Method and device for extracting key influence nodes of power flow section power adjustment

Abstract

The invention discloses a method and a device for extracting key influence nodes of power flow section power adjustment. The method comprises the following steps: acquiring state factor data of a power system; obtaining an optimal decision tree model based on the improved VIPER algorithm and the state factor data; obtaining each node in the optimal decision tree model; acquiring a corresponding relationship between each node and a physical node of the power system; according to a preset measurement index, evaluating the importance degree of each node on power flow section power adjustment of the power system; and extracting power flow section power adjustment key influence nodes from the physical nodes of the power system based on the corresponding relationship and the importance degree of each node on power flow section power adjustment of the power system. According to the improved VIPER algorithm based on the compression thought, the decision tree structure is analyzed, the importance degree of each node on power adjustment of the power flow section is evaluated, features or attributes which have great influence on a prediction result can be efficiently and accurately recognized, and then key influence nodes of power adjustment of the power flow section can be efficiently and accurately extracted.

Description

Technical Field

[0001] The invention relates to the field of electric power systems, and in particular to a method and a device for extracting key influencing nodes for power adjustment of a power flow section. Background Art

[0002] In the operation and control of power systems, power flow analysis is an important tool that can help power system operators understand the distribution of power flow, thereby ensuring the safe and reliable operation of the power system. Traditional power flow analysis methods mainly include power flow calculation and section power adjustment. Power flow calculation uses numerical algorithms such as the Newton-Raphson method and the PQ decomposition method to calculate the voltage and phase angle of each node in the power system, thereby obtaining the power flow distribution of the power system. Section power adjustment changes the power flow distribution by adjusting the generator output, transformer tap position, and reactive compensation equipment to control the power flow of the section. In power flow analysis, section power adjustment is a key issue. By adjusting the section power, the distribution of power flow can be controlled to meet the safety and economy requirements of the power system.

[0003] In power flow analysis, the extraction of key influencing nodes is an important research direction. Existing key node extraction methods mainly include sensitivity analysis, concentration analysis, network structure analysis and measured data driven method. The above methods can extract key influencing nodes in the power system to a certain extent, but there are still some shortcomings. The sensitivity analysis method usually requires a lot of computing resources, and the results are easily affected by changes in the system state and have poor stability. The concentration analysis method relies on the topological structure of the power system and cannot fully consider the dynamic changes in actual operation. The network structure analysis method focuses on static structural analysis and is difficult to reflect the real-time changes and complexity in actual operation. Although the measured data driven method can reflect the actual operation situation, it requires a large amount of historical data and real-time monitoring data, and it is difficult to predict future trends.

[0004] As the scale of power systems continues to expand and the operating environment becomes more complex, there is an urgent need to propose a more efficient and accurate key node extraction method. Summary of the invention

[0005] The present invention provides a method and device for extracting key influencing nodes of power regulation of a power section in order to solve the above technical problems, and can efficiently and accurately extract key influencing nodes of power regulation of a power section in order to solve the above technical problems.

[0006] In order to solve the above technical problems, the present invention provides a method for extracting key influencing nodes of power adjustment of power flow section, comprising:

[0007] Obtaining status factor data of the power system;

[0008] Obtaining an optimal decision tree model based on the improved VIPER algorithm and the state factor data;

[0009] Based on the best decision tree model, each node in the best decision tree model is obtained;

[0010] Obtain the corresponding relationship between each node and the physical node of the power system;

[0011] Evaluate the importance of each node to the power adjustment of the power flow section of the power system according to the preset metric indicators;

[0012] Based on the corresponding relationship and the importance of each node to the power regulation of the power flow section of the power system, the key influencing nodes of the power flow section power regulation are extracted from the physical nodes of the power system.

[0013] The beneficial effects of the present invention are:

[0014] The present invention obtains the best decision tree model and each node in the best decision tree model based on the improved VIPER algorithm and power system data, and extracts the key influencing nodes of the power flow section power adjustment from the physical nodes of the power system based on the corresponding relationship and the importance of each node to the power flow section power adjustment of the power system. The improved VIPER algorithm is based on the model compression idea of ​​the VIPER algorithm, which can reduce the computational complexity of obtaining the best decision tree model and improve the efficiency of computational analysis. The decision tree model is easy to understand and explain, and its nodes are used to characterize the characteristics or attributes of a power system, and the correlation between the characteristics can be better explored and utilized. By analyzing the structure of the decision tree and evaluating the importance of each node to the power flow section power adjustment of the power system, the characteristics or attributes that have a greater impact on the prediction result can be efficiently and accurately identified, and then the key influencing nodes of the power flow section power adjustment can be efficiently and accurately extracted.

[0015] As a preferred solution, the method of obtaining the best decision tree model based on the improved VIPER algorithm and the state factor data includes:

[0016] Based on the improved VIPER algorithm, a number of state-action data and category distributions are generated using the preset reinforcement learning model and state factor data;

[0017] Get an initial data set based on some state-action data and category distribution;

[0018] Based on the improved VIPER algorithm, the initial data set is iteratively updated and the preset decision tree model is trained based on the initial data set after each iteration until the iteration condition is met, and N classification models are obtained, where N is a positive integer;

[0019] Get the best decision tree model among N classification models.

[0020] As a preferred solution, the method of iteratively updating the initial data set based on the improved VIPER algorithm and training a preset decision tree model based on the initial data set after each iteration includes:

[0021] Based on the improved VIPER algorithm, each category in the initial data set is resampled to obtain the training data set;

[0022] Train the preset decision tree model based on the training data set to obtain the current classification model;

[0023] Control the current classification model to generate new state data under the initial state of random power flow;

[0024] Obtain new action data of the preset reinforcement learning model under new state data;

[0025] Obtain new state-action data based on the new state data and the new action data;

[0026] It is determined whether to merge the new state-action data into the initial data set according to the preset conditions, so as to update the initial data set and make the initial data set meet the preset state.

[0027] As a preferred solution, the loss function of each iteration is as follows:

[0028]

[0029] In the above formula, h t (s) refers to the proportion of actions taken at time t in the initial data set under state s, t∈{0,…,T-1}; is the action value function; is the state value function.

[0030] As a preferred solution, the step of determining whether to merge the new state-action data into the initial data set according to a preset condition includes:

[0031] Perform value assessment on the new state-action data to obtain the assessment result;

[0032] Calculate the retention probability of the new state-action data based on the evaluation results;

[0033] When it is determined that the retention probability is greater than the retention probability threshold, the new state-action data is merged into the initial data set.

[0034] As a preferred solution, the preset state includes: the data of each category in the initial data set is balanced and the category distribution of the initial data set is maintained.

[0035] Accordingly, in order to solve the above technical problems, the present invention also provides a device for extracting key influencing nodes of power adjustment of power flow section, comprising: a data acquisition module, a model acquisition module, a node evaluation module and a node extraction module;

[0036] Wherein, the data acquisition module is used for acquiring the state factor data of the power system;

[0037] The model acquisition module is used to acquire the best decision tree model based on the improved VIPER algorithm and the state factor data;

[0038] The node evaluation module is used to obtain each node in the optimal decision tree model based on the optimal decision tree model; obtain the corresponding relationship between each node and the physical node of the power system; and evaluate the importance of each node to the power adjustment of the power flow section of the power system according to the preset metric;

[0039] The node extraction module is used to extract the key influencing nodes of power flow section power adjustment from the physical nodes of the power system based on the corresponding relationship and the importance of each node to the power flow section power adjustment of the power system.

[0040] As a preferred solution, the model acquisition module includes: a data set generation unit, an iterative update unit and a model acquisition unit;

[0041] The data set generation unit is used to generate a number of state-action data and category distributions based on the improved VIPER algorithm using a preset reinforcement learning model and state factor data; and obtain an initial data set based on the number of state-action data and category distributions;

[0042] The iterative updating unit is used to iteratively update the initial data set based on the improved VIPER algorithm and train a preset decision tree model based on the initial data set after each iteration until an iteration condition is met to obtain N classification models, where N is a positive integer;

[0043] The model acquisition unit is used to acquire the best decision tree model from N classification models.

[0044] As a preferred solution, the iterative updating unit is used to iteratively update the initial data set based on the improved VIPER algorithm and train a preset decision tree model based on the initial data set after each iteration, including:

[0045] Based on the improved VIPER algorithm, each category in the initial data set is resampled to obtain the training data set;

[0046] Train the preset decision tree model based on the training data set to obtain the current classification model;

[0047] Control the current classification model to generate new state data under the initial state of random power flow;

[0048] Obtain new action data of the preset reinforcement learning model under new state data;

[0049] Obtain new state-action data based on the new state data and the new action data;

[0050] It is determined whether to merge the new state-action data into the initial data set according to the preset conditions, so as to update the initial data set and make the initial data set meet the preset state.

[0051] As a preferred solution, the loss function of each iteration is as follows:

[0052]

[0053] In the above formula, h t (s) refers to the proportion of actions taken at time t in the initial data set under state s, t∈{0,…,T-1}; is the action value function; is the state value function. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 : A flow chart of an embodiment of a method for extracting key influencing nodes for power adjustment of a power flow section provided by the present invention;

[0055] Figure 2 : A flow chart of the VIPER algorithm provided in an embodiment of the present invention;

[0056] Figure 3 : A schematic diagram of the structure of the improved VIPER algorithm provided by an embodiment of the present invention;

[0057] Figure 4 : A structural schematic diagram of an embodiment of the optimal decision tree model provided by the present invention;

[0058] Figure 5 : A structural schematic diagram of an embodiment of a device for extracting key influencing nodes for power adjustment of a power flow section provided by the present invention;

[0059] Figure 6 : A structural diagram of an embodiment of the model acquisition module provided by the present invention. DETAILED DESCRIPTION

[0060] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. 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.

[0061] In the operation and control of power systems, power flow analysis is an important tool that can help power system operators understand the distribution of power flow, thereby ensuring the safe and reliable operation of power systems. In power flow analysis, the extraction of key influencing nodes is an important research direction. Existing key node extraction methods mainly include sensitivity analysis, concentration analysis, network structure analysis and measured data driven method. Sensitivity analysis methods include voltage sensitivity analysis and power sensitivity analysis. By calculating the sensitivity of each node power or voltage to system changes, the nodes most sensitive to system changes are found. Concentration analysis methods include node concentration and current concentration. By calculating the electrical distance of the nodes or the concentration of current distribution, nodes with small electrical distance or high current concentration are identified. The network structure analysis method uses topological structure analysis and minimum cut set analysis to find nodes with high connectivity in the network or the greatest impact on network connectivity. The measured data driven method uses historical data analysis and real-time data monitoring to identify key nodes that frequently have problems in historical operation or fluctuate or are abnormal in real-time operation.

[0062] Although the above methods can extract key influencing nodes in the power system to a certain extent, they still have some shortcomings. Sensitivity analysis methods usually require a lot of computing resources, and the results are easily affected by changes in system status and have poor stability. The concentration analysis method relies on the topological structure of the power system and cannot fully consider the dynamic changes in actual operation. The network structure analysis method focuses on static structural analysis and is difficult to reflect the real-time changes and complexity in actual operation. Although the measured data-driven method can reflect the actual operation situation, it requires a large amount of historical data and real-time monitoring data, and it is difficult to predict future trends.

[0063] As the scale of power systems continues to expand and the operating environment becomes more complex, there is an urgent need to propose more efficient and accurate key node extraction methods. Future development directions may include intelligence and automation, combining artificial intelligence and machine learning technologies to achieve automatic analysis of power system operation data and key node extraction; real-time and dynamic, developing key node extraction methods that can reflect the dynamic changes of power systems in real time, and improving the response speed and adaptability of the system; comprehensiveness and multi-dimensionality, comprehensively considering the electrical characteristics, topological structure and operating data of the power system, and extracting and analyzing key nodes from multiple dimensions. Through research and development of related methods based on real-time operating data, the stability and efficiency of the power system can be further improved, providing a valuable scientific basis for the optimization of power system operation.

[0064] Embodiment 1

[0065] In order to solve the above technical problems, the present invention provides a method and device for extracting key influencing nodes of power regulation of a power section, which can efficiently and accurately extract key influencing nodes of power regulation of a power section.

[0066] Please refer to Figure 1 , which is a flow chart of an embodiment of a method for extracting key influencing nodes for power adjustment of a power flow section provided in an embodiment of the present invention.

[0067] like Figure 1 As shown, the method for extracting key influencing nodes of power adjustment of power flow section includes steps 101 to 106, which are specifically:

[0068] Step 101: Acquire state factor data of the power system;

[0069] Step 102: Obtaining an optimal decision tree model based on the improved VIPER algorithm and the state factor data;

[0070] Step 103: Obtain each node in the best decision tree model based on the best decision tree model;

[0071] Step 104: Obtaining the correspondence between each node and the physical node of the power system;

[0072] Step 105: Evaluate the importance of each node to the power adjustment of the power flow section of the power system according to the preset metric;

[0073] Step 106: Based on the corresponding relationship and the importance of each node to the power flow section power adjustment of the power system, extract the key influencing nodes of the power flow section power adjustment from the physical nodes of the power system.

[0074] The present invention obtains the best decision tree model and each node in the best decision tree model based on the improved VIPER algorithm and power system data, and extracts the key influencing nodes of the power flow section power adjustment from the physical nodes of the power system based on the corresponding relationship and the importance of each node to the power flow section power adjustment of the power system. The improved VIPER algorithm is based on the model compression idea of ​​the VIPER algorithm, which can reduce the computational complexity of obtaining the best decision tree model and improve the efficiency of computational analysis. The decision tree model is easy to understand and explain, and its nodes are used to characterize the characteristics or attributes of a power system, and the correlation between the characteristics can be better explored and utilized. By analyzing the structure of the decision tree and evaluating the importance of each node to the power flow section power adjustment of the power system, the characteristics or attributes that have a greater impact on the prediction result can be efficiently and accurately identified, and then the key influencing nodes of the power flow section power adjustment can be efficiently and accurately extracted.

[0075] In this embodiment, the improved VIPER algorithm is an improved algorithm based on the VIPER (Verifiability via IterativePolicy ExtRaction) algorithm, and the VIPER algorithm is an improved algorithm based on the DAGGER algorithm; the VIPER algorithm combines the ideas of model compression and imitation learning methods, uses the Q function of reinforcement learning to transform the DAGGER algorithm, and uses resampling for model compression, and trains a decision tree strategy with similar performance to the reinforcement learning model, thereby explaining the decision-making behavior of the reinforcement learning agent to a certain extent.

[0076] Among them, the imitation learning method mainly imitates the reinforcement learning model through a decision tree. The main problem involved in the reinforcement learning model is that the interpretability is very weak. It can basically be regarded as a black box model. If you want to obtain information from the reinforcement learning model and data, you can think of using other models with strong interpretability to imitate the reinforcement learning model. When you can have a high consistency with the original reinforcement learning model in terms of input and output, you can extract valuable knowledge from the model with strong interpretability. For the problem of selecting the power adjustment of the tidal section in the present invention, the decision tree model is completely in line with this idea, has natural interpretability, and is simple and easy to understand. A decision tree is a tree structure for classifying instances, which is divided into two parts: nodes and directed edges. Nodes are divided into two types: internal nodes and leaf nodes. Among them, the internal node represents a test condition of a feature or attribute (used to record different characteristics respectively), and the leaf node represents a category. Once a decision tree model is completed, it is easy to classify based on this decision tree model. The specific steps are to start from the root node, start testing a certain attribute of the instance, and match each instance to its corresponding child node (that is, an appropriately selected branch) according to the test structure; when a leaf node or another internal node is reached along the branch, the new test condition can be used to execute recursively until another leaf node is reached; when the leaf node is finally reached, the final classification result can be obtained.

[0077] Please refer to the following Figure 2 , Figure 2 A flowchart of the VIPER algorithm provided in an embodiment of the present invention.

[0078] like Figure 2 As shown, the process of the VIPER algorithm includes:

[0079] 1) First, the state-action data randomly generated by the reinforcement learning model is used as the initial data set D0;

[0080] 2) Resample the data set D0 based on the range of the Q value to obtain the training data set D0';

[0081] 3) Train the decision tree model on the resampled data set to obtain the i-th classification model; if all iterations have been completed, go to 5), otherwise continue;

[0082] 4) Generate new states based on the current classification model, and form new state-action data with the corresponding actions of the reinforcement learning model under these states, and merge them into the data set D0, and go to 2);

[0083] 5) Select the most appropriate classification model by evaluating all trained classification models in actual scenarios.

[0084] Among them, the generation and analysis of state-action data is an important step in the application of imitation learning in complex systems such as power systems. In order to convert the adjustment sequence data into state-action data suitable for imitation learning, the following steps need to be followed: first, define the state and action, then from the beginning of the adjustment sequence, record the current grid state as the initial state; for each adjustment action in the adjustment sequence, record the previous state, execute the action, and prepare the next state, and then combine the state before each adjustment action with the action itself into a state-action pair as an instance of training data; finally, perform necessary cleaning on the generated data, such as removing outliers and standardizing, to improve the efficiency and effect of subsequent model training. By converting the adjustment sequence into a state-action data pair, it is possible to capture the expert's operating experience in a form that can be understood by the machine learning model, providing strong support for subsequent model training and optimization.

[0085] Based on the DAGGER algorithm, the VIPER algorithm has two main improvements: one is the introduction of the Q function; the other is the addition of a resampling step. Through the Q function, the VIPER algorithm can help "experts" make optimal decisions for specific reinforcement learning models in complex environments; and the main purpose of resampling is to change the sampling rate of the signal to adapt to different emergency needs, thereby reducing the size of the decision tree nodes.

[0086] The theoretical analysis of the VIPER algorithm is consistent with the DAGGER algorithm. The theoretical analysis process of the VIPER algorithm is as follows:

[0087] Problem definition: Let (S, A, P, R) represent a finite-range Markov decision process MDP, where state S, action A, transition probability P: S×A×S→[0,1], and reward is Given a strategy π:S→A,t∈{0,…,T-1}, the action value function Q t (π) (s,a) and state value function V t (π) The formulas are as follows:

[0088]

[0089] Among them, V T (π) (s) = 0, assuming that there is an initial state s0∈S, then the state distribution at time step t is expressed as follows:

[0090]

[0091] The overall state distribution of strategy π is shown in the formula below; the cost representation of strategy π starting from s0 is shown in the formula below. The goal is to * : S→A and Q value Learn the best strategy in each situation.

[0092]

[0093] J(π)=-V0 (π) (s0);

[0094] The loss function based on Q value is expressed as l t (s,π), let g(s,π) be a 0-1 loss function, and is a convex upper bound of g(s,π), and the Q-DAGGER loss function is defined as in yes t The convex upper bounds of (s,π) are as follows:

[0095]

[0096] g(s,π)=Ⅱ[π(s)≠π * (s)];

[0097]

[0098] The loss function l for each iteration during training t (π) is defined as follows:

[0099]

[0100] Due to the non-concave nature of the decision tree loss function, there is no theoretically based online learning algorithm required by the DAGGER algorithm. However, a heuristic method based on the follow-the-leader algorithm can be used - in each iteration, a decision tree is trained on the aggregated dataset using a CART decision tree. Assume π * and In order to compress the model, instead of optimizing the loss function of CART, it is better to use Resample the samples (s, a)∈D0 with weights, and the sampling probability p((s, a)) is shown in the following formula.

[0101]

[0102] Because the resampling satisfies the following: The expectation of using CART to train a decision tree on D0' is the same as training with The expectation of the decision tree with loss function is equivalent, as shown in the following formula, so resampling is reasonable.

[0103]

[0104] The VIPER algorithm improves the DAGGER algorithm through Q functions and resampling techniques, optimizing the performance and explanatory power of the decision tree model. The VIPER algorithm first generates an initial state-action data set, then uses the extreme resampled data of the Q value to train the decision tree model, and continuously updates the data set and decision tree model in the iterative process, and finally selects the best model for practical application.

[0105] Furthermore, in step 101, the state factor data of the power system includes load demand, output of generator sets, topological structure of the power grid, line status, current, voltage, etc.

[0106] Further, in step 102, the obtaining of the best decision tree model based on the improved VIPER algorithm and the state factor data includes:

[0107] Based on the improved VIPER algorithm, a number of state-action data and category distributions are generated using the preset reinforcement learning model and state factor data;

[0108] Get an initial data set based on some state-action data and category distribution;

[0109] Based on the improved VIPER algorithm, the initial data set is iteratively updated and the preset decision tree model is trained based on the initial data set after each iteration until the iteration condition is met, and N classification models are obtained, where N is a positive integer;

[0110] Get the best decision tree model among N classification models.

[0111] In the reinforcement learning application of power system dispatch or control, the state space usually contains various state factor data of the power system, which are abstracted as state variables and used to construct the state space model of the power system. These state variables are the environmental information that the reinforcement learning agent can observe when making decisions. In this embodiment, based on the improved VIPER algorithm, a preset reinforcement learning model is used to generate state-action data based on the environmental information that the agent can observe when making decisions, and the generated state-action data is used as the initial data set.

[0112] In this embodiment, all trained classification models are evaluated in the actual power system section power adjustment, and the best decision tree model is selected as the optimal decision tree model. In addition, in actual scenarios such as load forecasting of power systems, risk assessment of financial markets, and decision support for medical diagnosis, relevant data sets can be collected according to the needs of actual scenarios, and the data sets can be divided into multiple subsets. Different subsets are used as test sets each time, and the remaining subsets are used as training sets. The training and testing process is repeated multiple times, and finally the average performance is taken as the final performance of the model.

[0113] Furthermore, the iterative updating of the initial data set based on the improved VIPER algorithm and training of a preset decision tree model based on the initial data set after each iteration include:

[0114] Based on the improved VIPER algorithm, each category in the initial data set is resampled to obtain the training data set;

[0115] Train the preset decision tree model based on the training data set to obtain the current classification model;

[0116] Control the current classification model to generate new state data under the initial state of random power flow;

[0117] Obtain new action data of the preset reinforcement learning model under new state data;

[0118] Obtain new state-action data based on the new state data and the new action data;

[0119] It is determined whether to merge the new state-action data into the initial data set according to the preset conditions, so as to update the initial data set and make the initial data set meet the preset state.

[0120] Furthermore, the loss function of each iteration is as follows:

[0121]

[0122] In the above formula, h t (s) refers to the proportion of actions taken at time t in the initial data set under state s, t∈{0,…,T-1}; is the action value function; is the state value function.

[0123] Further, judging whether to merge the new state-action data into the initial data set according to the preset conditions includes:

[0124] Perform value assessment on the new state-action data to obtain the assessment result;

[0125] Calculate the retention probability of the new state-action data based on the evaluation results;

[0126] When it is determined that the retention probability is greater than the retention probability threshold, the new state-action data is merged into the initial data set.

[0127] In this embodiment, when merging the new state-action data into the initial data set, attention needs to be paid to the integrity and consistency of the data, and the data set is merged according to the format and specification of the data set to ensure that the new data can be seamlessly connected with the original data.

[0128] Furthermore, the preset state includes: data of each category in the initial data set is balanced and the category distribution of the initial data set is maintained.

[0129] In this embodiment, the improved VIPER algorithm also considers balancing the category distribution and keeping the category distribution based on the VIPER algorithm. This is because the categories in the state-action data are very unbalanced, and its category distribution problem is shown in the VIPER algorithm. The improved VIPER algorithm of the present invention is the B&KDVIPER (Balance and Keep DistributionVIPER) algorithm. Among them, balancing the category distribution refers to adjusting the number of samples of each category in the data set by a certain method so that the number of samples of different categories is relatively balanced, which can avoid the model from being overly biased towards the majority class samples during the training process, thereby ignoring the important features of the minority class samples, causing the model to be biased during prediction. Keeping the category distribution means that in the process of adjusting the category distribution of the data set, the category distribution information of the original data set is kept unchanged as much as possible. This is because, although balancing the category distribution can improve the performance of the model, if the category distribution of the original data is excessively changed, it may cause the model to learn wrong category information, thereby affecting the generalization ability of the model. Therefore, the present invention determines whether to merge the new state-action data into the initial data set according to the preset conditions to update the initial data set and make the initial data set meet the preset state, and the preset state includes: the data of each category in the initial data set is balanced and the category distribution of the initial data set is maintained, thereby, the number of samples of different categories can be relatively balanced to avoid the model being overly biased towards the majority class samples during the training process and ignoring the important features of the minority class samples, resulting in deviations in the model when predicting, and when balancing the number of samples, avoid excessive changes in the category distribution of the original data, resulting in the model learning wrong category information, thereby affecting the generalization ability of the model. The improved VIPER algorithm optimizes data utilization efficiency by balancing and maintaining category distribution, and further improves the performance and interpretability of the decision tree model.

[0130] In this embodiment, the theoretical difference between the improved VIPER algorithm and the VIPER algorithm is that the loss function has a slight change, and the loss function of each iteration is shown in the following formula:

[0131]

[0132] Among them, h t (s) refers to the proportion of actions taken at time t in state s in the initial data set D, that is, the proportion of decisions or choices made by the agent based on the current state of the environment, directly collected from the data source without any preprocessing; that is, when calculating the loss function, an additional weight is added.

[0133] When resampling, the sampling probability of the improved VIPER algorithm becomes the following formula:

[0134]

[0135] From the theoretical analysis of the DAGGER algorithm, we know that for any δ>0, when When there is a strategy There is at least 1-δ probability that The improved VIPER algorithm also meets this conclusion.

[0136] Please refer to the following Figure 3 , Figure 3 A schematic diagram of the flow of the improved VIPER algorithm provided in an embodiment of the present invention.

[0137] like Figure 3 As shown, the process of improving the VIPER algorithm is as follows:

[0138] 1) First, a reinforcement learning model is used, using the environmental information that the agent can observe when making decisions, that is, randomly generated state-action data as the initial data set D;

[0139] 2) Use the range of Q values ​​as weights to resample each category in the initial data set D to obtain the training data set D';

[0140] 3) Train the decision tree on the resampled data set; if all iterations have been completed, go to 5), otherwise continue;

[0141] 4) In order to use a trained classification model to generate new states under the random initial state of the flow, and combine these new states with the actions corresponding to the reinforcement learning model under these states into new state-action data, the retention probability of each category is set according to the category distribution generated by the reinforcement learning model in the initial data set D, and the probability is used to determine whether to merge the state-action data into the initial data set D, and go to 2);

[0142] The specific steps of whether to merge the state-action data into the initial data set D are: first, define the retention probability threshold, then evaluate the value of the newly generated state-action data, calculate the retention probability based on the evaluation result, compare the retention probability with the threshold, and finally merge the state-action data that meets the retention conditions according to the format and specifications of the data set;

[0143] 5) Evaluate all trained decision tree models in actual scenarios and select the best decision tree model.

[0144] A decision tree is a machine learning algorithm that builds classification models by recursively splitting a dataset into smaller subsets. Each internal node represents a test or decision on an attribute, while each leaf node represents a class label.

[0145] In the power system, since the VIPER algorithm is not designed for power system analysis, it may not be directly feasible to directly use the improved VIPER algorithm to extract key influencing nodes. However, by drawing on the core idea of ​​identifying key nodes in the VIPER algorithm, that is, identifying nodes that have a significant impact on the performance of the power system by analyzing factors such as the location, connectivity, and influence of nodes in the network, it can be applied to the extraction of key influencing nodes of the power flow section. Although the nodes in the decision tree model are conceptually different from the physical nodes in the power system, the decision tree model can be used to identify the key factors or attributes that affect the performance of the power grid, and then infer that the physical nodes associated with these key factors may be key influencing nodes. Therefore, the present invention selects features or attributes that may affect the performance of the power system from the power system, collects historical data containing these features, performs necessary preprocessing, and uses the processed data to train a decision tree classification or regression model; by analyzing the structure of the decision tree, the features or attributes that have the greatest impact on the model prediction results are identified. These features or attributes are usually located at the top of the decision tree or near the root node because they can quickly divide the data set into subsets with different category labels; based on the results of the decision tree analysis, the features or attributes with significant impact are mapped back to the physical nodes in the power system; if a certain feature (such as generator power) has a significant impact on system performance, then the generator node associated with the feature may be the key influencing node.

[0146] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of an embodiment of the optimal decision tree model provided by the present invention. Figure 4 As shown, each node in the decision tree is related to the feature of generator power, and changes in generator power will significantly affect the power system.

[0147] Furthermore, in step 105, the preset metric index may be, but is not limited to, indexes such as degree centrality, closeness centrality, and betweenness centrality of the node.

[0148] The method for extracting key influencing nodes of power adjustment of power section of the present invention can efficiently and accurately extract key influencing nodes of power adjustment of power section of the power section. Furthermore, the extracted key influencing nodes of power adjustment of power section of the power section can provide reliable guidance direction for power adjustment of power section of the power system.

[0149] In the power system, power adjustment of the power flow section is a key link to ensure the stable and efficient operation of the power grid. However, the existing technology has shown significant limitations and shortcomings when making this adjustment.

[0150] First, many traditional adjustment methods rely too much on the personal experience and intuition of engineers. Although this method is intuitive and easy to understand, it is difficult to ensure the stability and consistency of its adjustment effect. Different engineers may give completely different adjustment plans based on their own experience and judgment, which leads to the unpredictability of the adjustment effect. Especially in the face of complex and changeable power system operating conditions, this method that relies on human experience is more challenging and difficult to adapt to rapidly changing needs.

[0151] Secondly, although the method based on the optimization algorithm can construct a complex mathematical model and solve the optimal solution through the algorithm, the computational complexity of this method is often very high. As the scale of the power grid expands and the number of nodes increases, the computational complexity will increase sharply, resulting in the solution process taking too long and even failing to meet the needs of real-time adjustment. In addition, the initial parameters are very sensitive to different settings of the optimization algorithm, and different parameter settings may lead to very different setting results, which further increases the uncertainty and complexity of the adjustment process.

[0152] More importantly, the adjustment methods in the existing technology often find it difficult to find a balance between adjustment accuracy and efficiency. Some methods may pursue high-precision adjustment results, but sacrifice the efficiency of adjustment, resulting in the adjustment process taking too long and failing to respond to changes in the power system in a timely manner. Other methods may pursue a high-efficiency adjustment process, but compromise on adjustment accuracy and fail to meet the power system's demand for high-precision adjustment. This trade-off between accuracy and efficiency makes it difficult for existing technologies to meet the comprehensive requirements of modern power systems for efficient, accurate, and real-time adjustment.

[0153] Finally, most of the adjustment methods in the existing technologies lack intelligent and automated features. They often require real-time human intervention and judgment, and cannot implement automated and intelligent adjustment strategies. This not only increases labor costs, but also reduces the efficiency and accuracy of the adjustment process. With the rapid development of smart grids and automation technologies, the power system has put forward higher requirements for the intelligence and automation level of the adjustment methods. Therefore, the deficiencies of the existing technologies in this regard also limit their application effects in the dispatching, operation and control of power systems.

[0154] In summary, the existing technology has the disadvantages of relying on manual experience, high calculation complexity, difficulty in balancing adjustment accuracy and efficiency, and lack of intelligence and automation when adjusting power in power flow sections.

[0155] Based on the method for extracting key influencing nodes of power adjustment of power section of the present invention, the key influencing nodes of power adjustment of power section of the power section extracted efficiently and accurately can be reliably applied to the power adjustment of power section of the power section, so that the power adjustment of power section of the power section can simultaneously meet the requirements of adjustment accuracy, efficiency and intelligence, and improve the safety of power system operation.

[0156] Accordingly, in order to solve the above technical problems, the embodiment of the present invention also provides a device for extracting key influencing nodes of power adjustment of power flow section. Figure 5 , which is a structural schematic diagram of an embodiment of a device for extracting key influencing nodes for power adjustment of a power flow section provided by the present invention.

[0157] like Figure 5 As shown, the device 20 for extracting key influencing nodes of power adjustment of power flow section includes: a data acquisition module 201, a model acquisition module 202, a node evaluation module 203 and a node extraction module 204;

[0158] Wherein, the data acquisition module 201 is used for acquiring the state factor data of the power system;

[0159] The model acquisition module 202 is used to acquire the best decision tree model based on the improved VIPER algorithm and the state factor data;

[0160] The node evaluation module 203 is used to obtain each node in the optimal decision tree model based on the optimal decision tree model; obtain the corresponding relationship between each node and the physical node of the power system; and evaluate the importance of each node to the power adjustment of the power flow section of the power system according to the preset metric;

[0161] The node extraction module 204 is used to extract the key influencing nodes of power flow section power adjustment from the physical nodes of the power system based on the corresponding relationship and the importance of each node to the power flow section power adjustment of the power system.

[0162] Please refer to Figure 6 , Figure 6 A schematic diagram of the structure of an embodiment of the model acquisition module provided by the present invention.

[0163] Furthermore, if Figure 6 As shown, the model acquisition module 202 includes: a data set generation unit 2021, an iterative update unit 2022 and a model acquisition unit 2023;

[0164] The data set generating unit 221 is used to generate a plurality of state-action data and category distributions based on the improved VIPER algorithm using a preset reinforcement learning model and state factor data; and obtain an initial data set based on the plurality of state-action data and category distributions;

[0165] The iterative updating unit 2022 is used to iteratively update the initial data set based on the improved VIPER algorithm and train a preset decision tree model based on the initial data set after each iteration until the iteration condition is met to obtain N classification models, where N is a positive integer;

[0166] The model acquisition unit 2023 is used to obtain the best decision tree model among N classification models.

[0167] Furthermore, the iterative updating unit is used to iteratively update the initial data set based on the improved VIPER algorithm and train a preset decision tree model based on the initial data set after each iteration, including:

[0168] Based on the improved VIPER algorithm, each category in the initial data set is resampled to obtain the training data set;

[0169] Train the preset decision tree model based on the training data set to obtain the current classification model;

[0170] Control the current classification model to generate new state data under the initial state of random power flow;

[0171] Obtain new action data of the preset reinforcement learning model under new state data;

[0172] Obtain new state-action data based on the new state data and the new action data;

[0173] It is determined whether to merge the new state-action data into the initial data set according to the preset conditions, so as to update the initial data set and make the initial data set meet the preset state.

[0174] Furthermore, the loss function of each iteration is as follows:

[0175]

[0176] In the above formula, h t (s) refers to the proportion of actions taken at time t in the initial data set under state s, t∈{0,…,T-1}; is the action value function; is the state value function.

[0177] The existing key node extraction methods are insufficient in terms of real-time and dynamics, and it is difficult to reflect the dynamic changes of the power system in real time without increasing the burden of computing resources; secondly, the stability and adaptability of the existing methods are insufficient, especially the sensitivity analysis method is easily affected by changes in the system state, while the concentration analysis method and the network structure analysis method cannot fully consider the dynamic changes in actual operation; in addition, most of the existing methods rely solely on electrical characteristics or topological structures, lack comprehensive analysis of power system operation data, and fail to extract and analyze key nodes from multiple dimensions; finally, although the measured data-driven method can use real-time monitoring methods such as SCADA systems to identify problem nodes in operation, it is insufficient in predicting future trends and lacks foresight.

[0178] Compared with the prior art, the embodiments of the present invention have the following beneficial effects:

[0179] By comprehensively considering the electrical characteristics, topological structure and real-time operation data of the power system, analyzing from multiple dimensions, and combining real-time monitoring data and prediction models, the present invention aims to provide an efficient, stable and adaptable key node extraction method. By imitating the reinforcement learning model through the decision tree, it is convenient to extract information from the model with strong interpretability, and can clearly display the results of different feature tests. It has good stability and predictability, thereby improving the safety and economy of the power system.

[0180] Specifically, the present invention greatly improves the adaptability and flexibility of the system by combining physical models and data-driven technology, can better cope with various changes and uncertainties in the power system, reduce dependence on a single physical model, and can improve the adaptability and analysis accuracy of the power system; the decision tree model is used to imitate the decision-making process of the reinforcement learning model to enhance the interpretability of the system, making the analysis results more transparent and interpretable, easy to understand and apply; the advanced algorithm optimization and resampling technology are adopted to significantly reduce the computational complexity of large-scale systems, improve computational efficiency, and achieve more efficient real-time analysis; make full use of the massive data generated by the smart grid, extract valuable information through data-driven methods, and enhance the accuracy and reliability of the analysis. In addition, the present invention further improves the model performance and data utilization efficiency through Q function optimization and category distribution balance, wherein, based on balancing category distribution and maintaining category distribution, it can ensure that the category distribution of newly generated data is reasonable and consistent, thereby improving the utilization effect of training data. Therefore, the present invention provides a new technical approach, which can not only greatly improve the efficiency and accuracy of steady-state operation analysis of power systems, but also provides a new technical approach and solution for solving the operation problems of complex power systems to help solve the steady-state operation analysis problems of complex power systems, thereby having important practical application value and broad promotion prospects. Through these advantages, the technical solution of the present invention has strong competitiveness and innovation potential in the field of steady-state operation analysis of power systems.

[0181] The specific embodiments described above further illustrate the purpose, technical solutions and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. It is particularly pointed out that for those skilled in the art, any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for extracting key influencing nodes of power adjustment of power flow section, characterized in that: include: Obtaining status factor data of the power system; Obtaining an optimal decision tree model based on the improved VIPER algorithm and the state factor data; Based on the best decision tree model, each node in the best decision tree model is obtained; Obtain the corresponding relationship between each node and the physical node of the power system; Evaluate the importance of each node to the power adjustment of the power flow section of the power system according to the preset metric indicators; Based on the corresponding relationship and the importance of each node to the power regulation of the power flow section of the power system, the key influencing nodes of the power flow section power regulation are extracted from the physical nodes of the power system.

2. A method for extracting key influencing nodes of power adjustment of power flow section according to claim 1, characterized in that: The method of obtaining the best decision tree model based on the improved VIPER algorithm and the state factor data includes: Based on the improved VIPER algorithm, a number of state-action data and category distributions are generated using the preset reinforcement learning model and state factor data; Get an initial data set based on some state-action data and category distribution; Based on the improved VIPER algorithm, the initial data set is iteratively updated and the preset decision tree model is trained based on the initial data set after each iteration until the iteration condition is met, and N classification models are obtained, where N is a positive integer; Get the best decision tree model among N classification models.

3. A method for extracting key influencing nodes of power adjustment of power flow section according to claim 2, characterized in that: The method of iteratively updating the initial data set based on the improved VIPER algorithm and training a preset decision tree model based on the initial data set after each iteration includes: Based on the improved VIPER algorithm, each category in the initial data set is resampled to obtain the training data set; Train the preset decision tree model based on the training data set to obtain the current classification model; Control the current classification model to generate new state data under the initial state of random power flow; Obtain new action data of the preset reinforcement learning model under new state data; Obtain new state-action data based on the new state data and the new action data; It is determined whether to merge the new state-action data into the initial data set according to the preset conditions, so as to update the initial data set and make the initial data set meet the preset state.

4. A method for extracting key influencing nodes of power adjustment of power flow section according to claim 2, characterized in that: The loss function for each iteration is as follows: In the above formula, h t (s) refers to the proportion of actions taken at time t in the initial data set under state s, t∈{0,…,T-1}; is the action value function; is the state value function.

5. The method for extracting key influencing nodes of power adjustment of power flow section according to claim 3, characterized in that: The step of determining whether to merge the new state-action data into the initial data set according to the preset conditions includes: Perform value assessment on the new state-action data to obtain the assessment result; Calculate the retention probability of the new state-action data based on the evaluation results; When it is determined that the retention probability is greater than the retention probability threshold, the new state-action data is merged into the initial data set.

6. A method for extracting key influencing nodes of power adjustment of power flow section according to claim 3, characterized in that: The preset state includes: data of each category in the initial data set is balanced and the category distribution of the initial data set is maintained.

7. A device for extracting key nodes affecting power adjustment of power flow section, characterized in that: include: Data acquisition module, model acquisition module, node evaluation module and node extraction module; Wherein, the data acquisition module is used for acquiring the state factor data of the power system; The model acquisition module is used to acquire the best decision tree model based on the improved VIPER algorithm and the state factor data; The node evaluation module is used to obtain each node in the optimal decision tree model based on the optimal decision tree model; obtain the corresponding relationship between each node and the physical node of the power system; and evaluate the importance of each node to the power adjustment of the power flow section of the power system according to the preset metric; The node extraction module is used to extract the key influencing nodes of power flow section power adjustment from the physical nodes of the power system based on the corresponding relationship and the importance of each node to the power flow section power adjustment of the power system.

8. The device for extracting key influencing nodes of power adjustment of power flow section according to claim 7, characterized in that: The model acquisition module includes: a data set generation unit, an iterative update unit and a model acquisition unit; The data set generation unit is used to generate a number of state-action data and category distributions based on the improved VIPER algorithm using a preset reinforcement learning model and state factor data; and obtain an initial data set based on the number of state-action data and category distributions; The iterative updating unit is used to iteratively update the initial data set based on the improved VIPER algorithm and train a preset decision tree model based on the initial data set after each iteration until an iteration condition is met to obtain N classification models, where N is a positive integer; The model acquisition unit is used to acquire the best decision tree model from N classification models.

9. The device for extracting key influencing nodes of power adjustment of power flow section according to claim 8, characterized in that: The iterative updating unit is used to iteratively update the initial data set based on the improved VIPER algorithm and train a preset decision tree model based on the initial data set after each iteration, including: Based on the improved VIPER algorithm, each category in the initial data set is resampled to obtain the training data set; Train the preset decision tree model based on the training data set to obtain the current classification model; Control the current classification model to generate new state data under the initial state of random power flow; Obtain new action data of the preset reinforcement learning model under new state data; Obtain new state-action data based on the new state data and the new action data; It is determined whether to merge the new state-action data into the initial data set according to the preset conditions, so as to update the initial data set and make the initial data set meet the preset state.

10. The device for extracting key influencing nodes of power adjustment of power flow section according to claim 8, characterized in that: The loss function for each iteration is as follows: In the above formula, h t (s) refers to the proportion of actions taken at time t in the initial data set under state s, t∈{0,…,T-1}; is the action value function; is the state value function.

Images