Real-time optimization method for power plant condenser vacuum based on data mining algorithm

Abstract

This invention relates to the field of data mining technology, specifically to a real-time optimization method for condenser vacuum in power plants based on data mining algorithms. The optimization method involves: acquiring time-series turbine load, circulating water temperature, circulating water flow rate, and condenser terminal differential pressure parameters from online monitoring instruments; constructing a neural network to fit the relationship between the time-series parameter data and the condenser vacuum; obtaining the importance ratio of feature changes for each parameter data; determining the initial number of particles for each feature based on the importance ratio; obtaining the vacuum prediction value for each particle at different circulating water volumes and calculating the initial velocity of the particle; constructing an objective function for the particle; and performing multi-dimensional iterative optimization based on the objective function to obtain the optimal circulating water flow rate for the current environment. This method finds the optimal circulating water flow rate within the range of data fluctuations, avoiding the waste of computational resources for condenser vacuum optimization due to frequent data changes.

Description

Technical Field

[0001] The invention relates to the field of data mining, and in particular to a real-time optimization method for a power plant condenser vacuum based on a data mining algorithm. Background Art

[0002] Condenser pressure is an important parameter for unit operation. Whether in the thermal design of the condenser or in the operation of the cold end equipment of the turbine, the condenser vacuum is required to have an optimal value. As the condenser vacuum increases, the unit output increases, but the power consumption of the circulating water pump and the vacuum pump also increases. When the difference between the turbine power and the circulating water pump and vacuum pump power reaches a maximum value, the condenser vacuum at this time is called the optimal vacuum. The condenser vacuum is closely related to the circulating water temperature, circulating water flow rate and vacuum pump working conditions. It can be seen that the condenser vacuum is the result of a complex multi-factor effect.

[0003] Condenser vacuum has always been an important focus of power plant management and consumption reduction work, because the condenser vacuum system is complex, prone to dynamic failures such as vacuum system leakage, condenser scaling, high heat load, and condenser performance is not easy to maintain. Therefore, real-time optimization of condenser vacuum is very important for condenser vacuum optimization operation and condenser vacuum system fault diagnosis. Summary of the invention

[0004] In order to solve the above technical problems, the purpose of the present invention is to provide a real-time optimization method for power plant condenser vacuum based on a data mining algorithm. The technical solution adopted is as follows:

[0005] An embodiment of the present invention provides a real-time optimization method for a power plant condenser vacuum based on a data mining algorithm, and the optimization method includes the following steps: based on an online monitoring instrument, obtaining time-series steam turbine load, circulating water temperature in a power plant, circulating water flow, and condenser terminal difference parameter data; constructing a multi-encoder neural network to fit the relationship between the time-series steam turbine load, circulating water temperature in a power plant, circulating water flow, condenser terminal difference parameter data and condenser vacuum; obtaining data volatility and feature importance of each parameter data; obtaining the change importance of each parameter data according to the data volatility and feature importance of the parameter data; obtaining the change importance ratio of each feature according to the change importance of the parameter data; obtaining the initial number of divided particles for each feature based on the feature change importance ratio; obtaining vacuum prediction values ​​of different circulating water volumes of each particle; obtaining the initialization speed of the particle based on the vacuum prediction values ​​of different circulating water volumes of each particle; constructing an objective function of the particle, and performing multi-dimensional space iterative optimization based on the objective function according to the initialization speed of each particle to obtain the optimal circulating water flow rate in the current environment.

[0006] Furthermore, the structure of the multi-encoder neural network is as follows: the neural network includes four encoders:

[0007] The input of the first encoder is the condenser terminal difference in the power plant, and the output is the terminal difference feature vector;

[0008] The input of the second encoder is the steam turbine load in the power plant, and the output is the load characteristic vector;

[0009] The input of the third encoder is the circulating water flow rate, and the output is the circulating water flow rate feature vector;

[0010] The input of the fourth encoder is the circulating water temperature, and the output is the circulating water temperature feature vector;

[0011] The first encoder, the second encoder, the third encoder, and the fourth encoder all use a fully connected network, and finally perform feature fusion on the terminal difference feature vector, the load feature vector, the circulating water flow feature vector, and the circulating water temperature feature vector, and the fusion operation uses the Add operation; the fused feature vector is obtained, and then the fused feature vector is input into the regression layer; the regression layer includes a fully connected layer and an activation function layer; the fully connected layer and the activation function layer are used to fit the relationship between the fused feature vector and the condenser vacuum, and finally output the condenser vacuum value; the activation function of the last activation function layer of the regression layer uses the Mish activation function.

[0012] Furthermore, the method for obtaining the data volatility of the parameter data is: obtaining the condenser end difference, circulating water temperature, and turbine load in the real-time operation process, and then calculating the variance of each type of data with a length of T, and the maximum variance is used as the data volatility. The larger the variance, the stronger the data volatility; for example, the condenser end difference is calculated once every 1 minute, and the variance of the condenser end difference data at the current moment and the continuous condenser end differences at the previous T-1 moments is calculated, and the maximum variance is used as the data volatility of the condenser end difference during this period of time.

[0013] Furthermore, the method for obtaining the feature importance of the parameter data is: using the BorutaShap method to obtain the importance score F of each input feature of the neural network; and then obtaining the importance of each feature:

[0014]

[0015] where F i is the importance score of the i-th feature, n is the number of features, and w i Indicates the importance of the i-th feature.

[0016] Furthermore, the calculation method of the importance of the parameter data change and the importance ratio of the feature change is:

[0017] Get the change importance U of each feature:

[0018] U=ew *σ

[0019] σ is the data volatility of the feature, w is the importance of the feature, and e is a natural constant, the purpose of which is to expand the value of the feature importance;

[0020] Feature change importance ratio:

[0021]

[0022] n is the number of features, U i represents the importance of the change of the i-th feature, P i is the change importance ratio of the i-th feature.

[0023] Furthermore, the number of initial partition particles K for each feature ′ The calculation method is:

[0024] K ′ =round(P i *K)

[0025] round is the rounding up function, P i is the change importance ratio of the i-th feature, and K is the maximum particle value in the single-dimensional space.

[0026] Furthermore, the calculation method of the initialization speed of the particle is as follows: the vacuum prediction value at each circulating water volume is converted into a matrix form, the matrix size is K′1*K′2*K′3, K′1, K′2, K′3 are the number of initial division examples of the three feature dimensions respectively, and each position in the matrix is ​​the current vacuum prediction value; then the 3D Sobel operator is used to obtain the vacuum prediction gradient of each position in each matrix, and then the average vacuum prediction gradient in the three directions of x, y, and z at each position is calculated, and the average vacuum prediction gradient in different directions at each position is used as the initialization speed of the particle.

[0027] Furthermore, the objective function of the particle is:

[0028]

[0029] min represents minimizing the objective function, is the maximum vacuum value, model(x1,x2,x3,x4) represents the vacuum prediction value of the neural network under the current input parameters, x1,x2,x3,x4 represent the condenser end difference, circulating water temperature, circulating water flow, and turbine load value respectively, τ represents the cost mapping coefficient of circulating water, D x3 , D′ x3 It represents the level difference between the searched circulating water flow and the current equipment circulating water flow, and θ represents the mapping coefficient of the circulating water flow level difference.

[0030] Furthermore, the particle speed and position update method is:

[0031] v i =ω*v i +c1*rand()*(pbest i -x i )+c2*rand()*(gbest i -x i )

[0032] x i =x i +v i

[0033] where v i 、x i Respectively represent the speed and position of the i-th particle; c1 and c2 are learning factors; rand() is a random number between 0 and 1; pbest i 、gbest i are the historical optimal position of the ith particle and the global optimal position of the group, respectively; ω represents the inertia factor sequence of the particle, which represents the inertia factor in different directions of the search space;

[0034] After the particle position is updated, the objective function value under all circulating water flow rates needs to be calculated.

[0035] Furthermore, the method for obtaining the inertia factor sequence of the particles is:

[0036] ω=[W1,W2,W3]

[0037] W1, W2, and W3 represent the inertia factors in the three directions of condenser end difference, circulating water temperature, and turbine load value respectively;

[0038]

[0039] w max is the maximum inertia weight, w min is the minimum inertia weight, w i represents the importance of the i-th feature, G k is the maximum number of iterations, and g is the current number of iterations.

[0040] The present invention has the following beneficial effects:

[0041] The embodiment of the present invention takes into account the volatility of data and finds an optimal circulating water flow value within the fluctuation range of data, thereby avoiding the waste of computing resources caused by optimizing the condenser vacuum due to frequent changes in data.

[0042] The multi-dimensional space weighted particle swarm used in the present invention effectively considers the importance of different feature dimensions, thereby being able to search for the optimal parameters more quickly. At the same time, the objective function is combined with the circulating water flow cost, which can maximize the reduction of the cost of real-time optimization of the condenser vacuum. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0044] Figure 1 A flow chart of a method for real-time optimization of a power plant condenser vacuum based on a data mining algorithm provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0045] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the real-time optimization method for power plant condenser vacuum based on data mining algorithm proposed by the present invention, its specific implementation method, structure, features and effects, in conjunction with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures, or characteristics in one or more embodiments may be combined in any suitable form.

[0046] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0047] The specific scheme of the real-time optimization method for the power plant condenser vacuum based on the data mining algorithm provided by the present invention is described in detail below with reference to the accompanying drawings.

[0048] See also Figure 1 , which shows a real-time optimization method for power plant condenser vacuum based on a data mining algorithm of the present invention, characterized in that the method comprises the following steps:

[0049] Step S001, based on the online monitoring instrument, obtaining the time series steam turbine load, circulating water temperature, circulating water flow rate, and condenser terminal difference parameter data in the power plant;

[0050] Condenser vacuum is closely related to circulating water temperature, circulating water flow, etc. Under certain conditions of unit load and cooling water temperature, unit condenser vacuum changes with the change of circulating water flow. Increasing the flow of circulating water pump increases condenser vacuum and unit output, but the power consumption of circulating water pump also increases at the same time. When the difference between the increment of turbine power and the increment of circulating water pump power caused by the increase of circulating water volume reaches the maximum value, the condenser vacuum at this time is called the optimal vacuum.

[0051] First, the power plant online monitoring system is used to obtain the turbine load, circulating water temperature, circulating water flow, and condenser end difference in the power plant. The power plant online monitoring system is a collection of sensor data in the power plant, which can obtain the parameters of each turbine in the power plant and the parameters to be monitored. The specific monitoring method is well known and will not be repeated here.

[0052] Step S002, constructing a multi-encoder neural network to fit the relationship between the time series steam turbine load, circulating water temperature in the power plant, circulating water flow, condenser terminal difference parameter data and condenser vacuum;

[0053] Further, based on the data mining algorithm, the mathematical relationship among the large amount of data is searched by the algorithm. The data mining algorithm preferably adopts the neural network technology, which has a strong fitting ability and performs better than most machine learning algorithms. A neural network is constructed to fit the relationship between the steam turbine load, the circulating water temperature in the power plant, the circulating water flow, the condenser terminal difference and the condenser vacuum. The circulating water temperature is the circulating water inlet temperature.

[0054] The neural network is a multi-encoder neural network:

[0055] The input of the first encoder is the condenser terminal difference in the power plant, and the output is the terminal difference feature vector.

[0056] The input of the second encoder is the steam turbine load in the power plant, and the output is the load characteristic vector.

[0057] The input of the third encoder is the circulating water flow rate, and the output is the circulating water flow rate feature vector.

[0058] The input of the fourth encoder is the circulating water temperature, and the output is the circulating water temperature feature vector.

[0059] The first encoder, the second encoder, the third encoder, and the fourth encoder all use a fully connected network, and finally perform feature fusion on the terminal difference feature vector, the load feature vector, the circulating water flow feature vector, and the circulating water temperature feature vector. The fusion operation is such as the Add operation, that is, addition. The fused feature vector is obtained, and then the fused feature vector is input into the regression layer. The regression layer includes a fully connected layer and an activation function layer. The fully connected layer and the activation function layer are used to fit the relationship between the fused feature vector and the condenser vacuum, and finally output the condenser vacuum value. The activation function of the last activation function layer of the regression layer adopts the Mish activation function.

[0060] The loss function is:

[0061] Loss = MSE(V,V′)

[0062] V, V′ are condenser vacuum and predicted vacuum values ​​respectively.

[0063] The optimization method of the neural network preferably uses the Adam optimizer with an adaptive learning rate.

[0064] Using three encoders can avoid the influence of different factors and make the network more robust.

[0065] At this point, the condenser vacuum can be predicted through the neural network.

[0066] Step S003, obtaining data volatility and feature importance of each parameter data; obtaining the change importance of each parameter data according to the data volatility and feature importance of the parameter data; obtaining the change importance ratio of each feature according to the change importance of the parameter data; obtaining the initial number of divided particles for each feature based on the feature change importance ratio;

[0067] In the network constructed in step S002, the condenser vacuum value is affected by the four input values ​​of the neural network. Therefore, in order to obtain the optimal vacuum, it is necessary to search for the optimal value in four-dimensional space. The present invention uses the following method to find the optimal circulating water flow rate:

[0068] The circulating water temperature is usually constant, mainly depending on the local climate conditions, and will not change in a short time. The condenser end difference is usually kept near a certain value, which can be used to characterize the degree of dirtiness of the copper tube on the cooling surface of the condenser, and is a certain value in most cases. The circulating water flow rate is a controllable variable that can control the vacuum of the condenser. The circulating water flow rate can be graded by changing the different combinations of water pumps. Implementers can make adjustments according to the conditions of the power plant equipment.

[0069] In the four-dimensional space optimization, the circulating water flow rate is a controllable variable, so the feature importance and data volatility are first obtained to find the most suitable circulating water flow rate.

[0070] For data volatility:

[0071] The condenser terminal difference, circulating water temperature, and turbine load in real-time operation are obtained, and then the variance of each data with a length of T is calculated. The variance is used as the volatility of the data. The larger the variance, the stronger the volatility of the data. For example, if the condenser terminal difference is calculated once every minute, the variance of the condenser terminal difference data at the current moment and the continuous condenser terminal difference data at the previous T-1 moments is calculated, and the maximum variance is used as the data volatility of the condenser terminal difference during this period.

[0072] For feature importance:

[0073] The present invention selects the BorutaShap method to obtain the importance of each feature. BorutaShap is a package feature selection method that combines the Boruta feature selection algorithm and the Shapley value. This combination is superior to the original permutation importance method in terms of speed and the quality of the generated feature subsets.

[0074] Finally, the BorutaShap method is used to obtain the importance score F of each input feature of the neural network.

[0075] Then get the importance of each feature:

[0076]

[0077] Among them, F i is the importance score of the i-th feature, n is the number of features, and w i Indicates the importance of the i-th feature.

[0078] At this point, the volatility of the condenser terminal difference, circulating water temperature, and turbine load data and the characteristic importance of the condenser terminal difference, circulating water temperature, and turbine load can be obtained. The characteristic importance needs to be calculated regularly because the characteristic importance calculation is affected by the value range of different characteristics in different periods.

[0079] Furthermore, the importance of changes in each feature is obtained:

[0080] U=e w *σ

[0081] σ is the data volatility of the feature, w is the importance of the feature, and e is a natural constant. The purpose is to expand the value of the feature importance. The greater the importance of the feature change, the greater the impact of the data change on the condenser vacuum change. Finally, the importance of the change of the condenser terminal difference, circulating water temperature, and turbine load is obtained.

[0082] Then, the multidimensional space weighted particle swarm algorithm is used to find the optimal control parameters, that is, the most suitable circulating water flow rate. The method is as follows:

[0083] First, obtain the variation range of each feature, which is:

[0084] [μ-σ,μ+σ]

[0085] μ and σ are the mean and maximum variance of the T sequence length when calculating data volatility.

[0086] When the difference of the variation range is 0, it indicates that the parameter is stable, so the search space will be smaller during the spatial optimization process. In special cases, when the difference of all feature variation ranges is 0, the enumeration method can be used to search for the optimal circulating water flow rate. When the enumeration method is used for optimization, the search can be flexibly performed based on the neural network. For example, when the difference of all feature variation ranges is 0, the input of the first encoder is the condenser terminal difference in the current power plant, and the output is the terminal difference feature vector.

[0087] The input of the second encoder is the turbine load in the current power plant, and the output is the load characteristic vector. The input of the fourth encoder is the current circulating water temperature, and the output is the circulating water temperature characteristic vector. Then the input of the third encoder is dynamically adjusted, and its input is different circulating water flow rates, and the output is multiple circulating water flow characteristic vectors. Then, based on the above four vectors, the vacuum value under different circulating water flow rates is obtained. Then the circulating water flow rate at the optimal vacuum is determined. That is, the first, second, and fourth encoders only need to be inferred once, and then the third encoder input can be dynamically adjusted to achieve optimization.

[0088] The present invention considers real-time optimization under the condition of continuous changes in condenser terminal difference, circulating water temperature, and turbine load. The reason for using the following method for optimization is that the circulating water flow corresponding to the optimal vacuum may be different due to the continuous changes in the above parameters. Adjusting the circulating water flow based only on the current condenser terminal difference, circulating water temperature, and turbine load will cause the optimal circulating water flow to change continuously. In actual operation, it is difficult to continuously adjust the circulating water flow in real time. That is, the circulating water volume of most power plants in my country cannot be continuously adjusted at present, and the circulating water volume can only be adjusted in stages by changing different combinations of water pumps. Therefore, all adjustable circulating water volumes are obtained and divided into levels, such as level 1 corresponding to a circulating water volume of 50L / h.

[0089] Then set a maximum particle value K in the single-dimensional space, and the empirical value of K is 10.

[0090] Then get the feature change importance ratio:

[0091]

[0092] Furthermore, the number of particles divided for each feature is obtained:

[0093] K′=round(P i *K)

[0094] That is, the greater the importance of the change of the feature, the more particles should be divided, so that the optimal circulating water value can be found more accurately and quickly. Round is the rounding function. Finally, K′1*K′2*K′3 particles are obtained, where K′1, K′2, and K′3 represent the number of particles divided by the condenser end difference, circulating water temperature, and turbine load, respectively.

[0095] Step S004, obtaining vacuum prediction values ​​of different circulating water volumes for each particle; obtaining the initialization speed of the particle based on the vacuum prediction values ​​of different circulating water volumes for each particle;

[0096] Then, the vacuum prediction value of different circulating water volumes is obtained for each particle point. Each particle point can be considered as a three-dimensional space point, and its coordinates are the condenser end difference, circulating water temperature, and turbine load value. Each particle point corresponds to the vacuum prediction value at different circulating water volumes.

[0097] The vacuum prediction value at each water circulation is converted into a matrix form, the size of which is K′1*K′2*K′3, and each position in the matrix is ​​the current vacuum prediction value. Then, the 3D Sobel operator is used to obtain the vacuum prediction gradient of each position in each matrix, and then the average vacuum prediction gradient in different directions (i.e., x, y, and z directions) is calculated for each position. The average vacuum prediction gradient in different directions at each position is used as the initialization velocity of the particle.

[0098] At this point, a reasonable number of particles can be obtained, avoiding the hyperparameter setting of the particle swarm in the traditional particle swarm algorithm. At the same time, each particle has its own initialization speed.

[0099] Step S005, constructing the objective function of the particles, performing multi-dimensional space iterative optimization based on the objective function according to the initialization speed of each particle, and obtaining the optimal circulating water flow rate in the current environment.

[0100] Update the velocity and position of each particle:

[0101] v i =ω*v i +c1*rand()*(pbest i -x i )+c2*rand()*(gbest i -x i )

[0102] xi =x i +v i

[0103] where v i 、x i Respectively represent the speed and position of the ith particle. c1 and c2 are learning factors, and their empirical values ​​are both 2. rand() is a random number between 0 and 1. pbest i 、gbest i are the historical optimal position of the ith particle and the global optimal position of the group, respectively. ω represents the inertia factor sequence of the particle, which represents the inertia factor in different directions of the search space, and its value is:

[0104] ω=[W1,W2,W3]

[0105] W1, W2, and W3 represent the inertia factors in the three directions of condenser end difference, circulating water temperature, and turbine load value respectively.

[0106]

[0107] w max is the maximum inertia weight, the empirical value is 1, w min is the minimum inertia weight, the empirical value is 0.3, w i Indicates the importance of the i-th feature. Ultimately, the greater the importance of the feature, the smaller its inertia weight, that is, the slower the update speed, because the more important the feature is, the more the change of the feature affects the optimal solution, so its weight should be smaller, making the particle move smaller. k is the maximum number of iterations, and the empirical value is 100. g is the current number of iterations.

[0108] After the particle position is updated, the objective function value under all circulating water flow rates needs to be calculated.

[0109] The objective function used by each particle is:

[0110]

[0111] min represents minimizing the objective function, is the maximum vacuum value, the empirical value is 100, and model(x1,x2,x3,x4) represents the vacuum prediction value of the neural network under the current input parameters. x1,x2,x3,x4 represent the condenser end difference, circulating water temperature, circulating water flow, and turbine load value respectively. τ represents the cost mapping coefficient of circulating water, which needs to be manually converted by power plant personnel. D x3 , D′ x3It indicates the level difference between the searched circulating water flow and the current equipment circulating water flow, because the circulating water flow is regulated by level, θ indicates the mapping coefficient, and the empirical value is 10. This item ensures that the adjusted circulating water flow is as small as possible and avoids frequent adjustments to the circulating water flow when the data volatility is large.

[0112] At present, the selection of the best vacuum of the condenser is to maximize the difference between the increase in the electric power of the turbine generator set and the increase in the power consumption of the circulating water pump under the premise of a certain turbine load and circulating water inlet temperature to determine the best circulating water volume and select the best vacuum of the condenser based on this. The biggest problem is that the cost of the circulating water itself is not taken into account in the treatment of the circulating water operation cost. The best vacuum of the condenser determined after considering the circulating water cost is the best vacuum in the true sense. This method of determining the best vacuum of the condenser can ensure the maximum economic benefits of the turbine operation rather than just the maximum energy conversion benefits.

[0113] At this point, the multidimensional space weighted particle swarm algorithm can be used to obtain the optimal position, and then the optimal control parameter at the position, that is, the circulating water flow rate, is obtained, and then the circulating water flow rate is adjusted to obtain the optimal vacuum value. The specific optimization process of the particle swarm algorithm is not described in detail in the present invention.

[0114] As for when to perform condenser vacuum optimization, etc., the implementer may manually set it based on the variation range, and the specific details of the present invention will not be discussed further.

[0115] It should be noted that the sequence of the embodiments of the present invention described above is for description only and does not represent the advantages and disadvantages of the embodiments. The above describes specific embodiments of this specification. Other embodiments are within the scope of the appended claims. In some cases, the actions or steps recorded in the claims can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0116] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.

[0117] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A real-time optimization method for power plant condenser vacuum based on data mining algorithm, characterized in that: The optimization method includes the following steps: Based on online monitoring instruments, obtain the time series steam turbine load, circulating water temperature, circulating water flow, and condenser terminal difference parameter data in the power plant; Constructing a multi-encoder neural network to fit the relationship between the time series steam turbine load, circulating water temperature in the power plant, circulating water flow, condenser terminal difference parameter data and condenser vacuum; Obtaining data volatility and feature importance of each parameter data; obtaining the change importance of each parameter data according to the data volatility and feature importance of the parameter data; obtaining the change importance ratio of each feature according to the change importance of the parameter data; obtaining the initial number of divided particles for each feature based on the feature change importance ratio; Obtaining vacuum prediction values ​​of different circulating water volumes for each particle; obtaining initialization speeds of the particles based on the vacuum prediction values ​​of different circulating water volumes for each particle; Constructing the objective function of the particles, performing multi-dimensional space iterative optimization based on the objective function according to the initialization speed of each particle, and obtaining the optimal circulating water flow rate in the current environment; The objective function of the particle is: min represents minimizing the objective function, is the maximum vacuum value, model(x1,x2,x3,x4) represents the vacuum prediction value of the neural network under the current input parameters, x1,x2,x3,x4 represent the condenser end difference, circulating water temperature, circulating water flow, and turbine load value respectively, τ represents the cost mapping coefficient of circulating water, D x3 , D′ x3 It represents the level difference between the searched circulating water flow and the current equipment circulating water flow, and θ represents the mapping coefficient of the circulating water flow level difference.

2. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 1 is characterized in that: The structure of the multi-encoder neural network is: The neural network contains four encoders: The input of the first encoder is the condenser terminal difference in the power plant, and the output is the terminal difference feature vector; The input of the second encoder is the steam turbine load in the power plant, and the output is the load characteristic vector; The input of the third encoder is the circulating water flow rate, and the output is the circulating water flow rate feature vector; The input of the fourth encoder is the circulating water temperature, and the output is the circulating water temperature feature vector; The first encoder, the second encoder, the third encoder, and the fourth encoder all use a fully connected network, and finally perform feature fusion on the terminal difference feature vector, the load feature vector, the circulating water flow feature vector, and the circulating water temperature feature vector, and the fusion operation uses the Add operation; the fused feature vector is obtained, and then the fused feature vector is input into the regression layer; the regression layer includes a fully connected layer and an activation function layer; the fully connected layer and the activation function layer are used to fit the relationship between the fused feature vector and the condenser vacuum, and finally output the condenser vacuum value; the activation function of the last activation function layer of the regression layer uses the Mish activation function.

3. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 1 is characterized in that: The method for obtaining the data volatility of the parameter data is: The condenser end difference, circulating water temperature, and turbine load in real-time operation are obtained, and then the variance of each data with a length of T is calculated. The maximum variance is used as the volatility of the data. The larger the variance, the stronger the volatility of the data. The condenser terminal difference is calculated once every minute. The variance of the condenser terminal difference data at the current moment and the continuous condenser terminal difference at the previous T-1 moments is calculated, and the maximum variance is taken as the data volatility of the condenser terminal difference during this period.

4. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 1 is characterized in that: The method for obtaining the feature importance of the parameter data is: The importance score F of each input feature of the neural network is obtained using the BorutaShap method; Then get the importance of each feature: where F i is the importance score of the i-th feature, n is the number of features, and w i Indicates the importance of the i-th feature.

5. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 1 is characterized in that: The calculation method of the importance of the parameter data change and the importance ratio of the feature change is: Get the change importance U of each feature: U=e w *σ σ is the data volatility of the feature, w is the importance of the feature, and e is a natural constant, the purpose of which is to expand the value of the feature importance; Feature change importance ratio: n is the number of features, U i represents the importance of the change of the i-th feature, P i is the change importance ratio of the i-th feature.

6. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 1 is characterized in that: The calculation method of the initial number of particles K′ for each feature is: K′=round(P i *K) round is the rounding up function; P i is the change importance ratio of the i-th feature, and K is the maximum particle value in the single-dimensional space.

7. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 1 is characterized in that: The calculation method of the particle initialization speed is: The vacuum prediction value at each circulating water volume is converted into a matrix form, the size of which is K′1*K′2*K′3, where K′1, K′2, and K′3 are the numbers of initial division examples in the three feature dimensions respectively; each position in the matrix is ​​the current vacuum prediction value; then the 3D Sobel operator is used to obtain the vacuum prediction gradient of each position in each matrix, and then the average vacuum prediction gradient in the three directions of x, y, and z at each position is calculated, and the average vacuum prediction gradient in different directions at each position is used as the initialization speed of the particle.

8. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 1 is characterized in that: The particle speed and position update method is: V i =ω*v i +c1*rand()*(pbest i -x i )+c2*rand()*(gbest i -x i ) x i =x i +v i where v i 、x i Respectively represent the speed and position of the i-th particle; c1 and c2 are learning factors; rand() is a random number between 0 and 1; pbest i 、gbest i are the historical optimal position of the ith particle and the global optimal position of the group, respectively; ω represents the inertia factor sequence of the particle, which represents the inertia factor in different directions of the search space; After the particle position is updated, the objective function value under all circulating water flow rates needs to be calculated.

9. The real-time optimization method for power plant condenser vacuum based on data mining algorithm according to claim 8, characterized in that: The method for obtaining the inertia factor sequence of the particles is: ω=[W1,W2,W3] W1, W2, and W3 represent the inertia factors in the three directions of condenser end difference, circulating water temperature, and turbine load value respectively; w max is the maximum inertia weight, w min is the minimum inertia weight, w i represents the importance of the i-th feature, G k is the maximum number of iterations, and g is the current number of iterations.

Images