A thermal nonlinear system identification method, system, medium, device and terminal

Abstract

The present application belongs to the technical field of thermal process nonlinear system identification of thermal power units, and discloses a thermal nonlinear system identification method, system, medium, equipment and terminal, obtains a sample set for model thermal nonlinear system identification, divides the sample set into two parts of a training set and a test set; the training set sample is used as an input sample for model training, training and reasoning are performed through a fuzzy width learning system; three parameters of a fuzzy rule Nr, a fuzzy subsystem Nf and an enhanced node Ne of the fuzzy width learning system are searched through a node fitness particle swarm algorithm; and prediction is performed on the optimal fuzzy width learning system structure searched through optimization according to the test set sample. The present application has the advantages of simple and efficient calculation process, simultaneous use for neural network approximation and fuzzy reasoning, fast determination of the optimal model structure, effective improvement of the model nonlinear system identification capability, and strong application value in the nonlinear system identification task.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of thermal process nonlinear system identification of thermal power generating units, and particularly relates to a thermal nonlinear system identification method, system, medium, equipment and terminal. BACKGROUND

[0002] At present, system identification is one part of the field of modern control science and engineering, and is used for researching and establishing a technical method of a system based on a mathematical model. In terms of practice, system identification needs to select an effective model which can well fit the input-output relationship in the actual system according to the input and output. Essentially, system identification is a kind of optimization problem, and the identification method can convert the identification problem into a parameter estimation problem by establishing a system parameter model. The main difficulty of the system identification method is that it is difficult to be used for nonlinear systems, and good identification effect can be achieved for linear systems. In actual life, most systems are nonlinear due to the influence of various factors, so nonlinear system identification is an important and difficult task, and has attracted widespread attention from researchers.

[0003] Nonlinear systems are closely related to daily production and life, and the current research on nonlinear system identification is still in the development stage. Establishing an effective model to describe nonlinear systems is the basis for researching nonlinear problems, and the common research method for nonlinear systems mainly ignores some nonlinear conditions or uses linear relationships to approximate nonlinear relationships, so as to solve the problem by using linear system identification theory. With the rapid development of science and technology, the linear system identification theory gradually cannot meet the requirements of control precision of complex control systems. Many complex industrial processes, including thermal production processes of thermal power plants, are sometimes seriously affected by the dynamic characteristics identification of the object due to the random changes of a large number of random characteristics of the on-site equipment or the random changes of the external influencing factors of the process. How to use system engineering methods to analyze the basic characteristics of the object, the external disturbance characteristics and the equipment disturbance characteristics in the process of dynamic characteristic identification of the object, and to seek the dynamic characteristic data which can best represent the dynamic characteristics of the object, and then to accurately analyze the main dynamic characteristics of the object by using intelligent optimization algorithms is the main work content of the present application.

[0004] With the continuous development of artificial intelligence technology, methods represented by fuzzy neural network models have extremely wide and far-reaching development in pattern recognition, machine learning, intelligent computing and the like. The fuzzy neural network model has strong nonlinear approximation capability. With the help of fuzzy set theory and nonlinear approximator, the fuzzy width learning system can have the characteristics of fuzzy reasoning and self-adaptation, and can fit the complex input-output relationship in the nonlinear system. However, the existing fuzzy neural network models for nonlinear system identification generally have the characteristics of low calculation efficiency, poor identification effect and weak generalization ability, and cannot fully meet the precision requirements of nonlinear system identification.

[0005] The fuzzy width learning system is a novel fuzzy neural network. Compared with the existing fuzzy neural network, the fuzzy width learning system inherits the advantages of the flat structure of the width learning system, and can effectively improve the performance of the model through the horizontal expansion of the fuzzy subsystem and the enhanced node. Thanks to the structural advantages, the fuzzy width learning system can effectively fit the input-output relationship of the nonlinear system, thereby fully improving the recognition ability of the model. Although the fuzzy width learning system has achieved competitive results in nonlinear system identification, the model contains multiple hyperparameters that need to be adjusted, which directly affects the stability of the fuzzy width learning system in nonlinear system identification, and the performance of the model needs to be further improved. Therefore, when facing increasingly complex nonlinear systems, the nonlinear system identification effect of the fuzzy width learning system is not optimal without determining the optimal structure of the fuzzy width learning system.

[0006] Through the above analysis, the problems and defects of the prior art are:

[0007] (1) The existing fuzzy neural network model for nonlinear system identification generally has low computational efficiency, poor recognition effect, weak generalization ability and other characteristics, and cannot fully meet the accuracy and speed requirements of nonlinear system identification.

[0008] (2) The fuzzy width learning system has better approximation ability than other neural fuzzy models, but when the fuzzy width learning system model is used for nonlinear system identification, multiple hyperparameters need to be adjusted to obtain the best performance. At present, there is still a lack of heuristic algorithm for parameter optimization of the fuzzy width learning system, so it cannot meet the accuracy and speed requirements in actual engineering applications.

[0009] (3) When facing complex nonlinear systems with disturbances, the nonlinear system identification effect of the fuzzy width learning system is not optimal without determining the optimal structure of the fuzzy width learning system. SUMMARY

[0010] In view of the model identification problem of the existing power plant thermal nonlinear system, the present application provides a thermal nonlinear system identification method, system, medium, equipment and terminal, especially a thermal nonlinear system identification method, system, medium, equipment and terminal based on NSFPSO-FBLS, which aims to overcome the shortcomings of the fuzzy width learning system in solving the thermal nonlinear system identification task.

[0011] The application is achieved by a thermal nonlinear system identification method, which comprises the following steps: constructing an input-output relationship of a thermal nonlinear system by a fuzzy width learning system, updating weights of the fuzzy width learning system by using a ridge regression algorithm, and searching for an optimal fuzzy width learning system structure by using an NSFPSO algorithm (node fitness particle swarm optimization algorithm), so as to realize identification of the thermal nonlinear system.

[0012] Further, the thermal nonlinear system identification method comprises the following steps:

[0013] Step one: simulation is performed on a main steam pressure system and a main steam temperature system, a sample set for model thermal nonlinear system identification is obtained, and the sample set is divided into a training set and a test set;

[0014] Step two: training set samples are used as input samples for model training, reasoning is performed by a fuzzy subsystem in the fuzzy width learning system, a ridge regression algorithm is used to update weights of the fuzzy width learning system, and three parameters of the fuzzy width learning system, i.e., a fuzzy rule Nr, a fuzzy subsystem Nf and an enhanced node Ne, are searched by using the NSFPSO algorithm, so as to determine an optimal model structure;

[0015] Step three: according to test set samples, prediction is performed by using the optimal fuzzy width learning system structure searched and optimized, a root mean square error RMSE index is used to evaluate the prediction effect, and a model test time is evaluated.

[0016] Further, in the step one, 500 training samples and 200 test samples are respectively generated according to different nonlinear system formulas; the 500 training samples are used to train an FBLS model, and the 200 test samples are used to test the FBLS model.

[0017] Further, the fuzzy width learning system in the step two comprises a TSK fuzzy subsystem, an enhanced layer and an output layer;

[0018] An input sample number and dimension of the TSK fuzzy subsystem are respectively N and m, and are expressed as:

[0019]

[0020] An output of the fuzzy subsystem is wherein is a weighted intensity based on a Gaussian kernel function, is a result of an s-th fuzzy rule;

[0021] An output of the enhanced layer is H=(H1, H2,..., H l); where l(q = 1, 2,..., l) denotes the number of enhanced node groups; the output of the qth enhanced node group is denoted as H q = ε(ZW eq + b q ); W eq and b q are randomly generated matrices and bias terms; ε represents a nonlinear activation function of the enhanced layer;

[0022] The output of the output layer is where W df is a weight matrix of the fuzzy subsystem to the output layer, and W h is a weight matrix of the enhanced layer to the output layer.

[0023] Further, the step two is that the training set samples are taken as input samples for model training, and are trained and inferred by the fuzzy width learning system; and the fuzzy rule Nr, the fuzzy subsystem Nf, and the enhanced node Ne of the fuzzy width learning system are searched by the node fitness particle swarm optimization algorithm, including:

[0024] (1) initialization of population operation, population size N, learning factor c1 and c2, inertia weight w, and velocity v; where the position information x of different particles includes the number of fuzzy rules, the number of TSK fuzzy subsystems, and the number of enhanced nodes;

[0025] (2) for the tth iteration, the fitness function of the NSFPSO is calculated; the fitness of each particle is calculated by using the node sensitive fitness function according to the given information of the particle; the current individual optimal value pbest and the population optimal value gbest are obtained by comparing the fitness value of each particle and the current optimal global fitness value; the node sensitive fitness function can be expressed as:

[0026] sum = Nr + Nf + Ne + 1

[0027]

[0028]

[0029] where the fitness function fitness is composed of two indicators, the mean square error RMSE and the node sensitivity parameters Nr, Nf, and Ne; Y and are the expected output value and the actual output value of the fuzzy width learning system respectively; α, β, ε, γ, and θ are the weighting coefficients of the fitness function;

[0030] (3) determine whether the iteration termination condition is reached - determine whether the maximum number of iterations is reached, if the maximum number of iterations is not reached, the particle position information and the velocity information are obtained;

[0031] (4) update the position information and velocity information of each particle, wherein the position information is updated as:

[0032] w t = (w ini -w end )(G max -g) / G max +w end ;

[0033] wherein G max is the maximum number of iterations, w ini is the initial inertia weight, w end is the inertia weight corresponding to the maximum number of iterations, and the velocity information is updated as:

[0034] v i = w x v i + c1 x rand x (pbest i -x i ) + c2 x rand x (gbest i -x i ) ;

[0035] wherein v i is the velocity of the particle, i = 1, 2,..., N is the number of particles; w is the inertia factor for balancing the local search ability and the global search ability; and rand is a random number between 0 and 1.

[0036] (5) update the position information x of the t+1 iteration according to the velocity information obtained in step (3):

[0037] x t+1 = x t + v t+1 ;

[0038] (6) train the fuzzy width learning system by the ridge regression algorithm according to the particle position information obtained in step (5), and return to step (2).

[0039] (7) according to step (3), if the maximum number of iterations is reached, output the best particle information - the best parameter combination, and finally select the optimal number of nodes as the output result.

[0040] Further, the initial particle population number is 30, the learning factors c1 and c2 are both 2.1, the inertia weights w ini and w end are 0.9 and 0.3 respectively, and the initial values of the fitness weighting parameters a, b, e, g and q are 0.2, 0.5, 0.3, 0.2 and 0.8 respectively.

[0041] In the step (6), the training of the fuzzy width learning system by the ridge regression algorithm further comprises:

[0042] 1) a random initialization operation of randomly generating a fuzzy subsystem deblurring weight

[0043] 2) obtaining a deblurring output of the fuzzy subsystem according to the formula wherein F ip is an output of the fuzzy subsystem, is a weighted intensity based on a Gaussian kernel function, is a deblurring random weight, is a result of a fuzzy rule;

[0044] 3) calculating the deblurring output of different fuzzy subsystems according to the formula

[0045] 4) calculating an output of the enhancement layer according to the formula H q = epsilon (ZW eq + b q );

[0046] 5) calculating a formula of the fuzzy width learning system FBLS weight trained according to the ridge regression algorithm W FBLS = (Z, H) + Y = (lambda I + (Z, H) T (Z, H)) -1 (Z, H) T Y calculates the weight of the fuzzy width learning system, wherein Y is an expected output of the model.

[0047] wherein the number of nodes of the initial fuzzy subsystem is 4, the number of nodes of the fuzzy rule is 10, and the number of nodes of the enhancement node is 6.

[0048] Another purpose of the present application is to provide a thermal nonlinear system identification system applying the thermal nonlinear system identification method, which comprises:

[0049] a sample set acquisition module, configured to obtain a sample set for model thermal nonlinear system identification, and divide the sample set into two parts of a training set and a test set;

[0050] a model training module, configured to take the training set sample as an input sample of model training, and train and infer by a fuzzy width learning system;

[0051] a parameter search module, configured to search three parameters of a fuzzy rule Nr, a fuzzy subsystem Nf and an enhancement node Ne of the fuzzy width learning system by a node fitness particle swarm algorithm;

[0052] ​A system prediction module is configured to predict, according to the test set samples, using the optimal fuzzy width learning system structure obtained through the optimized search.

[0053] Another object of the present application is to provide a computer device comprising a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to cause the processor to perform the steps of the thermal nonlinear system identification method.

[0054] Another object of the present application is to provide a computer-readable storage medium storing a computer program, the computer program being executed by a processor to cause the processor to perform the steps of the thermal nonlinear system identification method.

[0055] Another object of the present application is to provide an information data processing terminal for implementing the thermal nonlinear system identification system.

[0056] In combination with the above technical solutions and the technical problems solved, the technical solution of the present application has the following advantages and positive effects:

[0057] The present application provides a thermal nonlinear system identification method based on NSFPSO-FBLS, which constructs the input-output relationship of the nonlinear system through the fuzzy width learning system, updates the weight of the fuzzy width learning system using the ridge regression algorithm, and searches for the optimal fuzzy width learning system structure using the NSFPSO algorithm.

[0058] Compared with the existing technology, the fuzzy width learning system provided by the present application eliminates redundant operations, has a relatively simple network structure, and has high computational efficiency.

[0059] The fuzzy width learning system provided by the application has simple structure, removes redundant structure, has strong fitting capacity, high calculation efficiency, and can effectively improve the nonlinear system identification precision. The NSFPSO algorithm of the application is used for searching the optimal fuzzy width learning system structure, the node sensitive fitness parameter is calculated, the algorithm convergence process is accelerated, the optimal model structure can be found at a faster speed, the training process is more stable, and the generalization ability of the model in the nonlinear system identification task is improved. The ridge regression algorithm of the application is a regularized weight updating algorithm, can relieve overfitting and underfitting problems, has fast operation speed, does not need iterative updating, can perform global parameter updating, and obtains a satisfactory fuzzy width learning system.

[0060] After the optimal structure of the NSFPSO-FBLS model of the application is obtained, the RMSE mean square error and running time of the BLS width learning system on the same test sample are compared, it can be seen that the NSFPSO-FBLS can obtain smaller RMSE and running time, the NSFPSO-FBLS has more accurate identification ability and higher calculation efficiency, and the NSFPSO-FBLS can use less node quantity to obtain satisfactory performance, and can effectively reflect the input-output mapping relationship of the nonlinear system.

[0061] The thermal nonlinear system identification method provided by the application has the advantages of high identification precision, strong stability and high calculation efficiency, and can realize fast and effective approximation of the input-output mapping of the nonlinear system.

[0062] The application provides a node fitness particle swarm optimization algorithm for effectively searching the optimal structure of a fuzzy width learning system, which can ensure that the parameter searching process converges to the global optimal value quickly. The algorithm can be extended to the optimal structure search of other width learning system variant structures, and ensure that the model obtains the optimal parameter combination.

[0063] The model provided by the application has low calculation complexity, simple and efficient calculation and reasoning process, can be used for neural network approximation and fuzzy reasoning at the same time, has the advantages of quickly determining the optimal model structure and quickly updating the model weight, can effectively improve the identification effect of the model in the thermal nonlinear system, and can be popularized to other nonlinear system identification tasks. Therefore, the method provided by the application has strong application value in the nonlinear system identification task.

[0064] The technical scheme of the present application fills the technical blank in the industry at home and abroad: there are great differences in different fields of nonlinear system identification tasks, the generalization ability of the existing neural network model is limited, and a universal model cannot be used for different nonlinear system identification tasks. The nonlinear system identification method provided by the present application has strong generalization ability and can be widely expanded to other nonlinear system identification fields, fully compensating for the incompatibility of models in different nonlinear system identification fields, thereby filling the gap of insufficient nonlinear system identification models.

[0065] Whether the technical scheme of the present application solves the technical problems that people have been eager to solve but have always failed to succeed: as one of the main sources of electricity in China, thermal power generation has become the basis for application in the field of process control. Establishing a more accurate system identification model and more quickly and effectively approximating the real output is the main demand of enterprises in production scheduling, prediction and simulation. On the one hand, the performance of the existing nonlinear system identification algorithm based on the neural network model is limited, and it cannot approximate the nonlinear system with large disturbance and high complexity; on the other hand, the existing identification method has poor real-time performance and cannot obtain the output of the controller in effective time, which has high difficulty in actual engineering application. The present application provides a new method for thermal nonlinear system identification, which overcomes the difficulty of accurately fitting the complex nonlinear relationship between input and output of thermal system, ensures the real-time performance of the model while effectively fitting the real output of the nonlinear system, and promotes the practical application of the neural network model. BRIEF DESCRIPTION OF DRAWINGS

[0066] In order to more clearly illustrate the technical scheme of the embodiments of the present application, the drawings needed in the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor on the basis of these drawings.

[0067] Figure 1 is a flow chart of the thermal nonlinear system identification method provided by the embodiments of the present application;

[0068] Figure 2 is a principle diagram of the thermal nonlinear system identification method provided by the embodiments of the present application;

[0069] Figure 3 is a principle diagram of the fuzzy width learning system structure provided by the embodiments of the present application;

[0070] Figure 4 is a specific flow chart of step 2 provided by the embodiments of the present application;

[0071] Figure 5 is a specific flow chart of step 26 provided by the embodiments of the present application;

[0072] Figure 6 is a convergence graph of an iterative fitness function based on NSFPSO-FBLS provided by the embodiment of the present application;

[0073] Figure 7 is a comparison graph based on NSFPSO-FBLS output, BLS output and real system output data provided by the embodiment of the present application;

[0074] Figure 8 is a comparison graph based on NSFPSO-FBLS and BLS prediction error provided by the embodiment of the present application. DETAILED DESCRIPTION

[0075] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and should not be used to limit the present application.

[0076] In view of the problems in the prior art, the present application provides a thermal nonlinear system identification method, system, medium, equipment and terminal, which will be described in detail below with reference to the accompanying drawings.

[0077] In order to enable those skilled in the art to fully understand how the present application is specifically implemented, this part is an explanatory embodiment for explaining and describing the technical solution of the claims.

[0078] As shown in Figure 1 the thermal nonlinear system identification method provided by the embodiment of the present application comprises the following steps:

[0079] S101, generating a sample set by using a given nonlinear system, and dividing the sample set into two parts of a training set and a test set;

[0080] S102, taking the training set samples as input samples for model training, and training by using a fuzzy width learning system;

[0081] S103, searching three parameters of a fuzzy rule Nr, a fuzzy subsystem Nf and an enhanced node Ne of the fuzzy width learning system by using a node fitness particle swarm algorithm;

[0082] S104, predicting by using an optimal fuzzy width learning system structure searched and optimized according to the test set samples.

[0083] Compared with the prior art, the fuzzy width learning system provided by the embodiment of the present application removes redundant operations, has a relatively simple network structure, is high in calculation efficiency, can quickly update the weight of the entire network by calculating the consequent parameters of the fuzzy subsystem and the weight of the enhanced node through the ridge regression algorithm, and further searches the super parameter of the fuzzy width learning system in combination with the node fitness particle swarm optimization algorithm (NSFPSO algorithm) to obtain the optimal fuzzy width learning system structure, thereby solving the problems that the super parameter of the fuzzy width learning system is difficult to adjust and the identification effect on a complex nonlinear system is poor, effectively improving the fitting precision of the nonlinear system, and stabilizing the training process of the model as a whole.

[0084] As a preferred embodiment, as shown in Figure 2 The NSFPSO-FBLS-based thermal nonlinear system identification method provided by the embodiment of the present application specifically comprises the following steps:

[0085] Step 1: generating a sample set by using a given nonlinear system, and dividing the sample set into a training set and a test set;

[0086] The nonlinear system used can be specifically described as the following formula:

[0087]

[0088] The training and the test use different generated sample sets, wherein the training and the test can be respectively expressed as:

[0089]

[0090]

[0091] Wherein y is an output variable, and u is an input variable. According to different nonlinear system formulas, 500 training samples and 200 test samples are generated. The 500 training samples are used to train the FBLS model, and the 200 test samples are used to test the FBLS model.

[0092] The embodiment of the present application can fully evaluate the performance of the model in convergence ability and prediction ability by using different nonlinear systems to generate training and test samples.

[0093] Step 2: taking the training set samples as input samples for model training, training and reasoning through the fuzzy width learning system, and searching the fuzzy rule Nr, the fuzzy subsystem Nf and the enhanced node Ne of the fuzzy width learning system (FBLS) through the node fitness particle swarm optimization algorithm (NSFPSO);

[0094] The fuzzy width learning system is a novel type of fuzzy neural network that has good nonlinear approximation ability, fast learning speed, few node parameters, and high computational efficiency.

[0095] like Figure 3 The diagram shown illustrates the schematic of a fuzzy width learning system structure used in one embodiment of the present invention. The fuzzy width learning system includes a Takagi–Sugeno–Kang (TSK) fuzzy subsystem, an enhancement layer, and an output layer. The TSK fuzzy subsystem consists of multiple TSK fuzzy systems and is used to perform fuzzy transformations within the fuzzy width learning system, including fuzzification, fuzzy inference, and defuzzification processes. In the enhancement layer, the defuzzified fuzzy subsystem is input for nonlinear transformation, resulting in a stronger approximation capability of the model. In the output layer, the outputs of the fuzzy subsystem and the enhancement layer are input to the output layer for global weight updates. In this invention, the number of input samples for the fuzzy width learning system is N, and the dimension of each sample is m.

[0096] The number of input samples and the dimension of the TSK fuzzy subsystem are N and m, respectively, which can be expressed as follows:

[0097]

[0098] Where x i (i = 1, 2, ..., N) represents the i-th input sample, x ij (j = 1, 2, ..., m) represents the j-th input feature of the i-th input sample. For the p-th fuzzy subsystem, S p (s=1,2,...,S p Fuzzy rules can be expressed as:

[0099]

[0100] Where the fuzzy set A p By S p Composition of fuzzy rules. In the TSK fuzzy subsystem, the result of the s-th fuzzy rule can be expressed in the following form:

[0101]

[0102] in This represents a random parameter used to adjust the input x. i FBLS uses a Gaussian kernel function as the membership partitioning function, which takes the following form:

[0103]

[0104] in Indicates the cluster center. denotes the width of the Gaussian kernel function. For the s-th fuzzy rule in the p-th fuzzy subsystem, the weighted intensity based on the Gaussian kernel function can be further denoted as:

[0105]

[0106] The output of the i-th sample in the fuzzy subsystem is wherein is the weighted intensity based on the Gaussian kernel function, is the result of the s-th fuzzy rule;

[0107] For k (p = 1, 2,..., k) TSK fuzzy subsystems, Z = (Z1, Z2,..., Zk) can be obtained. k wherein Z p = (Z 1p , Z 2p ,..., Z Np ) T denotes the output of all samples in the p-th fuzzy subsystem.

[0108] The output of the enhanced layer is H = (H1, H2,..., Hl). l );

[0109] wherein l (q = 1, 2,..., l) denotes the number of enhanced node groups. The output of the q-th enhanced node group can be denoted as H q = ε (ZW eq + b q ). W eq and b q are randomly generated matrix and bias term. ε represents the nonlinear activation function of the enhanced layer;

[0110] The output of the output layer is

[0111] wherein W df is the weight matrix of the fuzzy subsystem to the output layer, and W h is the weight matrix of the enhanced layer to the output layer.

[0112] The fuzzy width learning system provided by the embodiment of the present application has simple structure, removes redundant structure, has strong fitting capacity, high calculation efficiency, and can effectively improve the nonlinear system identification precision.

[0113] In the embodiment of the present application, the number of nodes of the initial fuzzy subsystem is set to 4, the number of nodes of the fuzzy rule is set to 10, and the number of nodes of the enhanced node is set to 6.

[0114] Particle swarm optimization (PSO) is a global optimization algorithm based on bionics. As a random search algorithm, it can realize parallel search with memory according to the adaptive information of individuals. Since PSO algorithm only relies on the speed of particles for updating, it has simpler calculation principle, less parameter amount and simpler implementation process compared with other heuristic algorithms. In the process of using fuzzy width learning system for nonlinear system identification, since multiple node parameters need to be adjusted, using grid search and other methods for parameter search will be very time-consuming, and stable parameter combination cannot be obtained within a limited time. Therefore, in order to obtain the optimal model structure, the PSO algorithm based on node sensitivity (NSFPSO) is introduced, which is improved on the basis of the original PSO algorithm according to the node characteristics of the fuzzy width learning system. According to the sensitivity difference of different nodes to the performance index, the improved fitness function can avoid the situation of easily falling into local optimum as much as possible, and ensure the faster convergence of PSO algorithm. This method can find the optimal structure of the fuzzy width learning system in a shorter time, improve the system identification accuracy of the fuzzy width learning system, and effectively reduce the time loss caused by parameter search.

[0115] As Figure 4 shown, it is a specific flow chart of step 2 in one embodiment of the application, and specifically includes the following steps:

[0116] Step 21: initialize the population operation, the population size is N, the learning factor is c1 and c2, the inertia weight is w, and the speed is v. The position information x of different particles includes the number of fuzzy rules, the number of TSK fuzzy subsystems and the number of enhanced nodes;

[0117] Step 22: for the tth iteration, calculate the fitness function of NSFPSO. According to the given information of particles, the fitness of each particle is calculated by using the node sensitive fitness function. By comparing the fitness value of each particle with the current optimal global fitness value, the current individual optimal value pbest and the population optimal value gbest are obtained. The node sensitive fitness function can be expressed as:

[0118] sum = Nr + Nf + Ne + 1

[0119]

[0120]

[0121] Where the fitness function fitness is composed of two indicators, the mean square error RMSE and the node sensitivity parameters Nr, Nf and Ne. Y and are the expected output value and the actual output value of the fuzzy width learning system respectively, and a, b, e, g and q are the weighting coefficients of the fitness function respectively.

[0122] In one embodiment, the initial particle population number is 30, the learning factors c1 and c2 are both 2.1, the inertia weight w ini and w end are 0.9 and 0.3 respectively, and the initial values of the fitness weighting parameters a, b, e, g and q are 0.2, 0.5, 0.3, 0.2 and 0.8 respectively. By selecting more appropriate parameter values, the convergence speed of the model can be improved and the prediction accuracy can be improved.

[0123] Step 23: Determine whether the iteration termination condition is reached, i.e., whether the maximum number of iterations is reached, if the maximum number of iterations is not reached, obtain the particle position information and the speed information;

[0124] Step 24: Update the position information and the speed information of each particle, wherein the position information is updated as:

[0125] w t =(w ini -w end )(G max -g) / G max +w end

[0126] wherein G max is the maximum number of iterations, w ini is the initial inertia weight, and w end is the inertia weight corresponding to the maximum number of iterations. The speed information is updated as:

[0127] v i =w×v i +c1×rand×(pbest i -x i )+c2×rand×(gbest i -x i )

[0128] wherein v i is the speed of the particle, i = 1, 2,..., N is the number of particles, w is the inertia factor for balancing the local search ability and the global search ability, and rand is a random number between 0 and 1;

[0129] Step 25: Update the position information x of the t+1 iteration according to the speed information obtained in step 23:

[0130] x t+1 =x t +v t+1

[0131] Step 26: Based on the particle position information obtained in Step 25, train the fuzzy width learning system using the ridge regression algorithm, and then return to Step 22.

[0132] In wide-area networks, the pseudoinverse can be considered a very convenient method for solving the output layer weights of neural networks. Unlike existing general methods for calculating inverses, such as iterative methods, orthogonal inverses, and singular value decomposition, the pseudoinverse can be solved by solving the least-squares estimate of linear equations and by searching for output weights through minimizing training error, thus effectively solving for high-dimensional training samples. Ridge regression algorithms are a special case of the pseudoinverse, primarily solving the L2 norm regularization problem, which can effectively avoid overfitting. Therefore, using ridge regression algorithms can quickly and efficiently solve for the weights of neural networks without consuming a large amount of iterative training time.

[0133] like Figure 5 The diagram shown is a flowchart of step 26 in one embodiment of the present invention. In the global weight update of the fuzzy width learning system, the fuzzy subsystem and enhancement layer weights are trained using the ridge regression algorithm, specifically including the following steps:

[0134] Step 261: Random initialization operation, randomly generate defuzzification weights for the fuzzy subsystem.

[0135] Step 262: According to the formula Obtain the deblurred output of the fuzzy subsystem, where F ip It is the output of the fuzzy subsystem. It is a weighted intensity based on the Gaussian kernel function. It is to defuzzify random weights. It is the result of fuzzy rules;

[0136] Step 263: According to the formula Calculate the defuzzified output of different fuzzy subsystems;

[0137] Step 264: According to formula H q =ε(ZW eq +b q Calculate the output of the enhancement layer;

[0138] The final output can be composed of the output of the fuzzy subsystem, the output of the enhancement layer, and the corresponding weights W. df (weights of defuzzified output) and W h(weights of enhancement layer) representation, the final output can be represented as:

[0139]

[0140] Step 265: The fuzzy width learning system FBLS weight is trained according to the ridge regression algorithm, which can be described as minimizing the optimization objective function of two 2-norms, and the optimization objective function is described using the argmin function:

[0141]

[0142]

[0143] Where the plus inverse of the output Z of the fuzzy subsystem and the output H of the enhancement layer can be represented using the above limit form.

[0144] W FBLS =[Z,H] + Y=(λI+[Z,H][Z,H] T ) -1 [Z,H] T Y

[0145] Where I represents the unit matrix, λ is a balance coefficient for balancing the complexity of the formula, and W FBLS represents the final weight of the FBLS model, and the weight update can be performed using the above form provided by the application.

[0146] In the embodiments of the application, λ is set to 2 -30 , and the optimal weight can be obtained.

[0147] The ridge regression algorithm provided in the embodiments of the application is a regularized weight update algorithm, which can alleviate the problems of overfitting and underfitting, has fast operation speed, does not need to be updated iteratively, can perform global parameter update, and obtains a satisfactory fuzzy width learning system.

[0148] Step 27: According to step 23, if the maximum number of iterations is reached, the best particle information, i.e., the best parameter combination, is output. Finally, the optimal number of nodes is selected as the output result.

[0149] The NSFPSO algorithm provided in the embodiments of the application is used to search for an optimal fuzzy width learning system structure, accelerates the convergence process of the algorithm by calculating the node sensitive fitness parameter, can find the optimal model structure at a faster speed, makes the training process more stable, and improves the generalization ability of the model in the nonlinear system identification task.

[0150] Step 3: according to the test set samples, an optimal fuzzy width learning system structure obtained through optimized search is used for prediction.

[0151] The thermal nonlinear system identification system provided by the embodiment of the application comprises:

[0152] A sample set acquisition module is configured to obtain a sample set for model thermal nonlinear system identification, and divide the sample set into a training set and a test set.

[0153] A model training module is configured to use the training set samples as input samples for model training, and train and infer through a fuzzy width learning system.

[0154] A parameter search module is configured to search for three parameters of the fuzzy width learning system, i.e., a fuzzy rule Nr, a fuzzy subsystem Nf and an enhanced node Ne, through a node fitness particle swarm algorithm.

[0155] A system prediction module is configured to use an optimal fuzzy width learning system structure obtained through optimized search to perform prediction according to the test set samples.

[0156] In order to prove the creativity and technical value of the technical scheme of the application, this part is an application embodiment of the technical scheme of the claim on a specific product or related technology.

[0157] The application embodiment is mainly used for an engineer of an automatic generation control (AGC) control system of a power plant unit, designs an advanced boiler steam pressure control system in combination with a model predictive control algorithm, and thus improves key operation indexes of the unit such as main steam pressure and boiler-turbine coordination. The AGC boiler uses an NSFPSO-FBLS method to dynamically identify a steam pressure object model, and uses a generalized predictive control (GPC) algorithm to control a dynamic characteristic model of a controlled object. In the boiler steam pressure control system, the control algorithm uses a GPC type predictive control algorithm of the generalized predictive control, forces a boiler main control PID to be a step-up signal, keeps a power control loop of a turbine main control as automatic customization, and can obtain an input-output transfer function. Then, a sample set can be generated according to the input-output, and a dynamic robust identification is realized in combination with a nonlinear system identification method and a self-correcting mechanism of the application.

[0158] The embodiment of the application has achieved some positive effects in the research and development or use process, and indeed has great advantages compared with the prior art. The following content is described in combination with data and graphs of an experimental process.

[0159] The effectiveness of the method provided by the application is verified through several groups of experiments.

[0160] (1) Iterative convergence analysis of NSFPSO-FBLS

[0161] The NSFPSO-FBLS model provided by the embodiment of the present application is used to perform iterative calculation on the fitness value, and the fitness value generated in each iteration is recorded, as shown in Table 1. Figure 6 As shown in the figure, it is an iteration convergence diagram based on the NSFPSO-FBLS, and from the experimental results, it can be seen that the NSFPSO-FBLS model proposed in the embodiment of the present application can quickly converge and can converge to a satisfactory result, ensuring the system identification accuracy of the model while using as few node parameters as possible to obtain the optimal model structure.

[0162] (2) Comparison and analysis of BLS width learning system and expected output

[0163] The NSFPSO-FBLS model proposed in the embodiment of the present application is used to obtain the optimal fuzzy width learning system structure, and 200 test samples are used for testing to obtain the final test output. Similarly, the BLS width learning system is used to train and test the same samples, and the actual output and the expected output of the two models are analyzed by comparison. As shown in Table 1, Figure 7 As shown in the figure, it is a comparison diagram based on the NSFPSO-FBLS output, the BLS output and the real system output data, and from the experimental results, it can be seen that the NSFPSO-FBLS model proposed in the embodiment of the present application can obtain more accurate output results when predicting the output of a nonlinear system, which is significantly better than the output results of the original BLS width learning system.

[0164] (3) Comparison and analysis of prediction error of BLS width learning system

[0165] The NSFPSO-FBLS model proposed in the embodiment of the present application is used to obtain the optimal fuzzy width learning system structure, and the BLS width learning system is used to obtain the actual output and the expected output, i.e., the prediction error, by using 200 test samples for testing. As shown in Table 1, Figure 8 As shown in the figure, it is a comparison diagram based on the NSFPSO-FBLS prediction error and the BLS prediction error, and the experimental results fully demonstrate that the method proposed in the present application has smaller prediction error, further verifying the effectiveness and feasibility of the method in the embodiment of the present application. The overall performance comparison of the nonlinear system identification method is shown in Table 1.

[0166] Table 1 Overall performance comparison of nonlinear system identification method

[0167] Model RMSE Number of nodes Run time (s) BLS 3.526e-7 31 0.23 NSFPSO-FBLS 2.331e-7 24 0.21

[0168] As shown in Table 1, after obtaining the optimal structure using the NSFPSO-FBLS model proposed in the embodiments of the present application, the RMSE mean square error and the running time are compared with the BLS width learning system on the same test sample, it can be seen that the NSFPSO-FBLS proposed in the embodiments of the present application can obtain smaller RMSE and running time, which indicates that the NSFPSO-FBLS has more accurate recognition ability and higher calculation efficiency, and at the same time, the NSFPSO-FBLS can use less node quantity to obtain satisfactory performance, and can effectively reflect the input-output mapping relationship of the nonlinear system.

[0169] Compared with the prior art, the embodiments of the present application provide a thermal nonlinear system identification method based on NSFPSO-FBLS, the input-output relationship of the nonlinear system is constructed through the fuzzy width learning system, the weight of the fuzzy width learning system is updated using the ridge regression algorithm, and the optimal fuzzy width learning system structure is quickly searched using the NSFPSO algorithm, compared with the original fuzzy width learning system, the system identification ability of the method provided by the present application is higher, and the input-output mapping relationship of the system can be well described, and the method has the advantages of fast convergence speed, stable training process, strong generalization ability and high identification precision, and has high application value in nonlinear system identification.

[0170] It should be noted that the embodiments of the present application can be realized by hardware, software or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control code, such as provided on a carrier medium, such as a magnetic disk, CD or DVD-ROM, a programmable memory, such as a read-only memory (firmware), or a data carrier, such as an optical or electronic signal carrier. The devices of the present application and their modules can be realized by hardware circuits, such as very large scale integrated circuits or gate arrays, semiconductors, such as logic chips, transistors, etc., or programmable hardware devices, such as field programmable gate arrays, programmable logic devices, etc., can also be realized by software executed by various types of processors, and can also be realized by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0171] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any modification, equivalent replacement and improvement made by those skilled in the art within the technical range disclosed by the present application, as long as it is within the spirit and principles of the present application, should be covered within the protection scope of the present application.

Claims

1. A method for identifying a thermal nonlinear system, characterized in that, The thermal nonlinear system identification method includes: constructing the input-output relationship of the nonlinear system through a fuzzy width learning system, updating the weights of the fuzzy width learning system using a ridge regression algorithm, and searching for the optimal fuzzy width learning system structure using the NSFPSO algorithm, thereby realizing the identification of the thermal nonlinear system. The method for identifying thermal nonlinear systems includes the following steps: Step 1: Simulate the main steam pressure system and main steam temperature system to obtain a sample set for identifying the model's thermal nonlinear system, and divide the sample set into two parts: a training set and a test set. Step 2: Use the training set samples as input samples for model training, perform inference through the fuzzy subsystem in the fuzzy width learning system, and use the ridge regression algorithm to update the weights of the fuzzy width learning system; use the NSFPSO algorithm to search for the three parameters of the fuzzy rule Nr, fuzzy subsystem Nf, and enhancement node Ne of the fuzzy width learning system to determine the optimal model structure. Step 3: Based on the test set samples, use the optimized search for the optimal fuzzy width learning system structure to make predictions, use the mean squared error (RMSE) metric to evaluate the prediction performance, and at the same time evaluate the time required for model testing. The fuzzy width learning system in step two includes a TSK fuzzy subsystem, an enhancement layer, and an output layer. The number of input samples and the dimension of the TSK fuzzy subsystem are N and m, respectively, as follows: The output of the fuzzy subsystem is in For weighted intensity based on Gaussian kernel function, It is the result of the s-th fuzzy rule; The output of the enhancement layer is: H = (H1, H2, ..., H l ); where l (q = 1, 2, ..., l) represents the number of augmentation node groups; the output of the q-th augmentation node group is represented as H. q =ε(ZW eq +b q );W eq and b q The matrix and bias term are randomly generated; ε represents the nonlinear activation function of the enhancement layer; The output of the output layer is Among them, W df W is the weight matrix from the fuzzy subsystem to the output layer. h It is the weight matrix from the enhancement layer to the output layer.

2. The method for identifying thermal nonlinear systems as described in claim 1, characterized in that, In step one, 500 training samples and 200 test samples are generated according to different nonlinear system formulas; the 500 training samples are used to train the FBLS model, and the 200 test samples are used to test the FBLS model.

3. The method for identifying thermal nonlinear systems as described in claim 1, characterized in that, In step two, the training set samples are used as input samples for model training, and training and inference are performed through a fuzzy width learning system. The fuzzy rule Nr, fuzzy subsystem Nf, and augmenting node Ne of the fuzzy width learning system are searched using the node fitness particle swarm optimization algorithm, including: (1) Initialize the group operation, with a population size of N, learning factors of c1 and c2, inertia weight of w, and velocity of v; where the position information x of different particles includes the number of fuzzy rules, the number of TSK fuzzy subsystems, and the number of enhancement nodes; (2) For the t-th iteration, calculate the fitness function of NSFPSO; calculate the fitness of each particle using the node-sensitive fitness function based on the given particle information; obtain the current individual optimal value pbest and the population optimal value gbest by comparing the fitness value of each particle with the current best global fitness value; the node-sensitive fitness function can be expressed as: sum=Nr+Nf+Ne+1 The fitness function consists of two indices: mean squared error (RMSE) and node sensitivity parameters Nr, Nf, and Ne; Y and α, β, ε, γ, and θ are the expected and actual output values ​​of the fuzzy width learning system, respectively; α, β, ε, γ, and θ are the trade-off coefficients of the fitness function, respectively. (3) Determine whether the iteration termination condition has been met -- Determine whether the maximum number of iterations has been reached. If the maximum number of iterations has not been reached, obtain the particle position information and velocity information. (4) Update the position and velocity information of each particle, where the position information is updated as follows: w t =(w ini -w end )(G max -g) / G max +w end ; Among them, G max It is the maximum number of iterations, w ini It is the initial inertia weight, w end The inertia weight corresponds to the maximum number of iterations, and the velocity information is updated as follows: v i =w×v i +c1×rand×(pbest i -x i )+c2×rand×(gbest i -x i ); Among them, v i is the particle velocity, i = 1, 2, ..., N is the number of particles; w is the inertia factor, used to balance local search capability and global search capability; rand is a random number between 0 and 1; (5) Update the position information x for the (t+1)th iteration based on the velocity information obtained in step (3): x t+1 =x t +v t+1 ; (6) Based on the particle position information obtained in step (5), train the fuzzy width learning system using the ridge regression algorithm and return to step (2); (7) According to step (3), if the maximum number of iterations is reached, the best particle information -- the best parameter combination is output, and the optimal number of nodes is finally selected as the output result.

4. The method for identifying thermal nonlinear systems as described in claim 3, characterized in that, The initial particle population size is 30, the learning factors c1 and c2 are both 2.1, and the inertia weight w ini and w end The initial values ​​for fitness tradeoff parameters α, β, ε, γ, and θ are 0.9 and 0.3, respectively, and the initial values ​​for fitness tradeoff parameters α, β, ε, γ, and θ are 0.2, 0.5, 0.3, 0.2, and 0.8, respectively. In step (6), training the fuzzy width learning system using the ridge regression algorithm also includes: 1) Random initialization operation, randomly generating defuzzification weights for the fuzzy subsystem. 2) According to the formula Obtain the deblurred output of the fuzzy subsystem, where F ip It is the output of the fuzzy subsystem. It is a weighted intensity based on the Gaussian kernel function. It is to defuzzify random weights. It is the result of fuzzy rules; 3) According to the formula Calculate the defuzzified output of different fuzzy subsystems; 4) According to formula H q =ε(ZW eq +b q Calculate the output of the enhancement layer; 5) The formula for calculating the weights W of the FBLS system trained using the ridge regression algorithm. FBLS =(Z,H) + Y = (λI + (Z, H)) T (Z,H)) -1 (Z,H) T Y is used to calculate the weights of the fuzzy width learning system, where Y is the expected output of the model; The initial fuzzy subsystem has 4 nodes, the fuzzy rules have 10 nodes, and the enhancement nodes have 6 nodes.

5. A thermal nonlinear system identification system applying the thermal nonlinear system identification method as described in any one of claims 1 to 4, characterized in that, The thermal nonlinear system identification system includes: The sample set acquisition module is used to obtain a sample set for the identification of the model's thermal nonlinear system, and divides the sample set into two parts: a training set and a test set. The model training module is used to take training set samples as input samples for model training and to perform training and inference through a fuzzy width learning system. The parameter search module is used to search for three parameters of the fuzzy width learning system: the fuzzy rule Nr, the fuzzy subsystem Nf, and the enhancement node Ne, using the node fitness particle swarm algorithm. The system prediction module is used to make predictions based on test set samples by learning the system structure using the optimal fuzzy width obtained through optimized search.

6. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the thermal nonlinear system identification method as described in any one of claims 1 to 4.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the thermal nonlinear system identification method as described in any one of claims 1 to 4.

8. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the thermal nonlinear system identification system as described in claim 5.

Images