A power grid reactive power optimization intelligent regulation method, system, device and medium

Abstract

The application discloses a kind of power grid reactive power optimization intelligent regulation methods, systems, equipment and medium, comprising: collecting power grid operation data, using improved Newton-Raphson optimizer to optimize variational mode decomposition parameter to data is denoised, and by calculating historical position mean square deviation introduction guide learning strategy, obtain optimal data sequence;It is input into graph attention network-bidirectional long short-term memory network model training, capture space-time characteristics and obtain reactive power optimization model;Deployment model real-time analysis power grid data, adjust reactive power distribution.It realizes real-time, accurate, intelligent control to reactive power distribution of power grid, significantly enhances the stability and optimization efficiency of power grid operation.

Description

Technical Field

[0001] This invention relates to the field of power grid operation optimization technology, and in particular to a method, system, device and medium for intelligent control of reactive power optimization in power grids. Background Technology

[0002] The power grid is an indispensable energy hub and infrastructure in modern society, providing power and lighting for all social activities, including industrial manufacturing, agricultural production, commercial operations, and residential life. It is the cornerstone supporting the national economy and the operation of modern society. With the continuous increase in electricity demand, the power grid is constantly expanding, and its operation is becoming increasingly large and complex. Within this dynamic system, there is a wealth of valuable operational data, including operating parameters of various electrical equipment and electricity consumption patterns in different regions. Fully utilizing this data and extracting key information to achieve optimal reactive power distribution across the entire power grid is crucial for ensuring its economical and reliable operation.

[0003] However, current reactive power optimization and control methods for power grids mainly rely on manual experience and pre-set rules, resulting in significant drawbacks such as slow response speed, low efficiency, and poor adaptability. On the one hand, traditional control methods struggle to accurately capture the real-time operating status of the power grid, often proving ineffective in handling rapid changes in reactive power caused by factors such as load fluctuations. On the other hand, with the large-scale integration of new energy sources, the overall power grid structure exhibits significant intermittent and fluctuating characteristics, making it difficult for traditional methods to adapt to this new grid structure and effectively coordinate reactive power distribution among different power sources. Therefore, overcoming the bottlenecks of traditional control methods and achieving efficient, real-time reactive power optimization has become a pressing technical challenge. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method, system, device, and medium for intelligent control of reactive power optimization in power grids to address the problems of slow response, low decision-making efficiency, and poor global adaptability in existing power grid reactive power optimization methods when facing load fluctuations and intermittent access of new energy sources.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for intelligent control of reactive power optimization in power grids, comprising: Data sets are constructed by collecting power grid operation data, and noise reduction is performed on the dataset using variational mode decomposition. The parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer. The optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy to obtain the optimal data sequence after noise removal. The optimal data sequence is input into the constructed graph attention network-bidirectional long short-term memory network model for training. The spatial dependencies between nodes are captured by the multi-head graph attention layer, and the temporal features of the data are captured by the bidirectional long short-term memory network to obtain the trained reactive power optimization model. The trained reactive power optimization model is deployed to the power grid, and real-time power grid operation data is collected and analyzed. The reactive power allocation of the power grid is adjusted based on the analysis results.

[0007] As a preferred embodiment of the intelligent control method for reactive power optimization in the power grid described in this invention, the parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer, including: The objective function is constructed based on minimizing data envelopment entropy, and the variational mode decomposition parameters to be optimized are determined as the penalty factor and the number of decomposition layers. In the process of iteratively optimizing the parameters based on the objective function, the search direction is calculated using the Newton-Raphson search rule, and the individual with the best fitness value and the individual with the worst fitness value in the population are respectively replaced with the adjacent positions used to determine the search direction in the Newton method iterative formula to update the position of the population individuals and iterate cyclically until the convergence condition is met.

[0008] The beneficial effects of this preferred technical solution are that by introducing data envelopment entropy as an adaptive optimization objective and combining it with an improved Newton-Raphson search rule and a population extremum guidance strategy, the precise optimization of variational mode decomposition parameters is achieved, thereby effectively avoiding mode aliasing and significantly improving the noise reduction quality and feature extraction accuracy of complex power grid operation data.

[0009] As a preferred embodiment of the intelligent control method for reactive power optimization in the power grid described in this invention, the optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy, including: The mean squared error of the historical position of individuals in the population is calculated, and the empirical feedback parameters are obtained by combining the normalization coefficient. Compare the numerical values ​​of the empirical feedback parameters with the parameters to be designed. Perform the corresponding position update operation based on the comparison results; If the empirical feedback parameter is greater than the parameter to be designed, then the new individual position is calculated based on the optimal individual position of the population, combined with the tangent function value and the boundary value of the search space. If the empirical feedback parameter is not greater than the parameter to be designed, then a new individual location is randomly generated within the boundary of the search space.

[0010] The beneficial effect of this preferred technical solution is that by calculating the mean square error of the population's historical position, the current search state of the algorithm is dynamically evaluated, thereby adaptively guiding the algorithm to flexibly switch between local deep exploration and global random exploration. This effectively overcomes the defect of traditional optimization algorithms that are prone to getting trapped in local optima, and significantly improves the accuracy and convergence speed of optimization.

[0011] As a preferred embodiment of the intelligent control method for reactive power optimization in the power grid described in this invention, the step of training the graph attention network-bidirectional long short-term memory network model by inputting the optimal data sequence includes: The optimal input data sequence is normalized using min-max normalization to map the data to a preset interval.

[0012] As a preferred embodiment of the intelligent control method for reactive power optimization in the power grid described in this invention, the step of capturing the spatial dependencies between nodes through a multi-head graph attention layer includes: Perform a linear transformation on the input node features to obtain the transformed node features; Based on the transformed node features, the original attention scores between nodes are calculated using a feedforward neural network. The original attention scores are normalized using a normalized exponential function to obtain attention weights; The attention weights are used to perform a weighted summation of the feature information of neighboring nodes, and the node feature information is updated through a nonlinear activation function.

[0013] As a preferred embodiment of the intelligent control method for reactive power optimization in the power grid described in this invention, the step of capturing the temporal characteristics of data through a bidirectional long short-term memory network includes: The feature sequence output from the multi-head graph attention layer is input into a bidirectional long short-term memory network; The feature sequence is processed from both forward and backward directions using two independent long short-term memory network layers in a bidirectional long short-term memory network to capture contextual information in time series data.

[0014] As a preferred embodiment of the intelligent control method for reactive power optimization in the power grid described in this invention, the step of adjusting the reactive power distribution of the power grid operation based on the analysis results includes: The trained reactive power optimization model is deployed to the power grid dispatch automation energy management system in the power grid. The energy management system equipped with the model collects grid operation data in real time, inputs the operation data into the model for inference calculation, and regulates the grid based on the optimal reactive power allocation result output by the model.

[0015] Secondly, the present invention provides a smart control system for reactive power optimization in power grids, comprising: The preprocessing module is used to collect power grid operation data to construct a dataset and to perform noise reduction on the dataset using variational mode decomposition. The parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer. The optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy to obtain the optimal data sequence after noise removal. The construction and training module is used to input the optimal data sequence into the constructed graph attention network-bidirectional long short-term memory network model for training. It captures the spatial dependencies between nodes through the multi-head graph attention layer and captures the temporal features of the data through the bidirectional long short-term memory network to obtain the trained reactive power optimization model. The real-time reactive power optimization and control module is used to deploy the trained reactive power optimization model to the power grid, collect and analyze power grid operation data in real time, and adjust the reactive power distribution of the power grid operation based on the analysis results.

[0016] Thirdly, the present invention provides an electronic device, comprising: Memory, used to store programs; A processor is configured to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the intelligent control method for reactive power optimization in the power grid.

[0017] Fourthly, the present invention provides a computer-readable storage medium, comprising: when the program is executed by a processor, the steps of implementing the intelligent control method for reactive power optimization of the power grid.

[0018] The beneficial effects of this invention are as follows: By employing an improved Newton-Raphson optimizer to optimize the parameters of variational mode decomposition (VMD) and combining it with a guided learning strategy, this invention achieves efficient data denoising, accurately removing noise interference from power grid operation data and providing a high-quality data foundation for subsequent model training. By constructing a spatiotemporal hybrid model combining a graph attention network (GAT) and a bidirectional long short-term memory network (BiLSTM), it achieves synchronous and in-depth mining of the spatial dependencies and temporal characteristics of power grid data, thereby learning a more comprehensive and global representation of power grid operation characteristics. By establishing and deploying a data-driven intelligent control model, it achieves precise, efficient, and stable control of power grid reactive power, effectively improving the voltage stability, operational economy, and acceptance capacity of the power grid for new energy sources. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein: Figure 1 This is a basic flowchart illustrating a method for intelligent control of reactive power optimization in a power grid, as provided in one embodiment of the present invention. Figure 2 A flowchart illustrating an algorithm for a smart control method for reactive power optimization in a power grid, provided as an embodiment of the present invention. Figure 3 This is a data interaction timing diagram of a smart control method for reactive power optimization in a power grid, provided in one embodiment of the present invention. Detailed Implementation

[0020] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0021] Example 1, referring to Figure 1 As an embodiment of the present invention, a smart control method for reactive power optimization in a power grid is provided, comprising: S100: Collect power grid operation data to construct a dataset, and use variational mode decomposition to denoise the dataset. The parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer. The optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy to obtain the optimal data sequence after noise removal. S200: The optimal data sequence is input into the constructed graph attention network-bidirectional long short-term memory network model for training. The spatial dependencies between nodes are captured through the multi-head graph attention layer, and the temporal features of the data are captured through the bidirectional long short-term memory network to obtain the trained reactive power optimization model. S300: Deploy the trained reactive power optimization model to the power grid, collect and analyze power grid operation data in real time, and adjust the reactive power allocation of the power grid operation based on the analysis results.

[0022] It should be noted that modern power grids face multiple challenges during operation, including continuous load fluctuations and the intermittency and volatility brought about by the large-scale integration of new energy sources. This leads to rapid changes in the grid's operating status, making real-time and accurate monitoring difficult. Simultaneously, while massive amounts of grid operation data contain crucial information, traditional control methods struggle to effectively extract it, resulting in inefficient decision-making. Furthermore, traditional methods rely heavily on human experience and pre-set rules, exhibiting slow response times and poor adaptability. They are ill-suited to new grid structures and cannot achieve globally optimal reactive power allocation, thus posing challenges to the economic and reliable operation of the power grid.

[0023] Therefore, addressing the issues of slow response, low decision-making efficiency, and poor global adaptability in existing power grid reactive power optimization and control methods when facing load fluctuations and intermittent access of renewable energy, this paper proposes a solution through steps S100-S300. By introducing an improved Newton-Raphson optimizer to adaptively optimize the variational mode decomposition parameters, efficient noise reduction of power grid operation data is achieved. Furthermore, by combining a spatiotemporal hybrid model of graph attention network and bidirectional long short-term memory network (GAT-BiLSTM), the spatial dependence and temporal characteristics between power grid nodes are accurately captured, thereby achieving globally optimal reactive power allocation and significantly improving the stability, economy, and renewable energy acceptance capacity of the power grid.

[0024] Example 2, refer to Figures 2-3 As one embodiment of the present invention, based on the previous embodiment, a smart control method for reactive power optimization of a power grid is provided, comprising: like Figure 2 As shown, the algorithm initializes the parameters and the population, and determines the best individual in the current population. and the worst individual The main loop iteration process begins: In each iteration, the algorithm calculates the adaptive parameters. and auxiliary variables and And based on this, candidate individuals are calculated. , and To update the population position. By judging the random number and decision factor. The relative sizes of the individuals determine whether to use the NRSR rule or the TAO rule for individual updates, and a guided learning strategy is incorporated to further enhance the search capability. Finally, the better individuals are selected to participate in the next iteration and the extreme values ​​are updated until the maximum number of iterations is met. Given the condition, output the globally optimal individual. As the optimal parameters for variational mode decomposition.

[0025] like Figure 3As shown, the process is divided into two stages: In stage one (offline model training), the power grid operating equipment sends historical datasets to the VMD noise reduction and optimization module. After the parameters are optimized by the NRFO algorithm and the guided learning strategy is applied, the optimal data sequence is generated and input into the GAT-BiLSTM model for feature extraction and training. Finally, the trained model is deployed to the energy management system. In stage two (online real-time control), the energy management system collects power grid operating data in real time, calls the deployed model to perform inference calculations, and feeds back the obtained optimal reactive power allocation results to the power grid operating equipment, thereby realizing closed-loop intelligent control of the power grid reactive power.

[0026] In this embodiment of the application, the power grid operation data in step S100 includes real-time electrical status data such as voltage amplitude and phase angle of each node, as well as equipment parameter data. When constructing the dataset, the collected data is labeled and divided into a training set of 80% and a test set of 20%, providing a standardized data foundation for subsequent model training.

[0027] In this embodiment of the application, the noise reduction process in step S100 decomposes the original signal into a series of signals with different center frequencies. modal components To achieve this, the core is to solve a constrained variational problem, namely, minimizing the sum of the estimated bandwidths of each modal component while ensuring that the sum of all modal components equals the original input signal. The mathematical model of this constrained variational problem can be expressed as: in, denoted by , where represents the center frequency of the k-th sub-signal, K is the total number of sub-signals decomposed, min indicates finding the minimum value, and st represents the constraint condition. This indicates taking the partial derivative with respect to time t. For the Dirac function, The imaginary unit, Pi The square operation represents the L2 norm. It represents the k-th modal component.

[0028] By introducing Lagrange multipliers and secondary penalty factor The constrained problem is transformed into an unconstrained augmented Lagrangian function problem for solution. The specific expression of the augmented Lagrangian function is as follows: in, Let Lagrange multipliers be the functions of the Lagrange multipliers. Let be the quadratic penalty factor, and ⟨·,·> represent the inner product operation. The above mathematical model is used for iterative solution, thereby effectively extracting key features and suppressing non-stationary noise. In this embodiment of the application, the parameters of the variational mode decomposition in step S100 are optimized by an improved Newton-Raphson optimizer, including: The objective function for optimization is constructed based on minimizing data envelopment entropy, and the variational mode decomposition parameters to be optimized are determined as the penalty factor and the number of decomposition layers. In an optional implementation, the data denoising and decomposition method in step S100 can also use Discrete Wavelet Transform (DWT) to perform multi-scale decomposition on the original power grid data, process high-frequency coefficients by setting a threshold to filter out noise, and finally perform wavelet reconstruction to obtain the denoised data sequence.

[0029] In an optional implementation, the data denoising and decomposition method in step S100 can also use empirical mode decomposition (EMD) to adaptively decompose the original power grid data into several intrinsic mode functions (IMFs), eliminate noise components by calculating the correlation coefficient between each IMF component and the original signal, and reconstruct the remaining effective components by linear superposition.

[0030] The objective function is solved using the Newton-Raphson search rule, and the derivative of the function is approximated using Taylor expansion to determine the search direction. For the objective function g(x), the search and update process of the standard Newton-Raphson method can be expressed as: in, Indicates the first The variable value after the next iteration. This represents the current value of the variable in the nth iteration. This represents the first derivative of the objective function g(x) at the current point. This represents the second derivative of the objective function g(x) at the current point.

[0031] This formula enhances the algorithm's global exploration capability and accelerates convergence by moving towards the next position in the gradient direction from an assumed initial value. To employ this search process, its second derivative needs to be determined using the Taylor expansion, specifically expressed as: in, and Representing higher-order terms, we can derive the expressions for its first and second derivatives as follows: The search process can then be transformed into: in, Represents a small increment of a variable. and This represents a higher-order infinitesimal term in a Taylor series expansion.

[0032] Specifically, the first derivative is derived using Taylor expansion. With the second derivative The difference approximation expression is obtained, and then a search update formula based on the fitness differences of individuals in the population is constructed; during the update process, an adaptive parameter is introduced. The algorithm balances the development and exploration capabilities at different iteration stages and utilizes a guiding term that includes a random factor. Guide the population to move towards areas of high fitness to avoid getting trapped in local optima; In the process of iteratively optimizing the parameters based on the objective function, the search direction is calculated using the Newton-Raphson search rule, and the individual with the best fitness value and the individual with the worst fitness value in the population are respectively replaced with the adjacent positions used to determine the search direction in the Newton method iterative formula to update the position of the population individuals and iterate cyclically until the convergence condition is met.

[0033] To achieve population-based search, the Newton-Raphson optimizer uses the best individual in the population. Replace the individual with the better fitness in the original adjacent position with the worst individual in the population. The detailed expression of the Newton-Raphson Search Rule (NRSR) for replacing individuals with poor fitness is as follows: in, This represents the worst individual in the population. This represents the best individual in the population. Indicates the current iteration number. Indicates the currently selected individual is in the [number]th position. The position at the next iteration, r is a random number between 0 and 1, and mean represents the average value. This indicates that random actions are considered during the optimization process to improve the Newton-Raphson optimizer and avoid local optima; Z is calculated as follows: After determining the search direction, individuals are further updated by combining gradient information and population diversity: in, This indicates the individual's new position after the update. , , As an intermediate transitional variable, and For two distinct individuals randomly selected from the population, This represents the maximum number of iterations, and δ is an adaptive parameter used to balance the algorithm's development and exploration capabilities; parameters This guides the population to move in the correct direction, and its expression is as follows: Where a and b are random numbers between 0 and 1. and These are two different individuals randomly selected from the population.

[0034] In this embodiment, a trap avoidance operator (TAO) is also introduced. By combining the current optimal solution, population mean information, and random perturbation to generate a new solution, the population diversity is further enhanced. The core objective of TAO is to avoid the algorithm getting trapped in local optima. The specific expression for generating a new solution is as follows: in, This represents the new solution generated after introducing the trap avoidance operator. and It is a random number between 0 and 3. and These are random numbers ranging from -1 to 1 and -0.5 to 0.5, respectively. Represents all individuals in the population. This represents the mean position of individuals in the current population; TAO enhances population diversity and avoids premature convergence by using random perturbations and mean information.

[0035] In an optional implementation, the hyperparameter optimization algorithm in step S100 can also use the VMD penalty factor and the number of decomposition layers as particle positions, and the data envelopment entropy as the fitness function, to find the parameter combination that optimizes fitness by simulating the cooperation and iterative updates of swarm intelligence in the search space.

[0036] In an alternative implementation, the hyperparameter optimization algorithm in step S100 can also construct a Gaussian process surrogate model to predict the objective function (data envelopment entropy), and use acquisition functions (such as desired improvement or upper confidence bound) to balance exploration and development, and find the optimal combination of hyperparameters for VMD with fewer iterations by iteratively evaluating and updating the model.

[0037] Furthermore, the final individual update method is determined and controlled in conjunction with decision factor (DF), specifically as follows: in, This indicates the position of the individual that will ultimately participate in the next iteration, and DF is the threshold for decision factors.

[0038] In this embodiment of the application, in step S100, the optimizer determines the population dispersion by calculating the root mean square error of historical positions and introduces a guided learning strategy, including: The mean squared error of the historical position of individuals in the population is calculated, and the empirical feedback parameters are obtained by combining the normalization coefficient. Compare the numerical values ​​of the empirical feedback parameters with the parameters to be designed. Perform the corresponding position update operation based on the comparison results; If the empirical feedback parameter is greater than the parameter to be designed, then the new individual position is calculated based on the optimal individual position of the population, combined with the tangent function value and the boundary value of the search space. If the empirical feedback parameter is not greater than the parameter to be designed, then new individual positions are randomly generated within the search space boundary. The position update calculation formula for the guided learning strategy is: in,, and Represents the upper and lower limits of an individual in a population. These are the parameters to be designed. This is an empirical feedback parameter, and its calculation formula is: Where Et represents the learning experience, std represents the function for calculating the standard deviation, and B is used to normalize the experience feedback parameters to prevent them from being affected by changes in the upper and lower limits.

[0039] Ultimately, this will guide the generation of individual learning strategies. Individuals generated with NRSR or TAO Perform fitness comparisons, select individuals with low fitness values ​​to participate in the next round of iterations, until the convergence condition is met or the maximum number of iterations is reached, and output the optimal parameter combination.

[0040] After obtaining the optimal penalty factor α and the number of decomposition levels K, variational mode decomposition (VMD) is performed on the original power grid operation dataset using this optimal parameter combination to obtain K modal components.

[0041] To filter out the effective components containing key features and remove noise from the K modal components, this embodiment calculates the data envelopment entropy of each modal component. Since noisy signals typically have higher complexity and uncertainty, their data envelopment entropy values ​​are relatively high.

[0042] A preset entropy threshold is set. Modal components with data envelopment entropy greater than the threshold are identified as noise components and removed. Modal components with data envelopment entropy less than or equal to the threshold are retained as valid components.

[0043] All retained effective components are linearly superimposed and reconstructed to obtain the optimal data sequence after noise removal, and this optimal data sequence is then input into the subsequently constructed GAT-BiLSTM model for training.

[0044] In this embodiment of the application, step S200, which involves training the constructed graph attention network-bidirectional long short-term memory network model by inputting the optimal data sequence, includes: The optimal input data sequence is normalized to improve the training speed of the model. The normalization process uses the min-max normalization method to map the data to a preset interval. The specific calculation formula is as follows: in, This represents the normalized data, and X represents the original input data. and These represent the minimum and maximum values ​​in the original dataset, respectively.

[0045] In this embodiment of the application, step S200, which involves capturing the spatial dependencies between nodes through a multi-head graph attention layer, includes: Perform a linear transformation on the input node features to obtain the transformed node features; Based on the transformed node features, the original attention scores between nodes are calculated using a feedforward neural network. The original attention scores are normalized using a normalized exponential function to obtain attention weights; The attention weights are used to perform a weighted summation of the feature information of neighboring nodes, and the node feature information is updated through a nonlinear activation function.

[0046] In specific calculations, the calculation steps for each single-head graph attention mechanism are as follows: Through the linear transformation matrix W and the parameters of the feedforward neural network Calculate the raw attention scores between nodes Furthermore, the LeakyReLU activation function is introduced to enhance nonlinear expressiveness, and its calculation method is as follows: in, Represents the parameters of the feedforward neural network transpose, This represents the feature vector of node i. For the parameters of the feedforward neural network, This represents a vector concatenation operation. This represents the characteristics of the neighbor node j after the transformation.

[0047] The original scores are normalized into attention weights using the Softmax normalization function. To ensure that the sum of weights within the neighborhood is 1, thus stabilizing model training, the calculation formula is as follows: in, Represents the set of neighboring nodes of node i. Let represent the original attention score of node i to its neighbor node k.

[0048] The attention weights are used to perform a weighted summation of the transformed features of neighboring nodes, and then processed by the ReLU activation function to obtain the updated node features. The calculation formula is as follows: In an optional implementation, the spatial feature extraction network in step S200 can also construct an adjacency matrix based on the power grid topology, perform spectral domain convolution operations on node features through graph convolution layers, and aggregate the feature information of neighboring nodes layer by layer to update the current node representation, thereby extracting the spatial dependency features between power grid nodes.

[0049] In an optional implementation, the spatial feature extraction network in step S200 can also use graph embedding technology (such as DeepWalk) to map the spatial structure of the power grid topology into a low-dimensional node feature matrix, and then combine it with a one-dimensional convolutional neural network (1D-CNN) or a recurrent neural network to extract local spatial and temporal features. Finally, the spatiotemporal information is fused through a fully connected layer.

[0050] In this embodiment of the application, step S200, which involves capturing the temporal characteristics of data using a bidirectional long short-term memory network, includes: The feature sequence output from the multi-head graph attention layer is input into a bidirectional long short-term memory network; By utilizing two independent long short-term memory (LSTM) layers in a bidirectional long short-term memory network, the feature sequences are processed from both forward and backward directions to capture contextual information in the time-series data. The bidirectional structure of BiLSTM excels at capturing complex temporal characteristics in power grid operation data, thereby obtaining a more coordinated and globally optimal reactive power compensation strategy, effectively improving the stability, economy, and renewable energy integration capabilities of the power grid.

[0051] In this embodiment, the Bidirectional Long Short-Term Memory (BiLSTM) network includes an input layer, a bidirectional LSTM hidden layer, and a fully connected output layer. The bidirectional LSTM hidden layer is used to extract time-series features of the input data from both forward and backward directions and then concatenate the features. To ensure the training effect of the model and suppress overfitting, this embodiment sets the number of hidden layer units of the bidirectional LSTM hidden layer to 64 and the dropout rate of the Dropout layer to 0.2. It should be noted that the above-mentioned number of hidden layer units and Dropout rate are only illustrative. Those skilled in the art can adaptively adjust the above parameters according to the scale of actual power grid operation data and the convergence situation during the training process. This application does not impose specific limitations in this regard.

[0052] In an optional implementation, in step S200, the time series prediction model can also normalize the input power grid time series data and construct a sequence using a sliding window. It can also assign temporal position information to the sequence through linear projection and position encoding, use a multi-layer self-attention mechanism to calculate the dependencies between each time point in the sequence in parallel to capture long-distance temporal features, and finally decode and output the prediction result through a fully connected layer.

[0053] In an optional implementation, in step S200, the time series prediction model can also input the preprocessed power grid time series data into the GRU network, control the forgetting degree of historical information by resetting the gate to capture short-term dependencies, and control the retention ratio of historical states by updating the gate to capture long-term time series dependencies, and finally output the extracted time series features.

[0054] In this embodiment of the application, step S300, which adjusts the reactive power distribution of the power grid operation based on the analysis results, includes: The trained reactive power optimization model is deployed to the power grid dispatch automation energy management system in the power grid. An energy management system equipped with the aforementioned model is used to collect real-time grid operation data and input the operation data into the model for inference calculations to analyze the reactive power demand and voltage status of the grid in real time.

[0055] Furthermore, after obtaining the optimal reactive power allocation result from the model output, the system can assist dispatchers in making manual adjustments based on the analysis results, or the energy management system can automatically issue instructions to achieve precise optimization of reactive power allocation in power grid operation.

[0056] In this embodiment, the reactive power optimization model further includes a splicing layer, a fully connected layer, and an output layer sequentially connected after the BiLSTM layer. The forward and backward time-series features output by the BiLSTM layer are spliced ​​dimensionally in the splicing layer. The spliced ​​feature vector is then input to the fully connected layer for feature mapping and nonlinear transformation. Finally, the output layer outputs the reactive power optimization control strategy. The output layer is configured as a Softmax layer or a linear regression layer depending on the specific optimization task type. When the reactive power optimization task is a discrete-level adjustment classification task, the output layer uses a Softmax layer to output the probability distribution of each control level. When the reactive power optimization task is a continuous reactive power regression task, the output layer uses a linear regression layer to output the specific reactive power compensation value.

[0057] Example 3 is an embodiment of the present invention. This embodiment differs from the first embodiment in that it provides a smart control system for reactive power optimization in a power grid.

[0058] It should be noted that the technical solution of the intelligent control system for reactive power optimization of the power grid is based on the same concept as the technical solution of the intelligent control method for reactive power optimization of the power grid described above. For details not described in detail in the technical solution of the intelligent control system for reactive power optimization of the power grid in this embodiment, please refer to the description of the technical solution of the intelligent control method for reactive power optimization of the power grid described above.

[0059] This embodiment of a power grid reactive power optimization intelligent control system includes: The preprocessing module is used to collect power grid operation data to construct a dataset and to perform noise reduction on the dataset using variational mode decomposition. The parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer. The optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy to obtain the optimal data sequence after noise removal. The construction and training module is used to input the optimal data sequence into the constructed graph attention network-bidirectional long short-term memory network model for training. It captures the spatial dependencies between nodes through the multi-head graph attention layer and captures the temporal features of the data through the bidirectional long short-term memory network to obtain the trained reactive power optimization model. The real-time reactive power optimization and control module is used to deploy the trained reactive power optimization model to the power grid, collect and analyze power grid operation data in real time, and adjust the reactive power distribution of the power grid operation based on the analysis results.

[0060] This embodiment also provides an electronic device applicable to a smart control method for reactive power optimization in a power grid, including: The system includes a memory and a processor. The memory stores computer-executable instructions, and the processor executes these instructions to implement a smart control method for reactive power optimization in a power grid, as proposed in the above embodiments.

[0061] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, it implements a power grid reactive power optimization and intelligent control method as proposed in the above embodiments.

[0062] The storage medium proposed in this embodiment belongs to the same inventive concept as the method for implementing intelligent control of reactive power optimization in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0063] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0064] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for intelligent control of reactive power optimization in power grids, characterized in that, include: Data sets are constructed by collecting power grid operation data, and noise reduction is performed on the dataset using variational mode decomposition. The parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer. The optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy to obtain the optimal data sequence after noise removal. The optimal data sequence is input into the constructed graph attention network-bidirectional long short-term memory network model for training. The spatial dependencies between nodes are captured by the multi-head graph attention layer, and the temporal features of the data are captured by the bidirectional long short-term memory network to obtain the trained reactive power optimization model. The trained reactive power optimization model is deployed to the power grid, and real-time power grid operation data is collected and analyzed. The reactive power allocation of the power grid is adjusted based on the analysis results.

2. The intelligent control method for reactive power optimization of power grids as described in claim 1, characterized in that: The parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer, including: The objective function is constructed based on minimizing data envelopment entropy, and the variational mode decomposition parameters to be optimized are determined as the penalty factor and the number of decomposition layers. In the process of iteratively optimizing the parameters based on the objective function, the search direction is calculated using the Newton-Raphson search rule, and the individual with the best fitness value and the individual with the worst fitness value in the population are respectively replaced with the adjacent positions used to determine the search direction in the Newton method iterative formula to update the position of the population individuals and iterate cyclically until the convergence condition is met.

3. The intelligent control method for reactive power optimization of power grids as described in claim 1 or 2, characterized in that: The optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy, including: The mean squared error of the historical position of individuals in the population is calculated, and the empirical feedback parameters are obtained by combining the normalization coefficient. Compare the numerical values ​​of the empirical feedback parameters with the parameters to be designed. Perform the corresponding position update operation based on the comparison results; In response to the empirical feedback parameter being greater than the design parameter, a new individual position is calculated based on the optimal individual position of the population, combined with the tangent function value and the search space boundary value. In response to the empirical feedback parameter not being greater than the parameter to be designed, new individual positions are randomly generated within the search space boundary.

4. The intelligent control method for reactive power optimization of power grids as described in claim 3, characterized in that: The step of training the graph attention network-bidirectional long short-term memory network model by inputting the optimal data sequence includes: The optimal input data sequence is normalized using min-max normalization to map the data to a preset interval.

5. The intelligent control method for reactive power optimization in power grids as described in claim 4, characterized in that: The method of capturing spatial dependencies between nodes through a multi-head graph attention layer includes: Perform a linear transformation on the input node features to obtain the transformed node features; Based on the transformed node features, the original attention scores between nodes are calculated using a feedforward neural network. The original attention scores are normalized using a normalized exponential function to obtain attention weights; The attention weights are used to perform a weighted summation of the feature information of neighboring nodes, and the node feature information is updated through a nonlinear activation function.

6. The intelligent control method for reactive power optimization of power grids as described in claim 5, characterized in that: The method of capturing the temporal characteristics of data through a bidirectional long short-term memory network includes: The feature sequence output from the multi-head graph attention layer is input into a bidirectional long short-term memory network; The feature sequence is processed from both forward and backward directions using two independent long short-term memory network layers in a bidirectional long short-term memory network to capture contextual information in time series data.

7. The intelligent control method for reactive power optimization in power grids as described in claim 6, characterized in that: The adjustment of reactive power distribution in power grid operation based on analysis results includes: The trained reactive power optimization model is deployed to the power grid dispatch automation energy management system in the power grid. The energy management system equipped with the model collects grid operation data in real time, inputs the operation data into the model for inference calculation, and regulates the grid based on the optimal reactive power allocation result output by the model.

8. A smart control system for reactive power optimization in a power grid, employing the method described in any one of claims 1-7, characterized in that, include: The preprocessing module is used to collect power grid operation data to construct a dataset and to perform noise reduction on the dataset using variational mode decomposition. The parameters of the variational mode decomposition are optimized by an improved Newton-Raphson optimizer. The optimizer determines the population dispersion by calculating the root mean square error of historical locations and introduces a guided learning strategy to obtain the optimal data sequence after noise removal. The construction and training module is used to input the optimal data sequence into the constructed graph attention network-bidirectional long short-term memory network model for training. It captures the spatial dependencies between nodes through the multi-head graph attention layer and captures the temporal features of the data through the bidirectional long short-term memory network to obtain the trained reactive power optimization model. The real-time reactive power optimization and control module is used to deploy the trained reactive power optimization model to the power grid, collect and analyze power grid operation data in real time, and adjust the reactive power distribution of the power grid operation based on the analysis results.

9. An electronic device, characterized in that, include: Memory, used to store programs; A processor for loading the program to perform the steps of the method as claimed in any one of claims 1-7.

10. A computer-readable storage medium storing a program, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1-7.

Images