the solution in the archive,

[0097]

[0098] Among them, z1 and z2 represent any two solutions in the population, I ε (z1,z2) represents the convergence index, f1 m (t) represents the mth target value of z1 at the tth iteration, represents the mth target value of z2 at the tth iteration. All the above target values ​​are normalized. The fitness value of z1 is calculated by the following formula:

[0099]

[0100] Among them, F(z1) represents the fitness value of z1, P represents the population, α is a constant with a value of 0.01, and the fitness values ​​of all solutions in the population are calculated by formula (31). The solutions in the population are sorted according to the fitness values, and the worst individuals in the convergence file will be deleted iteratively until the size of the convergence file is met. Finally, a convergence file with 50 solutions is obtained;

[0101] The non-dominated solutions in the population are added to the diversity archive. If the number of solutions in the diversity archive exceeds the archive size, a selection mechanism is used to redetermine the solutions that enter the diversity archive. The boundary solutions with the maximum or minimum target value enter the diversity archive first. Then, the similarity of the solutions is measured by the transfer-based density estimation and Euclidean distance. The transfer-based density estimation method calculates the transfer target value of the solution z1 by the following formula:

[0102]

[0103] Among them, f1 m (t) and Respectively represent the mth target value of z1 and z2 at the tth iteration, f1 m′ (t) represents the target value after the transfer of solution z1; based on the target value after the transfer, the solution with the greatest difference from the selected solution is iteratively determined by Euclidean distance to enter the diversity archive until the size of the diversity archive is met;

[0104] ⑥ Repeat the above steps ③-⑤, that is, determine the global optimal solution for updating the population, update the speed and position of the particles in the population, and use the updated population particles to maintain the convergence file and the diversity file, until the preset number of iterations reaches 30, and then output the Pareto optimization solution set corresponding to the particles in the diversity file;

[0105] (3) According to the utility function method, the optimal setting values ​​of the primary air inlet flow rate and the secondary air inlet flow rate are determined from the Pareto optimization solution set obtained. The calculation method is as follows:

[0106]

[0107] Among them, U i is the utility function value of the i-th solution in the optimization solution set, K represents the number of solutions in the optimization solution set, and f i m is the mth objective value of the ith solution in the optimization solution set, μ m is the weight value of the mth objective obtained according to the decision preference, the weight values ​​of the two objectives are set to 0.5, and the solution with the minimum utility function value is determined as the optimal setting value of the primary air inlet flow rate and the secondary air inlet flow rate; Figure 5 and Figure 6 The optimized setting results of the primary air inlet flow rate and the secondary air inlet flow rate are shown respectively. Figure 5 and Figure 6 X-axis: time, in minutes, Figure 5 Y axis: primary air inlet flow, unit is km 3 N / h, Figure 6 Y axis: secondary air inlet flow rate, unit is km 3 N / h; Using the optimized set values ​​of primary air inlet flow rate and secondary air inlet flow rate, the optimization results of combustion efficiency and nitrogen oxide emission concentration performance indicators are obtained, such as Figure 7 and Figure 8 , Figure 7 X-axis: time, in minutes, Figure 7 Y axis: combustion efficiency, in percentage, Figure 8 X-axis: time, in minutes, Figure 8 Y axis: NOx emission concentration, in mg / m 3 ; Figure 7The hollow circle is the combustion efficiency before optimization, and the asterisk is the combustion efficiency result after optimization. Figure 8 The hollow circles represent the nitrogen oxide emission concentration before optimization, and the asterisks represent the nitrogen oxide emission concentration after optimization; the results prove the effectiveness of this method.

Claims

1. Data-driven multi-objective operation optimization method for municipal solid waste incineration process, characterized by The following steps are involved: (1) The operating variables of the municipal solid waste incineration process, the optimization index, i.e., the nitrogen oxide emission concentration and the combustion efficiency, and multiple groups of operation data of other auxiliary variables are collected through the detection equipment to construct a sample data set, and the training set and the test set are divided; the other auxiliary variables are the flue gas temperature on the right side of the primary combustion chamber, the flue gas temperature on the left side of the primary combustion chamber, the temperature on the right side of the primary combustion chamber, the temperature on the left side of the primary combustion chamber, the right 1 air flow rate of the drying grate, the left 1 air flow rate of the drying grate, the right 1-1 air flow rate of the combustion section grate, the left 1-1 air flow rate of the combustion section grate, the amount of urea solution, the amount of lime feed, the amount of activated carbon storage bin feed, the main steam flow rate at the boiler outlet, the oxygen content of the inlet flue gas, and the average furnace temperature; (2) Establishing an optimization index evaluation model based on a self-organizing radial basis function neural network according to the training set to characterize the mapping relationship between the operating variables and the optimized target combustion efficiency and nitrogen oxide emission concentration in the solid waste incineration process, wherein the model input variables include the operating variables primary air inlet flow rate and secondary air inlet flow rate and other auxiliary variables, and the model output is the optimized index nitrogen oxide emission concentration and combustion efficiency, which is established as follows: ①The initial number of neurons in the hidden layer is 0; ② Calculate the network output of the training set samples, calculate the error between the network output value and the expected output value, determine the sample with the largest absolute error in the error vector, add a new hidden layer neuron in the self-organizing radial basis function neural network, and initialize the parameters of the newly added neuron based on the training sample corresponding to the maximum error value in the training set data, including the center and width of the neuron, to compensate for the current maximum absolute error; ③ Based on the second-order learning algorithm, the parameters of the self-organizing radial basis function neural network with a newly added hidden layer of neurons are adjusted to achieve the purpose of training the network; ④ Repeat the above steps ② and ③, that is, add a new hidden layer neuron based on the maximum network output error of the training sample and adjust the parameters of the self-organizing radial basis function neural network with a newly added hidden layer neuron, until the learning accuracy of the self-organizing radial basis function neural network reaches the expected training error or until the number of hidden layer neurons of the self-organizing radial basis function neural network reaches a preset number, the nitrogen oxide emission concentration index model and the combustion efficiency index model are constructed; (3) Design a multi-objective particle swarm optimization algorithm based on a dual-file mechanism for the multi-objective operation optimization method of the municipal solid waste incineration process, and obtain the Pareto optimal solution set of the operating variables. The above-mentioned nitrogen oxide emission concentration index model and combustion efficiency index model are used as the optimization objective functions of the multi-objective particle swarm optimization algorithm based on a dual-file mechanism to evaluate the quality of the solutions in the population, specifically: ① Parameter setting of multi-objective particle swarm optimization algorithm based on dual-archive mechanism: the particle swarm algorithm population size is set to 50, the external archives are divided into convergence archives and diversity archives, the size of each is set to 50, the maximum number of population iterations is set to 30, the maximum and minimum values ​​of inertia weight are set to 0.6 and 0.4 respectively, the maximum value of cognitive learning factor is 2.8, and the minimum value of social learning factor is 1.2; ② Initialize the population and external archives of the multi-objective particle swarm optimization algorithm based on the dual-archive mechanism: randomly generate an initial population (including the speed and position of the particles), the speed of each particle in the population is initialized to 0, and the initial position is randomly generated in the decision variable search space, wherein the decision variable search space is determined based on the training set, and the position of each particle represents an optimized solution of the operating variable, and the target value of the solution corresponding to the particle is calculated by the above-mentioned indicator model; establish two empty archives as the convergence archive and the diversity archive, add the solution in the population to the convergence archive, and add the non-dominated solution in the population to the diversity archive; ③ Determine the global optimal solution for updating the population: The solutions in the convergence archive and the diversity archive are used as parent solutions, and the offspring solutions are generated by simulating the genetic mating operation of binary crossover and polynomial mutation. The generated offspring solutions are used as the global optimal solution for updating the population; ④ Update the speed and position of particles in the population, where the inertia weight, cognitive learning factor and social learning factor are adaptively adjusted based on the population evolution information; ⑤ Use the updated population particles to maintain the convergence archive and diversity archive: add the solution in the population to the convergence archive, and ε The convergence index calculates the fitness value of the solutions in the archive, and sorts the solutions according to the fitness value. The worst individuals in the convergence archive will be iteratively deleted until the size of the convergence archive is met, and finally a convergence archive is obtained; the non-dominated solutions in the population are added to the diversity archive. If the number of solutions in the diversity archive exceeds the archive size, the selection mechanism will be used to redetermine the solutions entering the diversity archive, where the boundary solutions with the maximum or minimum target value enter the diversity archive first, and then, the similarity of the solutions is measured by density estimation based on transfer and Euclidean distance, and the solutions with the greatest difference from the selected solutions are iteratively determined to enter the diversity archive until the size of the diversity archive is met; ⑥ Repeat the above steps ③-⑤, that is, determine the global optimal solution for updating the population, update the speed and position of the particles in the population, and use the updated population particles to maintain the convergence file and the diversity file until the preset number of iterations is reached, and then output the Pareto optimization solution set corresponding to the particles in the diversity file; (3) Based on the utility function method, the optimal setting values ​​of the primary air inlet flow rate and the secondary air inlet flow rate are determined from the obtained Pareto optimization solution set.

2. The data-driven multi-objective operation optimization method for municipal solid waste incineration process according to claim 1 is characterized in that The following steps are involved: (1) The operating variables of the municipal solid waste incineration process, the optimization index, i.e., the nitrogen oxide emission concentration and the combustion efficiency, and multiple groups of operation data of other auxiliary variables are collected through the detection equipment to construct a sample data set, and the training set and the test set are divided; the other auxiliary variables are the flue gas temperature on the right side of the primary combustion chamber, the flue gas temperature on the left side of the primary combustion chamber, the temperature on the right side of the primary combustion chamber, the temperature on the left side of the primary combustion chamber, the right 1 air flow rate of the drying grate, the left 1 air flow rate of the drying grate, the right 1-1 air flow rate of the combustion section grate, the left 1-1 air flow rate of the combustion section grate, the amount of urea solution, the amount of lime feed, the amount of activated carbon storage bin feed, the main steam flow rate at the boiler outlet, the oxygen content of the inlet flue gas, and the average furnace temperature; (2) Establishing an optimization index evaluation model based on a self-organizing radial basis function neural network according to the training set to characterize the mapping relationship between the operating variables and the optimized target combustion efficiency and nitrogen oxide emission concentration during the solid waste incineration process, wherein the model input variables are Including the operating variables primary air inlet flow rate and secondary air inlet flow rate and other auxiliary variables, where and Represent the optimized operating variables primary air inlet flow rate and secondary air inlet flow rate respectively, and the model output y r is the optimization index of nitrogen oxide emission concentration and combustion efficiency, where r = 1, 2, y 1 represents the combustion efficiency, y 2 Represents the nitrogen oxide emission concentration. Both optimization index models are established based on the self-organizing radial basis function neural network in the same way as shown below: ①The initial number of neurons in the hidden layer is 0; ② Calculate the network output of the training set samples, Among them, y r (x l ) is the lth training sample x in the training set l The network output, w h is the connection weight between the hth hidden layer neuron and the output layer neuron, φ h (x l ) is the hth hidden layer output of the lth training sample in the training set, and the calculation formula is Among them, c h and σ h are the center and width of the hth hidden layer neuron (h=1,2,…,H), respectively, used to calculate the Euclidean distance, Calculate the error between the network output value and the expected output value, in, is the expected network output of the rth index of the lth training sample, and the network output error of all training samples is e=[e1,e2,...,e l ,...,the L ] (4) Among them, L represents the number of training samples, and its value is 810. Determine the sample with the largest absolute error in the error vector, it is max (p)=arg{max||e||} (5) The sample with the largest absolute error is determined to be the pth sample. A new hidden layer neuron is added to the self-organizing radial basis function neural network. The parameters of the newly added neuron, including the center and width of the neuron, are initialized based on the training sample corresponding to the maximum error value in the training set data to compensate for the current maximum absolute error. The center of the newly added kth neuron is initialized to c k =x p (6) Among them, x p is the training sample corresponding to the maximum error value, and the weight is set to the expected output of the training sample corresponding to the maximum error value The width of the newly added hidden layer neurons is determined by calculating the Euclidean distance between all hidden layer neurons of the self-organizing radial basis function neural network and the newly added kth hidden layer neuron. Among them, χ is a random number in the interval [0.4,0.6], dist(c k ,c i≠k ) means c k and c i≠k The Euclidean distance between represents the minimum Euclidean distance; ③ Based on the second-order learning algorithm, the parameters of the self-organizing radial basis function neural network with a new hidden layer of neurons are adjusted. The update rule of the algorithm is: D n+1 =D n -(Q n +λI) -1 g n (10) Among them, λ is the learning coefficient, its value is set to 0.01, I is the identity matrix, Q n is a Hessian-like matrix, g n is the gradient vector, n represents the number of network update iterations, and its value is set to 30, Δ n Represents the parameters of the network to be updated, including the center and width of the hidden layer neurons, the connection weights between the hidden layer neurons and the output nodes, and the Hessian matrix and gradient vector are calculated according to formulas (11) and (12) respectively. Among them, j l is the Jacobian matrix row vector ④ Repeat the above steps ② and ③, that is, add a new hidden layer neuron based on the maximum network output error of the training sample and adjust the parameters of the self-organizing radial basis function neural network with a newly added hidden layer neuron, until the learning accuracy of the self-organizing radial basis function neural network reaches the expected training root mean square error of 0.001 or until the number of hidden layer neurons of the self-organizing radial basis function neural network reaches a preset number of 20, the training process of the self-organizing radial basis function neural network is completed, and the nitrogen oxide emission concentration index model and the combustion efficiency index model are constructed, and the expressions are as follows; Among them, y 1 (x) represents the output of the optimization index evaluation model for combustion efficiency, y 2 (x) represents the output of the optimization index evaluation model for nitrogen oxide emission concentration, and x represents the input of the model; The trained combustion efficiency and nitrogen oxide emission concentration models are tested using test data, and the test samples of the test data set are used as the input of the combustion efficiency and nitrogen oxide emission concentration models, and the outputs are the predicted values ​​of combustion efficiency and nitrogen oxide emission concentration respectively; (3) Design a multi-objective particle swarm optimization algorithm based on a dual-file mechanism for the multi-objective operation optimization method of the municipal solid waste incineration process, and obtain the Pareto optimal solution set of the operating variables. The nitrogen oxide index model and the combustion efficiency index model are used as the optimization objective functions of the multi-objective particle swarm optimization algorithm based on a dual-file mechanism to evaluate the quality of the solutions in the population. Specifically: ① Parameter setting of multi-objective particle swarm optimization algorithm based on dual-archive mechanism: the particle swarm algorithm population size is set to 50, the external archives are divided into convergence archives and diversity archives, the size of each is set to 50, the maximum number of population iterations is set to 30, the maximum and minimum values ​​of inertia weight are set to 0.6 and 0.4 respectively, the maximum value of cognitive learning factor is 2.8, and the minimum value of social learning factor is 1.2; ② Initialize the population and external archives of the multi-objective particle swarm optimization algorithm based on the dual-archive mechanism: randomly generate the initial population, including the speed and position of the particles; v i (t)=[v i,1 (t),v i,2 (t)],z i (t)=[z i,1 (t),z i,2 (t)],i=1,2,...,50 (15) Where t=1,2,...,30 represents the number of iterations, v i (t) and z i (t) represent the speed and position of the i-th particle in the t-th iteration population, respectively. The speed of each particle in the population is initialized to 0, and the initial position is randomly generated in the decision variable search space, where the decision variable search space is determined based on the training set. The position of each particle represents an optimized solution of the operating variable, and the target value of the solution corresponding to the particle is calculated by the above-mentioned indicator model, that is, f i (t)=[f i 1 (t),f i 2 (t)],f1(t)=-y 1 (z i (t)),f2(t)=y 2 (z i (t)) (16) where f i 1 (t) represents the combustion efficiency target value of the i-th particle, f i 2 (t) represents the target value of nitrogen oxide emission concentration of the i-th particle, y 1 (z i (t)) represents the output of the optimization index evaluation model for combustion efficiency, y 2 (z i (t)) represents the output of the optimization index evaluation model for nitrogen oxide emission concentration, and the model input is z i (t); Establish two empty archives as the convergence archive and the diversity archive, add the solutions in the population to the convergence archive, and add the non-dominated solutions in the population to the diversity archive; ③ Determine the global optimal solution for updating the population: The solutions in the convergence archive and the diversity archive are used as parent solutions, and the offspring solutions are generated by simulating the genetic mating operation of binary crossover and polynomial mutation, with the parent solutions in the convergence archive as the parent solutions. and the parent solution in the diversity archive The detailed expressions of the simulated binary crossover operator are shown in formula (17) and formula (18): in, and Represents two offspring individuals generated by the crossover operation. The parameter β in the formula is calculated by the following formula Where μ is a random number uniformly distributed in the interval [0,1], and η c is the crossover distribution index, whose value is 20. After the crossover operation, the newly generated offspring individuals have a mutation probability p m Perform mutation operation, p m The value is 0.5, based on the offspring individual The expression of the mutation operator is shown in formula (20): Where j = 1, 2, u j and l j Represent the maximum and minimum boundaries of the j-th dimension search space, respectively, and are determined based on the training set. The parameter ξ i1,j Calculated by the following formula: where γ is a random number uniformly generated in the interval [0,1], and η m is the variation distribution index, whose value is 20. The final generated offspring solution is used as the global optimal solution of the updated population, that is, ④ Update the speed and position of particles in the population, Among them, v i (t+1) and z i (t+1) is the velocity and position of the ith particle at the t+1th iteration, p i (t) and g i (t) represents the individual optimal and global optimal solutions of the ith particle, r1 and r2 are random numbers between 0 and 1; when t = 1, the inertia weight ω i (t) Randomly generate cognitive learning factor c in the interval [0.4, 0.6] 1,i (t) and social learning factor c 2,i (t) is randomly generated in the interval [1.2, 2.8]; when t>1, the inertia weight and learning factor are adaptively adjusted based on the population evolution information. First, the evolution information of the particles in the population is calculated, and the calculation formula is as follows: Where M = 2 represents the number of targets. and Φ i (t) represent the evolutionary ability of the population and the evolutionary ability of the i-th particle in the population at the t-th iteration, respectively, and E i (t+1) represents the evolution efficiency of the i-th particle in the t+1th iteration of the population, and They represent the mth target value of the global optimal solution of the ith particle at the tth and t-1th iterations, respectively, m = 1, ..., M, and the particles in the population are sorted from small to large according to the target values ​​of each dimension. and They represent the mth target value of the second and second to last particle in the population at the tth iteration, respectively, and f i m (t) represents the mth target value of the ith particle in the population at the tth iteration. All the above target values ​​are normalized. Based on the population evolution information, i.e., population evolution ability, particle evolution ability and particle evolution efficiency, the self-organization parameter of the ith particle can be calculated by formula (26): Among them, Θ i (t+1) represents the self-organization parameter of the ith particle at the t+1th iteration, ε is a constant with a value of 0.01, and the inertia weight ω i (t), learning factor c 1,i (t) and c 2,i The adjustment strategy of (t) is shown in formulas (27), (28) and (29): Among them, ω min ,ω max Represent the maximum and minimum values ​​of the inertia weight, which are 0.6 and 0.4 respectively, c 1,max It represents the maximum value of cognitive learning factor, which is 2.8, c 2,min represents the minimum value of the social learning factor, which is 1.2, and T represents the maximum number of iterations of the population, which is 30; ⑤ Use the updated population particles to maintain the convergence archive and diversity archive: add the solution in the population to the convergence archive, and ε Convergence index calculates the fitness value of the solution in the archive, Among them, z1 and z2 represent any two solutions in the population, I ε (z1,z2) represents the convergence index, f1 m (t) represents the mth target value of z1 at the tth iteration, represents the mth target value of z2 at the tth iteration. All the above target values ​​are normalized. The fitness value of z1 is calculated by the following formula: Where F(z1) represents the fitness value of z1, P represents the population, α is a constant with a value of 0.01, and the fitness values ​​of all solutions in the population are calculated by formula (31). The solutions in the population are sorted according to the fitness values, and the worst individuals in the convergence file will be deleted iteratively to obtain a convergence file with 50 solutions; The non-dominated solutions in the population are added to the diversity archive. If the number of solutions in the diversity archive exceeds the archive size, a selection mechanism is used to redetermine the solutions that enter the diversity archive. The boundary solutions with the maximum or minimum target value enter the diversity archive first. Then, the similarity of the solutions is measured by the transfer-based density estimation and Euclidean distance. The transfer-based density estimation method calculates the transfer target value of the solution z1 by the following formula: Among them, f1 m (t) and Respectively represent the mth target value of z1 and z2 at the tth iteration, f1 m’ (t) represents the target value after the transfer of solution z1; based on the target value after the transfer, the solution with the greatest difference from the selected solution is iteratively determined by Euclidean distance to enter the diversity archive until the size of the diversity archive is met; ⑥ Repeat the above steps ③-⑤, that is, determine the global optimal solution for updating the population, update the speed and position of the particles in the population, and use the updated population particles to maintain the convergence file and the diversity file, until the preset number of iterations reaches 30, and then output the Pareto optimization solution set corresponding to the particles in the diversity file; (3) According to the utility function method, the optimal setting values ​​of the primary air inlet flow rate and the secondary air inlet flow rate are determined from the Pareto optimization solution set obtained. The calculation method is as follows: Among them, U i is the utility function value of the i-th solution in the optimization solution set, K represents the number of solutions in the optimization solution set, and f i m is the mth objective value of the ith solution in the optimal solution set, μ m The weight value of the mth objective is obtained according to the decision preference. The weight values ​​of the two objectives are set to 0.5, and the solution with the minimum utility function value is determined as the optimal setting value of the primary air inlet flow rate and the secondary air inlet flow rate.

Images