A topology optimization reconstruction method and system based on machine learning

Abstract

The present invention relates to the technical field of power system operation decision-making, and discloses a topology optimization reconstruction method and system based on machine learning. The method comprises: constructing a target distribution network reconstruction model corresponding to the distribution network; performing supervised learning on the initial XGBoost model based on historical reconstruction case data to obtain a target XGBoost model; performing line switch prediction on the real-time reconstruction case data based on the target XGBoost model to obtain the switch prediction value of each line in the distribution network and the prediction probability corresponding to each switch prediction value; screening each switch prediction value based on the prediction probability to obtain a screening result of the corresponding switch prediction value; solving the target distribution network reconstruction model based on the first switch prediction value for which all screening results are low-confidence decisions, and determining the first reconstruction strategy corresponding to the distribution network according to the first solution result and the second switch prediction value for which all screening results are high-confidence decisions. The present invention can significantly improve the efficiency of solving large-scale distribution network reconstruction problems.

Description

Technical Field

[0001] The present invention relates to the technical field of power system operation decision-making, and in particular to a topology optimization reconstruction method and system based on machine learning. Background Art

[0002] Power system operational decisions involve optimizing the allocation and scheduling of grid resources to ensure safe, stable, and economical system operation. Distribution network reconfiguration, a crucial operational decision-making tool, optimizes grid operation by adjusting the switch states within the distribution network and changing the network topology. The core of distribution network reconfiguration lies in selecting the optimal switch combination to achieve specific optimization goals while satisfying system constraints.

[0003] However, distribution network reconfiguration is a highly complex problem, especially in high-voltage distribution networks, where the system's operating state is dynamically influenced by multiple factors, including load demand, distributed generation output, and equipment failures. Distribution network reconfiguration is typically modeled as a mixed-integer quadratic programming (MIQP) problem, where the decision variables include both continuous variables (such as voltage and current) and 0-1 integer variables (such as switch states), making the solution complex and time-consuming. Traditional optimization methods can provide accurate solutions within a certain scale, but their computational time increases exponentially with increasing problem size, making it difficult to meet the demands of real-time decision-making. Furthermore, due to the uncertainty of grid load and distributed generation output, distribution network reconfiguration decisions require dynamic adjustment. Traditional optimization methods typically rely on static models and struggle to adapt to real-time changes, further complicating the problem.

[0004] Therefore, how to find a balance between complexity and real-time performance has become an urgent problem to be solved in distribution network reconstruction research. Summary of the Invention

[0005] The technical problem to be solved by the present invention is how to quickly adjust the distribution network reconstruction decision based on historical decision data and dynamically changing external parameters. In order to solve the above technical problem, the present invention provides a topology optimization reconstruction method and system based on machine learning.

[0006] In a first aspect, an embodiment of the present invention provides a topology optimization and reconstruction method based on machine learning, comprising:

[0007] Based on minimizing line power loss, constructing a target distribution network reconstruction model corresponding to the distribution network, wherein the decision variables of the target distribution network reconstruction model include line switch status;

[0008] Obtaining historical reconstruction case data of the distribution network, and performing supervised learning on an initial XGBoost model based on the historical reconstruction case data to obtain a target XGBoost model;

[0009] Acquire real-time reconstruction case data of the distribution network, perform line switching prediction on the real-time reconstruction case data based on the target XGBoost model, and obtain a switching prediction value of each line in the distribution network and a prediction probability corresponding to each switching prediction value;

[0010] Filtering each of the switch prediction values ​​based on the prediction probability to obtain a filtering result corresponding to the switch prediction value, wherein the filtering result is a high confidence decision or a low confidence decision;

[0011] The target distribution network reconstruction model is solved based on the first switch prediction value for all the screening results of low-confidence decisions, and the first reconstruction strategy corresponding to the distribution network is determined according to the first solution result and the second switch prediction value for all the screening results of high-confidence decisions.

[0012] Preferably, after determining the first reconstruction strategy corresponding to the distribution network, the method further includes:

[0013] Performing a power flow check on the first solution result based on the distribution network security constraint to obtain a first check result;

[0014] If the first verification result is failure, performing conflict rollback on all the second switch prediction values ​​to obtain all third switch prediction values ​​that cause conflicts;

[0015] Determining a fourth switch prediction value having the lowest confidence among all the third switch prediction values ​​based on the prediction probability corresponding to each of the third switch prediction values;

[0016] Solving the target distribution network reconstruction model based on the fourth switch prediction value and all the first switch prediction values ​​to obtain a second solution result;

[0017] Performing a power flow check on the second solution result based on the distribution network security constraint to obtain a second check result;

[0018] If the second verification result is successful, the first reconstruction strategy is adjusted based on the second solution result, all the second switch prediction values ​​and all the third switch prediction values ​​to obtain a second reconstruction strategy corresponding to the distribution network.

[0019] Preferably, the step of constructing a target distribution network reconstruction model corresponding to the distribution network based on minimizing line power loss includes:

[0020] With the goal of minimizing line power loss, an initial distribution network reconstruction model corresponding to the distribution network is constructed;

[0021] The initial distribution network reconstruction model is constrained to obtain a target distribution network reconstruction model corresponding to the distribution network, wherein the constraints include power load balance constraints, voltage stability constraints, line load capacity constraints, topology connectivity constraints and equipment capacity constraints.

[0022] Preferably, the target distribution network reconstruction model is characterized by the following formula:

[0023]

[0024] Among them, F represents the line power consumption, N represents the number of lines, represents the current of the i-th circuit, represents the resistance of the i-th circuit.

[0025] Preferably, the acquiring of historical reconstruction case data of the distribution network, and performing supervised learning on an initial XGBoost model based on the historical reconstruction case data to obtain a target XGBoost model, comprises:

[0026] Acquiring historical reconstruction case data of the distribution network, and performing feature extraction on the historical reconstruction case data to obtain an initial distribution network reconstruction data set;

[0027] Preprocessing the initial distribution network reconstruction dataset to obtain a target distribution network reconstruction dataset, wherein the preprocessing includes missing value filling, outlier removal and data standardization;

[0028] Solving the target distribution network reconstruction model based on the target distribution network reconstruction dataset to obtain a label for each sample in the target distribution network reconstruction dataset, wherein the label is a binary value of a circuit breaker state;

[0029] Using stratified sampling to divide the target distribution network reconstruction data set into a training set and a test set;

[0030] Training the initial XGBoost model based on the training set to obtain an intermediate XGBoost model, wherein the training process includes initializing the model, iteratively constructing a decision tree through gradient boosting, and preventing overfitting;

[0031] The intermediate XGBoost model is evaluated based on the test set, and the intermediate XGBoost model is optimized according to the obtained evaluation results to obtain a target XGBoost model, wherein the evaluation results include accuracy, precision, recall, F1 score and AUC, and the optimization process includes adjusting model hyperparameters.

[0032] Preferably, the acquiring of real-time reconstruction case data of the distribution network, and performing line switching prediction on the real-time reconstruction case data based on the target XGBoost model to obtain a switching prediction value of each line in the distribution network and a prediction probability corresponding to each switching prediction value, includes:

[0033] The following formula is used to represent the predicted probability corresponding to the switch prediction value:

[0034]

[0035] in, represents the predicted probability corresponding to the switch prediction value of the i-th line, represents the switch prediction value of the i-th line.

[0036] Preferably, screening each of the switch prediction values ​​based on the prediction probability to obtain a screening result corresponding to the switch prediction value includes:

[0037] Each of the switch prediction values ​​is screened based on a comparison result between the prediction probability and a preset confidence threshold to obtain a screening result corresponding to the switch prediction value.

[0038] Preferably, the filtering of each of the switch prediction values ​​based on the comparison result of the prediction probability and a preset confidence threshold to obtain the filtering result corresponding to the switch prediction value includes:

[0039] If the predicted probability is greater than or equal to a preset confidence threshold or the predicted probability is less than or equal to the complement of the preset confidence threshold, determining that the screening result corresponding to the switch prediction value is a high confidence decision;

[0040] If the predicted probability is less than the preset confidence threshold and the predicted probability is greater than the complement of the preset confidence threshold, it is determined that the screening result corresponding to the switch prediction value is a low confidence decision.

[0041] In a second aspect, an embodiment of the present invention provides a topology optimization and reconstruction system based on machine learning, comprising:

[0042] A first model building module is configured to build a target distribution network reconstruction model corresponding to the distribution network based on minimizing line power loss, wherein the decision variables of the target distribution network reconstruction model include line switch states;

[0043] A second model building module is configured to obtain historical reconstruction case data of the distribution network, and perform supervised learning on the initial XGBoost model based on the historical reconstruction case data to obtain a target XGBoost model;

[0044] a switch prediction module, configured to obtain real-time reconstruction case data of the distribution network, and perform line switch prediction on the real-time reconstruction case data based on the target XGBoost model, to obtain a switch prediction value for each line in the distribution network and a prediction probability corresponding to each switch prediction value;

[0045] a first confidence screening module, configured to screen each of the switch prediction values ​​based on the prediction probability to obtain a screening result corresponding to the switch prediction value, wherein the screening result is a high confidence decision or a low confidence decision;

[0046] A first strategy determination module is used to solve the target distribution network reconstruction model based on the first switch prediction value for all the screening results of low-confidence decisions, and determine the first reconstruction strategy corresponding to the distribution network according to the first solution result and the second switch prediction value for all the screening results of high-confidence decisions.

[0047] Preferably, it also includes:

[0048] A first verification module is configured to perform a power flow verification on the first solution result based on a distribution network security constraint to obtain a first verification result;

[0049] a conflict rollback module, configured to, if the first verification result is a failure, perform a conflict rollback on all the second switch prediction values ​​to obtain all third switch prediction values ​​that cause the conflict;

[0050] a second confidence screening module, configured to determine a fourth switch prediction value having the lowest confidence among all the third switch prediction values ​​based on the prediction probability corresponding to each of the third switch prediction values;

[0051] a model solving module, configured to solve the target distribution network reconstruction model based on the fourth switch prediction value and all the first switch prediction values ​​to obtain a second solution result;

[0052] A second verification module is configured to perform a power flow verification on the second solution result based on the distribution network security constraint to obtain a second verification result;

[0053] A second strategy determination module is used to adjust the first reconstruction strategy based on the second solution result, all the second switch prediction values ​​and all the third switch prediction values ​​if the second verification result is successful, to obtain a second reconstruction strategy corresponding to the distribution network.

[0054] Compared with the prior art, the topology optimization and reconstruction method and system based on machine learning in the embodiment of the present invention have the following beneficial effects: combining the XGBoost algorithm in machine learning to predict the probability of the optimal solution value of the 0-1 variable, effectively improving the solution efficiency and reliability of the distribution network reconstruction; performing switch prediction and screening processing based on the prediction probability on the real-time reconstruction case data, distinguishing different confidence decisions, directly locking the values ​​of high-confidence prediction variables, and optimizing the remaining low-confidence prediction variables by solving the target reconstruction model, which not only effectively compresses the solution space dimension of the branch and bound method, but also uses the probability prediction value to establish a feasible solution direction guide, fundamentally improving the problem of a sudden drop in solution efficiency caused by the explosion of the dependent variable dimension; the final determined reconstruction strategy reduces line power loss and optimizes the distribution network operation efficiency while ensuring the reliability and security of the reconstruction strategy. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 1 is a flow chart of a topology optimization and reconstruction method based on machine learning according to an embodiment of the present invention;

[0056] Figure 2 Schematic diagram of the process of obtaining the target XGBoost model according to an embodiment of the present invention;

[0057] Figure 3 1 is a schematic structural diagram of a topology optimization and reconstruction system based on machine learning according to an embodiment of the present invention;

[0058] Reference numerals:

[0059] 1. First model building module; 2. Second model building module; 3. Switch prediction module; 4. First confidence screening module; 5. First strategy determination module. DETAILED DESCRIPTION

[0060] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.

[0061] In the description of the present invention, it should be understood that the terms "first" and "second" etc. are used in the present invention to distinguish different objects rather than to describe a specific order.

[0062] In describing the present invention, it should be noted that, unless otherwise defined, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. Those skilled in the art will understand the specific meanings of the above terms in the present invention in specific circumstances.

[0063] like Figure 1 As shown, an embodiment of the present invention provides a topology optimization and reconstruction method based on machine learning, comprising the steps of:

[0064] S1. Based on minimizing line power loss, a target distribution network reconstruction model corresponding to the distribution network is constructed;

[0065] Specifically, step S1 includes:

[0066] 1) With the goal of minimizing line power loss, an initial distribution network reconstruction model corresponding to the distribution network is constructed;

[0067] The main goal of distribution network reconstruction is usually to minimize line power loss and improve the power supply reliability of the system.

[0068] The target distribution network reconstruction model is characterized by the following formula:

[0069]

[0070] Among them, F represents the line power consumption, N represents the number of lines, represents the current of the i-th circuit, represents the resistance of the i-th line. The decision variables of the target distribution network reconstruction model include the line switch status.

[0071] 2) Constrain the initial distribution network reconstruction model to obtain the target distribution network reconstruction model corresponding to the distribution network.

[0072] The distribution network reconstruction problem needs to meet a series of constraints, which ensure the stability and security of the power system and its compliance with external demands.

[0073] Specifically, the constraints include power load balance constraints, voltage stability constraints, line load capacity constraints, topology connectivity constraints, and equipment capacity constraints.

[0074] The following is a detailed description of each constraint:

[0075] 1) Power load balance constraints

[0076] During distribution network reconstruction, it is necessary to ensure a balance between power supply and load demand. For each node, its power supply and power demand must meet the following requirements:

[0077]

[0078] in, Indicates the The power supply of each node, Indicates the The load demand of each node, Indicates the number of nodes.

[0079] 2) Voltage stability constraints

[0080] The voltage of each node in the distribution network must be kept within a certain range to ensure the stability of the power system. The voltage is , then for each node, it must satisfy:

[0081]

[0082] in, Indicates the minimum allowed value of the node voltage, Indicates the maximum allowable value of the node voltage.

[0083] 3) Line load capacity constraints

[0084] The load of each line must be less than or equal to its maximum carrying capacity to avoid line overload. For each line, its load capacity constraint is:

[0085]

[0086] in, Indicates the The maximum load power of the line.

[0087] 4) Topological connectivity constraints

[0088] The distribution network must maintain connectivity to ensure that power can be transmitted from the power source to all load nodes. Specifically, the topology of the distribution network should satisfy the following constraints, that is, each load node can be connected to the power source node through a certain path:

[0089]

[0090] in, Representation and load nodes A collection of connected lines, Indicates line and load nodes Connection status, if the line and load nodes If connected, ,otherwise , Indicates the number of load nodes.

[0091] 5) Equipment capacity constraints

[0092] The load of each transformer and switchgear must not exceed its maximum capacity. The maximum capacity is , then there are the following constraints:

[0093]

[0094] in, Representation device The load, Indicates the number of devices.

[0095] S2. Obtain historical reconstruction case data of the distribution network, and perform supervised learning on the initial XGBoost model based on the historical reconstruction case data to obtain the target XGBoost model;

[0096] Traditional methods for solving distribution network reconstruction problems usually rely on specialized optimization solvers. However, when dealing with large-scale distribution network reconstruction problems, due to the large number of variables and constraints involved, and the objective function including nonlinear terms, relying solely on traditional solvers for solution often encounters challenges such as a surge in computational complexity and excessively long solution times, especially in scenarios with high real-time requirements or large-scale systems. Therefore, the present invention deeply combines machine learning models with operational optimization methods, utilizing the efficient prediction capabilities of machine learning to accelerate the solution process of distribution network reconstruction problems, thereby significantly improving computational efficiency.

[0097] Specifically, if Figure 2 As shown, step S2 includes:

[0098] S201, obtaining historical distribution network reconstruction case data, and performing feature extraction on the historical reconstruction case data to obtain an initial distribution network reconstruction data set;

[0099] When predicting the probability of the optimal solution of the 0-1 decision variable for distribution network reconstruction based on the XGBoost algorithm in machine learning, it is necessary to accurately label and reasonably divide the historical distribution network data to generate high-quality samples suitable for model training. Therefore, it is necessary to collect historical reconstruction case data and construct a A sample set of distribution network reconstruction problems.

[0100] Specifically, by extracting features from historical reconstruction case data, we can obtain The initial distribution network reconstruction dataset for the distribution network reconstruction problem. The initial distribution network reconstruction dataset includes samples, and the feature vector of each sample includes several extracted features.

[0101] In a specific embodiment, several extracted features include:

[0102] 1) Graph features in distribution network topology data

[0103] Node degree, common neighbors, Jaccard coefficient, Adamic-Adar index, and preferential connectivity. Specifically, the Jaccard coefficient is used to measure the similarity and difference between finite sample sets, and its value is equal to the ratio of the intersection and union of the sample sets. The Adamic-Adar index is used to measure the similarity between nodes in the distribution network. Its basic idea is to measure node similarity based on the number of common neighbors. The more common neighbors, the higher the node similarity. Preferential connectivity means that new nodes in the distribution network prefer to connect to nodes with existing high connectivity.

[0104] 2) Parameter characteristics in power system data

[0105] Power system load, transformer power factor, and transformer redundancy factor.

[0106] 3) Parameter characteristics in line data

[0107] Maximum allowable current, resistance, reactance.

[0108] 4) Measurement characteristics in line data

[0109] Active power, reactive power, voltage, current, and line connectivity status.

[0110] 5) Switchgear status characteristics in switchgear data

[0111] Switch status.

[0112] 6) Knife switch device status characteristics in switch knife switch data

[0113] Status of main mother knife switch and auxiliary mother knife switch.

[0114] 7) Parameter characteristics in transformer data

[0115] High voltage end resistor, medium voltage end resistor, low voltage end resistor.

[0116] 8) Measurement features in transformer data

[0117] High voltage end switch status, medium voltage end switch status, low voltage end switch status, high voltage end active power, medium voltage end active power, low voltage end active power.

[0118] Through feature extraction, comprehensive data support is provided for subsequent machine learning modeling and optimization solutions.

[0119] S202, preprocessing the initial distribution network reconstruction data set to obtain a target distribution network reconstruction data set;

[0120] To improve data quality and lay the foundation for machine learning model training, the initial distribution network reconstruction dataset needs to be preprocessed to obtain the target distribution network reconstruction dataset. Specifically, preprocessing includes missing value filling, outlier removal, and data normalization.

[0121] The following is a detailed description of each preprocessing operation:

[0122] 1) Missing value filling

[0123] Fill in the missing values ​​by interpolation. In the samples, the characteristics Value If it is missing, it can be filled by the mean of the feature in the remaining samples. The specific formula is as follows:

[0124]

[0125] in, Representation characteristics In the The values ​​in the sample, Represents the number of samples without missing values. Missing values ​​are filled by interpolation to ensure data integrity.

[0126] 2) Outlier removal

[0127] The quartile range (IQR) method is used to detect and remove outliers. The first quartile is Q1, the third quartile is Q3, and the interquartile range of the feature is IQR=Q3-Q1. According to the IQR principle, if the value of a sample is satisfy or , it is considered an outlier and removed. By removing outliers, we can avoid the adverse effects of extreme data on subsequent model training.

[0128] 3) Data standardization

[0129] Data normalization is used to eliminate the impact of different dimensions on model training and ensure that each feature is processed at the same scale. Specifically, the following formula is used for normalization:

[0130]

[0131] in, represents the normalized eigenvalue, Representation characteristics The mean of Representation characteristics Through standardization, the mean of all features is 0 and the standard deviation is 1, which can avoid the model training deviation caused by different dimensions and help improve the stability and convergence speed of the model.

[0132] S203, solving the target distribution network reconstruction model based on the target distribution network reconstruction dataset to obtain a label for each sample in the target distribution network reconstruction dataset;

[0133] Specifically, the label is a binary value representing the state of the line switch, 0 corresponds to closed and 1 corresponds to open. It can be defined as:

[0134]

[0135] in, represents the feature vector of the first sample, Indicates the label of the first sample, and the rest of the variables can be deduced similarly.

[0136] Furthermore, the generated labels need to not only be consistent with the feature vectors but also have physical interpretations. Therefore, the importance of features can be evaluated by calculating the correlation between the feature vectors and the labels.

[0137] S204, using stratified sampling to divide the target distribution network reconstruction data set into a training set and a test set;

[0138] In order to ensure the accuracy and robustness of model training, it is necessary to ensure data balance when dividing the target distribution network reconstruction dataset into training set and test set. If in the dataset, the sample with label 1 is positive and the sample with label 0 is negative, the number of the two samples is and , represents the number of positive samples, Represents the number of negative samples, then the distribution ratio of the label is:

[0139]

[0140] in, Indicates the ratio of positive samples or negative samples, Indicates the sample size.

[0141] In order to keep the label distribution of the training set and the test set consistent with the full data set, a stratified sampling method is used to sample positive and negative samples in equal proportions. Stratified sampling can effectively prevent the model from overfitting or underfitting a certain category during training. If the ratio of the training set is , the proportion of the test set is , the number of samples in the training set and test set are:

[0142]

[0143] in, represents the number of samples in the training set, Specifically, in this embodiment, the ratio of the training set is set to 0.7, and the ratio of the test set is set to 0.3.

[0144] S205, training the initial XGBoost model based on the training set to obtain an intermediate XGBoost model;

[0145] It should be noted that in the task of predicting the line switch status of distribution network reconstruction, the present invention uses the XGBoost model as the training core. This is because the XGBoost model has efficient computing performance, excellent prediction accuracy, and strong modeling capabilities for complex nonlinear relationships. Through random feature subset sampling and decision tree ensemble learning, it can effectively explore the deep correlation between distribution network characteristics and line switch status. It has good robustness and generalization ability and is particularly suitable for processing the high-dimensional feature space and dynamic operation scenarios of the distribution network. At the same time, the XGBoost model achieves an excellent balance between computing efficiency and model performance, can quickly adapt to the physical characteristics of the reconstruction problem, and ensure the reliability and stability of the prediction results.

[0146] To further improve the accuracy and adaptability of the initial XGBoost model, this step trains the initial XGBoost model based on the training set to obtain an intermediate XGBoost model, thereby enhancing the model's robustness and practical value. Specifically, the training process includes initializing the model, iteratively building a decision tree through gradient boosting, and preventing overfitting.

[0147] The following is a detailed description of each step of the training process:

[0148] 1) Initialize the model

[0149] For the binary classification task of line switch status prediction for distribution network reconstruction, the prediction values ​​of all samples are initialized to a constant value. Specifically, the following formula is used to calculate the initial prediction value:

[0150]

[0151] in, represents the initial prediction value, Indicates the proportion of positive samples.

[0152] 2) Training objective function

[0153] The objective function of the initial XGBoost model consists of a loss function and a regularization term, and the objective function is represented by the following formula:

[0154]

[0155] in, Represents the loss function, which is used to measure the predicted value and the true value The difference, Represents a regularization term, which is used to control the complexity of the model and prevent overfitting. Indicates the A decision tree, Indicates the number of decision trees.

[0156] Specifically, for the two-classification task, logarithmic loss is used, that is:

[0157]

[0158] in, Represents the predicted value The corresponding predicted probability.

[0159] Furthermore, the following formula is used to represent the regularization term:

[0160]

[0161] in, Indicates the number of leaf nodes, represents the leaf node weight, and represents the regularization coefficient.

[0162] 3) Iteratively construct a decision tree through gradient boosting

[0163] In each iteration Among them, there are:

[0164] i) For each sample , calculate the first-order gradient of the loss function and the second-order gradient :

[0165]

[0166] ii) Generate a decision tree

[0167] Based on the gradient information, build a decision tree , to minimize the objective function. Specifically, the decision tree generation process includes:

[0168] a) Feature Selection

[0169] Based on the greedy algorithm, the features and segmentation points that reduce the objective function the most are selected.

[0170] b) Calculate leaf node weights

[0171] For each leaf node , the weight is calculated using the following formula :

[0172]

[0173] in, Indicates that it belongs to a leaf node The sample collection.

[0174] c) Update the predicted value

[0175] Add the newly generated decision tree to the model and update the predicted values:

[0176]

[0177] in, Represents the learning rate, which is used to control the contribution of each decision tree.

[0178] 4) Prevent overfitting

[0179] Specifically, regularization and pruning are used to prevent overfitting, including the following methods:

[0180] 41) Introducing regularization terms

[0181] A regularization term is added to the objective function to limit the weight of leaf nodes and the structural complexity of the decision tree.

[0182] 42) Limit the maximum depth

[0183] Set the maximum depth of the decision tree to avoid generating overly complex trees.

[0184] 43) Set the minimum loss reduction threshold

[0185] Set the minimum loss drop threshold when splitting a node to avoid invalid splits.

[0186] 44) Subsampling

[0187] In each round of iteration, some samples or features are randomly selected for training to increase the generalization ability of the model.

[0188] S206. Evaluate the intermediate XGBoost model based on the test set, and optimize the intermediate XGBoost model according to the obtained evaluation results to obtain the target XGBoost model.

[0189] For the distribution network reconfiguration problem, the core of the evaluation phase is to analyze the switching prediction results of each line. The prediction results are compared with the actual results to construct a confusion matrix. The true positive (TP), false positive (FP), false negative (FN), and true negative (TN) of each category are defined as follows:

[0190] TP: The number of samples where the line switch is predicted to be on and is actually on;

[0191] FP: The number of samples where the line switch is predicted to be on but is not actually on;

[0192] FN: The number of samples where the circuit breaker is predicted to be closed but is not actually closed;

[0193] TN: The number of samples where the line switch is predicted to be closed and is actually closed.

[0194] Based on the above definitions, multiple evaluation results can be calculated. Specifically, the evaluation results include accuracy, precision, recall, F1 score, and AUC.

[0195] The calculation method for each evaluation result is described in detail below:

[0196] 1) Accuracy

[0197] Accuracy It indicates the proportion of correct predictions by the model, and the calculation formula is:

[0198]

[0199] Accuracy is a measure of the overall classification effect and is suitable for measuring the overall performance of the model.

[0200] 2) Accuracy

[0201] Accuracy It represents the proportion of true examples in the model's prediction of a certain category. The calculation formula is:

[0202]

[0203] A high precision means that the model makes fewer misjudgments in its predictions for that category.

[0204] 3) Recall

[0205] Recall It represents the proportion of samples that are correctly classified in a certain category. The calculation formula is:

[0206]

[0207] A high recall rate means that the model makes fewer missed predictions for this category.

[0208] 4) F1 score

[0209] The F1 score represents the harmonic average of precision and recall, and is suitable for comprehensively measuring the classification performance of the model for a certain category. The calculation formula is:

[0210]

[0211] 5) AUC

[0212] AUC (Area Under the Curve) measures the model's ability to distinguish samples of a certain category from samples of other categories by calculating the area under the ROC (Receiver Operating Characteristic Curve).

[0213] In order to comprehensively evaluate the model performance, an overall reliability score can be defined , the calculation formula is:

[0214]

[0215] in, Represents the precision, recall, F1 score and AUC after taking the macro average of all categories, and the weights satisfy Specifically, in this embodiment, each weight is set to 0.2, that is, the weight of each indicator is equal.

[0216] When the overall performance of the intermediate XGBoost model meets the first preset condition, the intermediate XGBoost model is considered to have good reliability and generalization ability; when some performance indicators do not meet the second preset condition, the intermediate XGBoost model needs to be optimized to obtain the target XGBoost model.

[0217] Specifically, the first preset condition of this embodiment is , the recall rate of each category and The second precondition is that the recall rate of positive samples is below a certain threshold. Furthermore, the optimization process includes adjusting model hyperparameters (increasing the number of decision trees, optimizing the sampling ratio of feature subsets), improving sample distribution (giving higher weights to difficult samples to balance the class distribution), and introducing more discriminative features to improve the model's classification performance.

[0218] S3. Obtain real-time reconstruction case data of the distribution network, and perform line switching prediction on the real-time reconstruction case data based on the target XGBoost model to obtain the switching prediction value of each line in the distribution network and the prediction probability corresponding to each switching prediction value;

[0219] Specifically, the following formula is used to represent the predicted probability corresponding to the switch prediction value:

[0220]

[0221] in, represents the predicted probability corresponding to the switch prediction value of the i-th line, represents the switch prediction value of the i-th line. It should be noted that In fact, the switch prediction value of the i-th line is the corresponding prediction probability when it is turned on.

[0222] S4. Filter each switch prediction value based on the prediction probability to obtain a filtering result of the corresponding switch prediction value;

[0223] Specifically, each switch prediction value is screened based on the comparison result between the prediction probability and the preset confidence threshold to obtain a screening result of the corresponding switch prediction value. The screening result is a high confidence decision or a low confidence decision.

[0224] Furthermore, if the predicted probability is greater than or equal to a preset confidence threshold or the predicted probability is less than or equal to the complement of the preset confidence threshold, the screening result of the corresponding switch prediction value is determined to be a high confidence decision. In this embodiment, the preset confidence threshold is 0.9. If the predicted probability is ≥ 0.9 or the predicted probability is ≤ 0.1, the screening result of the corresponding switch prediction value is determined to be a high confidence decision, and its binary classification result is directly fixed to 1 or 0.

[0225] Furthermore, if the predicted probability is less than a preset confidence threshold and the predicted probability is greater than the complement of the preset confidence threshold, the screening result of the corresponding switch prediction value is determined to be a low-confidence decision. In this embodiment, if 0.1 < predicted probability < 0.9, the screening result of the corresponding switch prediction value is determined to be a low-confidence decision, and its binary classification result is retained as the integer variable to be optimized in the target distribution network reconstruction model.

[0226] S5. Solve the target distribution network reconstruction model based on the first switch prediction value for all screening results with low confidence decisions, and determine the first reconstruction strategy corresponding to the distribution network based on the first solution result and the second switch prediction value for all screening results with high confidence decisions.

[0227] By fixing some variables with high-confidence probability results, the scale of the distribution network reconstruction problem is reduced, the search space of the branch-and-bound tree is reduced, and the solution of the target distribution network reconstruction model is accelerated. At the same time, the low-confidence probability results are used to provide feasible solution directions for the target distribution network reconstruction model, avoiding the search space explosion caused by random initialization.

[0228] Specifically, based on the first switch prediction value for all screening results with low confidence decisions, a solver is used to solve the target distribution network reconstruction model to obtain a first solution result, and based on the first solution result and the second switch prediction value for all screening results with high confidence decisions, a first reconstruction strategy corresponding to the distribution network is determined.

[0229] In order to ensure the feasibility of the final solution to the distribution network reconstruction problem, the embodiment of the present invention provides a topology optimization and reconstruction method based on machine learning, which further includes the following steps after step S5:

[0230] 1) Performing a power flow check on the first solution result based on the distribution network security constraint to obtain a first check result;

[0231] A power flow calculation tool is used to verify whether the first solution meets the distribution network safety constraints. If so, the first verification result is determined to be successful, otherwise, the first verification result is determined to be failed. Specifically, the distribution network safety constraints include line load limits and voltage limits.

[0232] 2) If the first verification result is failure, all second switch prediction values ​​are rolled back to obtain all third switch prediction values ​​that cause conflicts;

[0233] If the first verification result does not satisfy the distribution network safety constraint in the power flow calculation, all second switch prediction values ​​are conflict-rolled back to identify the third switch prediction value that causes the conflict.

[0234] 3) determining a fourth switch prediction value having the lowest confidence among all the third switch prediction values ​​based on the prediction probability corresponding to each third switch prediction value;

[0235] The third switch prediction value corresponding to the minimum prediction probability is characterized as the fourth switch prediction value with the lowest confidence among all the third switch prediction values.

[0236] 4) Solving the target distribution network reconstruction model based on the fourth switch prediction value and all first switch prediction values ​​to obtain a second solution result;

[0237] Based on the fourth switch prediction value and all the first switch prediction values, a solver is used to solve the target distribution network reconstruction model to obtain a second solution result.

[0238] 5) Performing a power flow check on the second solution result based on the distribution network security constraint to obtain a second check result;

[0239] A power flow calculation tool is used to verify whether the second solution result meets the distribution network safety constraint. If so, the second verification result is determined to be successful; otherwise, the second verification result is determined to be failed.

[0240] 6) If the second verification result is successful, the first reconstruction strategy is adjusted based on the second solution result, all second switch prediction values, and all third switch prediction values ​​to obtain a second reconstruction strategy corresponding to the distribution network.

[0241] If the second verification result satisfies the distribution network safety constraints in the power flow calculation, the first reconfiguration strategy is adjusted based on the second solution result, all second switch prediction values, and all third switch prediction values ​​to obtain a second reconfiguration strategy corresponding to the distribution network. It is understood that the second reconfiguration strategy includes the final switch state of each line in the distribution network.

[0242] It should be noted that steps 1) through 6) above are actually a power flow verification and conflict rollback mechanism specifically designed for this invention. For ease of understanding, the following describes this mechanism from two perspectives: triggering conditions and data processing. This mechanism is triggered in two situations: when the solver cannot find a feasible solution after fixing the variables, or when the solver's solution does not satisfy distribution network safety constraints during power flow calculations. Furthermore, the data processing process includes: first, using a power flow calculation tool to verify that the optimized solution satisfies distribution network safety constraints, such as power balance, power flow constraints, and voltage limits. If this verification fails, a conflict rollback is performed, identifying the fixed variables that caused the conflict, sorting them by confidence, and releasing the variable with the lowest confidence. Then, the solver optimization is re-invoked, and the power flow verification is repeated until a feasible solution is found or the computational time limit is reached, ensuring that the final reconstruction strategy is both further optimized and meets distribution network safety constraints.

[0243] In order to verify the effectiveness of a topology optimization and reconstruction method based on machine learning in an embodiment of the present invention, in one embodiment, a variety of typical large-scale distribution network reconstruction scenarios are combined to comprehensively evaluate the applicability and robustness of a topology optimization and reconstruction method based on machine learning in practical applications.

[0244] It should be noted that each scenario was modeled based on historical distribution network operating data to ensure the authenticity and representativeness of the experimental environment. Scenario 1 simulates the reconfiguration of a medium-sized distribution network with a system size of 158 nodes. This scenario has 9,638 constraints and weak nonlinear effects. Scenario 2 simulates the reconfiguration of a large distribution network with a system size of 642 nodes and 39,265 constraints. Scenario 3 simulates the reconfiguration of an extra-large distribution network with a system size of 1,090 nodes. This scenario has highly coupled constraints and a very high variable dimension. Specifically, the experimental results for each scenario are shown in Table 1.

[0245] Table 1 Experimental results in different scenarios

[0246]

[0247] Among them, the unit of time consumption is seconds, and the unit of energy consumption is kilowatt-hour. According to the analysis results, compared with traditional methods, the present invention significantly reduces the calculation time in all scenarios. In scenario one, the present invention saved about 270.4 seconds, and the saving percentage was about 25.7%; in scenario two, it saved about 981.8 seconds, and the saving percentage reached 30.4%, showing the most significant optimization effect; in scenario three, it saved about 1619.4 seconds, and the saving percentage was about 25.0%. Overall, the present invention can significantly improve computing efficiency, reduce time consumption, and is suitable for a variety of computing scenarios.

[0248] The embodiment of the present invention provides a topology optimization and reconstruction method based on machine learning, which combines the XGBoost algorithm in machine learning to predict the probability of the optimal solution value of 0-1 variables, thereby effectively improving the solution efficiency and reliability of distribution network reconstruction; performs switch prediction and screening processing based on prediction probability on real-time reconstruction case data, distinguishes different confidence decisions, directly locks the values ​​of high-confidence prediction variables, and optimizes the remaining low-confidence prediction variables by solving the target reconstruction model, which not only effectively compresses the solution space dimension of the branch and bound method, but also uses the probability prediction value to establish a feasible solution direction guide, fundamentally improving the problem of a sudden drop in solution efficiency caused by the explosion of dependent variable dimensions; the final determined reconstruction strategy reduces line power loss and optimizes the distribution network operation efficiency while ensuring the reliability and security of the reconstruction strategy.

[0249] Based on the above-mentioned topology optimization reconstruction method based on machine learning, such as Figure 3 As shown, an embodiment of the present invention provides a topology optimization and reconstruction system based on machine learning, including:

[0250] A first model building module 1 is configured to build a target distribution network reconstruction model corresponding to the distribution network based on minimizing line power loss, wherein the decision variables of the target distribution network reconstruction model include line switch states;

[0251] The second model building module 2 is used to obtain historical reconstruction case data of the distribution network, and perform supervised learning on the initial XGBoost model based on the historical reconstruction case data to obtain the target XGBoost model;

[0252] The switch prediction module 3 is used to obtain real-time reconstruction case data of the distribution network and perform line switch prediction on the real-time reconstruction case data based on the target XGBoost model to obtain the switch prediction value of each line in the distribution network and the prediction probability corresponding to each switch prediction value;

[0253] A first confidence screening module 4 is configured to screen each switch prediction value based on the prediction probability to obtain a screening result corresponding to the switch prediction value, wherein the screening result is a high confidence decision or a low confidence decision;

[0254] The first strategy determination module 5 is used to solve the target distribution network reconstruction model based on the first switch prediction value of all screening results for low-confidence decisions, and determine the first reconstruction strategy corresponding to the distribution network based on the first solution result and the second switch prediction value of all screening results for high-confidence decisions.

[0255] Furthermore, an embodiment of the present invention provides a topology optimization and reconstruction system based on machine learning, further comprising:

[0256] A first verification module is configured to perform a power flow verification on the first solution result based on a distribution network security constraint to obtain a first verification result;

[0257] a conflict rollback module, configured to, if the first verification result is a failure, perform a conflict rollback on all second switch prediction values ​​to obtain all third switch prediction values ​​that cause the conflict;

[0258] a second confidence screening module, configured to determine a fourth switch prediction value having the lowest confidence among all the third switch prediction values ​​based on the prediction probability corresponding to each third switch prediction value;

[0259] a model solving module, configured to solve a target distribution network reconstruction model based on the fourth switch prediction value and all first switch prediction values ​​to obtain a second solution result;

[0260] A second verification module is used to perform power flow verification on the second solution result based on the distribution network security constraint to obtain a second verification result;

[0261] The second strategy determination module is used to adjust the first reconstruction strategy based on the second solution result, all second switch prediction values ​​and all third switch prediction values ​​if the second verification result is successful, so as to obtain a second reconstruction strategy corresponding to the distribution network.

[0262] It should be noted that each module in the above-mentioned topology optimization and reconstruction system based on machine learning can be implemented in whole or in part through software, hardware, and a combination thereof. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the operations corresponding to the above modules. For the specific definition of a topology optimization and reconstruction system based on machine learning, please refer to the definition of a topology optimization and reconstruction method based on machine learning above. The two have the same functions and effects and will not be repeated here.

[0263] In summary, the embodiment of the present invention provides a topology optimization and reconstruction method and system based on machine learning, which combines the XGBoost algorithm in machine learning to predict the probability of the optimal solution value of 0-1 variables, effectively improving the solution efficiency and reliability of distribution network reconstruction; performing switch prediction and screening processing based on prediction probability on real-time reconstruction case data, distinguishing different confidence decisions, directly locking the values ​​of high-confidence prediction variables, and optimizing the remaining low-confidence prediction variables by solving the target reconstruction model, which not only effectively compresses the solution space dimension of the branch and bound method, but also uses the probability prediction value to establish a feasible solution direction guide, fundamentally improving the problem of a sudden drop in solution efficiency caused by the explosion of dependent variable dimensions; the final determined reconstruction strategy reduces line power loss and optimizes the distribution network operation efficiency while ensuring the reliability and security of the reconstruction strategy.

[0264] Each embodiment in this specification is described in a progressive manner, and the same or similar parts of each embodiment can be directly referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment. It should be noted that the various technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, not all possible combinations of the various technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0265] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and substitutions can be made without departing from the technical principles of the present invention. These improvements and substitutions should also be regarded as the scope of protection of the present invention.

Claims

1. A topology optimization and reconstruction method based on machine learning, characterized in that: include: Based on minimizing line power loss, constructing a target distribution network reconstruction model corresponding to the distribution network, wherein the decision variables of the target distribution network reconstruction model include line switch status; Obtaining historical reconstruction case data of the distribution network, and performing supervised learning on an initial XGBoost model based on the historical reconstruction case data to obtain a target XGBoost model; Acquire real-time reconstruction case data of the distribution network, and perform line switching prediction on the real-time reconstruction case data based on the target XGBoost model to obtain a switching prediction value of each line in the distribution network and a prediction probability corresponding to each switching prediction value; Filtering each of the switch prediction values ​​based on the prediction probability to obtain a filtering result corresponding to the switch prediction value, wherein the filtering result is a high confidence decision or a low confidence decision; Solving the target distribution network reconstruction model based on the first switch prediction values ​​for all the screening results that are low-confidence decisions, and determining a first reconstruction strategy corresponding to the distribution network according to the first solution result and the second switch prediction values ​​for all the screening results that are high-confidence decisions; Performing a power flow check on the first solution result based on the distribution network security constraint to obtain a first check result; If the first verification result is failure, performing conflict rollback on all the second switch prediction values ​​to obtain all third switch prediction values ​​that cause conflicts; Determining a fourth switch prediction value having the lowest confidence among all the third switch prediction values ​​based on the prediction probability corresponding to each of the third switch prediction values; Solving the target distribution network reconstruction model based on the fourth switch prediction value and all the first switch prediction values ​​to obtain a second solution result; Performing a power flow check on the second solution result based on the distribution network security constraint to obtain a second check result; If the second verification result is successful, the first reconstruction strategy is adjusted based on the second solution result, all the second switch prediction values ​​and all the third switch prediction values ​​to obtain a second reconstruction strategy corresponding to the distribution network.

2. The topology optimization and reconstruction method based on machine learning according to claim 1, characterized in that: The target distribution network reconstruction model corresponding to the distribution network is constructed based on minimizing line power loss, including: With the goal of minimizing line power loss, an initial distribution network reconstruction model corresponding to the distribution network is constructed; The initial distribution network reconstruction model is constrained to obtain a target distribution network reconstruction model corresponding to the distribution network, wherein the constraints include power load balance constraints, voltage stability constraints, line load capacity constraints, topology connectivity constraints and equipment capacity constraints.

3. The topology optimization and reconstruction method based on machine learning according to claim 2, characterized in that: The target distribution network reconstruction model is characterized by the following formula: Among them, F represents the line power consumption, N represents the number of lines, represents the current of the i-th circuit, represents the resistance of the i-th circuit.

4. The topology optimization and reconstruction method based on machine learning according to claim 1, characterized in that: The obtaining of historical reconstruction case data of the distribution network, and performing supervised learning on an initial XGBoost model based on the historical reconstruction case data to obtain a target XGBoost model, includes: Acquiring historical reconstruction case data of the distribution network, and performing feature extraction on the historical reconstruction case data to obtain an initial distribution network reconstruction data set; Preprocessing the initial distribution network reconstruction dataset to obtain a target distribution network reconstruction dataset, wherein the preprocessing includes missing value filling, outlier removal and data standardization; Solving the target distribution network reconstruction model based on the target distribution network reconstruction dataset to obtain a label for each sample in the target distribution network reconstruction dataset, wherein the label is a binary value of a circuit breaker state; Using stratified sampling to divide the target distribution network reconstruction data set into a training set and a test set; Training the initial XGBoost model based on the training set to obtain an intermediate XGBoost model, wherein the training process includes initializing the model, iteratively constructing a decision tree through gradient boosting, and preventing overfitting; The intermediate XGBoost model is evaluated based on the test set, and the intermediate XGBoost model is optimized according to the obtained evaluation results to obtain a target XGBoost model, wherein the evaluation results include accuracy, precision, recall, F1 score and AUC, and the optimization process includes adjusting model hyperparameters.

5. The topology optimization and reconstruction method based on machine learning according to claim 1, characterized in that: The acquiring of real-time reconstruction case data of the distribution network, and performing line switching prediction on the real-time reconstruction case data based on the target XGBoost model to obtain a switching prediction value of each line in the distribution network and a prediction probability corresponding to each switching prediction value, includes: The following formula is used to represent the predicted probability corresponding to the switch prediction value: in, represents the predicted probability corresponding to the switch prediction value of the i-th line, represents the switch prediction value of the i-th line.

6. The topology optimization and reconstruction method based on machine learning according to claim 1, characterized in that: The filtering of each of the switch prediction values ​​based on the prediction probability to obtain a filtering result corresponding to the switch prediction value includes: Each of the switch prediction values ​​is screened based on a comparison result between the prediction probability and a preset confidence threshold to obtain a screening result corresponding to the switch prediction value.

7. The topology optimization and reconstruction method based on machine learning according to claim 6, characterized in that: The step of screening each of the switch prediction values ​​based on the comparison result between the prediction probability and the preset confidence threshold to obtain a screening result corresponding to the switch prediction value includes: If the predicted probability is greater than or equal to a preset confidence threshold or the predicted probability is less than or equal to the complement of the preset confidence threshold, determining that the screening result corresponding to the switch prediction value is a high confidence decision; If the predicted probability is less than the preset confidence threshold and the predicted probability is greater than the complement of the preset confidence threshold, it is determined that the screening result corresponding to the switch prediction value is a low confidence decision.

8. A topology optimization and reconstruction system based on machine learning, characterized in that: include: A first model building module is configured to build a target distribution network reconstruction model corresponding to the distribution network based on minimizing line power loss, wherein the decision variables of the target distribution network reconstruction model include line switch states; A second model building module is configured to obtain historical reconstruction case data of the distribution network, and perform supervised learning on the initial XGBoost model based on the historical reconstruction case data to obtain a target XGBoost model; a switch prediction module, configured to obtain real-time reconstruction case data of the distribution network, and perform line switch prediction on the real-time reconstruction case data based on the target XGBoost model, to obtain a switch prediction value for each line in the distribution network and a prediction probability corresponding to each switch prediction value; a first confidence screening module, configured to screen each of the switch prediction values ​​based on the prediction probability to obtain a screening result corresponding to the switch prediction value, wherein the screening result is a high confidence decision or a low confidence decision; a first strategy determination module, configured to solve the target distribution network reconstruction model based on the first switch prediction values ​​for all the screening results of which are low-confidence decisions, and determine a first reconstruction strategy corresponding to the distribution network according to the first solution result and the second switch prediction values ​​for all the screening results of which are high-confidence decisions; A first verification module is configured to perform a power flow verification on the first solution result based on a distribution network security constraint to obtain a first verification result; a conflict rollback module, configured to, if the first verification result is a failure, perform a conflict rollback on all the second switch prediction values ​​to obtain all third switch prediction values ​​that cause the conflict; a second confidence screening module, configured to determine a fourth switch prediction value having the lowest confidence among all the third switch prediction values ​​based on the prediction probability corresponding to each of the third switch prediction values; a model solving module, configured to solve the target distribution network reconstruction model based on the fourth switch prediction value and all the first switch prediction values ​​to obtain a second solution result; A second verification module is configured to perform a power flow verification on the second solution result based on the distribution network security constraint to obtain a second verification result; A second strategy determination module is used to adjust the first reconstruction strategy based on the second solution result, all the second switch prediction values ​​and all the third switch prediction values ​​if the second verification result is successful, to obtain a second reconstruction strategy corresponding to the distribution network.

Images