Turbine unit prediction method based on density clustering and recurrent fuzzy neural network

Abstract

The present invention provides a turbine unit prediction method based on density clustering and recurrent fuzzy neural network, which belongs to the technical field. It solves the problem that the existing prediction method does not take into account the distribution density and clustering characteristics of historical input data, resulting in a decrease in the prediction accuracy of the nonlinear system output. Its technical solution is: including the following steps: S1: constructing a nonlinear system of a supercritical boiler turbine unit; S2: using the DBSCAN clustering algorithm to automatically determine the number of clusters, the number of local linear sub-models and the number of fuzzy rules, and designing fuzzy rules; S3: constructing a recurrent fuzzy neural network. The beneficial effect of the present invention is that the present invention uses the DBSCAN clustering algorithm to automatically determine the optimal number of fuzzy sets in the fuzzy neural network, while reducing the computational complexity, obtaining a higher system modeling prediction accuracy, thereby improving the control performance of the supercritical boiler turbine unit output.

Description

Technical Field

[0001] The invention relates to the technical field of industrial automation control, and in particular to a turbine unit prediction method based on density clustering and cyclic fuzzy neural network. Background Art

[0002] With the rapid development of the national economy and the development of high-energy-consuming industries, global energy demand has increased dramatically. However, 36% of global energy consumption still relies heavily on the combustion of coal. From the perspective of environmental protection and improvement, it is very necessary to study clean coal-fired power generation technology. Compared with traditional power generation technology, USC boiler turbine units can not only provide up to 46% net efficiency, but also reduce CO2 intensity by 30%. The main goal of USC is to achieve rapid tracking of electric power, main steam pressure and steam temperature, so it is crucial to accurately control the fuel flow, regulating valve volume and water flow of the turbine unit.

[0003] In order to improve the performance of boiler turbine units, fuzzy neural networks (FNNs) have been studied by combining the advantages of neural networks and fuzzy logic systems. In order to obtain better prediction accuracy of nonlinear systems, recursive fuzzy neural networks (RFNNs) have received widespread attention, and their feedback network loops are used to store past information. Therefore, RFNN can capture more dynamic responses of nonlinear systems and improve learning efficiency. The dynamic modeling of RFNN includes two stages: structure learning and parameter learning, which can be carried out simultaneously. Before structural learning, some parameters of RFNN are generally assigned in advance, such as the number of fuzzy sets, the initial position of membership functions, etc. Although the initial center and width of the membership function are predefined, these parameters will change during the iterative learning process. The parameter learning stage, in turn, affects the structural learning and prediction performance of RFNN. In order to solve the problem of dynamic system modeling and temperature control, recursive least squares (RLS) is used to identify the unknown parameters in RFNN, and the number of fuzzy sets is predefined based on expert experience. However, in the iterative learning process of RFNN, the center and width of the Gaussian membership function need to be trained. These predefined fuzzy sets and multiple parameters of the membership function reduce the adaptability and efficiency of the fuzzy neural network.

[0004] In general, too many fuzzy sets will reduce the flexibility and generalization ability of fuzzy neural networks. In order to construct a compact network, clustering methods can automatically determine the number of fuzzy sets and fuzzy rules. The existing optimization clustering methods mainly include k-means clustering, fuzzy C-means clustering, density-based clustering (DBSCAN), etc. Before the constructed neural network iterative learning, the obtained clustering results are used to determine the initial center and width of the Gaussian membership function, and the center of the membership function is replaced by the cluster center. However, the method of combining density-based clustering (DBSCAN) and recurrent fuzzy neural network (RFNN) for turbine unit output prediction has not been studied. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that the distribution density and clustering characteristics of historical input data are not taken into account in the prior art, resulting in a decrease in the prediction accuracy of the output of the nonlinear system, and a turbine unit prediction method based on density clustering and recurrent fuzzy neural network is proposed. According to the distribution density and clustering characteristics of the historical input data, the DBSCAN clustering algorithm is used to automatically determine the optimal number of fuzzy sets in the fuzzy neural network, thereby improving the generalization and efficiency of system modeling. At the same time, the present invention constructs a 7-layer recurrent fuzzy neural network (RFNN), improves the fuzzy membership function, and uses the cluster centroid obtained by the DBSCAN clustering algorithm as the center of the membership function, so that the learning process of the RFNN only needs to perform single-parameter iterative learning on the adaptive overlap coefficient. Compared with the FNN model, the DBSCAN-RFNN model of the present invention has the ability to make full use of past information, while reducing the computational complexity, obtaining a higher system modeling prediction accuracy, thereby improving the control performance of the supercritical boiler turbine unit output.

[0006] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a turbine unit prediction method based on density clustering and recurrent fuzzy neural network, comprising the following steps:

[0007] S1: Construct the nonlinear system of supercritical boiler turbine unit, further linearize the nonlinear system and divide it into several local linear sub-models;

[0008] S2: Use the DBSCAN clustering algorithm to automatically determine the number of clusters, the number of local linear sub-models and the number of fuzzy rules, and design fuzzy rules;

[0009] S3: A recurrent fuzzy neural network (RFNN) is constructed to predict the electric power, main steam pressure and steam temperature in different local linear sub-models, and the prediction output of the local linear sub-model is coupled to obtain the prediction output results of the supercritical boiler turbine unit.

[0010] The step S1 specifically comprises the following steps:

[0011] Step 1.1: Considering the complex nonlinear characteristics of the turbine unit, the supercritical (USC) boiler turbine unit is modeled as a nonlinear system, namely:

[0012] Y(t)=f[Y(t-1),...,Y(tn y ),

[0013] U(td),...,U(tn u ),

[0014] e(t-1),...,e(tn e )]+e(t) / △,

[0015] Where f(·) is an unknown nonlinear function, is the output vector of the nonlinear system at time t, n a is the number of output variables, is the input vector of the nonlinear system at time t, n b is the number of input variables, is zero-mean Gaussian white noise, d is the system delay time, n y 、n u and n e are the maximum lags in Y(t), U(t) and e(t), respectively, and Δ is the difference operator. Considering the computational cost and prediction accuracy, the maximum lag n is set y =n u =n e =3.

[0016] Step 1.2: Further linearize the nonlinear system into a controlled autoregressive integrated moving average (CARIMA) model:

[0017]

[0018] Among them, z -1 is the backward shift operator, ΔU(t-1)=U(t-1)-U(t-2) is the increment of the input vector at time t-1, and Denote the backshift operator z -1 The polynomial matrix of , and:

[0019]

[0020]

[0021] C(z -1 )=I,

[0022] Where I is the identity matrix;

[0023] Step 1.3: Collect actual sample data of 1000MW supercritical boiler turbine unit at different times;

[0024] Step 1.4: Take the electric power, main steam pressure and steam temperature as the control output Y(t) of the nonlinear system, and take the fuel flow, regulating valve amount and water flow as the control input U(t) of the system, then:

[0025] Y(t)=[y1(t), y2(t), y3(t)] T

[0026] U(t)=[u1(t),u2(t),u3(t)] T

[0027] n a =n b =3;

[0028] Among them, Y(t) is the output vector of the nonlinear system, U(t) is the output vector of the nonlinear system, y1(t) is the electric power at time t, y2(t) is the main steam pressure at time t, y3(t) is the steam temperature at time t, u1(t) is the fuel flow at time t, u2(t) is the regulating valve amount at time t, and u3(t) is the water flow at time t;

[0029] Step 1.5: In order to better approximate the nonlinear turbine group, the nonlinear system is divided into several local linear sub-models (local CARIMA models):

[0030]

[0031] Among them, Y i (t) is the output vector of the i-th local linear sub-model, ΔU i (t-1) is the increment of the input vector of the i-th local linear sub-model at time t-1, e i (t) is the zero-mean Gaussian white noise of the i-th local linear sub-model, and B i (z -1 )∈R 3×3 is the polynomial matrix of the ith local linear sub-model.

[0032] As a further optimization scheme of the turbine unit prediction method based on density clustering and recurrent fuzzy neural network provided by the present invention, the step S2 specifically comprises the following steps:

[0033] Step 2.1: Use the DBSCAN clustering algorithm to cluster the input data U(t) and automatically determine the number of clusters to be L;

[0034] Step 2.2: Determine that the number of local linear sub-models and the number of fuzzy rules are both L;

[0035] Step 2.2: Design L IF-THEN fuzzy rules R i :If u1(t) is u2(t) is u3(t) is Then the linear output of the i-th local linear sub-model is:

[0036]

[0037] Among them, R i is the i-th fuzzy rule, is the input variable u j The i-th fuzzy set, A iq ∈R 3×3 and B ik ∈R 3×3 is the polynomial matrix of the i-th local linear sub-model, i = 1, 2, …, L; j = 1, 2, 3; q = 1, 2, 3; k = 1, 2, 3.

[0038] As a further optimization scheme of the turbine unit prediction method based on density clustering and recurrent fuzzy neural network provided by the present invention, the step S3 specifically comprises the following steps:

[0039] Step 3.1: Construct a 7-layer recurrent fuzzy neural network (RFNN). The first layer uses U(t) as the input of RFNN and Y(t) as the output of RFNN.

[0040] Step 3.2: Design the improved fuzzy membership function in the second layer Take L cluster centers as the centers of the improved fuzzy membership function,

[0041]

[0042] in, is the input variable u j (t) and fuzzy membership function Center The distance between By fixing the cluster centroid The jth element in is determined, while the cluster centroid is fixed Determined by the DBSCAN clustering algorithm, η ij is the adaptive overlap coefficient for adjusting the width of the fuzzy membership function, and η ij The larger the value of the fuzzy membership function The greater the width;

[0043] Step 3.3: The third layer constructs the spatial intensity layer. Each node represents the spatial intensity of a fuzzy rule. Each fuzzy rule R i The spatial trigger intensity ω at time t i (t):

[0044]

[0045] Step 3.4: The fourth layer constructs the time intensity layer, and each node can recursively update information to form a feedback loop. is the output of the i-th recursive rule node at time t, that is is the temporal trigger strength, which depends not only on the spatial trigger strength ω at time t i (t), also depends on the time trigger strength at time t-1 The output of each recursive rule node is the spatial trigger intensity ω at time t i (t) and the time trigger intensity at time t-1 A linear combination of:

[0046]

[0047] Among them, β i is the dynamic feedback weight between (0, 1), which is used to measure the spatial trigger intensity ω at time t i (t) and the time trigger intensity at time t-1 The weight between

[0048] Step 3.5: The fifth layer constructs a sequential layer. The nodes in this layer are called subsequent nodes, which are used to adjust the weights of the time intensity nodes in the fourth layer. The output f of the subsequent node i (t) is:

[0049]

[0050] Among them, a ij and i It is a free parameter that needs to be adjusted during parameter learning;

[0051] Step 3.6: The sixth layer normalizes the output nodes:

[0052]

[0053] in, is the output f of the subsequent node i Normalized result of (t);

[0054] Step 3.7: The seventh layer predicts the electric power, main steam pressure and steam temperature in different local linear sub-models, and couples the prediction outputs of L local linear sub-models;

[0055] Step 3.8: Obtain the predicted output results for the supercritical boiler turbine unit

[0056] The meanings of the symbols in the formula of the present invention are shown in Table 1:

[0057]

[0058]

[0059]

[0060] Compared with the prior art, the present invention has the following beneficial effects:

[0061] (1) The present invention uses the DBSCAN clustering algorithm to automatically determine the optimal number of fuzzy sets in the fuzzy neural network based on the distribution density and clustering characteristics of historical input data, thereby improving the generalization degree and efficiency of system modeling.

[0062] (2) The present invention constructs a 7-layer recurrent fuzzy neural network (RFNN), improves the fuzzy membership function, and uses the cluster centroid obtained by the DBSCAN clustering algorithm as the center of the membership function, so that the learning process of the RFNN only needs to perform single-parameter iterative learning on the adaptive overlap coefficient. Compared with the FNN model, the DBSCAN-RFNN model of the present invention has the ability to make full use of past information, while reducing the computational complexity and obtaining higher system modeling prediction accuracy, thereby improving the control performance of the supercritical boiler turbine unit output. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0064] Figure 1 A flow chart of a turbine unit prediction method based on density clustering and recurrent fuzzy neural network is proposed for the present invention;

[0065] Figure 2 A flow chart of "constructing a nonlinear system of a supercritical boiler turbine unit, further linearizing the nonlinear system and dividing it into several local linear sub-models" for a turbine unit prediction method based on density clustering and recurrent fuzzy neural network proposed in the present invention;

[0066] Figure 3A flow chart of "automatically determining the number of clusters, the number of local linear sub-models and the number of fuzzy rules by using the DBSCAN clustering algorithm, and designing fuzzy rules" for a turbine unit prediction method based on density clustering and recurrent fuzzy neural network proposed in the present invention;

[0067] Figure 4 The present invention proposes a turbine unit prediction method based on density clustering and recurrent fuzzy neural network, which includes "constructing a recurrent fuzzy neural network (RFNN), predicting the electric power, main steam pressure and steam temperature in different local linear sub-models, coupling the prediction output of the local linear sub-models, and obtaining the prediction output result of the supercritical boiler turbine unit".

[0068] Figure 5 A schematic diagram of a supercritical coal-fired power plant for which the present invention proposes a turbine unit prediction method based on density clustering and recurrent fuzzy neural network;

[0069] Figure 6 The DBSCAN-RFNN structure diagram of the turbine unit prediction method based on density clustering and recurrent fuzzy neural network proposed by the present invention;

[0070] Figure 7 Output comparison diagram of the DBSCAN-RFNN model and the FNN model in the training set and test set of the turbine unit prediction method based on density clustering and recurrent fuzzy neural network proposed by the present invention;

[0071] in, Figure 7 (a) is a comparison curve of the actual data of the power plant power under the training set and the test set, the power output predicted by the DBSCAN-RFNN model, and the power output predicted by the FNN model;

[0072] Figure 7 (b) is a comparison curve of the actual data of the main steam pressure of the power plant under the training set and the test set, the main steam pressure output predicted by the DBSCAN-RFNN model, and the main steam pressure output predicted by the FNN model;

[0073] Figure 7 (c) is a comparison curve of the actual data of the power plant steam temperature in the training set and the test set, the steam temperature output predicted by the DBSCAN-RFNN model, and the steam temperature output predicted by the FNN model. DETAILED DESCRIPTION

[0074] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0075] Example 1

[0076] See also Figures 1 to 7 The technical solution provided in this embodiment is a turbine unit prediction method based on density clustering and recurrent fuzzy neural network. Taking into account the complex nonlinear characteristics of the turbine unit, the supercritical (USC) boiler turbine unit is modeled as a nonlinear system, that is:

[0077] Y(t)=f[Y(t-1),...,Y(tn y ),

[0078] U(td),...,U(tn u ),

[0079] e(t-1),...,e(tn e )]+e(t) / Δ,

[0080] Where f(·) is an unknown nonlinear function, is the output vector of the nonlinear system at time t, n a is the number of output variables, is the input vector of the nonlinear system at time t, n b is the number of input variables, is zero-mean Gaussian white noise, d is the system delay time, n y 、n u and n e are the maximum lags in Y(t), U(t) and e(t), respectively. Δ is the difference operator. Considering the computational cost and prediction accuracy, the maximum lag n is set. y =n u =n e =3;

[0081] The nonlinear system is further linearized into a controlled autoregressive integrated moving average (CARIMA) model:

[0082]

[0083] Among them, z -1 is the backward shift operator, ΔU(t-1)=U(t-1)U-(t-2) is the increment of the input vector at time t-1, and Denote the backshift operator z -1 The polynomial matrix of , and:

[0084]

[0085]

[0086] C(z -1 )=I,

[0087] Where I is the identity matrix;

[0088] Collect actual sample data of 1000MW supercritical boiler turbine unit at different times;

[0089] With electric power, main steam pressure and steam temperature as the control output Y(t) of the nonlinear system, and fuel flow, regulating valve amount and water flow as the control input U(t) of the system, then:

[0090] Y(t)=[y1(t), y2(t), y3(t)] T

[0091] U(t)=[u1(t),u2(t),u3(t)] T

[0092] n a =n b =3;

[0093] Among them, Y(t) is the output vector of the nonlinear system, U(t) is the output vector of the nonlinear system, y1(t) is the electric power at time t, y2(t) is the main steam pressure at time t, y3(t) is the steam temperature at time t, u1(t) is the fuel flow at time t, u2(t) is the regulating valve amount at time t, and u3(t) is the water flow at time t;

[0094] In order to better approximate the nonlinear turbine set, the nonlinear system is split into several local linear sub-models (local CARIMA models):

[0095]

[0096] Among them, Y i (t) is the output vector of the i-th local linear sub-model, ΔU i (t-1) is the increment of the input vector of the i-th local linear sub-model at time t-1, e i (t) is the zero-mean Gaussian white noise of the i-th local linear sub-model, and B i (z -1 )∈R 3×3 is the polynomial matrix of the i-th local linear sub-model;

[0097] Using the DBSCAN clustering algorithm, the input data U(t) is clustered and the number of clusters is automatically determined to be L;

[0098] Determine that the number of local linear sub-models and the number of fuzzy rules are both L;

[0099] Design L IF-THEN fuzzy rules R i:If u1(t) is u2(t) is u3(t) is Then the linear output of the i-th local linear sub-model is:

[0100]

[0101] Among them, R i is the i-th fuzzy rule, is the input variable u j The i-th fuzzy set, A iq ∈R 3×3 and B ik ∈R 3×3 is the polynomial matrix of the i-th local linear sub-model, i = 1, 2, ..., L; j = 1, 2, 3; q = 1, 2, 3; k = 1, 2, 3;

[0102] A 7-layer recurrent fuzzy neural network (RFNN) is constructed. The first layer uses U(t) as the input of RFNN and Y(t) as the output of RFNN.

[0103] Improved fuzzy membership function for the second layer design Take L cluster centers as the centers of the improved fuzzy membership function,

[0104]

[0105] in, is the input variable u j (t) and fuzzy membership function Center The distance between By fixing the cluster centroid The jth element in is determined, while the cluster centroid is fixed Determined by the DBSCAN clustering algorithm, η ij is the adaptive overlap coefficient for adjusting the width of the fuzzy membership function, and η ij The larger the value of the fuzzy membership function The greater the width;

[0106] The third layer constructs the spatial intensity layer. Each node represents the spatial intensity of a fuzzy rule. i The spatial trigger intensity ω at time t i (t):

[0107]

[0108] The fourth layer constructs the time intensity layer, where each node can recursively update information to form a feedback loop. is the output of the i-th recursive rule node at time t, that is is the temporal trigger strength, which depends not only on the spatial trigger strength ω at time t i (t), also depends on the time trigger strength at time t-1 The output of each recursive rule node is the spatial trigger intensity ω at time t i (t) and the time trigger intensity at time t-1 A linear combination of:

[0109]

[0110] Among them, β i is the dynamic feedback weight between (0, 1), which is used to measure the spatial trigger intensity ω at time t i (t) and the time trigger intensity at time t-1 The weight between

[0111] The fifth layer constructs the sequential layer. The nodes in this layer are called subsequent nodes, which are used to adjust the weights of the time intensity nodes in the fourth layer. The output f of the subsequent node i (t) is:

[0112]

[0113] Among them, a ij and i It is a free parameter that needs to be adjusted during parameter learning;

[0114] The sixth layer normalizes the output nodes:

[0115]

[0116] in, is the output f of the subsequent node i Normalized result of (t);

[0117] The seventh layer predicts the electric power, main steam pressure and steam temperature in different local linear sub-models, and couples the prediction outputs of L local linear sub-models;

[0118] Get the predicted output of supercritical boiler turbine unit

[0119] The meaning of the symbols in the formulas in this implementation is shown in Table 1:

[0120]

[0121]

[0122]

[0123] In order to verify the effect of this embodiment, MATLAB is used to perform numerical simulation verification.

[0124] Figure 7 The output comparison curve of the DBSCAN-RFNN model and the FNN model in the training set and test set in this embodiment is shown in Figure 2. The conventional technology uses a fuzzy neural network model to model the dynamic process of the boiler turbine system of the USC power plant. In order to compare the performance with the DBSCAN-RFNN model proposed in this embodiment, the FNN model is constructed using the same sample set and error function threshold. The actual output (solid line) and predicted output (dashed line) of the two models under training data and test data are shown in Figure 2. Figure 7 As shown, Figure 7 (a) is a comparison curve of the actual data of the power plant power under the training set and the test set, the power output predicted by the DBSCAN-RFNN model, and the power output predicted by the FNN model. Figure 7 (b) is a comparison curve of the actual data of the main steam pressure of the power plant under the training set and the test set, the main steam pressure output predicted by the DBSCAN-RFNN model, and the main steam pressure output predicted by the FNN model; Figure 7 (c) is a comparison curve of the actual data of the power plant steam temperature in the training set and the test set, the steam temperature output predicted by the DBSCAN-RFNN model, and the steam temperature output predicted by the FNN model. Figure 7 From (a) to (c), it can be seen that the DBSCAN-RFNN model proposed in this embodiment can better approximate the entire output variable than the FNN model, especially in the area marked by the dotted circle. Figure 7 In the test part (a) to (c), the online error between the predicted output of DBSCAN-RFNN and the actual output is relatively small. It can be seen that the DBSCAN-RFNN proposed in this embodiment can better model the nonlinear system with violent fluctuations of the USC power plant.

[0125] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A turbine unit prediction method based on density clustering and recurrent fuzzy neural network, characterized in that: The following steps are involved: S1: Construct the nonlinear system of supercritical boiler turbine unit, linearize the nonlinear system and divide it into several local linear sub-models; S2: Use the DBSCAN clustering algorithm to automatically determine the number of clusters, the number of local linear sub-models and the number of fuzzy rules, and design fuzzy rules; The step S1 comprises the following steps: Step 1.1: Considering the complex nonlinear characteristics of the turbine unit, the supercritical (USC) boiler turbine unit is modeled as a nonlinear system, namely: Y(t)=f[Y(t-1),...,Y(t-n y ), U(t-d),...,U(t-n u ), e(t-1),...,e(t-n e )]+e(t) / △, Where f(·) is an unknown nonlinear function, is the output vector of the nonlinear system at time t, n a is the number of output variables, is the input vector of the nonlinear system at time t, n b is the number of input variables, is zero-mean Gaussian white noise, d is the system delay time, n y 、n u and n e are the maximum lags in Y(t), U(t) and e(t), respectively, and △ is the difference operator; Step 1.2: Linearize the nonlinear system into a controlled autoregressive integrated moving average (CARIMA) model: Among them, z -1 is the backward shift operator, ΔU(t-1)=U(t-1)-U(t-2) is the increment of the input vector at time t-1, and Denote the backshift operator z -1 The polynomial matrix of , and: C(z -1 )=I, Where I is the identity matrix; Step 1.3: Collect actual sample data of 1000MW supercritical boiler turbine unit at different times; Step 1.4: Take the electric power, main steam pressure and steam temperature as the control output Y(t) of the nonlinear system, and the fuel flow, regulating valve amount and water flow as the control input U(t) of the system, then: Y(t)=[y1(t),y2(t),y3(t)] T , U(t)=[u1(t),u2(t),u3(t)] T , n a =n b =3, Among them, Y(t) is the output vector of the nonlinear system, U(t) is the output vector of the nonlinear system, y1(t) is the electric power at time t, y2(t) is the main steam pressure at time t, y3(t) is the steam temperature at time t, u1(t) is the fuel flow at time t, u2(t) is the regulating valve amount at time t, and u3(t) is the water flow at time t; Step 1.5: Split the nonlinear system into several local linear sub-models: Among them, Y i (t) is the output vector of the i-th local linear sub-model, △U i (t-1) is the increment of the input vector of the i-th local linear sub-model at time t-1, e i (t) is the zero-mean Gaussian white noise of the i-th local linear sub-model, and B i (z -1 )∈R 3×3 is the polynomial matrix of the i-th local linear sub-model; S3: Construct a recurrent fuzzy neural network (RFNN) to predict the electric power, main steam pressure and steam temperature in different local linear sub-models, couple the prediction output of the local linear sub-models, and obtain the prediction output results of the supercritical boiler turbine unit; The step S3 comprises the following steps: Step 3.1: Construct a 7-layer recurrent fuzzy neural network (RFNN). The first layer uses U(t) as the input of RFNN. As the output of RFNN; Step 3.2: Design the improved fuzzy membership function in the second layer Take L cluster centers as the centers of the improved fuzzy membership function, in, is the input variable u j (t) and fuzzy membership function Center The distance between By fixing the cluster centroid The jth element in is determined, while the cluster centroid is fixed Determined by the DBSCAN clustering algorithm, η ij is the adaptive overlap coefficient for adjusting the width of the fuzzy membership function, and η ij The larger the value of the fuzzy membership function The greater the width; Step 3.3: The third layer constructs the spatial intensity layer. Each node represents the spatial intensity of a fuzzy rule. Each fuzzy rule R i The spatial trigger intensity ω at time t i (t): Step 3.4: The fourth layer constructs the time intensity layer. Each node recursively updates information to form a feedback loop. is the output of the i-th recursive rule node at time t, that is is the temporal trigger strength, which depends not only on the spatial trigger strength ω at time t i (t), also depends on the time trigger strength at time t-1 The output of each recursive rule node is the spatial trigger intensity ω at time t i (t) and the time trigger intensity at time t-1 A linear combination of: Among them, β i is the dynamic feedback weight between (0, 1), which is used to measure the spatial trigger intensity ω at time t i (t) and the time trigger intensity at time t-1 The weight between Step 3.5: The fifth layer constructs a sequential layer. The nodes in this layer are called subsequent nodes, which are used to adjust the weights of the time intensity nodes in the fourth layer. The output f of the subsequent node i (t) is: Among them, a ij and i It is a free parameter that needs to be adjusted during parameter learning; Step 3.6: The sixth layer normalizes the output nodes: in, is the output f of the subsequent node i Normalized result of (t); Step 3.7: The seventh layer predicts the electric power, main steam pressure and steam temperature in different local linear sub-models, and couples the prediction outputs of L local linear sub-models; Step 3.8: Obtain the predicted output results for the supercritical boiler turbine unit 2. The turbine unit prediction method based on density clustering and recurrent fuzzy neural network according to claim 1 is characterized in that: The step S2 comprises the following steps: Step 2.1: Use the DBSCAN clustering algorithm to cluster the input data U(t) and automatically determine the number of clusters to be L; Step 2.2: Determine that the number of local linear sub-models and the number of fuzzy rules are both L; Step 2.2: Design L IF-THEN fuzzy rules R i :If u1(t) is u2(t) is u3(t) is Then the linear output of the i-th local linear sub-model is: Among them, R i is the i-th fuzzy rule, is the input variable u j The i-th fuzzy set, A iq ∈R 3×3 and B ik ∈R 3×3 is the polynomial matrix of the i-th local linear sub-model, i = 1, 2, …, L; j = 1, 2, 3; q = 1, 2, 3; k = 1, 2, 3.

Images