Flexible resource parameter identification method based on KAN neural network

Abstract

The invention provides a flexible resource parameter identification method based on a KAN neural network, and relates to the technical field of power grid resource parameter identification, and the method comprises the following steps: obtaining historical operation data of power grid resources, and identifying flexible resources in the power grid resources through an analytic hierarchy process; according to the identified flexible resources, constructing a flexible resource digital model through a digital twinning technology, and extracting a sample data set according to the flexible resource digital model; based on the extracted sample data set, constructing and training a resource parameter identification model through a KAN neural network; and acquiring operation data of the flexible resources in real time, and identifying flexible resource parameters by using the trained resource parameter identification model. According to the method, the accuracy of identifying the flexible resource parameters by the model can be improved, more accurate decision support can be provided for power grid operators by identifying the flexible resource parameters, and more reasonable decisions can be made in the aspects of resource scheduling, energy distribution and the like.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid resource parameter identification, and in particular to a flexible resource parameter identification method based on a KAN neural network. Background Art

[0002] With the continuous development of the energy industry, smart grid technology has been widely promoted and applied in my country. The introduction of this technology has not only injected new vitality into the power industry, but also built a more efficient, intelligent and sustainable power production, transmission, distribution and use system for us. The core concept of smart grid is to use advanced information technology and communication technology to realize the comprehensive digitalization and intelligent management of the power system. It can monitor and optimize all aspects of power production, transmission, distribution and use in real time to ensure the stability and security of power supply. Grid resources mainly refer to facilities and technologies used to generate, transmit and distribute electric energy. These resources ensure that electricity reaches end users from power stations through transmission networks. Grid resources can be divided into flexible resources and inflexible resources. Flexible resources refer to facilities that can quickly adjust output to cope with changes in grid demand. This type of resource is characterized by fast response speed, flexible scheduling, and can effectively support the stable operation of the grid during peak demand or other changes. Inflexible resources refer to grid facilities that take a long time to start and adjust output. These resources usually have high stability and reliability, but it is more difficult to cope with short-term and large changes in grid demand.

[0003] Flexible resource parameter identification is a key technology used to determine the key attributes and states of adjustable resources in the power grid so that these resources can be better utilized to support the stability and efficiency of the power grid. This identification usually involves extracting valuable information from a large amount of data to determine which resources can provide the necessary flexibility and regulation capabilities in power grid operations. Existing technologies often rely on traditional empirical judgments or relatively simple analysis methods when identifying and evaluating flexible resources in power grids. These methods lack sufficient accuracy and systematicity, which may lead to misjudgments in resource evaluation and classification, and fail to accurately identify resources with real regulation potential, thereby affecting the overall operating efficiency and responsiveness of the power grid.

[0004] Currently, no effective solution has been proposed for the problems in the related technologies. Summary of the invention

[0005] In view of this, the present invention provides a flexible resource parameter identification method based on KAN neural network to solve the above-mentioned problems.

[0006] In order to solve the above problems, the specific technical solutions adopted by the present invention are as follows:

[0007] A flexible resource parameter identification method based on KAN neural network, the method comprising the following steps:

[0008] S1. Obtain historical operation data of power grid resources and identify flexible resources in power grid resources through hierarchical analysis and evaluation method;

[0009] S2. Based on the identified flexible resources, a flexible resource digital model is constructed through digital twin technology, and a sample data set is extracted based on the flexible resource digital model;

[0010] S3, based on the extracted sample data set, construct and train the resource parameter identification model through the KAN neural network;

[0011] S4. Acquire the operation data of the flexible resources in real time, and identify the flexible resource parameters using the trained resource parameter identification model;

[0012] The process of obtaining historical operation data of power grid resources and identifying flexible resources in power grid resources through analytic hierarchy process includes the following steps:

[0013] S11. Preprocess the historical operation data of power grid resources, where the data preprocessing includes data cleaning and standardization;

[0014] S12. Construct an initial resource flexibility evaluation index system based on the preprocessed historical operation data;

[0015] S13, optimizing the initial resource flexibility evaluation index system by using the maximum irrelevance method to obtain a secondary resource flexibility evaluation index system;

[0016] S14, using particle swarm analytic hierarchy process to determine the weight of each evaluation index in the secondary resource flexibility evaluation index system;

[0017] S15. Based on the weight of each evaluation index, the flexibility of each power grid resource is evaluated by fuzzy comprehensive evaluation method;

[0018] S16. Based on the flexibility evaluation results, the grid resource indicators whose flexibility evaluation results are higher than the preset flexibility threshold are regarded as flexible resources.

[0019] Preferably, the optimization process of the initial resource flexibility evaluation index system by using the maximum irrelevance method to obtain the secondary resource flexibility evaluation index system includes the following steps:

[0020] S131. Calculate the average value of each evaluation index in the initial resource flexibility evaluation index system;

[0021] S132, calculating the variance of each evaluation indicator and the covariance between the evaluation indicators according to the average value of each evaluation indicator;

[0022] S133, calculating the correlation coefficient between the evaluation indicators according to the variance of the evaluation indicators and the covariance between the evaluation indicators;

[0023] S134, constructing a correlation matrix according to the correlation coefficient between each evaluation index, and calculating the complex correlation coefficient of each evaluation index;

[0024] S135. According to the complex correlation coefficient of each evaluation indicator, select the evaluation indicators that meet the preset coefficient threshold to construct a secondary resource flexibility evaluation indicator system.

[0025] Preferably, a correlation matrix is ​​constructed according to the correlation coefficients between each evaluation index, and the calculation formula for calculating the complex correlation coefficient of each evaluation index is:

[0026]

[0027] In the formula, G i represents the complex correlation coefficient of the i-th evaluation index; U -i represents the matrix formed by removing the i-th row and i-th column from the correlation matrix U; u i represents the correlation coefficient vector without the i-th row and i-th column in the correlation matrix U; T represents the matrix transpose.

[0028] Preferably, determining the weight of each evaluation index in the secondary resource flexibility evaluation index system by using the particle swarm analytic hierarchy process includes the following steps:

[0029] S141. Constructing a hierarchical structure model according to each evaluation index in the secondary resource flexibility evaluation index system;

[0030] S142. According to the hierarchical structure model, the relative importance of each evaluation indicator is evaluated and an initial judgment matrix is ​​constructed according to the relative importance evaluation results;

[0031] S143, optimizing the initial judgment matrix by using a particle swarm optimization algorithm to obtain an optimized judgment matrix;

[0032] S144. According to the optimization judgment matrix, the weight of each evaluation index in the secondary resource flexibility evaluation index system is calculated by the hierarchical analysis method.

[0033] Preferably, the initial judgment matrix is ​​optimized by a particle swarm optimization algorithm to obtain an optimized judgment matrix, which includes the following steps:

[0034] S1431, randomly generate a population of particles of relative importance, and initialize the positions and velocities of the particles of relative importance;

[0035] S1432, establishing an objective function with the goal of minimizing the judgment matrix, calculating the fitness value of each relatively important particle according to the objective function, and updating the optimal solution of the relatively important particles and the global optimal solution;

[0036] S1433, according to the optimal solution of the relative importance particle and the global optimal solution, update the speed and position of the particle, generate a new judgment matrix, and repeat step S1432 until the preset termination condition is met;

[0037] S1434. Output the judgment matrix corresponding to the global optimal solution as the optimized judgment matrix.

[0038] Preferably, based on the weights of the evaluation indicators, evaluating the flexibility of each power grid resource by a fuzzy comprehensive evaluation method comprises the following steps:

[0039] S151. Calculate the membership of each power grid resource on each evaluation index based on a preset membership function;

[0040] S152, constructing a weighted membership matrix according to the weight of each evaluation indicator and the corresponding membership degree;

[0041] S153. Calculate the flexibility score of each grid resource by using the fuzzy weighted average method.

[0042] Preferably, constructing a flexible resource digital model by digital twin technology according to the identified flexible resource, and extracting a sample data set according to the flexible resource digital model comprises the following steps:

[0043] S21. Establishing a corresponding physical model based on the historical operation data of each flexible resource and the physical characteristics and operation mechanism of the flexible resource;

[0044] S22. calibrating the physical model through a machine learning algorithm, and using the calibrated physical model for simulation to generate simulation data;

[0045] S23. Extract a sample data set from the generated simulation data.

[0046] Preferably, based on the extracted sample data set, constructing and training a resource parameter identification model through a KAN neural network includes the following steps:

[0047] S31, normalizing the extracted sample data set, and dividing the normalized sample data set into a training set and a test set;

[0048] S32, constructing a KAN neural network and determining the network structure of the KAN neural network, initializing weight parameters and spline function parameters of the KAN neural network;

[0049] S33, training the KAN neural network using the training set, and updating the weight parameters and spline function parameters of the KAN neural network by back propagation algorithm and elastic gradient descent method;

[0050] S34. After the KAN neural network training is completed, the trained KAN neural network is evaluated using the test set, and the network structure is adjusted according to the evaluation result to obtain a resource parameter identification model.

[0051] Preferably, training the KAN neural network using the training set and updating the weight parameters and spline function parameters of the KAN neural network by back propagation algorithm and elastic gradient descent method include the following steps:

[0052] S331, inputting the data in the training set into the KAN neural network, using the spline function and weight parameters and calculating the output value of the KAN neural network through forward propagation;

[0053] S332, using the loss function to calculate the difference between the output value and the actual value to obtain the network error;

[0054] S333, calculating the gradient of the spline function coefficient and the weight parameter by back propagation algorithm according to the network error;

[0055] S334, using the elastic gradient descent method to update the weights and spline function parameters of the KAN network according to the calculated gradient.

[0056] Preferably, updating the weights and spline function parameters of the KAN network according to the calculated gradient using the elastic gradient descent method comprises the following steps:

[0057] S3341, comparing and determining whether the gradient sign at the current moment and the gradient sign at the previous moment have changed according to the gradients of the spline function coefficients and the weight parameters;

[0058] S3342. If the sign of the gradient changes, dynamically adjust the learning step size according to the change of the gradient sign according to the preset rules;

[0059] S3343. Calculate the update amount of each weight and spline function parameter according to the adjusted learning step size and current gradient;

[0060] S3344. Update the weights and spline function parameters of the KAN network according to the calculated update amount.

[0061] The beneficial effects of the present invention are:

[0062] 1. The present invention can quickly and accurately identify flexible resources from complex power grid resources through the hierarchical analysis evaluation method, which helps power grid operators to better understand and manage the adjustable parts in the power grid and optimize resource allocation. The digital twin technology is used to build a flexible resource digital model, which can highly restore the operating status of real power grid resources and provide an accurate simulation environment. The sample data set is extracted through the digital model. This process has a high degree of automation, can reduce manual intervention, and improve the efficiency and accuracy of data processing. The KAN neural network has strong learning and generalization capabilities. Based on the extracted sample data set, a resource parameter identification model is constructed and trained, which can improve the accuracy of the model in identifying flexible resource parameters. Identifying flexible resource parameters can provide more accurate decision support for power grid operators and make more reasonable decisions in resource scheduling, energy allocation, etc.

[0063] 2. The present invention pre-processes the historical operation data of power grid resources to effectively remove abnormal data, fill missing values, and unify data formats and dimensions, thereby improving data quality and availability, constructing an initial resource flexibility evaluation index system and optimizing it using the maximum irrelevance method, which can ensure the independence of evaluation indicators and avoid redundancy between indicators, making the evaluation index system more scientific and streamlined, and using the particle swarm analytic hierarchy process to determine the weight of each evaluation indicator. This method combines the structure of the analytic hierarchy process and the global search capability of the particle swarm optimization algorithm, and can more objectively assign weights to each indicator, reduce the influence of subjective judgment, and effectively improve the accuracy and objectivity of power grid resource flexibility evaluation, providing strong support for the intelligent management and optimization of power grids.

[0064] 3. The present invention can eliminate the dimensional differences between different features by normalizing the extracted sample data set, so that the data can be processed on the same scale, and the KAN neural network is trained using the training set, and the network parameters are updated by the back propagation algorithm and the elastic gradient descent method. This method combines the ability of the back propagation algorithm to efficiently calculate gradients and the dynamic adjustment of the learning step size by the elastic gradient descent method, so that the model can converge faster and find a better solution. The elastic gradient descent method dynamically adjusts the learning step size by comparing the gradient signs at the current moment and the previous moment, and can adaptively adjust the learning speed during the training process, accelerate the convergence speed, and reduce the oscillation during the training process, effectively improving the performance and generalization ability of the model, and providing strong support for parameter identification of power grid resources. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. 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 creative work. In the drawings:

[0066] Figure 1 It is a flow chart of a flexible resource parameter identification method based on a KAN neural network according to an embodiment of the present invention. DETAILED DESCRIPTION

[0067] In order to enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of this application.

[0068] According to an embodiment of the present invention, a flexible resource parameter identification method based on a KAN neural network is provided.

[0069] The present invention will now be further described with reference to the accompanying drawings and specific embodiments. Figure 1 As shown, according to the flexible resource parameter identification method based on the KAN neural network according to an embodiment of the present invention, the method comprises the following steps:

[0070] S1. Obtain historical operation data of power grid resources and identify flexible resources in power grid resources through hierarchical analysis and evaluation method;

[0071] Specifically, the historical operating data of power grid resources usually comes from the power system's monitoring system, data acquisition and monitoring system or other related databases. These data include but are not limited to the grid's operating parameters such as voltage, current, power, frequency, as well as equipment status, fault records and other information.

[0072] As a preferred implementation, obtaining historical operation data of power grid resources and identifying flexible resources in power grid resources by using a hierarchical analysis evaluation method includes the following steps:

[0073] S11. Preprocess the historical operation data of power grid resources, where the data preprocessing includes data cleaning and standardization;

[0074] It should be noted that the purpose of data cleaning is to remove erroneous, duplicate or inconsistent data to ensure data quality and accuracy. Data standardization is an important part of data preprocessing, which aims to scale the data to a certain ratio so that it falls into a specific range, usually between 0 and 1 or -1 to 1, to eliminate the dimensional differences and order of magnitude differences between different features.

[0075] S12. Construct an initial resource flexibility evaluation index system based on the preprocessed historical operation data;

[0076] It should be noted that in determining the evaluation dimension of flexibility, flexibility is usually reflected in the ability of power grid resources to respond quickly to load changes, the wide range of regulation, and the stability of operation. Therefore, the evaluation dimensions can include response speed, regulation capability, operational stability, etc. Under each evaluation dimension, key indicators that can reflect the flexibility of power grid resources are selected. For example: response speed, regulation capability, operational stability, etc. Among them, the response speed can be obtained by calculating the time difference from the issuance of the instruction to the response of the resource; the regulation capability can be evaluated by testing the regulation range and accuracy of the resource under different loads; and the operational stability can be analyzed by counting the number of failures and maintenance records of the resource over a period of time.

[0077] S13, optimizing the initial resource flexibility evaluation index system by using the maximum irrelevance method to obtain a secondary resource flexibility evaluation index system;

[0078] As a preferred implementation, the initial resource flexibility evaluation index system is optimized using the maximum irrelevance method to obtain a secondary resource flexibility evaluation index system, which includes the following steps:

[0079] S131. Calculate the average value of each evaluation index in the initial resource flexibility evaluation index system;

[0080] S132, calculating the variance of each evaluation indicator and the covariance between the evaluation indicators according to the average value of each evaluation indicator;

[0081] It should be noted that variance measures the degree of dispersion of a single variable (here is the evaluation index), that is, the average of the squares of the deviations between its value and its mean. Covariance measures how two variables (here are two different evaluation indicators) change together. Positive values ​​indicate positive correlation, negative values ​​indicate negative correlation, and zero indicates no linear relationship.

[0082] S133, calculating the correlation coefficient between the evaluation indicators according to the variance of the evaluation indicators and the covariance between the evaluation indicators;

[0083] Specifically, the calculation formula for calculating the correlation coefficient between the evaluation indicators based on the variance of each evaluation indicator and the covariance between the evaluation indicators is:

[0084]

[0085] In the formula, R ij represents the correlation coefficient between the i-th evaluation index and the j-th evaluation index; S ij represents the covariance between the i-th evaluation index and the j-th evaluation index; S ii represents the variance of the i-th evaluation indicator; S jj represents the variance of the j-th evaluation metric.

[0086] S134, constructing a correlation matrix according to the correlation coefficient between each evaluation index, and calculating the complex correlation coefficient of each evaluation index;

[0087] Specifically, a correlation matrix is ​​constructed according to the correlation coefficients between each evaluation index, and the calculation formula for the complex correlation coefficient of each evaluation index is:

[0088]

[0089] In the formula, G i represents the complex correlation coefficient of the i-th evaluation index; U -i represents the matrix formed by removing the i-th row and i-th column from the correlation matrix U; u i represents the correlation coefficient vector without the i-th row and i-th column in the correlation matrix U; T represents the matrix transpose.

[0090] S135. According to the complex correlation coefficient of each evaluation indicator, select the evaluation indicators that meet the preset coefficient threshold to construct a secondary resource flexibility evaluation indicator system.

[0091] S14, using particle swarm analytic hierarchy process to determine the weight of each evaluation index in the secondary resource flexibility evaluation index system;

[0092] As a preferred implementation, the weights of each evaluation index in the secondary resource flexibility evaluation index system are determined by using the particle swarm analytic hierarchy process, including the following steps:

[0093] S141. Constructing a hierarchical structure model according to each evaluation index in the secondary resource flexibility evaluation index system;

[0094] S142. According to the hierarchical structure model, the relative importance of each evaluation indicator is evaluated and an initial judgment matrix is ​​constructed according to the relative importance evaluation results;

[0095] S143, optimizing the initial judgment matrix by using a particle swarm optimization algorithm to obtain an optimized judgment matrix;

[0096] As a preferred implementation, the initial judgment matrix is ​​optimized by a particle swarm optimization algorithm to obtain an optimized judgment matrix, which includes the following steps:

[0097] S1431, randomly generate a population of particles of relative importance, and initialize the positions and velocities of the particles of relative importance;

[0098] It should be noted that in particle swarm optimization, each particle represents a possible solution. In the application of the AHP, each particle can represent a judgment matrix. Therefore, the position of each particle represents a specific configuration of the judgment matrix, that is, an arrangement of the relative importance of each indicator in the AHP; a certain number of particles are randomly generated. The position of each particle (that is, the importance score of each evaluation indicator relative to other indicators) is initialized to a random value, usually within the allowed score range.

[0099] S1432, establishing an objective function with the goal of minimizing the judgment matrix, calculating the fitness value of each relatively important particle according to the objective function, and updating the optimal solution of the relatively important particles and the global optimal solution;

[0100] It should be noted that in the particle swarm optimization algorithm, in order to minimize the consistency ratio of the judgment matrix, the steps of establishing the objective function, calculating the fitness value, and updating the optimal solution are crucial. The objective function is usually defined as minimizing the consistency ratio (CR) of the judgment matrix. The consistency ratio is an indicator used in the hierarchical analysis method to evaluate the consistency of the judgment matrix. Ideally, CR should be less than 0.1. The objective function is expressed as f(X) = CR(X), where X is the judgment matrix represented by the particle.

[0101] The fitness value of each particle is the inverse of its objective function value. For each particle, if the current fitness value is higher than the highest fitness value found so far, the individual optimal position of the particle is updated to the current position. Among all particles, the particle with the highest fitness value is found and the global optimal position is updated to the position of the particle. The global optimal solution represents the best judgment matrix in the entire particle swarm.

[0102] S1433, according to the optimal solution of the relative importance particle and the global optimal solution, update the speed and position of the particle, generate a new judgment matrix, and repeat step S1432 until the preset termination condition is met;

[0103] S1434. Output the judgment matrix corresponding to the global optimal solution as the optimized judgment matrix.

[0104] S144. According to the optimization judgment matrix, the weight of each evaluation index in the secondary resource flexibility evaluation index system is calculated by the hierarchical analysis method.

[0105] Specifically, according to the optimization judgment matrix, the weights of each evaluation index in the secondary resource flexibility evaluation index system are calculated by the hierarchical analysis method, including the following steps:

[0106] The weight of each evaluation index is calculated using the optimized judgment matrix, which is usually done by calculating the eigenvector of the judgment matrix, where the eigenvector corresponding to the maximum eigenvalue is the weight vector; the calculation method of the weight vector is as follows:

[0107] Normalize each column of the judgment matrix, that is, divide each element by the sum of its column.

[0108] The vector obtained by averaging each row of the normalized matrix is ​​the weight vector.

[0109] S15. Based on the weight of each evaluation index, the flexibility of each power grid resource is evaluated by fuzzy comprehensive evaluation method;

[0110] As a preferred implementation, based on the weights of the evaluation indicators, the flexibility of each power grid resource is evaluated by a fuzzy comprehensive evaluation method, including the following steps:

[0111] S151. Calculate the membership of each power grid resource on each evaluation index based on a preset membership function;

[0112] It should be noted that membership function is a concept in fuzzy mathematics, which is used to describe the degree to which an element belongs to a set. In power grid resource evaluation, it is necessary to define a suitable membership function for each evaluation indicator. This function is usually designed according to the characteristics of the evaluation indicator and actual needs.

[0113] For example, for the indicator "power generation efficiency", the following membership function can be defined:

[0114] If the power generation efficiency is lower than a certain threshold, the membership is 0;

[0115] If the power generation efficiency is higher than a certain maximum value, the membership degree is 1;

[0116] If the power generation efficiency is between the two, the membership can be a linear increase or defined according to other logics.

[0117] For each power grid resource and each evaluation indicator, the membership is calculated using the corresponding membership function and the collected data. This calculation process may involve substituting actual data into the membership function for calculation, and the result is a value between 0 and 1, indicating the degree of membership of the power grid resource on this evaluation indicator.

[0118] S152, constructing a weighted membership matrix according to the weight of each evaluation indicator and the corresponding membership degree;

[0119] S153. Calculate the flexibility score of each grid resource by using the fuzzy weighted average method.

[0120] Specifically, the calculation of the flexibility score of each grid resource by the fuzzy weighted average method includes the following steps:

[0121] For each power grid resource, the weighted membership of all evaluation indicators is summed to obtain the total weighted membership of the power grid resource.

[0122] The total weighted membership is divided by the number of evaluation indicators to obtain the flexibility score of the power grid resource.

[0123] S16. Based on the flexibility evaluation results, the grid resource indicators whose flexibility evaluation results are higher than the preset flexibility threshold are regarded as flexible resources.

[0124] S2. Based on the identified flexible resources, a flexible resource digital model is constructed through digital twin technology, and a sample data set is extracted based on the flexible resource digital model;

[0125] As a preferred implementation, according to the identified flexible resources, a flexible resource digital model is constructed by digital twin technology, and a sample data set is extracted according to the flexible resource digital model, including the following steps:

[0126] S21. Establishing a corresponding physical model based on the historical operation data of each flexible resource and the physical characteristics and operation mechanism of the flexible resource;

[0127] Specifically, based on the historical operating data of each flexible resource, such as power output, energy storage, charge and discharge rates, response time, operating efficiency, fault records, etc., the data covers different operating conditions and grid demand scenarios to provide a more comprehensive understanding of the dynamic characteristics of the resources.

[0128] In-depth analysis of the historical operating data of each flexible resource to identify the operating mode and characteristics of the resource. This includes statistical analysis, time series analysis, spectrum analysis, etc. to reveal hidden patterns and associations in the data;

[0129] Have an in-depth understanding of the physical characteristics and operating mechanisms of flexible resources, including the capacity, energy conversion efficiency, response speed, stability, etc. of the resources. At the same time, it is also necessary to understand how the resources respond to grid demand, how they interact with other resources, and how they behave under different operating conditions;

[0130] Based on the above analysis, a physical model is established, which should accurately describe the dynamic behavior of flexible resources, including their power output, energy storage and release, response speed, etc. The model includes a set of differential equations, difference equations or algebraic equations that describe how the state variables of the resource (such as energy storage, temperature, etc.) change over time.

[0131] S22. calibrating the physical model through a machine learning algorithm, and using the calibrated physical model for simulation to generate simulation data;

[0132] It should be noted that machine learning algorithms include linear regression, support vector machine (SVM), neural network, etc.; the selected machine learning algorithm is trained using the prepared data set. In this process, the algorithm will learn the inherent laws and patterns of the flexible resource operation data, so that it can predict the dynamic behavior of the resource. During the training process, it may be necessary to adjust the parameters and structure of the model to achieve the best prediction performance. After the training is completed, the prediction results of the machine learning model are used to calibrate the physical model. This usually involves adjusting the parameters of the physical model or correcting some of its assumptions so that the output of the physical model is more consistent with the prediction of the machine learning model. The goal of calibration is to reduce the difference between the physical model and the actual operation data.

[0133] S23. Extract a sample data set from the generated simulation data.

[0134] It should be noted that a sample data set is extracted from the generated simulation data. This sample data set includes input features (such as grid demand, resource status, environmental conditions, etc.) and corresponding output labels (resource parameters that the model needs to identify, such as power output, efficiency, response time, etc.).

[0135] S3, based on the extracted sample data set, construct and train the resource parameter identification model through the KAN neural network;

[0136] It should be noted that KAN neural network (Kolmogorov-Arnold Networks) is an innovative neural network architecture that has made significant improvements on the traditional multi-layer perceptron (MLPs), mainly by transferring the activation function from the node to the edge, thereby improving the performance and interpretability of the network. In KAN, all weight parameters are replaced by univariate spline functions, which can be adaptively adjusted according to the training data.

[0137] As a preferred implementation, based on the extracted sample data set, constructing and training a resource parameter identification model through a KAN neural network includes the following steps:

[0138] S31, normalizing the extracted sample data set, and dividing the normalized sample data set into a training set and a test set;

[0139] It should be noted that normalization is a data preprocessing technique used to adjust the scale of data so that it falls within a specific range (usually between 0 and 1) in order to better adapt to machine learning algorithms. Commonly used normalization methods include minimum-maximum normalization.

[0140] Dividing the dataset into training and test sets is a key step in evaluating the performance of machine learning models. The training set is used to train the model, while the test set is used to evaluate the generalization ability of the model. Usually, the ratio of training set to test set can be 70:30 or 80:20.

[0141] S32, constructing a KAN neural network and determining the network structure of the KAN neural network, initializing weight parameters and spline function parameters of the KAN neural network;

[0142] It should be noted that determining the network structure of the KAN neural network includes: determining the input and output dimensions of the network according to application requirements, and determining the network level and number of nodes.

[0143] In KANs, traditional weights are viewed as spline functions on the edges. Therefore, initializing these functions is a key step. Spline functions can be linear splines, cubic splines, or higher-order spline functions, and the parameters of each spline function (such as node positions, slopes, etc.) need to be initialized. The initial settings of these parameters can be random or preset based on prior knowledge.

[0144] S33, training the KAN neural network using the training set, and updating the weight parameters and spline function parameters of the KAN neural network by back propagation algorithm and elastic gradient descent method;

[0145] As a preferred implementation, training the KAN neural network using the training set and updating the weight parameters and spline function parameters of the KAN neural network by the back propagation algorithm and the elastic gradient descent method include the following steps:

[0146] S331, inputting the data in the training set into the KAN neural network, using the spline function and weight parameters and calculating the output value of the KAN neural network through forward propagation;

[0147] It should be noted that in KAN, each connection or edge is not a simple weight, but a spline function that can automatically adjust its shape according to the input. These spline functions adaptively adjust their parameters, such as slope or curvature, according to the input of the edge to transmit information in a way that best suits the data. The network calculates the output of each layer through these spline functions until the last layer. The output of the spline function is passed to the nodes of the next layer and serves as the input of the nodes of the next layer. This process continues until the output layer is reached. Ultimately, the output layer of the network will generate predictions or classification results, which are obtained based on a series of transformations from input to output of the entire network.

[0148] S332, using the loss function to calculate the difference between the output value and the actual value to obtain the network error;

[0149] It should be noted that the loss function uses the mean square error (MSE), which is the average of the squares of the differences between the predicted values ​​and the actual values, comparing the network's output values ​​with the corresponding actual values. Specifically, it involves comparing each predicted value with its corresponding true value and applying the loss function to calculate the difference between the two. After applying the loss function to all the predictions for the entire dataset, the total error is accumulated. This total error is the target that needs to be minimized during the optimization process, and it reflects the performance of the model on the entire training set.

[0150] S333, calculating the gradient of the spline function coefficient and the weight parameter by back propagation algorithm according to the network error;

[0151] Specifically, according to the network error, the gradient of the spline function coefficients and weight parameters is calculated by the back propagation algorithm, which includes the following steps:

[0152] Based on the obtained network error, we start from the output layer and work backwards to the input layer to calculate the influence of each parameter on the total error, i.e. the gradient of the parameter. For the spline function parameters in KAN, this means that we need to calculate the contribution of each spline function to the error under given input and current parameters.

[0153] When calculating gradients, the chain rule is applied, which is a basic rule in calculus for calculating the derivatives of composite functions. In backpropagation, the output of each layer is a function of the input of the next layer, so the gradient calculation of each layer depends on the gradient of the next layer.

[0154] For each coefficient of the spline function, calculate their sensitivity to the final output error. These gradients indicate how to adjust these coefficients to most effectively reduce the overall error;

[0155] The gradients computed over all training examples are accumulated for batch updates, which helps stabilize the learning process and speed up convergence.

[0156] S334, using the elastic gradient descent method to update the weights and spline function parameters of the KAN network according to the calculated gradient.

[0157] As a preferred embodiment, updating the weights and spline function parameters of the KAN network according to the calculated gradient using the elastic gradient descent method includes the following steps:

[0158] S3341, comparing and determining whether the gradient sign at the current moment and the gradient sign at the previous moment have changed according to the gradients of the spline function coefficients and the weight parameters;

[0159] It should be noted that the gradient value of each parameter in the previous training step is stored during the training process. In each iteration, the current gradient of each parameter (including the coefficients of the spline function and the network weights) is calculated by the back propagation algorithm. For each parameter, the sign of the current gradient value is compared with the sign of the stored previous gradient value. The sign can be determined by calculating the positive and negative of the gradient value, such as using a sign function. If the sign of the current gradient is different from the sign of the previous gradient, it means that the sign of the gradient has changed, which can be detected by comparing whether the product of the two signs is less than zero; after each iteration, the stored gradient value is updated to the current gradient value for use in the next iteration.

[0160] S3342. If the sign of the gradient changes, dynamically adjust the learning step size according to the change of the gradient sign according to the preset rules;

[0161] Specifically, if the sign of the gradient changes, dynamically adjusting the learning step size according to the preset rules based on the change of the gradient sign includes the following steps:

[0162] Check whether the sign of the gradient has changed, which usually means that the signs of the gradients before and after are opposite, such as from positive to negative or from negative to positive.

[0163] If the gradient sign is opposite, this usually indicates that the previous learning step size may be too large, causing the optimization process to oscillate around the minimum value and fail to stabilize. Therefore, the learning step size can be reduced to reduce the amplitude of parameter updates and help the model converge more smoothly;

[0164] If the gradient sign remains the same, consider keeping the current learning step size or increasing it appropriately according to the training situation to speed up convergence;

[0165] Dynamically adjusting the learning step size includes: reducing the learning step size and increasing the learning step size;

[0166] Among them, to reduce the learning step size, you can multiply the current learning step size by a factor less than 1 (such as 0.5 or 0.8). This can effectively reduce the update amount of each iteration and reduce the fluctuation of parameter updates;

[0167] Increase the learning step size. If the training process is stable and the error continues to decrease, you can increase the learning step size appropriately by multiplying it by a factor greater than 1 (such as 1.1 or 1.2) to try to speed up convergence.

[0168] In each iteration, the learning step size is adjusted according to the change in the gradient sign and the preset rules, and then the new learning step size is used to update the weights and spline function parameters.

[0169] S3343. Calculate the update amount of each weight and spline function parameter according to the adjusted learning step size and current gradient;

[0170] It should be noted that for each parameter, the adjusted learning step size is multiplied by the current gradient. This product is the update amount of the parameter, which indicates the direction and distance the parameter should move.

[0171] S3344. Update the weights and spline function parameters of the KAN network according to the calculated update amount.

[0172] It should be noted that the calculated update is applied to each weight and spline function parameter. If the gradient is positive, the parameter will be reduced; if the gradient is negative, the parameter will be increased, thereby reducing the overall error of the network.

[0173] S34. After the KAN neural network training is completed, the trained KAN neural network is evaluated using the test set, and the network structure is adjusted according to the evaluation result to obtain a resource parameter identification model.

[0174] It should be noted that the trained KAN neural network is applied to the test set. The test set is data that is not used in the training process and can provide an independent evaluation of the network performance. The network predicts the output of the test data and compares the predicted results with the actual values ​​to calculate evaluation indicators such as accuracy, recall, and precision.

[0175] Analyze the performance of the network on the test set in detail, identify the types of data on which the model performs well or poorly, and adjust the network structure, including increasing or decreasing the number of layers, adjusting activation functions, modifying optimization algorithms, etc. Train and evaluate again based on the adjusted network structure to observe whether the optimization has effectively improved the performance of the model. This process requires multiple iterations until satisfactory accuracy and generalization ability are achieved.

[0176] S4. Acquire the operation data of the flexible resources in real time, and identify the flexible resource parameters using the trained resource parameter identification model;

[0177] Specifically, obtaining the operation data of the flexible resource in real time and identifying the flexible resource parameters using the trained resource parameter identification model includes the following steps:

[0178] Deploy the trained resource parameter identification model to the server or edge device so that it can receive and process real-time data;

[0179] The operating data of flexible resources is acquired in real time and input into the resource parameter identification model. The model performs the reasoning process based on the input data to identify the parameters of flexible resources. The model outputs the identification results, which can be parameters such as the status, type, and performance indicators of the resource. The identified resource parameters are applied to the decision support system to provide a basis for resource management.

[0180] In summary, with the help of the above-mentioned technical scheme of the present invention, the present invention can quickly and accurately identify flexible resources from complex power grid resources through the hierarchical analysis evaluation method, which helps power grid operators to better understand and manage the adjustable parts in the power grid and optimize resource allocation. The digital twin technology is used to build a flexible resource digital model, which can highly restore the operating status of real power grid resources and provide an accurate simulation environment. The sample data set is extracted through the digital model. This process has a high degree of automation, can reduce manual intervention, and improve the efficiency and accuracy of data processing. The KAN neural network has strong learning and generalization capabilities. Based on the extracted sample data set, a resource parameter identification model is constructed and trained, which can improve the accuracy of the model in identifying flexible resource parameters. Identifying flexible resource parameters can provide more accurate decision support for power grid operators and make more reasonable decisions in resource scheduling, energy allocation, etc. The present invention preprocesses the historical operation data of power grid resources to effectively remove abnormal data, fill missing values, and unify data formats and dimensions, thereby improving data quality and availability, constructing an initial resource flexibility evaluation index system and optimizing it using the maximum irrelevance method, which can ensure the independence of evaluation indicators and avoid redundancy between indicators, making the evaluation index system more scientific and streamlined, and using the particle swarm analytic hierarchy process to determine the weight of each evaluation indicator. This method combines the structure of the analytic hierarchy process and the global search capability of the particle swarm optimization algorithm, and can more objectively assign weights to each indicator, reduce the influence of subjective judgment, and effectively improve the accuracy and objectivity of power grid resource flexibility evaluation, providing strong support for the intelligent management and optimization of power grids. The present invention normalizes the extracted sample data set to eliminate the dimensional differences between different features, so that the data is processed on the same scale, uses the training set to train the KAN neural network, and updates the network parameters through the back propagation algorithm and the elastic gradient descent method. This method combines the ability of the back propagation algorithm to efficiently calculate gradients and the elastic gradient descent method to dynamically adjust the learning step size, so that the model can converge faster and find a better solution. The elastic gradient descent method dynamically adjusts the learning step size by comparing the gradient signs at the current moment and the previous moment, and can adaptively adjust the learning speed during the training process, accelerate the convergence speed, and reduce the oscillation during the training process, effectively improving the performance and generalization ability of the model, and providing strong support for parameter identification of power grid resources.

[0181] It will be appreciated by those skilled in the art that embodiments of the present invention may be provided as methods, systems or computer program products. Therefore, the present invention may take the form of a complete hardware embodiment, a complete software embodiment or an embodiment combining software and hardware. Moreover, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, optical storage, etc.) containing computer-usable program code.

[0182] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A flexible resource parameter identification method based on KAN neural network, characterized in that: The method comprises the following steps: S1. Obtain historical operation data of power grid resources and identify flexible resources in power grid resources through hierarchical analysis and evaluation method; S2. Based on the identified flexible resources, a flexible resource digital model is constructed through digital twin technology, and a sample data set is extracted based on the flexible resource digital model; S3, based on the extracted sample data set, construct and train the resource parameter identification model through the KAN neural network; S4. Acquire the operation data of the flexible resources in real time, and identify the flexible resource parameters using the trained resource parameter identification model; The method of obtaining historical operation data of power grid resources and identifying flexible resources in power grid resources by using a hierarchical analysis evaluation method includes the following steps: S11, performing data preprocessing on historical operation data of power grid resources, wherein the data preprocessing includes data cleaning and standardization processing; S12. Construct an initial resource flexibility evaluation index system based on the preprocessed historical operation data; S13, optimizing the initial resource flexibility evaluation index system by using the maximum irrelevance method to obtain a secondary resource flexibility evaluation index system; S14, using particle swarm analytic hierarchy process to determine the weight of each evaluation index in the secondary resource flexibility evaluation index system; S15. Based on the weight of each evaluation index, the flexibility of each power grid resource is evaluated by fuzzy comprehensive evaluation method; S16. Based on the flexibility evaluation results, the grid resource indicators whose flexibility evaluation results are higher than the preset flexibility threshold are regarded as flexible resources.

2. According to claim 1, a flexible resource parameter identification method based on KAN neural network is characterized in that: The optimization process of the initial resource flexibility evaluation index system by using the maximum irrelevance method to obtain the secondary resource flexibility evaluation index system comprises the following steps: S131. Calculate the average value of each evaluation index in the initial resource flexibility evaluation index system; S132, calculating the variance of each evaluation indicator and the covariance between the evaluation indicators according to the average value of each evaluation indicator; S133, calculating the correlation coefficient between the evaluation indicators according to the variance of the evaluation indicators and the covariance between the evaluation indicators; S134, constructing a correlation matrix according to the correlation coefficient between each evaluation index, and calculating the complex correlation coefficient of each evaluation index; S135. According to the complex correlation coefficient of each evaluation indicator, select the evaluation indicators that meet the preset coefficient threshold to construct a secondary resource flexibility evaluation indicator system.

3. The method for identifying flexible resource parameters based on KAN neural network according to claim 2 is characterized in that: The correlation matrix is ​​constructed according to the correlation coefficient between each evaluation index, and the calculation formula for calculating the complex correlation coefficient of each evaluation index is: In the formula, G i represents the complex correlation coefficient of the i-th evaluation index; U -i represents the matrix formed by removing the i-th row and i-th column from the correlation matrix U; u i Represents the correlation coefficient vector in the correlation matrix U without the i-th row and i-th column; T stands for matrix transpose.

4. The method for identifying flexible resource parameters based on KAN neural network according to claim 2 is characterized in that: The method of using the particle swarm analytic hierarchy process to determine the weight of each evaluation index in the secondary resource flexibility evaluation index system includes the following steps: S141. Constructing a hierarchical structure model according to each evaluation index in the secondary resource flexibility evaluation index system; S142. According to the hierarchical structure model, the relative importance of each evaluation indicator is evaluated and an initial judgment matrix is ​​constructed according to the relative importance evaluation results; S143, optimizing the initial judgment matrix by using a particle swarm optimization algorithm to obtain an optimized judgment matrix; S144. According to the optimization judgment matrix, the weight of each evaluation index in the secondary resource flexibility evaluation index system is calculated by the hierarchical analysis method.

5. The method for identifying flexible resource parameters based on KAN neural network according to claim 4 is characterized in that: The optimization process of the initial judgment matrix by the particle swarm optimization algorithm to obtain the optimized judgment matrix comprises the following steps: S1431, randomly generate a population of particles of relative importance, and initialize the positions and velocities of the particles of relative importance; S1432, establishing an objective function with the goal of minimizing the judgment matrix, calculating the fitness value of each relatively important particle according to the objective function, and updating the optimal solution of the relatively important particles and the global optimal solution; S1433, according to the optimal solution of the relative importance particle and the global optimal solution, update the speed and position of the particle, generate a new judgment matrix, and repeat step S1432 until the preset termination condition is met; S1434. Output the judgment matrix corresponding to the global optimal solution as the optimized judgment matrix.

6. The method for identifying flexible resource parameters based on KAN neural network according to claim 2 is characterized in that: The method of evaluating the flexibility of each power grid resource by a fuzzy comprehensive evaluation method based on the weight of each evaluation index includes the following steps: S151. Calculate the membership of each power grid resource on each evaluation index based on a preset membership function; S152, constructing a weighted membership matrix according to the weight of each evaluation indicator and the corresponding membership degree; S153. Calculate the flexibility score of each grid resource by using the fuzzy weighted average method.

7. The method for identifying flexible resource parameters based on KAN neural network according to claim 1, characterized in that: The method of constructing a flexible resource digital model by using digital twin technology according to the identified flexible resource and extracting a sample data set according to the flexible resource digital model includes the following steps: S21. Establishing a corresponding physical model based on the historical operation data of each flexible resource and the physical characteristics and operation mechanism of the flexible resource; S22. calibrating the physical model through a machine learning algorithm, and using the calibrated physical model for simulation to generate simulation data; S23. Extract a sample data set from the generated simulation data.

8. The method for identifying flexible resource parameters based on KAN neural network according to claim 1, characterized in that: The method of constructing and training a resource parameter identification model based on the extracted sample data set by using a KAN neural network includes the following steps: S31, normalizing the extracted sample data set, and dividing the normalized sample data set into a training set and a test set; S32, constructing a KAN neural network and determining the network structure of the KAN neural network, initializing weight parameters and spline function parameters of the KAN neural network; S33, training the KAN neural network using the training set, and updating the weight parameters and spline function parameters of the KAN neural network by back propagation algorithm and elastic gradient descent method; S34. After the KAN neural network training is completed, the trained KAN neural network is evaluated using the test set, and the network structure is adjusted according to the evaluation result to obtain a resource parameter identification model.

9. The method for identifying flexible resource parameters based on KAN neural network according to claim 8, characterized in that: The method of training the KAN neural network using the training set and updating the weight parameters and spline function parameters of the KAN neural network by the back propagation algorithm and the elastic gradient descent method includes the following steps: S331, inputting the data in the training set into the KAN neural network, using the spline function and weight parameters and calculating the output value of the KAN neural network through forward propagation; S332, using the loss function to calculate the difference between the output value and the actual value to obtain the network error; S333, calculating the gradient of the spline function coefficient and the weight parameter by back propagation algorithm according to the network error; S334, using the elastic gradient descent method to update the weights and spline function parameters of the KAN network according to the calculated gradient.

10. The method for identifying flexible resource parameters based on KAN neural network according to claim 9, characterized in that: The method of updating the weights and spline function parameters of the KAN network according to the calculated gradient using the elastic gradient descent method comprises the following steps: S3341, comparing and determining whether the gradient sign at the current moment and the gradient sign at the previous moment have changed according to the gradients of the spline function coefficients and the weight parameters; S3342. If the sign of the gradient changes, dynamically adjust the learning step size according to the change of the gradient sign according to the preset rules; S3343. Calculate the update amount of each weight and spline function parameter according to the adjusted learning step size and current gradient; S3344. Update the weights and spline function parameters of the KAN network according to the calculated update amount.

Images