Transient voltage stability assessment method based on disturbance signal energy and decision tree combination

Abstract

The application discloses a transient voltage stability evaluation method based on disturbance signal energy and decision tree combination, which comprises the following steps: obtaining voltage quantity of wide area measurement information of each node in a power system, converting the voltage quantity into corresponding disturbance signal energy value, constructing a model for evaluating transient voltage stability based on an ID3 decision tree algorithm, obtaining voltage dynamic time sequence by using PMU measurement of each node in the power system after short-term large disturbance of the power system, calculating voltage disturbance signal energy obtained by converting each voltage time sequence, comparing the voltage disturbance signal energy with a disturbance signal energy split threshold value of each node in the constructed decision tree, and obtaining an evaluation result. The application realizes evaluation on transient voltage stability of the power system, greatly reduces training time cost of the model, improves the problem that existing methods are prone to confusing voltage instability and voltage stability in different forms, and improves evaluation accuracy of the model on transient voltage stability of the power system.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of power grid safety monitoring, and particularly relates to a transient voltage stability evaluation method based on disturbance signal energy and decision tree combination. BACKGROUND

[0002] With the large-scale grid-connection of new energy power generation equipment and the wide application of power electronic equipment in power systems, the scale and complexity of power grids increase, and the voltage regulation characteristics and reactive power distribution of power systems have changed significantly. This change makes the system more prone to transient voltage instability when disturbed, which seriously threatens the safe and stable operation of power systems. After a transient short-circuit fault occurs in a power system, the system voltage will drop temporarily and rise rapidly after the fault is removed. Whether the system voltage can recover to stability after the fault is removed is an important task for evaluating the transient voltage stability of the system. To solve this type of problem, domestic and foreign scholars have conducted a large number of researches. At present, the main methods can be divided into two categories: one is the analysis method based on physical modeling, which relies on detailed model parameters of the system, such as time-domain simulation method, power flow analysis method, energy function method, etc.; the other is the data-driven method based on wide-area measurement information, such as the transient evaluation method based on voltage Shapelet time series trajectory feature learning. These methods have their own advantages and disadvantages, and further research and optimization are still needed for the evaluation needs of specific scenarios.

[0003] However, with the continuous expansion of power grid interconnection, it becomes increasingly difficult to obtain detailed models and accurate parameters of actual systems, which limits the use of the first type of method. The second type can only use wide area measurement system (WAMS) data to learn voltage time series trajectory features to evaluate transient stability, but when facing many high-dimensional time series data sets and large sample capacity in actual applications, the model training time cost of shapelet-based time series trajectory feature learning is huge, which seriously affects the efficiency of time series feature learning and construction of transient voltage evaluation model. Moreover, the voltage shapelet-based time series trajectory feature learning is based on voltage time series curve shape for model training, while there are many shapes of transient voltage instability, such as sustained low voltage, overvoltage, divergent oscillation, etc. The single optimal shapelet sequence obtained by the voltage shapelet-based time series trajectory feature learning method cannot well distinguish between various transient voltage instability and transient voltage stability, resulting in a decrease in evaluation accuracy.

[0004] In recent years, signal energy method has been widely applied in mechanical, power, electric power and other fields, such as structural health monitoring, vibration analysis, fault diagnosis, energy management system and the like. The theory of signal energy method is to extract a certain signal from the time domain simulation results of the system under study, define the signal energy as the integral of the square of the transient component of the signal with respect to the simulation time, and the corresponding disturbance signal energy as the integral of the square of the difference between the transient component and the steady component of the signal with respect to the simulation time. In this theory, the disturbance signal energy is used to measure the energy accumulation of the signal after being disturbed in a period of time, which is a one-dimensional scalar data, and the time cost of its characteristic learning is very small. Combined with the two-class machine learning method such as decision tree, it provides a new idea for using data-driven method based on wide-area measurement information to evaluate the transient voltage stability of power system.

[0005] In summary, there is an urgent need for a transient voltage stability evaluation technology based on the combination of disturbance signal energy and decision tree, which is applied to the research on safe and stable operation of power system, so as to realize the transient voltage stability evaluation of the system after fault removal quickly and accurately with greatly reduced time cost of training model. SUMMARY

[0006] In order to overcome the defects and deficiencies existing in the prior art, the present application provides a transient voltage stability evaluation method based on the combination of disturbance signal energy and decision tree. The voltage quantity of the wide-area measurement information of each node in the system is obtained, which is converted into the corresponding disturbance signal energy value, and then the ID3 decision tree algorithm is used to construct a two-class evaluation model which can be used to evaluate the transient voltage stability of the system. The voltage curve can be predicted whether it can recover to stability after a few seconds (generally 8-10s in engineering) after fault removal, so as to realize the evaluation of the transient voltage stability of the system, and more time is obtained for the operation personnel to take relevant control measures. The decision tree model training based on the disturbance signal energy greatly reduces the training time cost, and the transient voltage instability and stability of the system are evaluated by the energy size, which can avoid the decline of evaluation accuracy caused by different voltage instability forms.

[0007] In order to achieve the above purpose, the technical scheme adopted by the present application is as follows:

[0008] The present application provides a transient voltage stability evaluation method based on the combination of disturbance signal energy and decision tree, comprising the following steps:

[0009] For the selected power system, the simulation case is generated by numerical simulation, the time sequence voltage data of each node of the power system after fault removal is obtained as a data sample set, and each sample is normalized;

[0010] According to the voltage time sequence sample data, the difference between the transient voltage amplitude after fault removal and the steady-state voltage value of the corresponding node is calculated, the disturbance signal energy accumulated by each node in a set time period after fault removal is obtained based on the signal energy method, the time sequence data set is converted into a disturbance energy data set, the disturbance energy data set is sorted and averaged to obtain a disturbance signal energy splitting threshold candidate set;

[0011] The disturbance energy data set is taken as an original data object, and the disturbance signal energy splitting threshold candidate set is taken as a condition, and classification learning is performed based on the ID3 decision tree algorithm to establish a decision tree model as a classification evaluation model of transient voltage stability, and each node in the tree corresponds to a disturbance signal energy splitting threshold of a node;

[0012] When a short-term large disturbance occurs in the power system, dynamic time sequences are obtained based on PMU measurement of each node in the power system, the disturbance signal energy obtained by converting each dynamic time sequence is calculated, and the disturbance signal energy splitting threshold of each node in the decision tree model is compared to obtain an evaluation result.

[0013] As a preferred technical solution, the simulation case is generated by numerical simulation, time sequence voltage data of each node of the power system after fault removal is obtained as a data sample set, and each sample is normalized, and the specific steps include:

[0014] n simulation cases are generated by transient numerical simulation, one simulation case is one sample, one sample includes d nodes, and the voltage amplitude trajectory of each node is extracted;

[0015] Taking the fault removal time as the starting point, a given time window length T and a sampling time Δt are given, the normalized value time sequence of the voltage of each node is extracted using the time window, the length of each sequence is m=T / Δt, and the d-dimensional transient voltage time sequence data set S of the power system after fault removal is obtained raw , S raw ={S raw1 ,S raw2 ,…,S rawd}, n is the number of samples, d is the number of nodes included in the power system, and m is the length of a single time sequence;

[0016] Each data in the time sequence data set sample S raw is divided by the reference voltage U n of the corresponding node to obtain the normalized voltage time sequence data set S={S1,S2,…,S d}={S raw1 / U n ,S raw2 / U n ,…,S rawd / Un};

[0017] The i-dimensional data set of the voltage time series data set S is a data set composed of n time series with length m, that is:

[0018] S i ={U i,1 ,U i,2 ,...,U i,n},i=1,2,...d

[0019] Wherein, U i,j ={u i,j (t1),u i,j (t2),…,u i,j (t m )},j=1,2,…,n,u i,j (t1) represents the voltage of node i in the jth sample at the first measurement point.

[0020] As a preferred technical solution, the signal energy method is used to obtain the disturbance signal energy accumulated by each node in each sample within a set time period after fault removal, and the time series data set is converted into a disturbance energy data set. The specific steps include:

[0021] The transient voltage time series data set is converted into a voltage disturbance signal energy set based on the signal energy method. Each time series is converted into disturbance signal energy generated within T time period, which is represented as:

[0022]

[0023] Wherein, u(0) is the initial value of voltage stability, m is the number of measurement points contained in the time period considered in the time series, Δt is the time interval between two consecutive measurements, u(t i ) is the voltage amplitude data corresponding to the i-th measurement point;

[0024] The time series data set corresponding to all nodes of the voltage time series data set S is converted to obtain the disturbance energy data set, which is represented as:

[0025]

[0026] Wherein, each element in the matrix in the disturbance energy data set is a scalar, and each element E i,j represents the disturbance signal energy generated by the voltage of a node in a sample corresponding scene in a given time series.

[0027] As a preferred technical solution, the disturbance energy data set is sorted and averaged to obtain a disturbance signal energy splitting threshold candidate set, which specifically includes:

[0028] E i,j The perturbation signal energy set E i The energy values of the n samples are sorted in ascending order to obtain E i sort E i,1 , E i,2 , …, E i,n , the average value between each two adjacent elements in E i,1 is calculated to obtain the split threshold candidate set of the node perturbation signal energy The split threshold candidate set of the voltage perturbation signal energy of each node of the power system under the determined sample number is obtained, and is expressed as:

[0029]

[0030] Among them, the split threshold candidate set of the voltage perturbation signal energy E mean _ sort has n-1 rows, indicating that n-1 mean value data samples are obtained by taking the mean value of each two adjacent elements in n samples, each column represents the same system node under different samples, and each element is a scalar, which is used as the condition for the decision tree model based on conditional information entropy to divide the original data.

[0031] As a preferred technical solution, the disturbance energy data set is taken as the original data object, the disturbance signal energy split threshold candidate set is taken as the condition, the ID3 decision tree algorithm is used for classification learning, and a decision tree model is established as a classification evaluation model for transient voltage stability. Specifically, it includes:

[0032] According to the stable and unstable labels of the disturbance signal energy set E

[0033]

[0034] Among them, Ent(E) represents the information entropy of the disturbance signal energy set E. For each node bus in a d-dimensional power system, , a candidate energy split threshold is selected from E , and is recorded as When the disturbance signal energy set E is divided, the disturbance signal energy set E is divided into two subsets, which are recorded as E less , E greater , E less Each element in E greaterEach element in is greater than

[0035] Statistical analysis of subset E less and subset E greater The number of stable and unstable samples is used to calculate the conditional information entropy after classification. The calculation expression is as follows:

[0036]

[0037] Where, p less and p greater E respectively less and E greater The proportion of the sample size in E;

[0038] pass The information gain obtained by partitioning the dataset is:

[0039]

[0040] Split threshold candidate set from the energy set of the perturbation signal The node with the largest information gain among the n-1 candidate energy splitting thresholds is selected as the optimal energy splitting threshold for node i. The optimal energy splitting threshold for all nodes in the power system is obtained. By comparing the information gain corresponding to the optimal energy splitting threshold of multiple nodes, the optimal energy splitting threshold for all nodes in the power system is found.

[0041] The next decision tree node is selected by maximizing the information gain and its corresponding energy splitting threshold. After the original data samples are divided into two branches according to the energy splitting threshold of the decision tree node, the corresponding rows of samples that are not in the branch in the original perturbation signal energy set E should be deleted according to the sample situation in the new branch, and the corresponding perturbation signal energy set E' should be updated again. This is used as the original perturbation energy dataset for the next decision tree node to be screened.

[0042] The decision tree classifies the perturbation signal energy into a candidate threshold E using multi-layered if-then rules. mean _sort After input, the data in each dimension is compared starting from the root node. The optimal attribute is selected as the node based on information gain. The data is recursively segmented step by step until each subset contains only samples of the same type or reaches a preset threshold. The leaf nodes store the labels of samples of the same type or the labels of most samples. A decision tree model is established as a classification and evaluation model for transient voltage stability.

[0043] As a preferred technical solution, the calculation of the disturbance signal energy converted from each dynamic time sequence is compared with the disturbance signal energy split threshold of each node in the decision tree model to obtain an evaluation result, and specifically includes:

[0044] Based on the voltage dynamic time sequence obtained by each node PMU measurement in the power system, the voltage disturbance signal energy converted from each voltage time sequence is calculated, compared with the disturbance signal energy split threshold of each node in the decision tree, and the path search and matching are performed from top to bottom, and when the path reaches the terminal (leaf) node, the evaluation result is given by the class label in the terminal node. If it is determined that the power system is unstable, an alarm signal is sent; if it is determined that the power system is stable, the monitoring is continued.

[0045] Compared with the prior art, the present application has the following advantages and beneficial effects:

[0046] (1) The present application combines the data-driven method based on the wide-area measurement system with the disturbance signal energy, converts the original two-dimensional trajectory time series data set into a single-dimensional flat energy data set, greatly reduces the training time cost of constructing the data-driven transient voltage stability evaluation model, and improves the evaluation model construction efficiency.

[0047] (2) The present application is based on the fact that the voltage disturbance signal energy can fully reflect the difference in the accumulated energy of the node voltage during transient voltage instability and stability, and the data-driven method based on a large amount of sample data can learn the characteristics of different transient voltage instability and stability, so as to evaluate whether the subsequent node voltage can recover to stability within a short time (1s) after fault removal according to the learned stable characteristics; and the disturbance signal energy is combined with the decision tree algorithm, which has clear physical interpretability, and also improves the problem that the original voltage time sequence shapelet sequence method based on voltage trajectory is easy to confuse different forms of voltage instability and voltage stability, and improves the evaluation accuracy of the model for transient voltage stability of the power system. BRIEF DESCRIPTION OF DRAWINGS

[0048] Figure 1 It is a flowchart of the transient voltage stability evaluation method based on the disturbance signal energy and the decision tree combination of the present embodiment;

[0049] Figure 2 It is an architecture diagram of the decision tree model of the present embodiment;

[0050] Figure 3 It is a topological graph of the IEEE standard 39-node system of the present embodiment;

[0051] Figure 4 It is a classification evaluation model of transient voltage stability constructed by the present embodiment;

[0052] Figure 5(a) is a schematic diagram of the voltage amplitude trajectory and voltage disturbance signal energy accumulation under transient voltage stabilization conditions in this embodiment;

[0053] Figure 5(b) is a schematic diagram of the voltage amplitude trajectory and voltage disturbance signal energy accumulation under transient voltage instability in this embodiment. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0055] Example

[0056] like Figure 1 As shown, this invention provides a transient voltage stability assessment method based on a combination of disturbance signal energy and decision tree, comprising the following steps:

[0057] S1. For a given power system, considering its various possible operating conditions and accident types, a large number of simulation cases are generated through numerical simulation. The time-series voltage data of each node of the power system after fault clearance is obtained as a data sample set, and each sample is normalized.

[0058] (1) For the power system under study, considering its various possible operating conditions and accident types, n simulation cases are generated through transient numerical simulation. Each simulation case is a sample, and each sample contains d nodes.

[0059] This example, based on a standard power system operating scenario, assumes system power flow convergence and randomly sets the system load between 85% and 125% of the standard scenario. The randomness of wind power output is simulated using the Monte Carlo method. Synchronous generators randomly set their output between 90% and 100% of their rated capacity. The fault setting scheme involves randomly setting three-phase short-circuit faults at 10%, 30%, 60%, and 90% of the distance from the beginning of all transmission lines. The fault occurrence time is set to 0 seconds, and the fault clearing time is fixed at 0.1 seconds.

[0060] (2) Under the above conditions, data samples are generated and the voltage amplitude trajectories of each node are extracted through transient numerical simulation. Starting from the fault clearing time, given a time window length of T and a sampling time of Δt, the voltage U time series of each node is extracted from each sample using the time window. The length of each sequence is m = T / Δt, thus obtaining the d-dimensional (i.e., containing d nodes) transient voltage time series dataset S of the power system after fault clearing. raw S raw ={S raw1 ,S raw2 ,…,Srawd}, n is the number of samples, d is the number of nodes included in the power system, and m is the length of a single time series data;

[0061] Specifically, taking the fault removal time as the starting point, a given time window length T, the measurement interval Δt is the same as the electromechanical transient simulation step 0.01s, the time window is used to extract the per-unit time series of the voltage of each node, the length of each sequence is m=T / Δt, T is taken as 0.8s, and m=80 is obtained. 800 data samples are randomly generated by transient numerical simulation, which are used as the data sample set S raw , Divide each data in the time series data set sample S raw by the reference voltage U n of the corresponding node to obtain the normalized voltage time series data set S={S1, S2, …, S 39}.

[0062] Divide each data in the time series data set sample S raw by the reference voltage U n of the corresponding node to obtain the normalized voltage time series data set S={S1, S2, …, S d}={S raw1 / U n ,S raw2 / U n ,…,S rawd / U n}. The i-th dimension (1≤i≤d) of the voltage time series data set S, the single-dimensional data set (i.e. the transient voltage time series data set of a specific node) is a data set composed of n time series with a length of m, that is:

[0063] S i ={U i,1 ,U i,2 ,...,U i,n},i=1,2,...d

[0064] Wherein, U i,j ={u i,j (t1),u i,j (t2),…,u i,j (t m )},j=1,2,…,n,u i,j (t1) represents the value of the voltage of node i in the jth sample at the first measurement point.

[0065] S2, calculating the difference between the transient voltage amplitude after fault removal and the steady-state voltage value of the corresponding node; and using the signal energy method to obtain the disturbance signal energy accumulated by each node in the set time period after fault removal, thereby converting the time series data set into a disturbance energy data set, further sorting and averaging the disturbance energy data set to obtain a disturbance signal energy splitting threshold candidate set, specifically including:

[0066] (1) According to the signal energy method, the transient voltage time series data set S obtained in step S1 is converted into a voltage disturbance signal energy set. Each time series can be converted into the disturbance signal energy generated in the T time period according to the following formula:

[0067]

[0068] Where u(0) is the voltage steady-state initial value, m is the number of measurement points contained in the time period considered in the time series, Δt is the time interval between two consecutive measurements, i.e. the step, u(t i ) is the voltage amplitude data corresponding to the i-th measurement point;

[0069] After n (preferably 800) time series are converted, the disturbance signal energy set E i of a single node is obtained: i,1 E i,2 , …,E i,800}. The time series data set corresponding to all nodes in S is processed in the same way to generate the disturbance signal energy set E, thereby completing the transformation of the original trajectory type time series data set to the flat energy data set. The disturbance signal energy set E is represented as:

[0070]

[0071] Each element in the above formula matrix is a scalar. Each element E i,j represents the disturbance signal energy generated by the voltage of a node in a given time series in a sample corresponding scenario;

[0072] (2) Taking each element E i,j in the disturbance signal energy set E as the feature attribute of the corresponding node i in the corresponding sample j, the energy values of the 800 samples in the disturbance signal energy set E i of a single node are sorted from small to large to obtain E i sort ={E’ i,1 ,E’ i,2 …,E’ i,800}. The average value between two adjacent elements in E’ i,1 is calculated to obtain the splitting threshold candidate set of the disturbance signal energy of the node Similarly, after performing the same processing on all nodes in the power system, a candidate set E for the splitting threshold of the voltage disturbance signal energy of each node under a certain number of samples is obtained. mean_sort As shown in the following formula:

[0073]

[0074] Candidate set E for splitting threshold of voltage disturbance signal energy mean_sort There are 799 rows in the table, representing 799 mean data samples obtained by taking the mean of each pair of adjacent samples from 800 samples. Each column represents the same system node under different samples, and each element is a scalar. These values ​​will be used as the conditions for the subsequent decision tree algorithm based on conditional information entropy to divide the original data.

[0075] S3. Using the disturbance energy dataset obtained in step S2 as the original data object, and the candidate set of disturbance signal energy splitting thresholds as conditions, the ID3 decision tree algorithm is used for classification learning to establish a decision tree model. The established decision tree model is the classification and evaluation model for transient voltage stability. Each node in the tree corresponds to the disturbance signal energy splitting threshold of a certain node. The specific steps include:

[0076] (1) Taking the disturbance signal energy set E obtained in step S2 as the original data object, since the evaluation of transient voltage stability in this invention is a binary classification problem, the data samples in the disturbance signal energy set E are divided into two categories, stable and unstable, according to the stable and unstable labels of the disturbance signal energy set E, denoted as C1 and C2 respectively, and their proportions in the disturbance signal energy set E are p1 and p2 respectively. Calculate the original information entropy of the disturbance signal energy set E:

[0077]

[0078] Here, Ent(E) is called the information entropy of E. For each node bus in a d-dimensional power system, from Choose one of the candidate energy splitting thresholds (denoted as ). When dividing E, Divide E into two subsets, denoted as Ei, Ej, less E greater (where E) less Each element in the matrix is ​​less than or equal to the specified value in dimension i. And E greater Each element in is greater than );

[0079] Specifically, step S2 yields the candidate set E for the energy set splitting threshold of the perturbation signal. mean_sort These are the conditions for partitioning the original data. For example... Figure 3As shown, for each node bus in the given 39-node power system, from Choose one of the candidate energy splitting thresholds (denoted as ). When dividing E, Divide E into two subsets, denoted as Ei, Ej, less E greater (where E) less Each element in the matrix is ​​less than or equal to the specified value in dimension i. And E greater Each element in is greater than ).

[0080] (2) To compute subset E less and E greater The information entropy is used to count the number of stable and unstable samples in the two subsets respectively, and the subset E is calculated. less and E greater The information entropy of each element is determined, and then the conditional information entropy after classification is calculated. The calculation expression is as follows:

[0081]

[0082] Where, p less and p greater E respectively less and E greater The proportion of the sample size in E;

[0083] Therefore, through After dividing the dataset into partitions and obtaining the conditional information entropy, the corresponding information gain is calculated as follows:

[0084]

[0085] From the three formulas in step S3 The node with the largest information gain among the 799 candidate energy splitting thresholds is selected as the optimal energy splitting threshold for node i. Similarly, the optimal energy splitting threshold for all nodes in the power system can be obtained. By comparing the information gain corresponding to the optimal energy splitting thresholds of multiple nodes, the optimal energy splitting threshold for all nodes in the power system can be found. The next decision tree node is selected by maximizing the information gain and its corresponding energy splitting threshold. less and E greater The corresponding This will be further used to update the original information entropy and corresponding information gain of the disturbance signal energy set E;

[0086] (3) Figure 2 As shown, the decision tree classifies the perturbation signal energy splitting threshold candidate set E through multi-layered if-then rules.mean_sort After input, the model starts from the root node to compare each dimension data, selects the optimal attribute (voltage node) as the node according to the information gain, recursively divides the data step by step, until each subset contains only the same sample or reaches the preset threshold. Finally, the leaf node stores the label of the same sample or the label of the majority sample. As shown in Figure 4 The decision tree model constructed in this example has three internal nodes, respectively corresponding to the disturbance energy splitting thresholds of nodes 5, 11 and 7 of the given power system. The class label on the leaf node represents the stability judgment result, -1 and 1 represent instability and stability respectively. The accuracy of the constructed classification model is tested by k-fold cross-validation, k is 10, and the cross-validation accuracy is 99.8%, which shows that the model has ideal classification performance.

[0087] S4, after a short-term large disturbance occurs in the power system, the voltage dynamic time series of each node PMU in the power system is obtained, the voltage disturbance signal energy converted from each voltage time series is calculated, and the disturbance signal energy splitting threshold of each node in the decision tree constructed in step S3 is compared, the path search and matching are performed from top to bottom, and when the path reaches the terminal (leaf) node, the evaluation result is given by the class label in the terminal node. If it is judged that the power system is unstable, an alarm signal is immediately sent out; if the stability judgment result of the power system is stable, continuous monitoring is performed.

[0088] (1) The same random fault scene setting as in step S1 is adopted, and 1 line maintenance is randomly set in the system line to simulate N-1 fault, and 400 transient simulation cases are randomly generated again as test samples.

[0089] (2) The voltage time series of nodes 5, 11 and 7 within 0.8s after fault removal are collected by the wide area measurement system, then the disturbance signal energy generated by the sequence and the steady-state value is calculated, and the energy splitting threshold in the model is compared from top to bottom, until the leaf node is encountered, then the stability judgment result can be obtained. The overall accuracy of the test evaluation result is 99.25%, among which the misjudgment rate (the probability of evaluating stability as instability) is 0.5%, and the omission rate (the probability of evaluating instability as stability) is 0.25%, which proves that the classification model can be reliably and effectively used for online monitoring. As shown in FIG. 5(a) and FIG. 5(b), the typical system node voltage response curve change diagrams of two types of samples evaluated as transient voltage stability and transient voltage instability in this example are shown, which intuitively presents the voltage response characteristics under different conditions. The local enlarged view is a physical diagram of the cumulative disturbance signal energy, which more clearly shows the difference between transient voltage stability and instability.

[0090] (3) In order to reflect the superiority of the present application in model construction efficiency and evaluation accuracy, the method of the present application is compared with a transient voltage stability evaluation method based on voltage shapelet time sequence feature learning using consistent parameters and environment, and the performance comparison result is shown in Table 1:

[0091] Table 1 Comparison of training time and accuracy of two methods

[0092]

[0093] In terms of model training time, compared with the method based on shapelet, the model training time of the method of the present application is reduced by two orders of magnitude, is greatly shortened to within 300s, and greatly reduces the training time cost of the model. In terms of accuracy, compared with the method based on shapelet, the method of the present application reduces the omission rate and the misjudgment rate, and improves the overall accuracy.

[0094] The above embodiments are preferred embodiments of the present application, but the embodiments of the present application are not limited by the above embodiments, and any changes, modifications, substitutions, combinations, simplifications made without departing from the spirit and principles of the present application should be equivalent replacement methods, and are included in the protection scope of the present application.

Claims

1. A transient voltage stability assessment method based on the combination of disturbance signal energy and decision tree, characterized in that, The method comprises the following steps: The simulation case is generated by numerical simulation for the selected power system, time sequence voltage data of each node of the power system after fault removal is obtained as a data sample set, and each sample is normalized; According to the voltage time sequence sample data, the difference between the transient voltage amplitude after fault removal and the voltage steady-state value of the corresponding node is calculated, the disturbance signal energy accumulated by each node within a set time period after fault removal is obtained based on the signal energy method, the time series data set is converted into a disturbance energy data set, the disturbance energy data set is sorted and averaged, and a disturbance signal energy splitting threshold candidate set is obtained, which specifically comprises: According to the disturbance signal energy set E and the stable and unstable labels, the data samples in the disturbance signal energy set E are divided into two categories, namely stable and unstable, denoted as C1 and C2, and the proportions of C1 and C2 in the disturbance signal energy set E are p1 and p2 respectively, and the original information entropy of the disturbance signal energy set E is calculated: ； wherein Ent(E) represents the information entropy of the set E of disturbance signal energies, for each node bus in the d-dimensional power system, a candidate energy splitting threshold, denoted as when splitting the set E of disturbance signal energies, the set E of disturbance signal energies is split into two subsets, denoted as E less , E greater , E less each element in E is less than or equal to each element in E greater is greater than ; The number of stable samples and unstable samples in subsets E less and E greater is counted respectively, and the conditional information entropy after classification is calculated, and the expression is as follows: ； wherein, and are E less and E greater the proportion of the number of samples in E By The information gain resulting from the division of the data set is: ； Splitting threshold candidate set from disturbance signal energy set Select the maximum information gain from the n-1 candidate energy splitting thresholds as the optimal energy splitting threshold of node i Get the optimal energy splitting threshold of all nodes in the power system, and find the optimal energy splitting threshold of all nodes in the power system by comparing the size of the information gain corresponding to the optimal energy splitting threshold of multiple nodes The next decision tree node is selected by maximizing the information gain and the corresponding energy splitting threshold, and after the original data samples are divided into two branches according to the energy splitting threshold of the decision tree node, the corresponding rows of the samples in the original disturbance signal energy set E that are not in the branch should be deleted, and the corresponding disturbance signal energy set E' is updated as the original disturbance energy data set for the next decision tree node selection; The decision tree classifies by multiple layers of if-then rules, and splits the candidate set of disturbance signal energy thresholds After input, the dimensional data is compared starting from the root node, the optimal attribute is selected as the node according to the information gain, the data is recursively segmented step by step, until each subset contains only the same samples or reaches the preset threshold, the leaf node stores the label of the same sample or the label of the majority sample, and a decision tree model is established as a classification evaluation model of transient voltage stability. The disturbance energy data set is taken as the original data object, the disturbance signal energy splitting threshold candidate set is taken as the condition, and the ID3 decision tree algorithm is used for classification learning to establish a decision tree model as a transient voltage stability classification evaluation model, and each node in the tree corresponds to a disturbance signal energy splitting threshold of a node. When a short-term large disturbance occurs in the power system, the dynamic time series is obtained based on the PMU measurement of each node in the power system, the disturbance signal energy obtained by converting each dynamic time series is calculated, and the disturbance signal energy splitting threshold of each node in the decision tree model is compared to obtain an evaluation result.

2. The method for transient voltage stability assessment based on disturbance signal energy and decision tree combination according to claim 1, characterized in that, The simulation case is generated by numerical simulation for the selected power system, time sequence voltage data of each node of the power system after fault removal is obtained as a data sample set, and each sample is normalized; n simulation cases are generated by transient numerical simulation, one simulation case is one sample, and each node includes d nodes, and the voltage amplitude trajectory of each node is extracted; Starting from the fault clearing time, given a time window length of T and a sampling time of Δt, the per-unit time series of voltage at each node is extracted using the time window. The length of each series is m = T / Δt, resulting in a d-dimensional transient voltage time series dataset S of the power system after fault clearing. raw S raw = { S raw1 , S raw2 ,…, S rawd }, n is the number of samples, d in the d-dimensional power system is the number of nodes in the power system, and m is the length of a single time series data. A sample S of a time series dataset is obtained raw Each data in the sample S is divided by the reference voltage U of the corresponding node n to obtain a normalized voltage time series dataset S = {S1, S2, …, S d} = {S raw1 / U n , S raw2 / U n , …, S rawd / U n} The i-dimensional data set of the voltage time sequence data set S is a data set composed of n time series with a length of m, that is: ； wherein U i,j = {u i,j (t1), u i,j (t2),…, u i,j (t m )}, j = 1, 2, …, n, u i,j (t1) denotes the value of the voltage of node i in the jth sample at the 1st measurement point.

3. The method for transient voltage stability assessment based on disturbance signal energy and decision tree combination according to claim 2, characterized in that, The transient voltage time sequence data set is converted into a voltage disturbance signal energy set based on the signal energy method, each time series is converted into a disturbance signal energy generated within a T time period, and is represented as: The time sequence data set corresponding to all nodes of the voltage time sequence data set S is converted to obtain a disturbance energy data set, which is represented as: ； wherein u(0) is the voltage steady-state initial value, m is the number of measurement points contained in the time period considered in the time series trace, At is the time interval between two consecutive measurements, u(t i ) is the voltage amplitude data corresponding to the i-th measurement point; ​ ； wherein each element in the matrix in the perturbation energy dataset is a scalar, each element E i,j represents the perturbation signal energy generated by the voltage of a certain node in the scene corresponding to a sample in the given time series.

4. The method for transient voltage stability assessment based on disturbance signal energy and decision tree combination according to claim 3, characterized in that, The sorting and mean value processing of the disturbance energy data set obtains a candidate set of disturbance signal energy splitting thresholds, and specifically includes: E i,j The perturbation signal energy set E i The energy values of n samples are sorted in ascending order to obtain E i sort ={E ’ i,1 , E ’ i,2 …, E ’ i,n}, and the average value between two adjacent elements in E ’ i,1 is calculated to obtain the splitting threshold candidate set of the node perturbation signal energy The splitting threshold candidate set of the voltage perturbation signal energy of each node of the power system under the determined number of samples is obtained, which is represented as: ； Wherein, the split threshold candidate set E of the voltage disturbance signal energy mean_sort Among them, there are n-1 rows, which represent n-1 mean data samples obtained by taking the mean value of every two adjacent samples from n samples, each column represents the same system node under different samples, and each element is a scalar, which is the condition for the decision tree model based on conditional information entropy to divide the original data.

5. The method for transient voltage stability assessment based on disturbance signal energy and decision tree combination according to claim 1, characterized in that, The comparison of the disturbance signal energy converted from each dynamic time sequence with the disturbance signal energy splitting threshold of each node in the decision tree model obtains an evaluation result, and specifically includes: Based on the voltage dynamic time sequences obtained from the PMU measurement of each node in the power system, the voltage disturbance signal energy converted from each voltage time sequence is calculated, compared with the disturbance signal energy splitting threshold of each node in the decision tree, and a path search and matching are performed from top to bottom. When the path reaches the terminal node, the evaluation result is given by the class label in the terminal node. If it is determined that the power system is unstable, an alarm signal is sent. If it is determined that the power system is stable, the monitoring is continued.

Images