Improved modeling method for coordinated control system of ultra-supercritical unit

Abstract

The invention provides an improved modeling method for an ultra-supercritical unit coordinated control system, and relates to the technical field of ultra-supercritical unit coordinated control systems.Firstly, a new data-driven experiment modeling method is designed, and the capacity of extracting long-term dependency relationships and local features is combined with an improved transformer network of ResNet50 and a hyper-parameter optimization scheme; and the modeling accuracy of the CCS under the conditions of peak regulation and rapid power change is ensured. Then, for improvement of the perception domain and the complex feature learning ability, an improved neural network based on transformer convolution operation and a ResNet50 self-attention mechanism is provided, and the convergence of the network is proved by using a Lyapunov function. And finally, an HHBO algorithm based on a reactor optimizer and Harris eagle optimization is provided, the convergence speed and the optimization capability of the algorithm are improved, and it is ensured that the data-driven modeling method can achieve good performance under the condition that the large power change in the load range is fast.

Description

Technical Field

[0001] The invention relates to the technical field of coordinated control systems for ultra-supercritical units, and in particular to an improved modeling method for coordinated control systems for ultra-supercritical units. Background Art

[0002] Renewable energy is a new type of power generation with low pollution and sustainable development. However, the stability of the power grid is threatened by the regulation of new energy sources such as wind power and solar energy. The safety and stability of thermal power units are the basis for maintaining the reliability of the power grid. Thermal power units can smooth the power grid fluctuations caused by the randomness and intermittency of renewable energy by participating in the deep frequency regulation and peak regulation of the power grid. Therefore, it is very important to achieve stable and rapid regulation of thermal power units within a large load range. Ultra-supercritical units are widely used due to their high cycle thermal efficiency and low pollutant emissions. Ultra-supercritical units track load instructions externally and maintain the energy balance and parameter stability of the boiler-turbine system internally. Therefore, a coordinated control system (hereinafter referred to as CCS) is used to adjust the dynamic characteristics difference between the boiler and the turbine. A straight-through boiler is used to adapt to high working fluid parameters. However, due to the fact that steam-water mixing hinders fast and stable dynamic response, the unit exhibits multivariable coupling, severe nonlinearity and high complexity.

[0003] Therefore, it is necessary to conduct dynamic modeling research on thermal power units within a large load range in order to design control strategies. Summary of the invention

[0004] In view of the deficiencies of the prior art, the present invention provides an improved modeling method for a coordinated control system of an ultra-supercritical unit;

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

[0006] An improved ultra-supercritical unit coordinated control system modeling method comprises the following steps:

[0007] Step 1: Build an improved neural network based on transformer convolution operation and ResNet50 self-attention mechanism;

[0008] Step 1.1: Preprocess the ultra-supercritical unit operation data to obtain a neural network training data set;

[0009] The preprocessing includes determining the data length and encoding the position, and specifically includes the following steps:

[0010] Step 1.1.1: Aiming at the model identification problem of coordinated control system CCS, the sampling data embedding method is improved and the sliding window is used to segment the data;

[0011] The sliding window size is denoted by S lIndicates that the data input into the neural network is as follows:

[0012] data(n),data(n-1),…,data(nS l +1),n=S l ,S l +1,…,N

[0013] in, n For the n sampling points, N is the total number of samples, and the specific form of a single data consists of the input at the current moment and the output at the last moment, as shown in the following formula:

[0014]

[0015] Among them, X(·) is a column vector, which represents the input at the current moment, and y(·) represents the output at the last moment;

[0016] Step 1.1.2: Encode the location;

[0017] The position is encoded using sine and cosine functions of different frequencies, as shown in the following formula:

[0018]

[0019] Among them, p(m,i) is the encoded data, m is a parameter, representing from 0 to S l -1, i represents an odd or even number, α is the correction factor, d l Indicates the new dimension to which the data is promoted.

[0020] Step 1.2: Design an encoder network with a layered stacked structure;

[0021] The encoder network includes a multi-head attention layer, a feedforward neural network layer, and a convolutional layer, which are used to mine long-term relationships between input data, improve training efficiency and accuracy, and enhance the network's perception domain;

[0022] Step 1.2.1: Design a multi-head attention layer;

[0023] The attention value of each head is calculated by scaling the dot product attention, i.e., dot multiplication operation, as shown in the following formula:

[0024]

[0025] Q=X f W Q ,K=X f W K ,V=X f W V

[0026] Among them, X f , Q, K and V represent the input after dimensionality increase, g(·) normalizes the feature vector so that the sum of the vector elements is 1, d is the dimension of Q, K, V to keep the gradient stable, and W Q , W K and W V are the weight matrices of Q, K, and V, respectively, and their dependence on history is determined by the function g(·);

[0027] According to the number of heads in the multi-head attention layer, the input information is divided into different information subspaces, focusing on the features between the inputs from different angles, as shown in the following formula:

[0028] p(Q,K,V)=c(h1,h2,…,h m )·W

[0029] h i =f(X f W i Q ,X f W i K ,X f W i V ),i=1,2,…,m

[0030] Among them, h i is the output of a single attention, c(·) concatenates the output of each head, W is the multi-head attention weight matrix, W i Represents the value of the i-th attention weight matrix; the output of the multi-head attention layer will be sent in parallel to the feed-forward neural network layer FFN for synchronous training;

[0031] Step 1.2.2: Design the feed-forward neural network layer FFN;

[0032] The feed-forward neural network layer FFN consists of two fully connected layers and one activation layer; the exponential linear unit ELU activation function is used for nonlinear transformation;

[0033] Step 1.2.3: Design the convolutional layer;

[0034] The convolution layer includes one-dimensional convolution, activation function ELU and maximum pooling operation; the encoder network is stacked in layers, each encoder layer contains only one one-dimensional convolution operation, and each layer uses residual connection operation. The data after position encoding and dimension increase is processed in the network to obtain the long-term dependency between the inputs, forming the encoder network, and finally the output data is transmitted to the decoder network;

[0035] Step 1.3: Design and improve the residual network;

[0036] Specifically, an improved residual network combining ReNet50 and multi-head attention mechanism is proposed, that is, the multi-head attention mechanism is embedded in the residual network; the main part of the improved residual network consists of four convolutional layers, and the multi-head attention layer is used to replace the original three-dimensional convolution operation; ResNet50 and the attention layer are used for feature extraction and similarity calculation to obtain local dependencies;

[0037] Step 1.4: Design the decoder network structure;

[0038] The decoding network is used for information integration; the long-term dependency and local feature mapping in the input are provided by the encoder network and the improved residual network respectively. The historical trend relationship between these features and the output is sent to the attention layer to capture the dependency between input and output. The following error function is used to represent the training results and actual data:

[0039]

[0040] Among them, B s is the number of training data, is the model output value, and y is the actual output value;

[0041] The dynamic adjustment of network parameters is based on the first-order moment and second-order moment formula of the gradient, as shown in the following formula:

[0042] m t =β1·m t-1 +(1-β1)·g t

[0043]

[0044] In the formula, g t , m t and n t are the gradient, the first moment of the gradient, and the second moment of the gradient, respectively. β1 and β2 are the attenuation coefficients, which are 0.9 and 0.99, respectively;

[0045] Step 1.5: Use the Lyapunov function to improve the learning rate of the neural network to ensure that the output value of the neural network is close to the true output value;

[0046] Step 1.5.1: Define the Lyapunov function;

[0047] Using the output definition of the neural network That is, the error of the neural network, the Lyapunov function is defined as follows:

[0048]

[0049] The Lyapunov function is positive definite if it satisfies the following conditions:

[0050]

[0051] Step 1.5.2: Transform the Lyapunov function for the discrete system as shown below:

[0052] ΔV(t+1)=V(t+1)-V(t)

[0053] =e 2 (t+1)-e 2 (t)

[0054] =[e(t+1)+e(t)][e(t+1)-e(t)]

[0055] =Δe(t+1)·[2e(t)+Δe(t+1)]

[0056] In the formula, Δe(t) represents the change of error at time t;

[0057] Step 1.5.3: Determine the conditions that the learning rate satisfies;

[0058] The learning rate η satisfies the following conditions:

[0059]

[0060] Among them, α3 is a positive constant coefficient, H=[W i Q ,W i K ,W i V ,…,W];

[0061] Step 2: Propose a new HHBO algorithm based on Harris Hawk Optimization and Tent Chaotic Mapping to obtain the optimal solution to the neural network hyperparameter optimization problem;

[0062] Step 2.1: Individual location update;

[0063] The j-th update of individual c is as follows:

[0064]

[0065] in, and denote the position of the jth dimension of the previous generation and the random same generation, respectively. λ, γ, p1 and p2 are adaptive variables that change with the number of iterations. p is a random number uniformly distributed in [0,1]. f(·) represents the fitness function.

[0066] Step 2.2: Use Tent chaos mapping optimization algorithm to optimize population traversal capability;

[0067] The Tent chaotic mapping optimization algorithm diffuses the initial position of the population to the entire population space. When initializing the population position, the position of the first individual x(1) in the population is randomly generated. The normalization method is used to obtain the mapping value T(1) of x(1) under the Tent chaotic mapping. The Tent chaotic mapping is calculated by the following formula:

[0068]

[0069] Among them, c is the label of the individual in the population, representing the cth individual, T(·) is the mapping value, r1 is a random number uniformly distributed in [0,1]. The individuals are fully dispersed in the Tent chaotic space and then mapped to the population space for traversal ability improvement.

[0070] Step 2.3: Use the Harris Eagle algorithm optimization algorithm to correct the population iteration position;

[0071] The Harris Hawk optimization algorithm is used to optimize the hyperparameters. The optimization process includes the global exploration stage, the global exploration to local development stage, and the local development stage. The Harris Hawk position is used as the candidate solution, and the prey position is used as the best candidate solution. The specific process is as follows:

[0072] In the global exploration phase, Harris's hawk searches for prey with equal probability through tracking and detection strategies. When p < 0.5, the hawk moves according to other members and prey. When p ≥ 0.5, the hawk stays on a tree. The process is expressed by the following formula:

[0073]

[0074] Among them, U rand t is a randomly selected Harris hawk individual from the tth generation population, U prey represents the prey position, U m represents the average position of all eagles at present, u b , l b The population search is the upper and lower limits, and rand is a random number in [0,1];

[0075] In the transition phase from global exploration to local exploitation, it is divided into exploration and exploitation according to hunting habits. The energy of the prey gradually decreases during the escape process. The escape energy of the prey is:

[0076]

[0077] Among them, E0 is the initial escape energy of the prey, which is a random number between -1 and 1, t is the current evolutionary generation, and M is the maximum evolutionary generation of the population, when it enters the development stage;

[0078] During the development phase, the Harris Hawk will launch attacks, including soft encirclement, hard encirclement, progressive soft encirclement, and progressive hard encirclement. p is the probability of prey escaping, which ranges from 0 to 1. p <0.5 means there is a chance of escape, combining the prey escape energy |E| and the prey escape probability S p Determine the hunting strategy; when 0.5≤|E|<1 and S p ≥0.5, soft encirclement is performed; when |E|<0.5 and S p ≥0.5, perform hard encirclement; when 0.5≤|E|<1 and S p <0.5, progressive soft bracketing is performed. When |E|<0.5 and S p When the value is less than 0.5, progressive hard encirclement is performed; the Harris Hawk position update formula is as follows:

[0079]

[0080] Where F· is the fitness function, Le'vy is the flight strategy, S is a random vector of 0,1, and D is the dimension of the random vector; J is a random number in the interval (0,2), and Y and Z are intermediate variables, indicating the improved position information;

[0081] Step 3: Select model evaluation indicators;

[0082] The improved neural network improves the accuracy of the model through training and learning. The trained improved neural network is used to track the three subsystems of the CCS system, namely the electric power output system, the main steam pressure system and the temperature subsystem. The improved neural network is evaluated using an evaluation mechanism.

[0083] The evaluation mechanism is specifically as follows: the mean square error (MSE) is used to evaluate the model. The smaller the error, the better the performance of the improved neural network:

[0084]

[0085] Where N is the number of samples, is the model output value, is the actual value after normalization.

[0086] The beneficial effects of adopting the above technical solution are:

[0087] The present invention provides an improved modeling method for a coordinated control system of an ultra-supercritical unit. The present invention integrates deep learning and optimization algorithms. Based on historical operation data, a data-driven transformer neural network is developed in combination with ResNet50 to achieve accurate modeling of a large load range and rapid power changes. A new type of neural network model is established by using the structure of transformer and self-attention residual network to achieve long-term input dependency and local feature map extraction to obtain more accurate recognition results. The HHBO algorithm is committed to improving the convergence speed to obtain the optimal values ​​of certain parameters. The network and optimization algorithm proposed by the present invention can reduce the mean square error of the electric power output system, the main steam pressure system and the temperature subsystem to 1.34E-4, 1.54E-4 and 1.71E-4 respectively. The established model has ideal recognition accuracy under the conditions of 40% to 100% wide load range and rapid power change, which is of great significance for studying the modeling of CCS under low load and full working conditions. BRIEF DESCRIPTION OF THE DRAWINGS

[0088] Figure 1 The improved neural network structure diagram based on transformer convolution operation and ResNet50 self-attention mechanism of the present invention is shown;

[0089] Figure 2 The encoder network structure diagram of the present invention is shown;

[0090] Figure 3 The curve of the actual value of the electric power output and the output value in the implementation example 1 is shown;

[0091] Figure 4 The curve of the actual value and output value of the main steam pressure in the implementation case 1 is shown;

[0092] Figure 5 The curve of the actual value and output value of the separator outlet temperature system in Example 1 is shown; DETAILED DESCRIPTION

[0093] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0094] Embodiment 1:

[0095] In this embodiment, a continuous data set of a 1000MW ultra-supercritical unit under a wide range of operating conditions is used for training and experiments. The load variation range is from 650MW to 1050MW. The data set includes six features: total fuel flow, turbine opening degree, feed water flow, main steam pressure, output voltage and separator outlet temperature. Data is collected every 10 seconds, and a total of 15,000 sample data are collected, 12,000 sample data are used for neural network training, and 3,000 sample data are used for testing. The first three features are used as input and the last three features are used as output to train the neural network. The multi-input multi-output system can be divided into three multi-input single-output systems for modeling to obtain more accurate recognition results. Figure 1 This is the improved neural network structure diagram based on transformer convolution operation and ResNet50 self-attention mechanism of the present invention.

[0096] The data used in the present invention are from the operating data of a 1000MW ultra-supercritical unit. The specific implementation methods of the present invention are as follows:

[0097] An improved ultra-supercritical unit coordinated control system modeling method comprises the following steps:

[0098] Step 1: Build an improved neural network based on transformer convolution operation and ResNet50 self-attention mechanism;

[0099] Step 1.1: Preprocess the ultra-supercritical unit operation data to obtain a neural network training data set;

[0100] The preprocessing includes determining the data length and encoding the position, and specifically includes the following steps:

[0101] Step 1.1.1: Aiming at the model identification problem of coordinated control system CCS, the sampling data embedding method is improved and the sliding window is used to segment the data;

[0102] In order to effectively capture and process local features and patterns in the data, the historical data of ultra-supercritical units are processed in segments to improve the efficiency of neural network training and reduce the computational complexity. The sliding window size is denoted by S. l Indicates that the data input into the neural network is as follows:

[0103] data(n),data(n-1),…,data(nS l +1),n=S l ,S l +1,…,N

[0104] Among them, n is the nth sampling point, N is the total number of samples, and the specific form of a single data consists of the input at the current moment and the output at the last moment, as shown in the following formula:

[0105]

[0106] Among them, X(·) is a column vector, which represents the input at the current moment, and y(·) represents the output at the last moment;

[0107] Step 1.1.2: Encode the location;

[0108] In order to characterize the sequential relationship of the data, sine and cosine functions of different frequencies are used to encode the position, as shown in the following formula:

[0109]

[0110] Among them, p(m,i) is the encoded data, m is a parameter, representing from 0 to S l -1, i represents an odd or even number, α is the correction factor, which is equal to 0.01 in this network. The preprocessed samples are roughly between -1 and 1, d l Represents the new dimension to which the data is lifted to accommodate the attention mechanism.

[0111] Step 1.2: Since the straight-through boiler structure weakens the heat storage capacity of the unit and makes the CCS more sensitive to changes in input variables, an encoder network with a layered stacking structure is designed according to these dynamic characteristics of the unit, such as Figure 2 As shown;

[0112] The encoder network includes a multi-head attention layer, a feedforward neural network layer, and a convolutional layer, which are used to mine long-term relationships between input data, improve training efficiency and accuracy, and enhance the network's perception domain;

[0113] Step 1.2.1: Design a multi-head attention layer;

[0114] The attention mechanism is used to explore the relevant features between inputs. For CCS modeling, the long-term dependencies between inputs will be reflected through the multi-head attention layer. The attention value of each head is calculated by scaling the dot product attention, that is, the dot multiplication operation, as shown in the following formula:

[0115]

[0116] Q=X f W Q ,K=X f W K ,V=X f W V

[0117] Among them, X f , Q, K and V represent the input after dimensionality increase, g(·) normalizes the feature vector so that the sum of the vector elements is 1, d is the dimension of Q, K, V to keep the gradient stable, and W Q , W K and W V The weight matrices are Q, K, and V respectively, and the dependence on history is determined by the function g(·); for the CCS modeling problem, the query matrix Q of the encoder's initial attention layer consists only of the input data at the current moment. This improves the feature capture capability of the current input data.

[0118] According to the number of heads in the multi-head attention layer, the input information is divided into different information subspaces, focusing on the features between the inputs from different angles, as shown in the following formula:

[0119] p(Q,K,V)=c(h1,h2,…,h m )·W

[0120] h i =f(X f W i Q ,X f W i K ,X f W i V ),i=1,2,…,m

[0121] Among them, h i is the output of a single attention, c(·) concatenates the output of each head, W is the multi-head attention weight matrix, W i Represents the value of the i-th attention weight matrix;

[0122] Step 1.2.2: Design the feed-forward neural network layer FFN;

[0123] The feedforward neural network layer consists of two fully connected layers and an activation layer; feature combination is implemented in the fully connected layer to coordinate the nonlinear transformation of the activation layer. The exponential linear unit ELU activation function is used for nonlinear transformation. The ELU activation function is continuous and differentiable at each point to obtain faster training time and higher accuracy. The output of the multi-head attention layer will be sent in parallel to the feedforward neural network layer FFN for synchronous training to improve computational efficiency.

[0124] Step 1.2.3: Design the convolutional layer;

[0125] The straight-through boiler structure weakens the heat storage capacity of the unit, making the CCS more sensitive to changes in input variables. In view of these dynamic characteristics of the unit, a convolution layer is added for perceptual domain enhancement. The convolution layer includes one-dimensional convolution, activation function ELU and maximum pooling operation; the encoder network is stacked in layers, and each encoder layer contains only one one-dimensional convolution operation to increase the convergence speed. In order to prevent network degradation, each layer uses residual connection operations. The data after position encoding and dimensionality increase is processed in the network to obtain long-term dependencies between inputs, forming an encoder network, and finally the output data is transmitted to the decoder network; through these operations, the long-term dependencies of the CCS system operation data can be obtained, so that the trained neural network can better track changes in the system;

[0126] Step 1.3: Design and improve the residual network;

[0127] Thermal power units participate in peak load regulation to improve the renewable energy absorption capacity, and need to operate under rapid load changes. Therefore, the working fluid parameters frequently switch between subcritical and supercritical states, causing the dynamic characteristics of the unit to change at any time. This characteristic is further revealed in the input and output data. On this basis, an improved residual network combining ReNet50 and a multi-head attention mechanism is specifically proposed. Since the integrity of the information is protected, the learning objective is simplified. However, the rapid load changes make the model features more abstract and complex. Therefore, a multi-head attention mechanism is embedded in the residual network; the main part of the improved residual network consists of four convolutional layers, and the multi-head attention layer is used to replace the original three-dimensional convolution operation; ResNet50 and the attention layer are used for feature extraction and similarity calculation to obtain local dependencies;

[0128] Ultra-supercritical units need to face rapid changes in load when participating in the process of new energy consumption. Therefore, the operating characteristics of the units will change at any time, and it is necessary to further extract the features of the input data. In this embodiment, ReNet50 and multi-head attention mechanism are used to improve the residual network. ReNet50 ensures the integrity of the data and simplifies the learning objectives. At the same time, in order to cope with the impact of rapid load changes on the model, a multi-head attention mechanism is used to further extract features. By improving the residual network, the local dependency of the data is obtained.

[0129] Step 1.4: Design the decoder network structure;

[0130] The decoding network is used for information integration; the long-term dependency and local feature mapping in the input are provided by the encoder network and the improved residual network respectively. The historical trend relationship between these features and the output is sent to the attention layer to capture the dependency between input and output. The following error function is used to represent the training results and actual data:

[0131]

[0132] Among them, B s is the number of training data, is the model output value, and y is the actual output value;

[0133] The dynamic adjustment of network parameters is based on the first-order moment and second-order moment formula of the gradient, as shown in the following formula:

[0134] m t =β1·m t-1 +(1-β1)·g t

[0135]

[0136] In the formula, g t , m t and n t are the gradient, the first moment of the gradient, and the second moment of the gradient, respectively. β1 and β2 are the attenuation coefficients, which are 0.9 and 0.99, respectively;

[0137] Step 1.5: Use the Lyapunov function to improve the learning rate of the neural network to ensure that the output value of the neural network is close to the true output value;

[0138] Step 1.5.1: Define the Lyapunov function;

[0139] Using the output definition of the neural network That is, the error of the neural network, the Lyapunov function is defined as follows:

[0140]

[0141] The Lyapunov function is positive definite. In order to ensure the convergence of the neural network during training, the derivative of the function needs to be less than or equal to zero. By limiting the learning rate during neural network training, the derivative of the function satisfies the condition of being less than or equal to zero, and the following conditions are met:

[0142]

[0143] Step 1.5.2: Transform the Lyapunov function for discrete systems;

[0144] The derivative of the Lyapunov function is semi-negative. For discrete systems, a transformation is required, as shown in the following formula:

[0145] ΔV(t+1)=V(t+1)-V(t)

[0146] =e 2 (t+1)-e 2 (t)

[0147] =[e(t+1)+e(t)][e(t+1)-e(t)]

[0148] =Δe(t+1)·[2e(t)+Δe(t+1)]

[0149] In the formula, Δe(t) represents the change of error at time t;

[0150] Step 1.5.3: Determine the conditions that the learning rate satisfies;

[0151] Since CCS is a nonlinear system, for a stable nonlinear system, when the proposed network is used to model the system, the learning rate η satisfies the following conditions:

[0152]

[0153] Among them, α3 is a positive constant coefficient, H=[W i Q ,W i K ,W i V ,…,W], with the increase of training samples, the error e(t) keeps decreasing. Through further solution, it can be verified that ΔV(t+1)≤0, that is, it is verified that the neural network is convergent during the training process.

[0154] Through the above steps, an improved neural network based on transformer convolution operation and ResNet50 self-attention mechanism is constructed. The encoder structure and the improved residual network are combined to improve the neural network's long-term dependencies and local feature mapping. The Lyapunov function is used to improve the learning rate of the neural network to improve the training efficiency of the neural network.

[0155] Step 2: Propose a new HHBO algorithm based on Harris Hawk Optimization and Tent Chaotic Mapping to obtain the optimal solution to the neural network hyperparameter optimization problem;

[0156] The Tent chaotic mapping optimization algorithm and the Harris Eagle optimization algorithm are used to improve the traversal ability of the population and correct the iterative position of the population respectively to avoid the algorithm falling into the local optimum.

[0157] Step 2.1: Individual location update;

[0158] The j-th update of individual c is as follows:

[0159]

[0160] in, and denote the position of the jth dimension of the previous generation and the random same generation, respectively. λ, γ, p1 and p2 are adaptive variables that change with the number of iterations. p is a random number uniformly distributed in [0,1]. f(·) represents the fitness function.

[0161] Since some individuals keep their original positions in a single iteration, the convergence speed is slow. In addition, the randomly generated initialization population leads to poor traversal ability and even stagnation in the local optimum. Therefore, Harris Eagle optimization and Tent chaos mapping are used to improve the algorithm.

[0162] Step 2.2: Use Tent chaos mapping optimization algorithm to optimize population traversal capability;

[0163] The Tent chaotic mapping optimization algorithm diffuses the initial position of the population to the entire population space. When initializing the population position, the position of the first individual x(1) in the population is randomly generated. The normalization method is used to obtain the mapping value T(1) of x(1) under the Tent chaotic mapping. The Tent chaotic mapping is calculated by the following formula:

[0164]

[0165] Among them, c is the label of the individual in the population, representing the cth individual, T(·) is the mapping value, r1 is a random number uniformly distributed in [0,1]. The individuals are fully dispersed in the Tent chaotic space and then mapped to the population space for traversal ability improvement.

[0166] Step 2.3: Use the Harris Eagle algorithm optimization algorithm to correct the population iteration position;

[0167] In order to avoid falling into the local optimum during the hyperparameter optimization process, the Harris Hawk optimization algorithm is used to optimize the hyperparameters. The optimization process includes the global exploration stage, the global exploration to local development stage, and the local development stage. The Harris Hawk position is used as the candidate solution, and the prey position is used as the best candidate solution. The specific process is as follows:

[0168] In the global exploration phase, Harris's hawk searches for prey with equal probability through tracking and detection strategies. When p < 0.5, the hawk moves according to other members and prey. When p ≥ 0.5, the hawk stays on a tree. The process is expressed by the following formula:

[0169]

[0170] Among them, U rand t is a randomly selected Harris hawk individual from the tth generation population, U prey represents the prey position, U m represents the average position of all eagles at present, u b , l bThe population search is the upper and lower limits, and rand is a random number in [0,1];

[0171] In the transition phase from global exploration to local exploitation, it is divided into exploration and exploitation according to hunting habits. The energy of the prey gradually decreases during the escape process. The escape energy of the prey is:

[0172]

[0173] Among them, E0 is the initial escape energy of the prey, which is a random number between -1 and 1, t is the current evolutionary generation, and M is the maximum evolutionary generation of the population, when it enters the development stage;

[0174] During the development phase, the Harris Hawk will launch attacks, including soft encirclement, hard encirclement, progressive soft encirclement, and progressive hard encirclement. p is the probability of prey escaping, which ranges from 0 to 1. p <0.5 means there is a chance of escape, combining the prey escape energy |E| and the prey escape probability S p Determine the hunting strategy; when 0.5≤|E|<1 and S p ≥0.5, soft encirclement is performed; when |E|<0.5 and S p ≥0.5, perform hard encirclement; when 0.5≤|E|<1 and S p <0.5, progressive soft bracketing is performed. When |E|<0.5 and S p When the value is less than 0.5, progressive hard encirclement is performed; the Harris Hawk position update formula is as follows:

[0175]

[0176] Where F· is the fitness function, Le'vy is the flight strategy, S is a random vector of 0,1, and D is the dimension of the random vector; J is a random number in the interval (0,2), and Y and Z are intermediate variables, indicating the improved position information;

[0177] Step 3: Select appropriate model evaluation metrics.

[0178] The improved neural network improves the accuracy of the model through continuous training and learning. The trained improved neural network is used to track the three subsystems of the CCS system: the electric power output system, the main steam pressure system and the temperature subsystem. The improved neural network is evaluated using an evaluation mechanism.

[0179] The evaluation mechanism is specifically as follows: the mean square error (MSE), a common indicator for evaluating the size of errors, is used to evaluate the model. The mean square error can intuitively measure the gap between the actual value and the output value.

[0180]

[0181] Where N is the number of samples, is the model output value, is the actual value after normalization.

[0182] In this embodiment, the initial position of the population is first diffused to the entire population space using the Tent chaotic mapping to improve the traversal ability of the population. Then, the group foraging modeling idea in Harris Hawk optimization is introduced to improve the position of the individual in a single iteration. The improved formula is as follows:

[0183]

[0184] Among them, r2, r3, r4, and r5 are random numbers uniformly distributed between [0, 1]. represents the optimal individual position of the population. The mean square error of tracking is used as the objective function of optimization, as shown in the following formula:

[0185]

[0186] Among them, l b and u b They represent the lower and upper bounds of the domain respectively. The parameter settings of the optimization algorithm are shown in Table 1. l is the size of the sliding window, B s Batch size for neural network training.

[0187] Table 1 Parameter settings of optimization algorithm

[0188]

[0189] Firstly, the improved transformer is trained using the operating data of the ultra-supercritical unit, and data-based methods such as long short-term memory network (LSTM), ResNet 50, convolutional neural network (CNN) and support vector machine (SVM) are selected for comparison. The training parameter settings of each model are shown in Table 2:

[0190] Table 2 Training parameter settings for each model

[0191]

[0192]

[0193] P E , P Tand T represent the output of the neural network, rbf is the Gaussian kernel function, the trained neural network is used to track the CCS system, 3000 pairs of actual operation data are selected for pattern recognition, and the superiority of the modeling method proposed in the present invention is verified. The errors of tracking control using different methods are shown in Table 3. The tracking curves of the electric power output, main steam pressure and separator outlet temperature of the method proposed in the present invention are shown in Table 3. Figure 3 , Figure 4 , Figure 5 shown.

[0194] Table 3 Tracking control errors of different methods

[0195]

[0196] From the above results, it can be seen that the general method can also track a certain output well, but it is difficult to track three outputs simultaneously with high precision. The method proposed in the present invention can take into account the tracking precision of the three outputs at the same time.

[0197] In summary, it is shown that a data-driven modeling method for coordinated control systems of ultra-supercritical units based on an improved transformer can achieve good performance under conditions of large load range and rapid power changes.

[0198] Embodiment 2

[0199] In this implementation case, data is collected from a 1000MW ultra-supercritical unit in southeast China under operating conditions with a load variation range of 400 to 1000MW. Data is collected every 10 seconds, and a total of 5500 sample data are collected. These sample data cover the operation of the device under most conditions. 4400 sample data are used for training and 1100 sample data are used for testing. The data used in this example comes from a 1000MW ultra-supercritical unit in the east;

[0200] In this embodiment, since the ultra-supercritical unit has to cope with the fluctuations caused by the new energy consumption process during operation, the operating characteristics of the unit are constantly changing. In order to better cope with this change, it is necessary to further extract features of the input data and improve the residual network using the multi-head attention mechanism. While improving the ability to cope with changes in load at any time, it obtains local features of the data and simplifies the learning objectives. Table 4 shows the parameter settings of the improved residual network.

[0201] Table 4 Parameter settings of improved residual network

[0202]

[0203] The decoder is designed to integrate the data feature relationships mined from the residual network and the encoder, and uses the multi-head attention mechanism to capture the relationship between the input and output of the neural network. The mean square error is used to evaluate the accuracy of the mapping relationship. The first-order and second-order moments of the gradient are used to adjust the network parameters during training.

[0204] After designing the neural network, it is necessary to analyze the convergence of the neural network and use the training error of the neural network to construct the Lyapunov function. Since the Lyapunov function is positive definite, the derivative of the function needs to be less than or equal to 0. The learning rate range under the condition of neural network convergence can be obtained, as shown in the following formula:

[0205]

[0206] The hyperparameters of the neural network directly affect the performance of the neural network. Therefore, the Harris Eagle optimization algorithm is used to optimize the hyperparameters of the network. In order to improve the traversal ability of the population and prevent the optimization algorithm from falling into the local optimum, the chaotic mapping is used to diffuse the initial position of the population to the entire population space.

[0207] Using the processed operating data of the ultra-supercritical unit, the transformer is trained and the electric power output, main steam pressure and separator outlet temperature are tracked. The LSTM, ResNet50, CNN and SVM methods are also used for comparison. Table 5 shows the tracking errors of the electric power output, main steam pressure and separator outlet temperature by different methods.

[0208] Table 5 Tracking control errors of different methods

[0209]

[0210] It can be seen from the above results that the existing method cannot guarantee the accuracy of the three outputs at the same time, and the method described in the present invention can effectively track the changes of the three outputs.

[0211] In summary, it is shown that a data-driven modeling method for coordinated control systems of ultra-supercritical units based on an improved transformer can show good performance on actual ultra-supercritical unit examples.

[0212] Embodiment 3

[0213] In this implementation case, a 1000MW ultra-supercritical unit of a thermal power plant in Liaoning is controlled, and data is collected under the operating conditions of load variation ranging from 500MW to 1050MW. The collection time interval is 10 seconds, and a total of 10,000 sample data are collected. 8,000 data samples are used to train the neural network, and 2,000 data samples are used for testing. The data used in this example comes from a 1000MW ultra-supercritical unit of a thermal power plant in Liaoning;

[0214] The data collected from an ultra-supercritical unit in a power plant in Liaoning was processed to train an improved transformer, and the trained network was used to track and control the output power, main steam pressure and separator outlet temperature. LSTM, ResNet50, CNN and SVM methods were used for comparison. Table 6 shows the tracking and control errors of different methods.

[0215] Table 5 Tracking control errors of different methods

[0216]

[0217] From the above results, it can be seen that compared with the existing methods, the data-driven modeling method of the coordinated control system of ultra-supercritical units based on the improved transformer can effectively track the three outputs.

[0218] In summary, the data-driven modeling method of ultra-supercritical unit coordinated control system based on the improved transformer has good performance in CCS modeling.

[0219] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. An improved modeling method for coordinated control system of ultra-supercritical units, characterized in that: The following steps are involved: Step 1: Build an improved neural network based on transformer convolution operation and ResNet50 self-attention mechanism; Step 2: Propose a new HHBO algorithm based on Harris Hawk Optimization and Tent Chaotic Mapping to obtain the optimal solution to the neural network hyperparameter optimization problem; Step 3: Select model evaluation indicators; The improved neural network improves the accuracy of the model through training and learning. The trained improved neural network is used to track the three subsystems of the CCS system: the electric power output system, the main steam pressure system and the temperature subsystem. The improved neural network is evaluated using an evaluation mechanism.

2. The improved ultra-supercritical unit coordinated control system modeling method according to claim 1, characterized in that: The step 1 comprises the following steps: Step 1.1: Preprocess the ultra-supercritical unit operation data to obtain a neural network training data set; Step 1.2: Design an encoder network with a layered stacked structure; The encoder network includes a multi-head attention layer, a feedforward neural network layer, and a convolutional layer, which are used to mine long-term relationships between input data, improve training efficiency and accuracy, and enhance the network's perception domain; Step 1.3: Design and improve the residual network; Specifically, an improved residual network combining ReNet50 and multi-head attention mechanism is proposed, that is, the multi-head attention mechanism is embedded in the residual network; the main part of the improved residual network consists of four convolutional layers, and the multi-head attention layer is used to replace the original three-dimensional convolution operation; ResNet50 and the attention layer are used for feature extraction and similarity calculation to obtain local dependencies; Step 1.4: Design the decoder network structure; The decoding network is used for information integration; the long-term dependency and local feature mapping in the input are provided by the encoder network and the improved residual network respectively. The historical trend relationship between these features and the output is sent to the attention layer to capture the dependency between input and output. The following error function is used to represent the training results and actual data: Among them, B s is the number of training data, is the model output value, and y is the actual output value; The dynamic adjustment of network parameters is based on the first-order moment and second-order moment formula of the gradient, as shown in the following formula: m t =β1·m t-1 +(1-β1)]g t In the formula, g t , m t and n t are the gradient, the first moment of the gradient, and the second moment of the gradient, respectively. β1 and β2 are the attenuation coefficients, which are 0.9 and 0.99, respectively; Step 1.5: Use the Lyapunov function to improve the learning rate of the neural network to ensure that the output value of the neural network is close to the true output value.

3. The improved ultra-supercritical unit coordinated control system modeling method according to claim 2, characterized in that: The preprocessing described in 1.1 includes determining the data length and encoding the position, and specifically includes the following steps: Step 1.1.1: Aiming at the model identification problem of coordinated control system CCS, the sampling data embedding method is improved and the sliding window is used to segment the data; The sliding window size is denoted by S l Indicates that the data input into the neural network is as follows: data(n),data(n-1),…,data(n-S l +1),n=S l ,S l +1,…,N Among them, n is the nth sampling point, N is the total number of samples, and the specific form of a single data consists of the input at the current moment and the output at the last moment, as shown in the following formula: Among them, X(·) is a column vector, which represents the input at the current moment, and y(·) represents the output at the last moment; Step 1.1.2: Encode the location; The position is encoded using sine and cosine functions of different frequencies, as shown in the following formula: Among them, p(m,i) is the encoded data, m is a parameter, representing from 0 to S l -1, i represents an odd or even number, α is the correction factor, d l Indicates the new dimension to which the data is promoted.

4. The improved ultra-supercritical unit coordinated control system modeling method according to claim 2, characterized in that: The step 1.2 comprises the following steps: Step 1.2.1: Design a multi-head attention layer; The attention value of each head is calculated by scaling the dot product attention, i.e., dot multiplication operation, as shown in the following formula: Q=X f W Q ,K=X f W K ,V=X f W V Among them, X f , Q, K and V represent the input after dimensionality increase, g(·) normalizes the feature vector so that the sum of the vector elements is 1, d is the dimension of Q, K, V to keep the gradient stable, and W Q , W K and W V are the weight matrices of Q, K, and V, respectively, and their dependence on history is determined by the function g(·); According to the number of heads in the multi-head attention layer, the input information is divided into different information subspaces, focusing on the features between the inputs from different angles, as shown in the following formula: p(Q,K,V)=c(h1,h2,…,h m )·W h i =f(X f W i Q ,X f W i K ,X f W i V ),i=1,2,…,m Among them, h i is the output of a single attention, c(·) concatenates the output of each head, W is the multi-head attention weight matrix, W i Represents the value of the i-th attention weight matrix; the output of the multi-head attention layer will be sent in parallel to the feed-forward neural network layer FFN for synchronous training; Step 1.2.2: Design the feed-forward neural network layer FFN; The feed-forward neural network layer FFN consists of two fully connected layers and one activation layer; the exponential linear unit ELU activation function is used for nonlinear transformation; Step 1.2.3: Design the convolutional layer; The convolution layer includes one-dimensional convolution, activation function ELU and maximum pooling operation; the encoder network is stacked in layers, each encoder layer contains only one one-dimensional convolution operation, and each layer uses residual connection operation. The data after position encoding and dimensionality increase is processed in the network to obtain the long-term dependency between the inputs, forming an encoder network, and finally the output data is transmitted to the decoder network.

5. The improved ultra-supercritical unit coordinated control system modeling method according to claim 2, characterized in that: The step 1.5 comprises the following steps: Step 1.5.1: Define the Lyapunov function; Using the output definition of the neural network That is, the error of the neural network, the Lyapunov function is defined as follows: The Lyapunov function is positive definite if it satisfies the following conditions: Step 1.5.2: Transform the Lyapunov function for a discrete system as shown below: ΔV(t+1)=V(t+1)-V(t) =e 2 (t+1)-e 2 (t) =[e(t+1)+e(t)][e(t+1)-e(t)] =Δe(t+1)·[2e(t)+Δe(t+1)] In the formula, Δe(t) represents the change of error at time t; Step 1.5.3: Determine the conditions that the learning rate satisfies; The learning rate η satisfies the following conditions: Among them, α3 is a positive constant coefficient, H=[W i Q ,W i K ,W i V ,…,W].

6. The improved ultra-supercritical unit coordinated control system modeling method according to claim 1, characterized in that: The step 2 comprises the following steps: Step 2.1: Individual location update; The j-th update of individual c is as follows: in, and denote the position of the jth dimension of the previous generation and the random same generation, respectively. λ, γ, p1 and p2 are adaptive variables that change with the number of iterations. p is a random number uniformly distributed in [0,1]. f(·) represents the fitness function. Step 2.2: Use Tent chaos mapping optimization algorithm to optimize population traversal capability; The Tent chaotic mapping optimization algorithm diffuses the initial position of the population to the entire population space. When initializing the population position, the position of the first individual x(1) in the population is randomly generated. The normalization method is used to obtain the mapping value T(1) of x(1) under the Tent chaotic mapping. The Tent chaotic mapping is calculated by the following formula: Among them, c is the label of the individual in the population, representing the cth individual, T(·) is the mapping value, r1 is a random number uniformly distributed in [0,1], and the individual is fully dispersed in the Tent chaotic space and then mapped to the population space for traversal ability improvement; Step 2.3: Use the Harris Eagle algorithm optimization algorithm to correct the population iteration position; The Harris Hawk optimization algorithm is used to optimize the hyperparameters. The optimization process includes the global exploration stage, the global exploration to local development stage, and the local development stage. The Harris Hawk position is used as the candidate solution, and the prey position is used as the best candidate solution. The specific process is as follows: In the global exploration phase, Harris's hawks searched for prey with equal probability using two strategies: tracking and detection. . When p≥0.5, the eagle moves according to other members and prey. When p≥0.5, the eagle stays on a tree. The following formula expresses the process: Among them, U rand t is a randomly selected Harris hawk individual from the tth generation population, U prey represents the prey position, U m represents the average position of all eagles at present, u b , l b The population search is the upper and lower limits, and rand is a random number in [0,1]; In the transition phase from global exploration to local exploitation, it is divided into exploration and exploitation according to hunting habits. The energy of the prey gradually decreases during the escape process. The escape energy of the prey is: Among them, E0 is the initial escape energy of the prey, which is a random number between 1 and 1, t is the current evolutionary generation, and M is the maximum evolutionary generation of the population, when it enters the development stage; During the development phase, the Harris Hawk will launch attacks, including soft encirclement, hard encirclement, progressive soft encirclement, and progressive hard encirclement. p is the probability of prey escaping, which ranges from 0 to 1. p <0.5 means there is a chance of escape, combining the prey escape energy |E| and the prey escape probability S p Determine the hunting strategy; when 0.5≤|E|<1 and S p ≥0.5, soft encirclement is performed; when |E|<0.5 and S p ≥0.5, perform hard encirclement; when 0.5≤|E|<1 and S p <0.5, progressive soft bracketing is performed. When |E|<0.5 and S p When the value is less than 0.5, progressive hard encirclement is performed; the Harris Hawk position update formula is as follows: Where F is the fitness function, Le'vy is the flight strategy, S is a random vector of 0,1, and D is the dimension of the random vector; J is a random number in the interval (0, 2), and Y and Z are intermediate variables, representing the improved position information.

7. The improved ultra-supercritical unit coordinated control system modeling method according to claim 1, characterized in that: The evaluation mechanism described in step 3 is as follows: the mean square error (MSE) is used to evaluate the model. The smaller the error, the better the performance of the improved neural network: Where N is the number of samples, is the model output value, is the actual value after normalization.

Images