Rapid dynamic response method for ultra-supercritical unit based on multiple models

Abstract

The invention discloses an ultra-supercritical unit rapid dynamic response method based on multiple models, and relates to the technical field of thermal power generation. According to the method, a nonlinear control method which unifies internal model control and generalized predictive control is provided, and nonlinear control over an ultra-supercritical unit (USC) system is effectively achieved: firstly, through internal model control based on LSTM-Transform, the USC system achieves rapid convergence to be close to an expected value, and the response speed of a load is remarkably increased; secondly, the model is further optimized by enhancing a composite weighted human learning optimization network (CWHLO), and the accuracy of the model is improved; and finally, a generalized predictive control (CWHLO-GPC) algorithm based on an enhanced composite weighted human learning optimization network is adopted to control the USC system, so that high-precision load tracking performance is realized, and the problem of slow dynamic response of the ultra-supercritical unit system is solved.

Description

Technical Field

[0001] The invention belongs to the technical field of thermal power generation, and in particular relates to a fast dynamic response method of an ultra-supercritical unit based on multiple models. Background Art

[0002] With the massive access of renewable energy sources such as wind power, solar power and other random and uncertain energy sources to the power grid, the traditional power system is facing unprecedented challenges. As an advanced technology in the field of thermal power generation, ultra-supercritical unit power generation technology needs to have fast load response capabilities to achieve deep peak regulation, promote the consumption of new energy and maintain the stability of the power grid frequency as a high proportion of renewable energy continues to be connected to the power grid. Compared with traditional thermal power units, as ultra-supercritical units operate for a long time under high-voltage environment and are large-scale, the complex characteristics of ultra-supercritical units such as strong nonlinearity, strong coupling, and load disturbance make the control of the units more difficult.

[0003] Generalized Predictive Control (GPC) is a model-based rolling optimization control strategy that can optimize the system by adjusting the control variables according to the objective function. It has been widely used in ultra-supercritical unit control systems due to its strong robustness and ability to overcome system delays. However, in actual control systems, in addition to overshoot and robustness, response speed is also an important indicator. At present, researchers combine generalized predictive control with complementary technologies such as Kalman filtering, neural networks, and iterative learning to improve control performance. Although the improved GPC method can enable ultra-supercritical units to obtain better load tracking performance, it still cannot solve the problem of slow dynamic response of ultra-supercritical units. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, an object of the present invention is to provide a fast dynamic response method for ultra-supercritical units based on multiple models, aiming to solve the problem of slow dynamic response of existing ultra-supercritical units.

[0005] The technical solution of the present invention is:

[0006] A multi-model-based ultra-supercritical unit rapid dynamic response method comprises the following steps:

[0007] Step 1: Build an IMC model based on LSTM-Transformer to obtain the ideal input and output of the USC unit model;

[0008] Step 2: Based on the ideal input and output of the USC unit model, the USC operating load range is divided into n local areas, and an enhanced composite weighted human learning optimization network CWHLO is constructed to model and process each local area to obtain the fuzzy model of the entire USC unit;

[0009] Step 3: Based on the enhanced composite weighted human learning optimization network CWHLO, design the CWHLO-GPC control algorithm to obtain the optimal local controller increment for each local area;

[0010] Step 4: Based on the optimal local controller increment, a global controller is designed to control the USC unit model to achieve the goal of fast dynamic response of the USC unit.

[0011] Furthermore, the step 1 comprises the following steps:

[0012] Step 1.1: Construct an LSTM-Transformer model to pre-process the input and output of the USC unit operation, and train the LSTM-Transformer model;

[0013] Step 1.2: Build an IMC model based on the LSTM-Transformer model to obtain the ideal input and output of the USC unit model;

[0014] The IMC model includes a forward model and a reverse model, the forward model serves as an internal model of the IMC model, and the reverse model serves as an internal model controller; in the forward model, an LSTM-Transformer model is used to predict future outputs based on historical inputs and outputs; in the reverse model, an LSTM-Transformer model is also used, but the input is the future expected output of the ultra-supercritical unit and the current actual output, and the output is the control quantity; through the combined action of the forward model and the reverse model, a rapid dynamic response of the ultra-supercritical unit is achieved.

[0015] Furthermore, the step 2 comprises the following steps:

[0016] Step 2.1: Use the K-shape clustering algorithm to divide the USC operating load range into n local areas, and establish a fuzzy membership function for each local area to obtain the local area fuzzy membership;

[0017] Step 2.2: Normalize the fuzzy membership of the local area to obtain the activation intensity of each local area;

[0018] Step 2.3: Use enhanced human learning optimization HLO algorithm to perform local area modeling to obtain a local area model;

[0019] (1) A local area model is designed by using a USC unit in the form of a control autoregressive integral moving average to obtain the output of the local area model; a Z-transform is performed on the output of the local area model using a zero-order holder to obtain a local area model of the USC unit;

[0020] (2) Using the enhanced HLO algorithm to search the model coefficients of the divided local area models in batches to simulate the local area operation process, that is, to identify the unknown parameters of the local area model;

[0021] (3) In order to ensure a wider exploration space and faster search speed, the learning operator is randomly selected in the enhanced HLO algorithm;

[0022] (4) Using the random learning operator of the enhanced HLO algorithm to expand the knowledge space at the beginning of the search;

[0023] (5) After a certain number of random searches, an individual knowledge database is obtained;

[0024] (6) Individuals learn from other individuals to obtain a social knowledge database and obtain individual solutions by executing social learning operators;

[0025] Step 2.4: Based on the activation intensity of the local area, the local area model is fuzzified to obtain the fuzzy model of the entire USC unit.

[0026] Furthermore, the step 3 comprises the following steps:

[0027] Step 3.1: Embed the generalized predictive control GPC into the enhanced composite weighted human learning optimization network to design a cost function for each local area;

[0028] Step 3.2: Calculate the local controller increment according to the local area cost function;

[0029] Step 3.3: Continuously iterate to seek the minimum local area cost function and obtain the optimal local controller increment.

[0030] Further, the step 4 comprises the following steps:

[0031] Step 4.1: Design a local controller based on the optimal local controller increment and the characteristics of GPC rolling optimization and predictive control;

[0032] Step 4.2: Compound the local controllers to obtain the controller of the entire USC unit;

[0033] Step 4.3: Add the controller of the entire USC unit to the USC unit fuzzy model established in step 2, so as to adjust the input of the USC unit model so that the output can accurately track the desired trajectory, thereby achieving the purpose of rapid dynamic response of the ultra-supercritical unit.

[0034] Furthermore, the step 1.1 includes the following steps:

[0035] Step 1.1.1: Collect the long time series data of USC unit operation required for building the LSTM-Transformer model, normalize the data, and divide it into training set, validation set, and test set;

[0036] Step 1.1.2: Define the LSTM-Transformer model structure, where the LSTM layer is used to capture the temporal dependency of the sequence, and the Transformer layer is used to process the position information in the sequence and capture the global dependency;

[0037] Step 1.1.3: Train the LSTM-Transformer model using the training set, and then adjust the hyperparameters of the LSTM-Transformer model as needed;

[0038] Step 1.1.4: Use the validation set to verify the performance of the LSTM-Transformer model.

[0039] Step 1.1.5: Use the test set to evaluate the accuracy and generalization ability of the LSTM-Transformer model.

[0040] Step 1.1.6: Iteratively optimize the LSTM-Transformer model based on actual application requirements and performance considerations.

[0041] Furthermore, the enhanced human learning optimization (HLO) algorithm includes a random learning operator, an individual learning operator, and a social learning operator, and the combination of these three learning operators enables rapid search and accurate identification of model coefficients.

[0042] Furthermore, in step 2.3, an adaptive mechanism is used to enhance the human learning optimization HLO algorithm, and the formula is as follows:

[0043]

[0044] In the formula, I L (R) and S L (R) are the values ​​of the individual learning operator and the social learning operator after the Rth iteration, and are the initial values ​​of the individual learning operator and the social learning operator, and are the final values ​​of the individual learning operator and the social learning operator respectively, α is the shape coefficient, ns is the cutoff point, R represents the current iteration number, and epoch represents the total iteration number.

[0045] Compared with the prior art, the present invention has the following beneficial effects:

[0046] The invention proposes a nonlinear control method which unifies internal model control (IMC) and generalized predictive control (GPC), and effectively realizes nonlinear control of an ultra supercritical unit (USC) system. Firstly, through internal model control based on LSTM-Transformer, the USC unit realizes rapid convergence to the vicinity of an expected value, and significantly improves the response speed of the load. Secondly, the model is further optimized by enhancing a composite weighted human learning optimization network (CWHLO), and the accuracy of the model is improved. Finally, by adopting a generalized predictive control (CWHLO-GPC) algorithm based on an enhanced composite weighted human learning optimization network to control the USC unit, high-precision load tracking performance is achieved, and the problem of slow dynamic response of the ultra supercritical unit (USC) system is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 Schematic diagram of the process of the ultra-supercritical unit rapid dynamic response method based on multiple models in this embodiment;

[0048] Figure 2 Schematic diagram of a multi-model structure of a fast dynamic response method of an ultra-supercritical unit based on a multi-model in this embodiment;

[0049] Figure 3 This is an enhanced CWHLO network structure diagram of the ultra-supercritical unit fast dynamic response method based on multiple models in this implementation mode;

[0050] Figure 4 Schematic diagram of enhanced CWHLO algorithm of the ultra-supercritical unit fast dynamic response method based on multiple models in this embodiment;

[0051] Figure 5 Schematic diagram of local area activation weights of the ultra-supercritical unit rapid dynamic response method based on multiple models in this implementation mode;

[0052] Figure 6Schematic diagram of the enhanced CWHLO-GPC algorithm of the ultra-supercritical unit fast dynamic response method based on multiple models in this embodiment;

[0053] Figure 7 This is a graph showing the three output training results of the load, main steam pressure, and intermediate point temperature of the IMC model based on LSTM-Transformer in Example 1 of the present invention;

[0054] Figure 8 This is a training result diagram of three inputs of water supply, coal supply, and valve opening of the IMC model based on LSTM-Transformer in Example 1 of the present invention;

[0055] Fig. 9 This is a verification result diagram of the three outputs of load, main steam pressure, and intermediate point temperature of the IMC model based on LSTM-Transformer in Example 1 of the present invention;

[0056] Fig.10 This is a verification result diagram of three inputs of water supply, coal supply, and valve opening of the IMC model based on LSTM-Transformer in Example 1 of the present invention;

[0057] Fig.11 is a diagram showing changes in the middle point of a local area of ​​USC operating load in Example 1 of the present invention;

[0058] Fig.12 This is a comparison diagram of the operating performance of the three outputs of load, main steam pressure and intermediate point temperature under load reduction of the USC unit in Example 1 of the present invention;

[0059] Fig.13 This is a comparison diagram of the operating performance of the three inputs of water supply, coal supply and valve opening under load reduction of the USC unit in Example 1 of the present invention;

[0060] Fig.14 This is a comparison diagram of the operating performance of the three outputs of load, main steam pressure and intermediate point temperature under load increase of the USC unit in Example 1 of the present invention;

[0061] Fig.15 This is a comparison chart of the operating performance of the three inputs of water supply, coal supply and valve opening under load increase of the USC unit in Example 1 of the present invention;

[0062] Fig.16 This is a comparison diagram of the operating performance of the three outputs of load, main steam pressure and intermediate point temperature under load reduction of the USC unit in Example 2 of the present invention;

[0063] Fig.17 This is a comparison diagram of the operating performance of the three inputs of water supply, coal supply and valve opening under load reduction of the USC unit in Example 2 of the present invention;

[0064] Fig.18 This is a comparison diagram of the operating performance of the three outputs of load, main steam pressure and intermediate point temperature under load increase of the USC unit in Example 3 of the present invention;

[0065] Fig.19 This is a comparison chart of the operating performance of the three inputs of water supply, coal supply and valve opening under load increase of the USC unit in Example 3 of the present invention; DETAILED DESCRIPTION

[0066] To facilitate understanding of the present application, the present application will be described more comprehensively below with reference to the relevant drawings.

[0067] Thermal power generation is one of the main ways of global power supply at present. In order to solve the shortcomings of traditional thermal power units, the present invention provides a fast dynamic response method for ultra-supercritical units based on multiple models. The basic idea of ​​the present invention is as follows: construct an IMC model based on LSTM-Transformer for data processing, construct an enhanced composite weighted human learning optimization network based on the processed input and output to simulate the nonlinear operation process of the ultra-supercritical unit, and propose a CWHLO-GPC control algorithm to design a global controller, so as to control the ultra-supercritical unit model, so as to achieve the purpose of fast dynamic response of the ultra-supercritical unit, realize high-precision load tracking performance, and solve the problem of slow dynamic response of the ultra-supercritical unit (USC) system.

[0068] Figure 1 FIG. 1 is a flow chart of the ultra-supercritical unit rapid dynamic response method based on multiple models in this embodiment. Figure 1 As shown, the ultra-supercritical unit rapid dynamic response method based on multiple models comprises the following steps:

[0069] Step 1: Build an IMC model based on LSTM-Transformer, such as Figure 2 As shown, it is used to obtain the ideal input and output of the USC unit model;

[0070] At present, the commonly used ultra-supercritical unit model is a three-input-three-output model, that is, the coal feed u1, water feed u2 and valve opening u3 are selected as inputs, and the main steam pressure y1, intermediate point temperature y2 and load y3 are selected as outputs. Since there is a significant time delay in the ultra-supercritical unit USC, the output of the USC unit is affected by both the current input and the past variables. In order to make the internal model control more accurate, a composite neural network that is good at processing long sequence data, namely LSTM-Transformer, is used to establish a model that reflects the feedforward and feedback operation status of the USC unit to quickly obtain the control quantity.

[0071] Step 1.1, build an LSTM-Transformer model to process the long time series data of three inputs of coal feed u1, water feed u2 and valve opening u3 and three outputs of main steam pressure y1, intermediate point temperature y2 and load y3 during the operation of the USC unit, and train the LSTM-Transformer model;

[0072] (1) The long time series data during the operation of the USC units required for establishing the LSTM-Transformer model are collected and normalized to accelerate the convergence of the model. Then, the processed long time series data are divided into training set, validation set and test set.

[0073] (2) Define the LSTM-Transformer model structure, where the LSTM layer is used to capture the temporal dependency of the sequence, and the Transformer layer is used to process the position information in the sequence and capture global dependencies.

[0074] The LSTM layer includes: input layer, hidden layer, unit and output layer, as well as forget gate, input gate and output gate, among which,

[0075] ① Forget Gate

[0076] f t =σ[W f ·(h t-1 ,x t )+b f (1)

[0077] ② Input gate

[0078]

[0079] ③Output gate

[0080]

[0081] In formulas (1), (2), and (3), x t 、h t , C t and t They represent the states of the input layer, hidden layer, unit and output layer at time t, respectively. t-1 , C t-1 represents the hidden layer and the unit state at time t-1, is the intermediate state of the unit at time t, f t ,I t are the states of the forget gate and input gate at time t, f, I and o are the forget gate, input gate and output gate respectively, Wf , W o , W I and W c are the forget gate, input gate, output gate and unit weight matrix, b f , b o , b I and b c They are the forget gate, input gate, output gate and unit bias respectively, and σ and tanh are activation functions.

[0082] In the Transformer layer, there are position encoders, self-attention mechanism layers, and feedforward network layers, where:

[0083] ①Position encoder

[0084] Fixed position encoding (PositionalEncoding, PE) can set the posth row element of the position encoding matrix PE to:

[0085]

[0086] In formula (4), PE (pos,2a) is the element at row pos and column 2a, PE (pos,2a+1) is the element at row pos and column 2a+1, where pos is the position, a is the dimension, and d is the model It is the dimension of PE.

[0087] ②Self-attention mechanism layer

[0088] The self-attention mechanism can be expressed as:

[0089]

[0090] In formula (5), X represents the encoder input; W represents the hidden matrix between layers; d k represents the dimension of the attention mechanism layer; Q, K, and V represent query (the position of the current position in the attention mechanism), key (the position of all other positions), and value (the value of all other positions), respectively; W Q , W K , W V Represent the hidden matrices of Q, K, and V respectively; K T Represents the transposed matrix of K.

[0091]

[0092] In formula (6), the parameter matrix where headi represents the i-th sub-value, and there are a total of h sub-values.

[0093] ③ Feedforward network layer

[0094] The calculation formula of the feed-forward network layer (FFN) is expressed as:

[0095] FFN(x)=max(0,x r W1+b1)W2+b2 (7)

[0096] In formula (7), x r is the input of the feedforward network layer; W1 and W2 are parameter matrices, and b1 and b2 are bias vectors.

[0097] (3) Use the training set to train the LSTM-Transformer model, and then adjust the hyperparameters of the LSTM-Transformer model according to actual needs. The specific training process is as follows: First, the long time series data of the USC unit operation process is normalized and preprocessed to obtain the training set, and then the output is obtained by LSTM neural network training; then, the output is encoded by the position encoder in the Transformer model to encode the output data of the LSTM neural network. The encoded data matrix captures the relationship between different positions in the input sequence through the self-attention mechanism layer and performs nonlinear transformation and mapping of features through the fully connected feedforward layer. Finally, the output of the LSTM-Transformer model is obtained by normalization through the residual link.

[0098] (4) Use the validation set to verify the LSTM-Transformer model and use the mean square error (MSE) function to verify its model performance.

[0099] (5) Use the test set to evaluate the accuracy and generalization ability of the LSTM-Transformer model.

[0100] (6) Iteratively optimize the LSTM-Transformer model based on actual application requirements and performance considerations.

[0101] Step 1.2: Construct an IMC model based on the LSTM-Transformer model, including a forward model and a reverse model. The forward model serves as the internal model of the IMC model, and the reverse model serves as the internal model controller. In the forward model, the LSTM-Transformer model is used to predict future outputs based on historical inputs and outputs. In the reverse model, the LSTM-Transformer model is also used, but the input is the expected future output of the ultra-supercritical unit and the current actual output, and the output is the control quantity. Through the combined action of the forward model and the reverse model, the rapid dynamic response of the ultra-supercritical unit can be achieved.

[0102] Considering the ultra-supercritical unit (USC) has a fast water feed rate and a slow coal grinding rate, the expected output [y r1 (k),y r2 (k),y r3 (k)], the actual outputs [y1(k-1), y2(k-1), y3(k-1)] and [u1(k-1), u2(k-1), u3(k-1)] are taken as inputs, and [u1(k), u2(k), u3(k)] is taken as outputs. In the forward model, [u1(k), u2(k), u3(k), y1(k-1), y2(k-1), y3(k-1)] are taken as inputs, and [y1(k), y2(k), y3(k)] is taken as outputs.

[0103] Among them, y r1 (k),y r2 (k),y r3 (k) are the expected values ​​of the main steam pressure y1, the intermediate point temperature y2 and the load y3 at the moment k; y1(k-1), y2(k-1), y3(k-1) are the actual values ​​of the main steam pressure y1, the intermediate point temperature y2 and the load y3 at the moment k-1; y1(k), y2(k), y3(k) are the predicted values ​​of the main steam pressure y1, the intermediate point temperature y2 and the load y3 at the moment k; u1(k-1), u2(k-1), u3(k-1) are the actual values ​​of the coal feed u1, the water feed u2 and the valve opening u3 at the moment k-1; u1(k), u2(k), u3(k) are the predicted values ​​of the coal feed u1, the water feed u2 and the valve opening u3 at the moment k.

[0104] Step 2: Based on the ideal input and output of the IMC model obtained in step 1, Figure 3 , 4 As shown, an enhanced composite weighted human learning optimization network (CWHLO) is constructed to simulate the nonlinear operation process of ultra-supercritical units;

[0105] Step 2.1, use the K-shape clustering algorithm to divide the USC operating load range into n local areas, and establish a fuzzy membership function for each local area to evaluate the relationship between the load and the local area;

[0106] In this embodiment, the K-shape clustering algorithm is used to divide the USC operating load range into five local areas: (1) "high"; (2) "medium-high"; (3) "medium"; (4) "medium-low"; and (5) "low".

[0107] The K-shape clustering algorithm uses a shape-based distance (SBD) to cluster time series data. First, n cluster centers are determined according to specific task requirements, the SBD between the data in the time series and each current cluster center is calculated, and cluster assignment and update are completed. The cluster assignment and update steps are repeated until the clustering result no longer changes significantly or the preset maximum number of iterations is reached, and clustering is completed to determine the local area.

[0108] like Figure 5 As shown in the figure, taking load as a linguistic variable, the local fuzzy membership function of the USC operating load range is determined as:

[0109]

[0110] In formula (8), c i 、c i+1 They represent the membership of y to the i-th and i+1-th local regions respectively, and the values ​​of both parameters are between 0 and 1; y is the value of the language variable; center i and center i+1 are the center points of the i-th and i+1-th local areas respectively; the fuzzy membership function is used to evaluate the relationship between the load at a certain moment and n local areas, c i 、c i+1 The closer the value is to 1, the higher the degree of membership to the local area.

[0111] Step 2.2, normalize the fuzzy membership of the local area to obtain the activation intensity of each local area;

[0112]

[0113] In formula (9), Indicates c i The result obtained after normalization; n represents the number of local areas;

[0114] Will It is considered as a scheduler that determines the activation intensity of local regions.

[0115] Step 2.3, using enhanced human learning optimization HLO algorithm to perform local area modeling;

[0116] (1) The i-th local area model is designed by using the USC unit in the form of controlled autoregressive integrated moving average (CARIMA) to obtain the output of the i-th local area model:

[0117]

[0118] In formula (10), is the output of the i-th local area model, Y(tj) and U(t-j+1) represent the delayed output and input of the i-th local area model, respectively. ij , B ij are the polynomial matrices Y(tj) and U(t-j+1) of the i-th local area model, and V(t) represents white noise with zero mean.

[0119] The local area model of the USC unit is obtained by performing Z transform on equation (10) using a zero-order holder:

[0120]

[0121] In formula (11), y i1 (t), y i2 (t), y i3 (t) are the main steam pressure y1, intermediate point temperature y2 and load y3 of the ith local area of ​​the USC unit at time t; U i (t)=[u i1 (t)u i2 (t)u i3 (t)] T is the i-th local controller, u i1 (t),u i2 (t),u i3 (t) are the values ​​of the three inputs of coal supply u1, water supply u2 and valve opening u3 in the i-th local area of ​​the USC unit at time t; A i (Z -1 ), B i (Z -1 ) is the parameter to be identified, where A i (Z -1 ) is a 1×3 matrix, B i (Z -1 ) is a 3×3 matrix; C i (Z -1 ) is a 3×3 identity matrix, V(t) is a 3×1 zero-mean white noise matrix, Δ=1-Z -1 represents the difference operator, Z -1 It represents the inverse transform of Z, which is equivalent to subtracting one from t in the time domain and is obtained by the Plath transform, i.e., the Z transform.

[0122] For A in formula (10), i (Z -1 ), B i (Z -1) for further explanation.

[0123] A i (Z -1 )=[A i1 (Z -1 )A i2 (Z -1 )A i3 (Z -1 )] (12)

[0124] In formula (12), an example is given, in It is an unknown parameter that needs to be continuously optimized and identified through the enhanced HLO algorithm.

[0125]

[0126] In formula (13), an example is given, in It is an unknown parameter that needs to be continuously optimized and identified through the enhanced HLO algorithm.

[0127] (2) Using the enhanced HLO algorithm to search the model coefficients of the divided local area models in batches to simulate the local area operation process, that is, to identify the unknown parameters of the three-input-one-output model;

[0128] If the three-input-three-output model is directly identified, 36 parameters need to be identified at the same time, namely the unknown parameters mentioned in equations (12) and (13). The calculation amount is large and the identification accuracy is low. In order to avoid the above problems, parameter identification is now performed in batches, that is, the unknown parameters of the three-input-one-output model are identified. For example, first identify A i1 (Z -1 ), B i11 (Z -1 ), B i12 (Z -1 ), B i13 (Z -1 ), there are 3 unknown parameters in each equation, a total of 12, described as follows:

[0129]

[0130] In formula (14), x k Represents a set of identification parameters.

[0131] Each individual in the enhanced HLO algorithm can simulate three kinds of human learning behaviors through three learning operators, namely, random learning operator, individual learning operator and social learning operator, to continuously iterate and find the optimal solution. When the optimal solution is found, the parameter identification is completed.

[0132] After all individuals generate new candidate solutions, the fitness value is calculated according to the predefined fitness function to evaluate the fitness of each individual. In addition, the fitness function is also used to update the personal knowledge database (IKD) and the social knowledge database (SKD) for subsequent searches to summarize and update the experience for better learning in the following steps. The personal knowledge database (IKD) and the social knowledge database (SKD) are two key knowledge storage methods in the human learning optimization algorithm (HLO), which simulate the personal experience and social information used by humans in the learning and decision-making process. For the update of IKD, if the fitness value of the newly generated solution is better than the worst solution in IKD, or the number of solutions in the current IKD is less than the predefined value, the newly generated solution is stored in IKD. For the update of SKD, if the fitness value of the optimal solution of the current generation is better than the worst solution in SKD, or the number of current solutions in SKD is less than the predefined number, the optimal solution of the current generation is saved in SKD.

[0133] When the algorithm is iteratively updated, if the fitness value of the solution changes very little in several consecutive iterations or meets the preset stop condition, the algorithm stops running and outputs the identified parameters.

[0134] The fitness function used in the enhanced HLO algorithm is In the formula, RMSE represents the root mean square error, L is the number of data, is the estimated value obtained by the HLO algorithm, Y r (j) is the expected value.

[0135] (3) In order to ensure a wider exploration space and faster search speed, the learning operator is randomly selected in the enhanced HLO algorithm;

[0136]

[0137] In formula (15), x represents the learning operator, RLO represents the random learning operator, ILO represents the individual learning operator, SLO represents the social learning operator, r4 is a random number between [0, 1], and p1 and p2 represent the probability of the enhanced HLO algorithm executing a certain search operator.

[0138] (4) The random learning operator (RLO) of the enhanced HLO algorithm is used to expand the knowledge space at the beginning of the search. The search result is

[0139]

[0140] In formula (16), xb represents the search results of the random learning operator, are the maximum and minimum search ranges of individual b; r1 is a random number between [0, 1].

[0141] (5) After a certain number of random searches, the individual knowledge database (IKD) is obtained;

[0142] IKD = [ikd1 ikd2…ikd l …ikd N ] T (17)

[0143]

[0144] In formulas (17) and (18), N represents the number of individuals in the individual knowledge database, ikd l represents the knowledge database of the lth individual, represents the jth element of the pth optimal solution of the lth individual, T and M represent the size of IKD and the dimension of the solution respectively.

[0145] Generate new recognition results by executing the Individual Learning Operator (ILO)

[0146]

[0147] In formula (19), represents the area where the best solution may exist, represents the jth element of the pth optimal solution of the lth individual, sk qj represents the jth element of the qth optimal solution in the social knowledge base (SKD); r2 is a random number between [-1, 1], I L represents the personal learning factor, x lj represents the jth solution of the lth individual.

[0148] (6) Individuals learn from other individuals and obtain a social knowledge database (SKD):

[0149]

[0150] In formula (20), sdk q represents the optimal solution of the qth individual, sk qj is the jth parameter of the optimal solution for the qth individual, S and M represent the population size and the dimension of the solution, respectively.

[0151] By executing the Social Learning Operator (SLO)

[0152] x i =sk qj +S L·r3·X SLO (twenty one)

[0153] In formula (20), represents the jth element of the pth optimal solution of the lth individual, sk qj represents the jth element of the qth optimal solution in the social knowledge base (SKD); S L is the social learning factor, r3 is a random number between [-1, 1], x i represents the solution of the ith individual.

[0154] (7) Considering that the search ranges of ILO and SLO are respectively L and S L OK, I L and S L Too large a value may lead to excessive focus on the convergence speed of the algorithm. On the contrary, I L and S L Too small may lead to excessive focus on the algorithm's search accuracy. Therefore, an adaptive mechanism is used to enhance the human learning optimization HLO algorithm. The formula is as follows:

[0155]

[0156] In formula (22), I L (R) and S L (R) are the values ​​of ILO and SLO after the Rth iteration, and are the initial values ​​of ILO and SLO, and are the final values ​​of ILO and SLO respectively, α is the shape coefficient, ns is the cutoff point, R represents the current iteration number, and epoch represents the total number of iterations.

[0157] Step 2.4: Fuzzify the local model established by the enhanced human learning optimization HLO algorithm to obtain the fuzzy model of the entire USC unit. According to equations (9) and (10), the fuzzy model of the entire USC unit can be expressed as

[0158]

[0159] In formula (23), is the output of the fuzzy model of the entire USC unit, is the output of the ith local area model of the USC unit, c i (t) is the fuzzy membership of the i-th local region, It is the result obtained by normalizing the fuzzy membership of the i-th local area.

[0160] Step 3: Based on the enhanced composite weighted human learning optimization network proposed in step 2, Figure 6 As shown, the CWHLO-GPC control algorithm is designed;

[0161] Step 3.1: Based on step 2, embed the generalized predictive control GPC into the CWHLO network and design a cost function for each local area. The cost function of the USC unit can be written as:

[0162]

[0163] In formula (24), J is the cost function of the USC unit, Q y and Q u is the weighting coefficient of the prediction error and the control increment, is the predicted value of the i-th local area, Y ri is the reference trajectory of the ith local region, N y and N u is the maximum cost calculation range for forecast errors and control variables, is the predicted value of the i-th local area at time t+j, Y ri (t+j) is the reference trajectory of the i-th local area at time t+j, ΔU i (t+j-1) Controller of the i-th local area at time t+j.

[0164] In order to reduce the amount of calculation, the Cauchy inequality is used to simplify equation (24) to obtain the upper bound of the cost function J:

[0165]

[0166] In formula (25), Local area cost function J i The sum is the upper bound of the cost function J, minimizing J i Essentially equivalent to minimizing J.

[0167] Step 3.2: According to the local area cost function J in step 3.1 i Calculate the local controller increment as:

[0168]

[0169] In formula (26), ΔU i (t) is the local controller increment, H i and G i Satisfies the Diophantine equation.

[0170] The Diophantine equation is shown below:

[0171]

[0172] Step 3.3: Continuously iterate to find the minimum local area cost function and obtain the optimal local controller increment Δ′U i (t);

[0173] Continuously iterate to find the minimum local area cost function J i , when the local area cost function J i When the minimum value is reached, according to equations (25) and (26), we can know that the minimum J i It is essentially equal to minimizing J, and the optimal local controller increment Δ′U is obtained at this time i (t):

[0174]

[0175] In formula (27), P = [1, 1, 1, 0, ..., 0], and P is the scheduler.

[0176] Step 4: Based on the local controller increment designed by the CWHLO-GPC control algorithm in step 3, a global controller is designed to control the ultra-supercritical unit model to achieve the purpose of rapid dynamic response of the ultra-supercritical unit.

[0177] Step 4.1: The optimal local controller increment Δ′U obtained in step 3 i (t) and the characteristics of GPC rolling optimization and predictive control, design the local controller U i (t) is

[0178] U i (t) = U i (t-1)+Δ′U i (t) (28)

[0179] Step 4.2, compound the local controllers designed in step 4.1 to obtain the controller of the entire USC unit;

[0180]

[0181] In formula (29), U(t) is the controller of the entire USC unit.

[0182] Step 4.3, add the controller U(t) of the entire USC unit obtained to the fuzzy model of the USC unit established in step 2, so as to adjust the input of the USC unit model so that the output can accurately track the desired trajectory, thereby achieving the purpose of rapid dynamic response of the ultra-supercritical unit.

[0183] Example 1

[0184] This embodiment tests the method of the present invention on a 1000W MUSC unit. In order to enhance the practicality of the experiment, this experiment selected 5000 sets of operating data of a 1000W MUSC unit for training through the LSTM-Transformer-based IMC model. These data cover almost all dynamic characteristics of the operation process. The training results are as follows: Figure 7 , 8 In addition, 2000 sets of data were selected for LSTM-Transformer-based IMC model verification. The verification results are shown in Fig. 9 , 10 shown.

[0185] The experiment uses a K-Shape clustering algorithm to divide the entire USC operating load area. The change process of the middle point of the local area of ​​USC operating load is as follows: Fig.11 Then, the experiment uses the enhanced HLO network to establish the local model of the USC unit according to the trained input and output of the IMC model, and the unknown parameter identification results in the local model are processed by composite weighting to obtain the unknown parameter identification results in the overall model as follows:

[0186] 1) The unknown parameter identification results of the model with three inputs of coal feed rate u1, water feed rate u2 and valve opening u3 and one output of main steam pressure y1 are as follows:

[0187] A1=1-2.0866z -1 +1.2632z -2 -0.1698z -3

[0188] B 11 =-0.003+0.005z -1 -0.0006z -2

[0189] B 12 =0.00006-0.0005z -1 -0.000003z -2

[0190] B 13 =0.0026-0.0018z -1 -0.0015z -2

[0191] 2) The unknown parameter identification results of the model with three inputs of coal supply u1, water supply u2 and valve opening u3 and one output of intermediate point temperature y2 are as follows:

[0192] A2=1-1.3244z -1 +0.3202z -2-0.0181z -3

[0193] B 21 =0.0063-0.0150z -1 +0.0106z -2

[0194] B 22 =0.0006-0.0012z -1 +0.0003z -2

[0195] B 23 =-0.0426-0.0492z -1 -0.0915z -2

[0196] 3) The unknown parameter identification results of the model with three inputs of coal supply u1, water supply u2 and valve opening u3 and one output of load y3 are as follows:

[0197] A3=1-0.9688z -1 +0.1822z -2 -0.2050z -3

[0198] B 31 =0.0353-0.0738z -1 +0.0526z -2

[0199] B 32 =-0.0036+0.0011z -1 -0.0083z -2

[0200] B 33 =-0.3401-0.5562z -1 -0.2215z -2

[0201] Finally, based on the above model, the experiment combined the local controllers designed based on GPC according to the dispatcher 00 to control the entire USC unit. The operating performance of the USC unit under a wide range of load reduction (8MW-600MW) is shown in the figure. Fig.12 , 13 The operating performance of the USC unit under a wide range of load increases (600MW-940MW) is as follows: Fig.14 , 15 ; The numerical comparison results are shown in Table 1 and Table 2.

[0202] Table 1 Comparison of IAE under wide range of load changes

[0203]

[0204] Table 2 Comparison of stabilization time under large range load changes

[0205]

[0206] It can be seen from the results that the method of the present invention can quickly adjust the control variables, obtain the output of the USC unit in a short time, and accurately track the desired trajectory. Although there is a slight overshoot of the feedwater during the controller switching process, the impact of the overshoot of the feedwater is minimal in the USC unit. In addition, the two inputs and three outputs all show satisfactory performance, indicating that the method of the present invention is relatively good.

[0207] Compared with Fuzzy-GPC, the IAE of this method is improved by an average of 74.36%, and compared with the traditional GPC, it is improved by an average of 87.93%. In terms of stabilization time, compared with Fuzzy-GPC, it is improved by an average of 13.73%, and compared with the traditional GPC, it is improved by an average of 20.51%. The method of the present invention shows its superiority, because the internal model controller is used as a pre-regulator to provide a rapidly changing control quantity, so that the output of the USC unit can quickly approach the desired trajectory. On this basis, CWHLO-GPC is further controlled, and the output of the USC unit can accurately track the desired trajectory by fine-tuning the control variables. For Fuzzy-GPC, although the output of the USC unit can quickly track the desired trajectory, the adjustment speed is significantly slower than the method of the present invention. For traditional GPC, when the load exceeds a certain range, the USC unit control model established based on the GPC method will be mismatched, resulting in the inability to effectively adjust the control quantity to accurately track the expected trajectory.

[0208] Example 2

[0209] This embodiment tests the method of the present invention on a 660WM USC unit. In order to enhance the practicality of the experiment, this experiment selected 5000 sets of operating data of a 660WMUSC unit for training through the LSTM-Transformer-based IMC model. These data cover almost all dynamic characteristics during operation. In addition, 2000 sets of data were selected for verification of the LSTM-Transformer-based IMC model.

[0210] The experiment used a K-Shape clustering algorithm to divide the entire USC operating load area. Then, the experiment used the enhanced HLO network to establish the local model of the USC unit based on the trained input and output obtained from the IMC model. The unknown parameter identification results in the local model were processed by composite weighting to obtain the unknown parameter identification results in the overall model as follows:

[0211] 1) The unknown parameter identification results of the model with three inputs of coal feed rate u1, water feed rate u2 and valve opening u3 and one output of main steam pressure y1 are as follows:

[0212] A1=1-2.1223z -1 +1.2145z -2 -0.1588z -3

[0213] B 11 =-0.003+0.006z -1 -0.0007z -2

[0214] B 12 =0.00008-0.0005z -1 -0.000005z -2

[0215] B 13 =0.0025-0.0017z -1 -0.0018z -2

[0216] 2) The unknown parameter identification results of the model with three inputs of coal supply u1, water supply u2 and valve opening u3 and one output of intermediate point temperature y2 are as follows:

[0217] A2=1-1.3165z -1 +0.3174z -2 -0.0179z -3

[0218] B 21 =0.006-0.0148z -1 +0.0102z -2

[0219] B 22 =0.0006-0.0011z -1 +0.0002z -2

[0220] B 23 =-0.0416-0.0486z -1 -0.0901z -2

[0221] 3) The unknown parameter identification results of the model with three inputs of coal supply u1, water supply u2 and valve opening u3 and one output of load y3 are as follows:

[0222] A3=1-0.9537z -1 +0.1786z-2 -0.2002z -3

[0223] B 31 =0.0346-0.0712z -1 +0.0537z -2

[0224] B 32 =-0.0031+0.0013z -1 -0.0078z -2

[0225] B 33 =-0.3326-0.5541z -1 -0.2203z -2

[0226] Finally, based on the above model, the experiment combined the local controllers designed based on GPC according to the dispatch program to control the entire USC unit. The operating performance of the USC unit under a large range of load reduction (620MW-420MW) is shown in the figure. Fig.16 , 17 The numerical comparison results are shown in Tables 3 and 4.

[0227] Table 3 Comparison of IAE under wide range of load changes

[0228]

[0229] Table 4 Comparison of stabilization time under large range load changes

[0230]

[0231] It can be seen from the results that the proposed control strategy can quickly adjust the control variables, obtain the output of the USC unit in a short time, and accurately track the desired trajectory. Although there is a slight overshoot of the feedwater during the controller switching process, the impact of the feedwater overshoot in the USC unit is minimal. In addition, the two inputs and three outputs all show satisfactory performance, indicating that the proposed fast dynamic response method is relatively good.

[0232] Compared with Fuzzy-GPC, the IAE of this method is improved by 69.53% on average, and by 80.61% on average compared with traditional GPC. In terms of stabilization time, it is improved by 13.42% on average compared with Fuzzy-GPC, and by 18.56% on average compared with traditional GPC.

[0233] Example 3

[0234] This embodiment tests the method of the present invention on a 350WM USC unit. In order to enhance the practicality of the experiment, this experiment selected 5000 sets of operating data of a 350WMUSC unit for training through the LSTM-Transformer-based IMC model, and these data almost cover all dynamic characteristics during the operation process. In addition, 2000 sets of data were selected for verification of the LSTM-Transformer-based IMC model.

[0235] The experiment used a K-Shape clustering algorithm to divide the entire USC operating load area. Then, the experiment used the enhanced HLO network to establish the local model of the USC unit based on the trained input and output obtained from the IMC model. The unknown parameter identification results in the local model were processed by composite weighting to obtain the unknown parameter identification results in the overall model as follows:

[0236] 1) The unknown parameter identification results of the model with three inputs of coal feed rate u1, water feed rate u2 and valve opening u3 and one output of main steam pressure y1 are as follows:

[0237] A1=1-2.0322z -1 +1.2321z -2 -0.1709z -3

[0238] B 11 =-0.004+0.003z -1 -0.0009z -2

[0239] B 12 =0.0001-0.0003z -1 -0.000005z -2

[0240] B 13 =0.0024-0.0019z -1 -0.0013z -2

[0241] 2) The unknown parameter identification results of the model with three inputs of coal supply u1, water supply u2 and valve opening u3 and one output of intermediate point temperature y2 are as follows:

[0242] A2=1-1.3342z -1 +0.3215z -2 -0.0204z -3

[0243] B 21 =0.0061-0.0162z -1 +0.0108z -2

[0244] B 22 =0.0005-0.0011z -1 +0.0005z -2

[0245] B 23 =-0.0442-0.0501z -1 -0.0889z -2

[0246] 3) The unknown parameter identification results of the model with three inputs of coal supply u1, water supply u2 and valve opening u3 and one output of load y3 are as follows:

[0247] A3=1-0.9431z -1 +0.1921z -2 -0.2113z -3

[0248] B 31 =0.0376-0.0724z -1 +0.0518z -2

[0249] B 32 =-0.0038+0.0012z -1 -0.0088z -2

[0250] B 33 =-0.3362-0.5638z -1 -0.2256z -2

[0251] Finally, based on the above model, the experiment combined the local controllers designed based on GPC according to the dispatch program to control the entire USC unit. The operating performance of the USC unit under a wide range of load changes (200MW-330MW) is shown in the figure. Fig.18 , 19 The numerical comparison results are shown in Tables 5 and 6.

[0252] Table 5 Comparison of IAE under wide range of load changes

[0253]

[0254] Table 6 Comparison of stabilization time under wide range of load changes

[0255]

[0256] It can be seen from the results that the proposed control strategy can quickly adjust the control variables, obtain the output of the USC unit in a short time, and accurately track the desired trajectory. Although there is a slight overshoot of the feedwater during the controller switching process, the impact of the feedwater overshoot in the USC unit is minimal. In addition, the two inputs and three outputs all show satisfactory performance, indicating that the proposed fast dynamic response method is relatively good.

[0257] Compared with Fuzzy-GPC, the IAE of this method is improved by an average of 70.11%, and compared with traditional GPC, it is improved by an average of 85.16%. In terms of stabilization time, compared with Fuzzy-GPC, it is improved by an average of 12.72%, and compared with traditional GPC, it is improved by an average of 22.64%.

[0258] It should be understood that, inspired by the technical concept of the present invention, those skilled in the art may make various improvements or changes based on the above content without departing from the content of the present invention, which still fall within the protection scope of the present invention.

Claims

1. A multi-model based ultra-supercritical unit rapid dynamic response method, characterized in that: The following steps are involved: Step 1: Build an IMC model based on LSTM-Transformer to obtain the ideal input and output of the USC unit model; Step 2: Based on the ideal input and output of the USC unit model, the USC operating load range is divided into n local areas, and an enhanced composite weighted human learning optimization network CWHLO is constructed to model and process each local area to obtain the fuzzy model of the entire USC unit; Step 3: Based on the enhanced composite weighted human learning optimization network CWHLO, design the CWHLO-GPC control algorithm to obtain the optimal local controller increment for each local area; Step 4: Based on the optimal local controller increment, a global controller is designed to control the USC unit model to achieve the goal of fast dynamic response of the USC unit.

2. The method for rapid dynamic response of ultra-supercritical units based on multiple models according to claim 1, characterized in that: The step 1 further comprises the following steps: Step 1.1: Build an LSTM-Transformer model to pre-process the input and output of the USC unit operation, and train the LSTM-Transformer model; Step 1.2: Build an IMC model based on the LSTM-Transformer model to obtain the ideal input and output of the USC unit model; The IMC model includes a forward model and a reverse model, the forward model serves as an internal model of the IMC model, and the reverse model serves as an internal model controller; in the forward model, an LSTM-Transformer model is used to predict future outputs based on historical inputs and outputs; in the reverse model, an LSTM-Transformer model is also used, but the input is the future expected output of the ultra-supercritical unit and the current actual output, and the output is the control quantity; through the combined action of the forward model and the reverse model, a rapid dynamic response of the ultra-supercritical unit is achieved.

3. The method for rapid dynamic response of ultra-supercritical units based on multiple models according to claim 1, characterized in that: The step 2 further comprises the following steps: Step 2.1: Use the K-shape clustering algorithm to divide the USC operating load range into n local areas, and establish a fuzzy membership function for each local area to obtain the local area fuzzy membership; Step 2.2: Normalize the fuzzy membership of the local area to obtain the activation intensity of each local area; Step 2.3: Use enhanced human learning optimization HLO algorithm to perform local area modeling to obtain a local area model; (1) A local area model is designed by using a USC unit in the form of a control autoregressive integral moving average to obtain the output of the local area model; a Z-transform is performed on the output of the local area model using a zero-order holder to obtain a local area model of the USC unit; (2) Using the enhanced HLO algorithm to search the model coefficients of the divided local area models in batches to simulate the local area operation process, that is, to identify the unknown parameters of the local area model; (3) In order to ensure a wider exploration space and faster search speed, the learning operator is randomly selected in the enhanced HLO algorithm; (4) Using the random learning operator of the enhanced HLO algorithm to expand the knowledge space at the beginning of the search; (5) After a certain number of random searches, an individual knowledge database is obtained; (6) Individuals learn from other individuals to obtain a social knowledge database and obtain individual solutions by executing social learning operators; Step 2.4: Based on the activation intensity of the local area, the local area model is fuzzified to obtain the fuzzy model of the entire USC unit.

4. The method for rapid dynamic response of ultra-supercritical units based on multiple models according to claim 1, characterized in that: The step 3 further comprises the following steps: Step 3.1: Embed the generalized predictive control GPC into the enhanced composite weighted human learning optimization network to design a cost function for each local area; Step 3.2: Calculate the local controller increment according to the local area cost function; Step 3.3: Continuously iterate to seek the minimum local area cost function and obtain the optimal local controller increment.

5. The method for rapid dynamic response of ultra-supercritical units based on multiple models according to claim 1, characterized in that: The step 4 further comprises the following steps: Step 4.1: Design a local controller based on the optimal local controller increment and the characteristics of GPC rolling optimization and predictive control; Step 4.2: Compound the local controllers to obtain the controller of the entire USC unit; Step 4.3: Add the controller of the entire USC unit to the USC unit fuzzy model established in step 2, so as to adjust the input of the USC unit model so that the output can accurately track the desired trajectory, thereby achieving the purpose of rapid dynamic response of the ultra-supercritical unit.

6. The method for rapid dynamic response of ultra-supercritical units based on multiple models according to claim 2, characterized in that: The step 1.1 further comprises the following steps: Step 1.1.1: Collect the long time series data of USC unit operation required for building the LSTM-Transformer model, normalize the data, and divide it into training set, validation set, and test set; Step 1.1.2: Define the LSTM-Transformer model structure, where the LSTM layer is used to capture the temporal dependency of the sequence, and the Transformer layer is used to process the position information in the sequence and capture the global dependency; Step 1.1.3: Train the LSTM-Transformer model using the training set, and then adjust the hyperparameters of the LSTM-Transformer model as needed; Step 1.1.4: Use the validation set to verify the performance of the LSTM-Transformer model. Step 1.1.5: Use the test set to evaluate the accuracy and generalization ability of the LSTM-Transformer model. Step 1.1.6: Iteratively optimize the LSTM-Transformer model based on actual application requirements and performance considerations.

7. The method for rapid dynamic response of ultra-supercritical units based on multiple models according to claim 3, characterized in that: The enhanced human learning optimization (HLO) algorithm includes a random learning operator, an individual learning operator, and a social learning operator. The combination of these three learning operators enables rapid search and accurate identification of model coefficients.

8. The method for rapid dynamic response of ultra-supercritical units based on multiple models according to claim 3, characterized in that: In step 2.3, an adaptive mechanism is used to enhance the human learning optimization HLO algorithm, and the formula is as follows: Among them, I L (R) and S L (R) are the values ​​of the individual learning operator and the social learning operator after the Rth iteration, and are the initial values ​​of the individual learning operator and the social learning operator, and are the final values ​​of the individual learning operator and the social learning operator respectively, α is the shape coefficient, ns is the cutoff point, R represents the current iteration number, and epoch represents the total iteration number.

Images