Power distribution network state estimation model precision-generalization capability coordinated optimization method based on evolutionary intermediate domain

Abstract

The invention relates to a power distribution network state estimation model precision-generalization capability coordinated optimization method based on an evolutionary intermediate domain. The method specifically comprises the following steps: considering the measurement precision and time scale of multi-source data and the synchronism of state estimation calculation, and respectively preprocessing day-ahead and intra-day measurement data; the method comprises the following steps: taking measurement data of distributed power supply output and power distribution network operation as a total data set for training, and establishing a power distribution network multi-layer output single-hidden-layer state estimation initial model by adopting a semi-supervised learning method; on the basis of an SSL method, a pseudo marking method is proposed, a state estimation initial model is adopted to add pseudo labels for unmarked measurement data, pseudo measurement data is generated, and the problem that training model data is less is solved; designing a progressive antagonism pseudo-mark model to obtain a power distribution network SE model with relatively high precision; proposing a double-calibration generalization model; and proposing a power distribution network SE model precision-generalization capability coordination optimization method considering the evolution intermediate domain.

Description

Technical Field

[0001] The present invention relates to a coordinated optimization method for the accuracy and generalization capability of a distribution network state estimation model based on an evolutionary intermediate domain. Background Art

[0002] Building a new power system with clean energy as the main body is the foundation for the sustainable development of future society. The power industry must actively transform the way of energy production and consumption, optimize the energy structure, and increase the development of renewable energy. When a high proportion of distributed power sources represented by photovoltaic and wind power are connected to the distribution network, due to their inherent intermittent and volatile characteristics, it increases the difficulty of distribution network state estimation and optimal scheduling.

[0003] Specifically, the high proportion of distributed power access will lead to the following problems:

[0004] 1) The output of distributed power sources is easily affected by weather and human factors, and the large fluctuations increase the difficulty of prediction;

[0005] 2) Insufficient measurement level and insufficient measurement accuracy lead to missing historical data or increased errors;

[0006] 3) Traditional distribution network models are slow to solve and have poor ability to handle high penetration scenarios;

[0007] 3) It is difficult to summarize a probability model from massive data. It needs to be analyzed and modeled based on a large amount of historical sampling data. However, due to the insufficient measurement level of the distribution network, there are often problems such as missing data and insufficient measurement accuracy. The true value of the parameters is uncertain, making it difficult to directly establish an accurate probability model.

[0008] 4) It is difficult to ensure both the accuracy and generalization ability of the distribution network state estimation model.

[0009] In the face of these technical shortcomings of existing technologies, it is necessary to design a new solution to meet actual usage needs. Summary of the invention

[0010] The present invention takes into account that source-load uncertainty and communication unreliability of multi-source measurement data seriously affect the accuracy and generalization ability of the distribution network state estimation model. A new cyclic learning framework is adopted to substitute the measurement data set and the pseudo-measurement data set into the ACN and DCDG modules respectively to screen out clean data that meet the two modules, generate interpolation of the clean set of measurement data and pseudo-measurement data, and mix it with the original data set to obtain an evolvable intermediate domain to obtain the optimal gap between the measurement data set and the pseudo-measurement data set, thereby improving the SE accuracy of the distribution network while ensuring the generalization ability of the SE model.

[0011] The present invention adopts the following technical means to achieve: a coordinated optimization method of distribution network state estimation model accuracy-generalization ability based on evolutionary intermediate domain, comprising the following steps:

[0012] Step 1: Considering the measurement accuracy, communication error, time scale and synchronization of state estimation calculation of multi-source data, pre-process the measurement data of the day before and within the day respectively;

[0013] Step 2: Using the output of distributed generation and the measured data of distribution network operation as the training data set, a semi-supervised learning (SSL) method is used to establish the initial model of distribution network multi-layer output single hidden layer state estimation;

[0014] Step 3: Based on the SSL method, a pseudo labeling (APL) method is proposed. The state estimation initial model is used to add pseudo labels to the unlabeled measurement data to generate pseudo measurement data, thus solving the problem of insufficient training model data.

[0015] Step 4: A progressive adversarial pseudo-labeling model (ACN) is designed, which includes a pseudo-labeling model and a discriminator. The accuracy of pseudo-measurement data is improved through the mutual confrontation between the pseudo-labeling model and the discriminator, the total data set is updated, and a higher-precision distribution network SE model is obtained through further training.

[0016] Step 5: To avoid the overfitting problem of the training model, a dual calibration generalization model with style confusion training and label diversity regularization is proposed to improve the generalization performance of the SE training model;

[0017] Step 6: In order to balance the contradiction between the accuracy and generalization ability of the distribution network state estimation model, a distribution network SE model accuracy-generalization ability coordinated optimization method considering the evolutionary intermediate domain is proposed.

[0018] In step 1, the measurement accuracy, time scale and synchronization of state estimation calculation of multi-source data are considered, and the measurement data of the day before is preprocessed as follows:

[0019] Before using the distribution network measurement data for state estimation, the data needs to be preprocessed. Calculate ν measurement values ​​respectively Mean μ and standard deviation σ, and analyze the νth measurement data separately Does it satisfy the formula Where M is the bad data discrimination threshold. If the above formula is satisfied, the measured data is bad data, and finally the average value of the remaining data is taken as the measured data at the SE moment.

[0020] In step 2, the output of distributed power sources and the measured data of distribution network operation are used as the training data set, and the semi-supervised learning method is used to establish the initial model of distribution network multi-layer output single hidden layer state estimation:

[0021] First, assume that a set of original data samples {x Θ ,y Θ},in is the input vector of the model, represents the predicted target vector, l is the measurement number, and r is the dimension of the data; the output of a single hidden layer neural network of hidden nodes is expressed as:

[0022]

[0023] In the formula, x Θ is the input vector, λ j is the weight value of the model; ω j is the target vector y Θ The weight of j is the bias of the jth hidden layer node; g(·) is the activation function tanh(·) based on hyperbolic tangent. Compared with the traditional sigmoid(·) activation function, tanh(·) converges faster;

[0024] The core goal of neural network optimization is to minimize the actual output c j and the target value y Θ The error between them makes the following equation true:

[0025]

[0026] Therefore, as long as the weight value λ corresponding to the given input vector j , and the hidden layer unit bias value b j , then the training of a single hidden layer neural network can be transformed into solving a linear system, and the corresponding weight value ω of the output vector to be solved j It can be obtained by solving the above τ linear equations;

[0027] In the distribution network state prediction, each set of data includes the node injection power, branch power and the corresponding system voltage amplitude and phase angle; the possible operating state of the estimated system is determined by the interval information of the measurement and a set of mapping relationships. The mathematical relationship can be expressed as:

[0028]

[0029] Where E(·) is the interval set of the state variables to be determined; h(·) is the nonlinear mapping function between the state variables and the model input; and Z(·) is the pseudo-measurement uncertainty set of the input.

[0030] Step 3: Based on the semi-supervised learning method, a pseudo-labeling method is proposed. The state estimation initial model is used to add pseudo-labels to the unlabeled measurement data to generate pseudo-measurement data, solving the problem of less training model data:

[0031] Define distribution network measurement and unlabeled measurement datasets:

[0032]

[0033] In the formula, x and y are scalars; Measure data for the system; It is an unlabeled measurement data set and has no corresponding output;

[0034] The pre-processed day-ahead and day-intraday labeled measurement data are trained to obtain the distribution network SE initial model; the unlabeled measurement data are substituted into the trained initial model and the APL method is used to obtain the distribution network SE initial model. Add pseudo-label Q to generate pseudo-measurement data That is, generate fake data that is as close to the real data distribution as possible.

[0035] In step 4, the APL model adds pseudo labels to the unlabeled data to deceive the discriminator, and the discriminator is updated to become more sensitive to distinguish between real and fake data. This adversarial training process implements the model by optimizing the loss function of the two models.

[0036] In step 4, the loss function of the progressive adversarial pseudo-labeling model is:

[0037]

[0038] In the formula, Measure data for the system; After being processed by the encoder A() is the defined function; for The posterior distribution of ; for The mathematical expectation of

[0039] The loss function is:

[0040]

[0041] In the formula, ACN is a progressive pseudo-labeling model; for The generated samples of for The mathematical expectation of for The mathematical expectation of

[0042] The loss function is:

[0043]

[0044] In the formula, It is pseudo measurement data; After being processed by the encoder for The generated samples of for The mathematical expectation of for The mathematical expectation of

[0045] According to the loss function, progressive training is performed so that the game between the generator and the discriminator gradually reaches an equilibrium state. The objective function for the progressive adversarial pseudo-labeling model to reach an equilibrium state is defined as the minimum deviation between the loss function of the pseudo-measurement data and the measurement data:

[0046] minf1=(T3-T2)-T1(9)

[0047] In equilibrium, the generator is able to generate data realistic enough to fool the discriminator, while the higher-precision discriminator has difficulty distinguishing between real and fake data.

[0048] In step 5, in order to avoid the overfitting problem of the training model, a dual calibration generalization model with style confusion training and label diversity regularization is proposed to improve the generalization performance of the SE training model;

[0049] The dual calibration model uses a low-cost style confusion training method. A style confuser is inserted in each subnet to confuse the features between the measured data and the pseudo-measurement data. The measured and pseudo-measurement data samples are divided into N sample sets, and the mean and variance are calculated respectively:

[0050]

[0051] Randomly shuffle the sample data in the sample set dimension to obtain a new mean and variance In order to obtain different characteristics of the sample set, {μ,σ} and The linear expression of the relationship can obtain the transformation parameters {β,γ}:

[0052] Use transformation parameters to transfer the original features to the new space:

[0053]

[0054] In the formula, ∈ is a fixed constant that can be ignored, and They are the features of each sample set before and after the conversion. Feature confusion can enrich the style of the feature layer.

[0055] Label diversity regularization is used for the remaining data, aiming to improve the generalization ability and robustness of the model by increasing the diversity and uncertainty of labels. Maximizing the global entropy of label diversity can achieve model generalization by reducing assumptions and biases, balancing label distribution, avoiding overfitting, and enhancing the robustness of the model. Therefore, the ultimate goal of the generalized model can be summarized as follows:

[0056]

[0057] In the formula, is the clean set in the pseudo-measurement data, and These are clean samples selected from subnet 1 and subnet 2 respectively; Classification loss of pseudo-measurements; To maximize the global entropy of label diversity, label diversity regularization is achieved.

[0058] In step 6, in order to balance the contradiction between the accuracy and generalization ability of the distribution network state estimation model, a distribution network SE model accuracy-generalization ability coordinated optimization method considering the evolutionary intermediate domain is proposed:

[0059] Through a new cyclic learning framework, positive feedback of the progressive adversarial pseudo-labeling model and the double-calibration generalization model is realized, and a cyclically updateable intermediate domain is added to connect the measurement dataset and the pseudo-measurement dataset and gradually adjust the gap between the datasets to achieve the ideal effect.

[0060] The proposed cyclic learning framework includes a progressive adversarial pseudo-labeling model and a dual calibration generalization model with style confusion and label diversity regularization. The measurement dataset and the pseudo measurement dataset are respectively substituted into the progressive adversarial pseudo-labeling model and the dual calibration generalization model to filter out clean data that conform to the two models, generate interpolations of the clean sets of the measurement data and pseudo measurement data, and mix them with the original dataset to obtain an evolvable intermediate domain to adjust the gap between the measurement dataset and the pseudo measurement dataset.

[0061] By repeating the above process, the optimal gap between the measurement data set and the pseudo-measurement data set can be obtained, ensuring that the distribution network state prediction model has high accuracy and strong generalization ability.

[0062] The method requires the use of a pseudo-labeling module, a discriminator module, a progressive adversarial pseudo-labeling module, a dual-calibration generalization module, and a recurrent learning module.

[0063] The pseudo-labeling module uses the state estimation initial model to add pseudo-labels to the unlabeled measurement data, generating pseudo-measurement data to solve the problem of insufficient training model data.

[0064] The discriminator module is used to determine whether the pseudo measurement data generated by the pseudo label (APL) module meets the accuracy requirements;

[0065] The progressive adversarial pseudo-labeling module has: a containing module and a discriminator module, realizing the mutual confrontation between the pseudo-labeling module and the discriminator module to improve the accuracy of the pseudo-measurement data, update the total data set, and further train to obtain a higher-precision distribution network SE model;

[0066] Dual calibration generalization module: includes style confusion training and label diversity regularization to improve the generalization performance of SE training models;

[0067] Recurrent learning module with evolving intermediate domain: The recurrent learning framework includes ACN module and DCDG module. The measurement data set and pseudo-measurement data set are respectively substituted into the ACN module and DCDG module to filter out clean data that meet the requirements of the two modules, generate interpolation of clean sets of measurement data and pseudo-measurement data, and mix them with the original data set to obtain an evolving intermediate domain to obtain the optimal gap between the measurement data set and the pseudo-measurement data set, thereby improving the accuracy of the distribution network SE model while ensuring its generalization ability.

[0068] A device used in a coordinated optimization method for the accuracy-generalization capability of a distribution network state estimation model based on an evolving intermediate domain includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the coordinated optimization method for the accuracy-generalization capability of a distribution network state estimation model based on an evolving intermediate domain is implemented.

[0069] The beneficial effects of the present invention after adopting the above technical solution are:

[0070] 1) The ACN method proposed in this invention realizes the mutual confrontation between the APL module and the discriminator, gradually improving the accuracy of the SE model of the distribution network, which is significantly improved compared with the SE model based on DNN;

[0071] 2) Based on the ACN method, the DCDG method proposed in the present invention uses style confusion training and label diversity regularization method to increase label diversity, and obtains different decision boundaries through the two subnets in the module, which greatly improves the generalization ability of the SE model;

[0072] 3) The measurement data set and the pseudo-measurement data set are substituted into the ACN and DCDG modules respectively to filter out the clean data that meet the requirements of the two modules, generate interpolation of the clean sets of measurement data and pseudo-measurement data, and mix them with the original data set to obtain an evolvable intermediate domain to adjust the optimal gap between the measurement data set and the pseudo-measurement data set, which can ensure that the distribution network state prediction model has high accuracy and strong generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] The above and / or additional aspects and advantages of the present invention will become apparent and easily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0074] Figure 1 Shown is a schematic diagram of the workflow of the coordinated optimization method for the accuracy-generalization ability of the distribution network state estimation model based on the evolving intermediate domain.

[0075] Figure 2 The figure shows the configuration of active distribution network measurement device and its application.

[0076] Figure 3 Shown is the optimization flow chart of the distribution network state prediction model.

[0077] Figure 4 Shown is the generalization of double calibration with feature confusion training and label diversity regularization.

[0078] Figure 5 Shown is a recurrent learning framework with an evolving intermediate domain.

[0079] Figure 6 The figure shows the resource configuration diagram of IEEE 33-node system.

[0080] Figure 7 Shown are the daily load curve, daily operation curves of wind turbines and photovoltaic cells.

[0081] Figure 8 The figure shows the resource configuration diagram of the IEEE 118 node system.

[0082] Fig. 9 The figure shows a comparison of the average accuracy of pseudo labels generated by different methods.

[0083] Fig.10 The figure shows the comparison of the generalization ability and pseudo-label accuracy of different method models.

[0084] Fig.11 The figure shows the comparison of IEEE 33 system state estimation results using different methods.

[0085] Fig.12 The figure shows the comparison of IEEE 118 system state estimation results using different methods.

[0086] Fig.13 The figure shows the average error comparison of state estimation for IEEE 33 nodes with different noise interference conditions.

[0087] Fig.14 The figure shows the average error comparison of IEEE 118 node state estimation under different noise interference conditions. DETAILED DESCRIPTION

[0088] To make the purpose, technical solution and advantages of the embodiments of the present invention more clearly described, the following will be combined with the appended drawings of the embodiments of the present invention. Figure 1-14 , the technical solutions in the embodiments of the present invention are clearly and completely described. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0089] The above-mentioned coordinated optimization method of distribution network state estimation model accuracy and generalization ability based on the evolutionary intermediate domain includes the following process:

[0090] First, the current status of distribution network data measurement and data preprocessing

[0091] 1. Current status of distribution network data measurement

[0092] The goal of distribution network state estimation is to estimate the state of the distribution system with high precision, completeness and reliability by using appropriate algorithms based on the measurement information of the distribution system and supplemented by pseudo-measurement data. With the intelligent construction of active distribution networks, the distribution network information integration platform has realized the integration, interaction, analysis and application of various data sources. In addition to the data of DTU and AMI, the measurement data used by the active distribution network SE of the present invention also adds distributed power management system data. The grid-connected monitoring data of distributed power sources can improve the observability of the distribution network and provide strong support for the realization of high-precision state estimation and other advanced applications.

[0093] However, the lack of reliability of distribution network communication will profoundly affect the accurate capture and transmission efficiency of real-time data, thereby weakening the accuracy of the state prediction model, delaying the timeliness of fault identification and location, and posing a challenge that cannot be ignored to the overall stability and security of the distribution network. In view of this, effective processing of distribution network measurement data has become an indispensable basis for ensuring the efficient, stable and safe operation of the distribution network, and has far-reaching significance for improving the level of intelligent management of the power grid.

[0094] Configuration and application of active distribution network measurement devices Figure 2 shown.

[0095] 2. Data preprocessing

[0096] In view of the bad data caused by measurement device collection deviation, communication transmission error, etc. under actual operating conditions, the state estimation method needs to detect, identify and eliminate the bad data before and after state estimation to achieve accurate estimation of the operating state. Before using the distribution network measurement data for state estimation, the data needs to be preprocessed. Calculate the ν measurement values ​​separately Mean μ and standard deviation σ, and analyze the νth measurement data separately Does it satisfy the formula Where M is the bad data discrimination threshold. If the above formula is satisfied, the measured data is bad data, and finally the average value of the remaining data is taken as the measured data at the SE moment.

[0097] Second, the establishment of distribution network state prediction model

[0098] The construction of the single hidden layer prediction model of the multi-layer output of the system first assumes that a set of original data samples {x Θ ,y Θ},in is the input vector of the model, represents the predicted target vector, l is the number of measurements, and r is the dimension of the data. The output of the single hidden layer neural network of the hidden node is shown in formula (1):

[0099]

[0100] In the formula, x Θ is the input vector, λ j is the weight value of the model; ω j is the target vector y Θ The weight of j is the bias of the jth hidden layer node; g(·) is the activation function tanh(·) based on hyperbolic tangent. Compared with the traditional sigmoid(·) activation function, tanh(·) converges faster.

[0101] The core goal of neural network optimization is to minimize the actual output c j and the target value y Θ The error between them, that is, formula (2) is established:

[0102]

[0103] Therefore, as long as the weight value λ corresponding to the given input vector j , and the hidden layer unit bias value b j , then the training of a single hidden layer neural network can be transformed into solving a linear system, and the corresponding weight value ω of the output vector to be solved j It can be obtained by solving the above τ linear equations.

[0104] In the distribution network state prediction, each set of data includes node injection power, branch power and the corresponding system voltage amplitude and phase angle. The possible operating state of the estimated system can be determined through the measured interval information and a set of mapping relationships. The mathematical relationship can be expressed as:

[0105]

[0106] Where E(·) is the interval set of the state variables to be determined; h(·) is the nonlinear mapping function between the state variables and the model input; and Z(·) is the pseudo-measurement uncertainty set of the input.

[0107] Third, progressive adversarial pseudo-labeling

[0108] In order to solve the problems of insufficient and inaccurate distribution network measurement data, the semi-supervised learning (SSL) method introduced in the existing research aims to use a small amount of distribution network measurement data to add pseudo labels to a large amount of unlabeled measurement data to generate a pseudo measurement data set. Although this method can solve the problem of less training model data, when using pseudo measurement data, it is necessary to fully consider its reliability, error accumulation, model dependence, nonlinear relationship processing, and introduction of false information. Therefore, the present invention proposes a pseudo labeling (APL) module to add pseudo labels to unlabeled measurement data.

[0109] Define distribution network measurement and unlabeled measurement datasets:

[0110]

[0111] In the formula, x and y are scalars; Measure data for the system; It is an unlabeled measurement dataset (no corresponding output).

[0112] The pre-processed day-ahead and day-intraday labeled measurement data are trained to obtain the distribution network SE initial model. The unlabeled measurement data are substituted into the trained initial model and passed through the APL module to obtain Add pseudo-label Q to generate pseudo-measurement data (Generate fake data that is as close as possible to the real data distribution).

[0113] However, the accuracy of pseudo-measurement data seriously affects the accuracy of model training. Therefore, the present invention designs a progressive adversarial pseudo-labeling module (ACN), which includes an APL module and a discriminator. The pseudo-measurement data is judged by the discriminator, so that the APL module and the discriminator can improve performance by confronting each other. The APL module adds pseudo-labels to unlabeled data to deceive the discriminator, and the discriminator becomes more sensitive after being updated to distinguish true and false data. This adversarial training process realizes the module by optimizing the loss function of the two modules.

[0114] The loss function of the ACN module is:

[0115]

[0116] In the formula, Measure data for the system; After being processed by the encoder for The posterior distribution of ; for The mathematical expectation of ; A() is the defined function.

[0117] The loss function is:

[0118]

[0119] Where ACN is a progressive pseudo-labeling module; for The generated samples of for The mathematical expectation of for The mathematical expectation of .

[0120] The loss function is:

[0121]

[0122] In the formula, It is pseudo measurement data; After being processed by the encoder for The generated samples of for The mathematical expectation of for The mathematical expectation of .

[0123] According to the loss function, progressive training is performed so that the game between the generator and the discriminator gradually reaches a balance state. The objective function for the ACN module to reach a balance state is defined as the minimum deviation between the loss function of the pseudo-measurement data and the measurement data:

[0124] minf1=(T3-T2)-T1 (9)

[0125] In equilibrium, the generator is able to generate data realistic enough to fool the discriminator, while the higher-precision discriminator has difficulty distinguishing between real and fake data.

[0126] Figure 3 To optimize the accuracy of the distribution network state estimation model, a flow chart is developed. The pseudo-measurement data is mixed with the original data, and the distribution network SE model is retrained using the new data set. The accuracy of the state estimation model is continuously updated until the trained system SE model can meet the prediction accuracy.

[0127] Third, generalization of distribution network state prediction model

[0128] Compared with the traditional pseudo-labeling technology used in semi-supervised learning, the ACN module proposed in the present invention uses a progressive adversarial method to add pseudo-labels to unlabeled measurement data. It shows better labeling ability for unlabeled measurement data and can improve the estimation accuracy of the system to a certain extent, but it is also prone to overfitting, that is, the model has high recognition accuracy on the training set, but low recognition accuracy on the test set. Therefore, it is particularly important to improve the generalization performance of the SE training model. The present invention proposes a dual-calibration generalization module with style confusion training and label diversity regularization, which consists of two identical subnetworks, each parameterized. Figure 4 Illustration of double-calibration generalization with style confusion training and label diversity regularization.

[0129] Feature confusion training aims to randomly initialize each subnet and provide the same number of random sample sets to extract sample features respectively. The features of different sample sets can be confused to enrich the style of the feature layer. Label diversity regularization aims to improve the generalization ability and robustness of the model by increasing the diversity and uncertainty of labels.

[0130] Each subnet is calibrated by the peer network by updating the clean samples selected by each other. The two subnets will eventually have different decision boundaries, which will increase their noise tolerance to avoid overfitting to noisy pseudo labels.

[0131] 1. Dataset style confusion

[0132] In order to improve the generalization ability of the dual calibration generalization module, a low-cost style confusion training method is used in the dual calibration module. A style confuser is inserted in each subnet to confuse the features between the measured data and the pseudo-measurement data. Specifically, the measured and pseudo-measurement data samples are divided into N sample sets, and the mean and variance are calculated respectively:

[0133]

[0134] Randomly shuffle the sample data in the sample set dimension to obtain a new mean and variance In order to obtain different characteristics of the sample set, {μ,σ} and The linear expression of the relationship can obtain the transformation parameters {β,γ}:

[0135] Use transformation parameters to transfer the original features to the new space:

[0136]

[0137] In the formula, ∈ is a fixed constant that can be ignored, and They are the features of each sample set before and after the conversion. Feature confusion can enrich the style of the feature layer.

[0138] 2. Label diversity regularization

[0139] Only clean data in the pseudo-measurement data is retained to optimize the model, resulting in inefficient use of the remaining pseudo-measurement data. In order to solve this problem, the present invention adopts label diversity regularization for the remaining data, aiming to improve the generalization ability and robustness of the model by increasing the diversity and uncertainty of labels. Maximizing the global entropy of label diversity can achieve model generalization by reducing assumptions and biases, balancing label distribution, avoiding overfitting, and enhancing the robustness of the model. Therefore, the ultimate goal of the generalization module is summarized as follows:

[0140]

[0141] In the formula, is the clean set in the pseudo-measurement data, and These are clean samples selected from subnet 1 and subnet 2 respectively; Classification loss of pseudo-measurements; To maximize the global entropy of label diversity, label diversity regularization is achieved.

[0142] Fourth, the establishment of the evolutionary intermediate domain

[0143] There is a gap between the measured data set and the pseudo-measurement data set. If the gap is too large, the quality of the pseudo-label will be seriously affected, thereby reducing the accuracy of the training model. If the gap is too small, it will easily lead to overfitting of the training model and reduce the generalization performance of the SE model.

[0144] Therefore, the present invention further proposes the following solutions to solve the above problems: Figure 5 A new cyclic learning framework is shown, which realizes the positive feedback of ACN and DCDG modules, adds a cyclically updateable intermediate domain to connect the measurement dataset and the pseudo-measurement dataset and gradually adjusts the gap between the datasets to achieve the ideal effect.

[0145] The proposed cyclic learning framework includes a progressive adversarial pseudo-labeling module (ACN) and a dual calibration generalization module (DCDG) with style confusion and label diversity regularization. The measurement data set and the pseudo measurement data set are respectively substituted into the ACN module and the DCDG module to filter out the clean data that meets the two modules, generate interpolation of the clean set of measurement data and pseudo measurement data, and mix it with the original data set to obtain an evolvable intermediate domain to adjust the gap between the measurement data set and the pseudo measurement data set. Establishment of the evolving intermediate domain:

[0146]

[0147] In the formula, is an evolvable intermediate domain in the recurrent learning framework, and is a high quality clean set, and α is a weighted map with the same spatial resolution as the input data; E(α) is a scalar representing the average value of α.

[0148] By repeating the above process, the optimal gap between the measurement data set and the pseudo-measurement data set can be obtained, ensuring that the distribution network state prediction model has high accuracy and strong generalization ability.

[0149] Fifth, example analysis

[0150] Based on the above theoretical analysis, simulation verification is carried out in IEEE 33-node and IEEE 118-node radial networks to verify the effectiveness of the adversarial pseudo-labeling module, dual calibration generalization module and evolutionary intermediate domain. The topological structure, node load and branch impedance parameters of the IEEE 33-node and IEEE 118-node systems are referenced in the literature. The simulation software uses Matlab 2018a and python3.11.

[0151] Resource configuration in IEEE 33-node system Figure 6 As shown, the PV active output of nodes 9, 13, 17, and 21 is 0.08MW, the WT active output of nodes 25 and 29 is 0.35MW and 0.15MW respectively, and the capacity of the EV charging station connected to node 8 is 0.2MW. Considering the small coverage area of ​​the distribution network, all loads, photovoltaic power generation and wind power generation use the same time series curve, that is, the typical summer curve of a certain place, such as Figure 7 IEEE 118 node system resource configuration is shown in Figure 8 ,The node DG parameter settings are shown in Table 1. The EV charging station load connected to nodes 30, 65, and 104 is 0.15MW.

[0152] Table 1 Parameters of DG in IEEE 118-bus system

[0153] serial number DG Type Number of units Grid connection location Active power / MW DG1 PV 5 13、21、45、80、112 0.4 DG2 PV 3 52、66、74 0.6 DG3 WT 3 71、88、99 0.2

[0154] The measurement data for establishing the SE model of the present invention include branch power, node injection power, and the amplitude and phase angle of the distribution network voltage. The input of the distribution network SE model established by the present invention is the injection power and branch power of all nodes in the system, where the branch power data requires special equipment for measurement and transmission, and the output of the distribution network SE model is the voltage amplitude and voltage phase angle of all nodes.

[0155] 1. ACN module effect analysis

[0156] In order to demonstrate the effectiveness of the progressive adversarial pseudo-labeling method proposed in the present invention, the average accuracy of pseudo-labeling is compared with the deep neural network (DNN) method. Fig. 9 Comparison of the average accuracy of pseudo labels generated by different methods.

[0157] Depend on Fig. 9 It can be seen that the average accuracy of pseudo-labels of the method of the present invention and the DNN method increases with the increase of the number of iterations of sample training, and the pseudo-labeling effect is significant when the number of iterations is set to be greater than 40; when the number of iterations is set to be greater than 40, the average accuracy of pseudo-labels of the method of the present invention is increased by about 6% compared with the DNN method.

[0158] 2. Analysis of the effect of the cyclic learning framework

[0159] In order to verify that the cyclic learning framework with evolving intermediate domain proposed in the present invention can improve the model generalization ability while ensuring the accuracy of the distribution network SE model, the present invention compares the model generalization ability and pseudo-label accuracy of different methods. Fig.10 This is a comparison of the generalization ability and pseudo-label accuracy of different method models. In the figure, method one is the DNN method; method two is the ACN method; method three is the CLF method.

[0160] Depend on Fig.10 It can be seen that the average accuracy of pseudo labels and the generalization ability of the model of the DNN method are both low; the average accuracy of pseudo labels using the ACN method is greatly improved, but the generalization ability of the model is still poor; the average accuracy of pseudo labels using the CLF method is about 5% higher than that of the DNN method, and the generalization ability of the model is improved by about 7%. Therefore, the CLF proposed in the present invention can greatly improve the generalization ability of the model while ensuring the average accuracy of pseudo labels in advance.

[0161] 3. SE model accuracy verification

[0162] In order to verify the accuracy of the SE model of the distribution network proposed by the method of the present invention, the voltage states of the IEEE 33-node and IEEE 118-node systems at 12:00 are estimated respectively.

[0163] (1) IEEE 33-node system

[0164] Fig.11 The figure shows the comparison between the state estimation results and the actual values ​​of the IEEE 33-bus system using the DNN method and the CLF method. Fig.11 (a) is the comparison of voltage amplitude estimation results. Fig.11 (b) is the comparison of voltage phase angle estimation results.

[0165] Table 2 shows the comparison of various indicators of the state estimation results between the DNN method and the method of the present invention.

[0166] Table 2 Comparison of state estimation deviation of IEEE 33-node system

[0167]

[0168] In the comparison of the state estimation results of the voltage amplitude, the estimation results of the DNN method have a large deviation as a whole, and it is particularly obvious near nodes 22 to 25, with the maximum deviation reaching 0.0082pu. The results of the method of the present invention are almost consistent with the actual values, with a maximum deviation of 0.0031pu and an average error of only 0.0016pu.

[0169] In the comparison of the state estimation results of the voltage phase angle, the overall deviation of the estimation results of the DNN method is more obvious, and it is particularly prominent at nodes 7 to 14, with the maximum deviation reaching 0.0621 rad. The results of the method of the present invention are almost consistent with the actual value, with a maximum deviation of 0.0163 rad and an average error of only 0.0081 rad. Analysis shows that in the IEEE 33-node system, the state estimation method proposed in the present invention has a greatly improved effect compared with the DNN method.

[0170] (2) IEEE 118 node system

[0171] Fig.12 The figure shows the comparison between the state estimation results and the actual values ​​of the IEEE 118-node system using the DNN method and the CLF method. Fig.12 (a) is the comparison of voltage amplitude estimation results. Fig.12 (b) is the comparison of voltage phase angle estimation results.

[0172] Table 3 shows the comparison of various indicators of the state estimation results of the IEEE 118-node system using the DNN method and the method of the present invention.

[0173] Table 3 Comparison of IEEE 118-node system state estimation deviation

[0174]

[0175] In the comparison of the state estimation results of the voltage amplitude, the estimation results of the DNN method have a large deviation as a whole, and it is particularly obvious near the nodes with large voltage fluctuations. The maximum deviation reaches 0.0096pu. The results of the method of the present invention are almost consistent with the actual values, with a maximum deviation of 0.0036pu and an average error of only 0.0021pu.

[0176] In the comparison of the state estimation results of the voltage phase angle, the overall deviation of the estimation results of the DNN method is more obvious, with the maximum deviation reaching 0.0741 rad. The results of the method of the present invention are almost consistent with the actual value, with an average error of only 0.0081 rad. Analysis shows that in the IEEE 118-node system, the state estimation method proposed in the present invention has a greatly improved effect compared with the DNN method.

[0177] (4) SE model generalization ability test

[0178] In order to verify the generalization ability of the model proposed in this invention, the accuracy of the state estimation model is verified by adding different proportions of noise to the IEEE 33 and IEEE 118 node system test data. 1%, 2%, and 3% of measurement noise are added to the original test data set, respectively. Fig.13 , Fig.14 Comparison of average errors of distribution network state estimation at 12:00 under noise interference of different proportions.

[0179] Depend on Fig.13 , Fig.14 It can be obtained that when the DNN method is used to estimate the state of the distribution network, as the noise ratio of the measurement data increases, the estimation deviation gradually increases; compared with the DNN method, the method of the present invention can still maintain a low measurement deviation when the noise ratio of the measurement data increases, so the model proposed in the present invention has a strong generalization ability.

[0180] (5) SE efficiency analysis

[0181] The efficiency of distribution network state estimation is of great significance to the real-time and stability of the system, resource optimization and economic benefit improvement. Therefore, the present invention analyzes the application of different state estimation methods in different distribution network systems at a certain time section, as shown in Table 4.

[0182] Table 4 Comparison of state estimation time of different methods

[0183] system DNN methods CLF Method IEEE33 0.0462s 4.591s IEEE118 0.069s 6.236s

[0184] From the analysis of Table 4, it can be obtained that because the CLF method proposed in the present invention needs to consider the accuracy and generalization ability of the SE model at the same time, the state time is longer than that of the conventional DNN method, and the calculation time is slightly increased compared with the DNN method, but it is still maintained at a smaller order of magnitude, and the state estimation time meets the needs of distribution network scheduling.

Claims

1. A coordinated optimization method for the accuracy and generalization ability of the distribution network state estimation model based on the evolutionary intermediate domain. It is characterized by: The following steps are involved: Step 1: Considering the measurement accuracy, time scale and synchronization of state estimation calculation of multi-source data, pre-process the measurement data of the day before and within the day; Step 2: Using the output of distributed generation and the measured data of distribution network operation as the training data set, a semi-supervised learning method is used to establish the initial model of single hidden layer state estimation of multi-layer output of distribution network; Step 3: Based on the SSL method, a pseudo-labeling method is proposed. The state estimation initial model is used to add pseudo-labels to the unlabeled measurement data to generate pseudo-measurement data, solving the problem of insufficient training model data. Step 4: A progressive adversarial pseudo-labeling model is designed, which includes a pseudo-labeling model and a discriminator. The accuracy of the pseudo-measurement data is improved through the mutual confrontation between the pseudo-labeling model and the discriminator, the total data set is updated, and a higher-precision distribution network SE model is obtained through further training; Step 5: To avoid the overfitting problem of the training model, a dual calibration generalization model with style confusion training and label diversity regularization is proposed to improve the generalization performance of the SE training model; Step 6: In order to balance the contradiction between the accuracy and generalization ability of the distribution network state estimation model, a distribution network SE model accuracy-generalization ability coordinated optimization method considering the evolutionary intermediate domain is proposed.

2. The coordinated optimization method of distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1 is characterized by: In step 1, the measurement accuracy, time scale and synchronization of state estimation calculation of multi-source data are considered, and the measurement data of the day before is preprocessed as follows: Before using the distribution network measurement data for state estimation, the data needs to be preprocessed. Calculate ν measurement values ​​respectively Mean μ and standard deviation σ, and analyze the νth measurement data separately Does it satisfy the formula Where M is the bad data discrimination threshold. If the above formula is satisfied, the measured data is bad data, and finally the average value of the remaining data is taken as the measured data at the SE moment.

3. The coordinated optimization method of distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1 is characterized by: In step 2, the output of distributed power sources and the measured data of distribution network operation are used as the training data set, and the semi-supervised learning method is used to establish the initial model of distribution network multi-layer output single hidden layer state estimation: First, assume that a set of original data samples {x Θ ,y Θ },in is the input vector of the model, represents the predicted target vector, l is the measurement number, and r is the dimension of the data; the output of a single hidden layer neural network of hidden nodes is expressed as: In the formula, x Θ is the input vector, λ j is the weight value of the model; ω j is the target vector y Θ The weight of j is the bias of the jth hidden layer node; g(·) is the activation function tanh(·) based on hyperbolic tangent. Compared with the traditional sigmoid(·) activation function, tanh(·) converges faster; The core goal of neural network optimization is to minimize the actual output c j and the target value y Θ The error between them makes the following equation true: Therefore, as long as the weight value λ corresponding to the given input vector j , and the hidden layer unit bias value b j , then the training of a single hidden layer neural network can be transformed into solving a linear system, and the corresponding weight value ω of the output vector to be solved j It can be obtained by solving the above τ linear equations; In the distribution network state prediction, each set of data includes the node injection power, branch power and the corresponding system voltage amplitude and phase angle; the possible operating state of the estimated system is determined by the interval information of the measurement and a set of mapping relationships. The mathematical relationship can be expressed as: Where E(·) is the interval set of the state variables to be determined; h(·) is the nonlinear mapping function between the state variables and the model input; and Z(·) is the pseudo-measurement uncertainty set of the input.

4. The coordinated optimization method of distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1 is characterized by: Step 3: Based on the semi-supervised learning method, a pseudo-labeling method is proposed. The state estimation initial model is used to add pseudo-labels to the unlabeled measurement data to generate pseudo-measurement data, solving the problem of less training model data: Define distribution network measurement and unlabeled measurement datasets: In the formula, x and y are scalars; Measure data for the system; It is an unlabeled measurement dataset and has no corresponding output; The pre-processed day-ahead and day-intraday labeled measurement data are trained to obtain the distribution network SE initial model; the unlabeled measurement data are substituted into the trained initial model and the APL method is used to obtain the distribution network SE initial model. Add pseudo-label Q to generate pseudo-measurement data That is, generate fake data that is as close to the real data distribution as possible.

5. The coordinated optimization method for distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1, characterized in that: In step 4, the APL model adds pseudo labels to the unlabeled data to deceive the discriminator, and the discriminator is updated to become more sensitive to distinguish between real and fake data; This adversarial training process achieves the model by optimizing the loss function of the two models.

6. The coordinated optimization method for distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1, characterized in that: In step 4, the loss function of the progressive adversarial pseudo-labeling model is: In the formula, Measure data for the system; After being processed by the encoder A() is the defined function; for The posterior distribution of for The mathematical expectation of The loss function is: In the formula, ACN is a progressive pseudo-labeling model; for The generated samples of for The mathematical expectation of for The mathematical expectation of The loss function is: In the formula, It is pseudo measurement data; After being processed by the encoder for The generated samples of for The mathematical expectation of for The mathematical expectation of According to the loss function, progressive training is performed so that the game between the generator and the discriminator gradually reaches an equilibrium state. The objective function for the progressive adversarial pseudo-labeling model to reach an equilibrium state is defined as the minimum deviation between the loss function of the pseudo-measurement data and the measurement data: minf1=(T3-T2)-T1(9) In equilibrium, the generator is able to generate data realistic enough to fool the discriminator, while the higher-precision discriminator has difficulty distinguishing between real and fake data.

7. The coordinated optimization method of distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1, characterized in that: In step 5, in order to avoid the overfitting problem of the training model, a dual calibration generalization model with style confusion training and label diversity regularization is proposed to improve the generalization performance of the SE training model; The dual calibration model uses a low-cost style confusion training method. A style confuser is inserted in each subnet to confuse the features between the measured data and the pseudo-measurement data. The measured and pseudo-measurement data samples are divided into N sample sets, and the mean and variance are calculated respectively: Randomly shuffle the sample data in the sample set dimension to obtain a new mean and variance In order to obtain different characteristics of the sample set, {μ,σ} and The linear expression of the relationship can obtain the transformation parameters {β,γ}: Use transformation parameters to transfer the original features to the new space: In the formula, ∈ is a fixed constant that can be ignored, and They are the features of each sample set before and after the conversion. Feature confusion can enrich the style of the feature layer. Label diversity regularization is used for the remaining data, aiming to improve the generalization ability and robustness of the model by increasing the diversity and uncertainty of labels. Maximizing the global entropy of label diversity can achieve model generalization by reducing assumptions and biases, balancing label distribution, avoiding overfitting, and enhancing the robustness of the model. Therefore, the ultimate goal of the generalized model can be summarized as follows: In the formula, is the clean set in the pseudo-measurement data, and These are clean samples selected from subnet 1 and subnet 2 respectively; Classification loss of pseudo-measurements; To maximize the global entropy of label diversity, label diversity regularization is achieved.

8. The coordinated optimization method for distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1, characterized in that: In step 6, in order to balance the contradiction between the accuracy and generalization ability of the distribution network state estimation model, a distribution network SE model accuracy-generalization ability coordinated optimization method considering the evolutionary intermediate domain is proposed: Through a new cyclic learning framework, positive feedback of the progressive adversarial pseudo-labeling model and the double-calibration generalization model is realized, and a cyclically updateable intermediate domain is added to connect the measurement dataset and the pseudo-measurement dataset and gradually adjust the gap between the datasets to achieve the ideal effect. Through the proposed cyclic learning framework, which includes a progressive adversarial pseudo-labeling model and a dual-calibration generalization model with style confusion and label diversity regularization, the measurement dataset and the pseudo-measurement dataset are substituted into the progressive adversarial pseudo-labeling model and the dual-calibration generalization model respectively to filter out clean data that conforms to the two models, generate interpolations of the clean sets of the measurement data and pseudo-measurement data, and mix them with the original dataset to obtain an evolvable intermediate domain to adjust the gap between the measurement dataset and the pseudo-measurement dataset; By repeating the above process, the optimal gap between the measurement data set and the pseudo-measurement data set can be obtained, ensuring that the distribution network state prediction model has high accuracy and strong generalization ability.

9. The coordinated optimization method of distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1, characterized in that: The method requires the use of a pseudo-labeling module, a discriminator module, a progressive adversarial pseudo-labeling module, a dual-calibration generalization module, and a recurrent learning module.

10. The coordinated optimization method of distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 9, characterized in that: The pseudo-labeling module uses the state estimation initial model to add pseudo-labels to the unlabeled measurement data, generating pseudo-measurement data to solve the problem of insufficient training model data. The discriminator module is used to determine whether the pseudo measurement data generated by the pseudo label (APL) module meets the accuracy requirements; The progressive adversarial pseudo-labeling module has: a containing module and a discriminator module, realizing the mutual confrontation between the pseudo-labeling module and the discriminator module to improve the accuracy of the pseudo-measurement data, update the total data set, and further train to obtain a higher-precision distribution network SE model; Dual calibration generalization module: includes style confusion training and label diversity regularization to improve the generalization performance of SE training models; Recurrent learning module with evolving intermediate domain: The recurrent learning framework includes ACN module and DCDG module. The measurement data set and pseudo-measurement data set are respectively substituted into the ACN module and DCDG module to filter out clean data that meet the requirements of the two modules, generate interpolation of clean sets of measurement data and pseudo-measurement data, and mix them with the original data set to obtain an evolving intermediate domain to obtain the optimal gap between the measurement data set and the pseudo-measurement data set, thereby improving the accuracy of the distribution network SE model while ensuring its generalization ability.

11. The coordinated optimization method of distribution network state estimation model accuracy-generalization capability based on evolving intermediate domain according to claim 1, characterized in that: The invention comprises a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the method for coordinated optimization of the accuracy-generalization capability of a distribution network state estimation model based on an evolving intermediate domain as claimed in any one of claims 1 to 9 is implemented.

Images