A power distribution network typical operation scenario generation method based on a Gaussian mixture model

Abstract

The application discloses a kind of based on Gaussian mixture model's typical operation scene generation method of distribution network, first, the different types of data in the original data of distribution network operation scene are normalized in the internal of distribution network node, and text data in original data set is converted into discrete data for quantification;Again, through pearson correlation coefficient, the annual correlation coefficient between the output of distributed power supply and load and different influence factors is calculated, and the influence factors with high correlation degree are selected as the characteristic vector of distributed power supply and load respectively;Then, the Canopy algorithm is used to perform coarse clustering on the historical data of each power supply node and load typical operation scene, and the number of cluster centers is determined;Finally, based on Gaussian mixture model clustering, the typical operation scene is obtained.The application can improve the typicality of the generated scene, reduce the workload, enhance the application effect of Gaussian mixture model, and can also be used to generate new sample points when a large amount of scene data is needed for calculation and verification.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of scene generation, in particular to a power distribution network typical operation scene generation method based on Gaussian mixture model. BACKGROUND

[0002] With the continuous advancement of new power system construction, more and more distributed power sources and flexible loads are connected to the power distribution network. The randomness and uncertainty of distributed power output and flexible load will result in too many historical operation scenes of the power distribution network, which will bring too much computational burden and greater uncertainty to the accurate assessment, perception, tracing and governance optimization of the power supply quality of complex active power distribution network, and increase the complexity of planning and dispatching analysis.

[0003] Traditional typical scene generation mainly relies on manual experience, but with the increasing development of China's power grid, the voltage level is gradually enriched, the power grid structure is more complex, and new energy systems are continuously connected. It is difficult to meet the actual demand by relying on manual typical scene generation, and there are problems such as large workload, difficulty in meeting the needs of refined management, and the generated scenes are not typical.

[0004] As for the typical scene generation method, it mainly includes the following methods at present: K-means clustering, hierarchical clustering, DBSCAN clustering, etc. However, the traditional K-means clustering needs to preset the number of clusters, but in actual application, not only the number of typical operation scenes of different power distribution networks is different, but also it is difficult to estimate the number of clusters in advance. Hierarchical clustering and DBSCAN clustering do not need to preset the number of clusters, but the calculation complexity of hierarchical clustering is too high and there is a possibility of clustering in a chain shape, and DBSCAN clustering is difficult to identify clusters in data of different densities and difficult to cluster high-dimensional data, while the feature vectors of the power distribution network for clustering are often complex. In addition, the above clustering methods are all discriminant models, which directly find the optimal division in the sample space, and when a large number of operation scenes need to be calculated and verified, the data required for generating typical operation scenes cannot be obtained.

[0005] In summary, the existing technology has the following problems:

[0006] (1) It is difficult to meet the actual demand by relying on manual typical scene extraction, and there are problems such as large workload, subjectivity, difficulty in meeting the needs of refined management, and the extracted scenes are not typical.

[0007] (2) Gaussian mixture model clustering needs to preset the number of distributions.

[0008] (3) Discriminant model cannot obtain the required data for generating typical operation scenes when a large number of operation scene data are needed. SUMMARY

[0009] In view of the above problems, the purpose of the present application is to provide a power distribution network typical operation scene generation method based on Gaussian mixture model, which can improve the typicality of the generated scene and reduce the workload, and through the Canopy algorithm for rough clustering, the application effect of the Gaussian mixture model is enhanced, and when a large amount of scene data is needed for calculation and verification, it can also be used to generate new sample points. The technical scheme is as follows:

[0010] A power distribution network typical operation scene generation method based on Gaussian mixture model, comprising the following steps:

[0011] S1: For different types of data in the original data of the power distribution network operation scene, normalize in the power distribution network node, and convert the text data in the original data set into discrete data for quantization;

[0012] S2: Calculate the annual correlation coefficient between the output of the distributed power supply and the load and different influencing factors through the Pearson correlation coefficient, and select the influencing factors meeting the set requirements as the feature vectors of the distributed power supply and the load;

[0013] S3: Use the Canopy algorithm to perform rough clustering on the historical data of the typical operation scene of each power supply node and load, and determine the number of cluster centers;

[0014] S4: Cluster based on Gaussian mixture model to obtain the power supply typical operation scene and the load typical operation scene.

[0015] Further, the annual correlation coefficient between the output of the distributed power supply and the load and different influencing factors is calculated by the following formula:

[0016]

[0017] In the formula, and are the measured values of the two related variables at the jth time point in the Nth period, respectively; and are the average values of the two variables; x represents the influencing factors, including illumination radiation, temperature, relative humidity and wind speed; y represents the distributed power supply output and the load; cov(x (N) ,y (N) ) represents the covariance of x (N) and y (N) ; is the standard deviation of x (N) , is the standard deviation of y (N) ; n is the number of data points used for calculation in a period.

[0018] Further, the S3 specifically comprises:

[0019] S3.1: input a set List composed of original data of power grid operation scenarios, and set data distance thresholds T1 and T2, and T1>T2;

[0020] S3.2: randomly select a data point P from the set List, take the data point P as a first data center Canopy, and delete it from the set List;

[0021] S3.3: take a data point Q from the set List, calculate distances of the data point Q to all the data centers Canopy that have been generated; if the distance of the data point Q to a data center is less than the data distance threshold T2, add the data point Q to the data center Canopy, and delete it from the set List; if the distance of the data point Q to all the data centers Canopy is greater than the data distance threshold T1, take the data point Q as a new data center Canopy, and delete it from the set List; if the distance of the data point Q to a data center Canopy is between the data distance thresholds T2 and T1, add the data point Q to the data center Canopy, but do not delete it from the set List, and continue to participate in subsequent calculation;

[0022] S3.4: repeat the operation of S3.3 for other points in the set List until the set List is empty;

[0023] S3.5: calculate the number of coarse clustering centers kp of each power supply node, take the kp with the highest frequency of occurrence in the coarse clustering number as the best clustering number of the power supply typical operation scenario, and take the kp as an input parameter of the next Gaussian mixture model;

[0024] S3.6: calculate the number of coarse clustering centers kl of each load typical operation scenario, take the kl with the highest frequency of occurrence in the coarse clustering number as the best clustering number of the load typical operation scenario, and take the kl as an input parameter of the next Gaussian mixture model.

[0025] Further, the distance between two data points in S3.3 adopts a weighted Euclidean distance with a correlation coefficient as a weight, and when the data points are power supply nodes, the weight of each influencing factor is:

[0026]

[0027] In the formula, w m represents the weight of the mth influencing factor, |c m | and |c n | represent absolute values of correlation coefficients of the mth and nth influencing factors, and p represents the number of influencing factors;

[0028] The weighted Euclidean distance between two data points P and Q is expressed as:

[0029]

[0030] In the formula, w h g represents the weight of the h-th influencing factor. hP and g hQ These represent the values ​​of the h-th influencing factor at points P and Q, respectively.

[0031] Furthermore, S4 specifically includes:

[0032] S4.1: Substitute the optimal number of clusters kp for typical power supply operating scenarios into the Gaussian mixture model; first, initialize the mean, covariance matrix, and mixing coefficients of kp Gaussian distributions; the mean μ1, μ2, ..., μ kp Initialize the covariance matrices Σ1, Σ2, ..., Σ between (0,1). kp Initialized as a p-dimensional positive unit definite matrix with mixing coefficients Π1, Π2, ... Π kp All are initialized to 1 / kp;

[0033] S4.2: Calculate the E-step and M-step sequentially using the expectation-maximization method to determine the mean, covariance matrix, and mixing coefficients of the Gaussian mixture model;

[0034] E-step: For each data point D ai Calculate which distribution it belongs to, C1, C2, ..., C. kp The probability of:

[0035]

[0036] In the formula, Point D ai The probability of belonging to a certain distribution C, Π c N(D) represents the mixing coefficient of distribution C. ai μ c ,Σ c Let ) represent the probability density function of distribution C, expressed by the following formula:

[0037]

[0038] In the formula, Σ c μ represents the covariance matrix of distribution C; c Let C represent the mean of the distribution.

[0039] M-step: Using the probabilities estimated by E-step, calculate the parameters of the Gaussian mixture model for the next iteration;

[0040] The mean of distribution C is:

[0041]

[0042] where N represents the total number of data points;

[0043] The covariance matrix of distribution C is:

[0044]

[0045] where the mean of distribution C μ c Apply the updated value in this round;

[0046] The mixing coefficient of distribution C is:

[0047]

[0048] S4.3: Repeat the calculation of E-step and M-step until convergence, and then obtain the parameters of the Gaussian mixture model; wherein the mean of each Gaussian distribution μ is regarded as the center of the distribution, and is used as a typical operating scenario of the power supply of the power distribution network; the extraction method of the typical scenario of the load is the same;

[0049] S4.4: After obtaining m0 typical operating scenarios of the power supply and n0 typical operating scenarios of the load, respectively, the typical scenarios of the power supply and the load are combined two by two to obtain a total of m0*n0 typical operating scenarios of the power distribution network.

[0050] The beneficial effects of the present application are:

[0051] 1) The traditional typical operating scenario generation method relies on artificial experience, and the present application is data-driven, which can reduce artificial cost, improve the typicality of the scenario and reduce the workload.

[0052] 2) In view of the shortcoming that the Gaussian mixture model clustering needs to preset the number of distributions, the present application performs coarse clustering through the Canopy algorithm to obtain the number of distributions, which can enhance the application effect of the Gaussian mixture model.

[0053] 3) In view of the problem that the discriminant model cannot generate the required data according to the typical operating scenario when a large amount of operating scenario data is required, the Gaussian mixture model used in the present application is a generative model, which can not only reduce a large number of historical operating scenarios to several typical operating scenarios, but also generate new sample points for calculation. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 The flowchart of the present application based on the Gaussian mixture model of the power distribution network typical operating scenario generation method. DETAILED DESCRIPTION

[0055] The application proposes a power distribution network typical operation scenario generation method based on Gaussian mixture model. First, the correlation size of each influencing factor is calculated through Pearson correlation coefficient, and the high correlation degree is selected as the feature vector, and the correlation coefficient is used as the weight of each element in the feature vector. Then, Canopy algorithm is used for coarse clustering to determine the number of cluster centers. Finally, clustering is performed based on Gaussian mixture model to obtain the typical operation scenario. The application can reduce a large number of historical operation scenarios to several typical operation scenarios, which is convenient for power grid safety calculation, planning and dispatching.

[0056] The specific process is as follows:

[0057] S1: data preprocessing

[0058] Different types of data have their own specific physical meaning and different units and dimensions, and there may also be a difference in order of magnitude between different attributes. Therefore, the different types of data in the original data need to be normalized:

[0059]

[0060] In the formula, a i represents the i-th data in a certain type of data, min(a i ) and max(a i ) represent the minimum and maximum values of the type of data.

[0061] It is worth noting that, taking temperature as an example, the minimum temperature in the tropical region is higher than the maximum temperature in the polar region. Therefore, some types of data are not comparable between power distribution network nodes, and should be normalized within the power distribution network nodes, rather than all nodes. And in order to meet the needs of correlation analysis, the text data in the original data set needs to be converted into discrete data for quantification.

[0062] S2: feature vector selection based on Pearson correlation coefficient

[0063] The annual correlation coefficient between the output of the distributed power supply and the load and different influencing factors can be calculated from formula (2).

[0064]

[0065] In the formula, and are the measured values of the two related variables at the j-th time point in the N-th period; and respectively. In this study, x represents the influencing factors, such as light radiation, temperature, relative humidity, wind speed, etc.; y represents the distributed power output and load; cov(x (N) ,y (N) ) represents the covariance of x (N) and y (N) ; is the standard deviation of x (N) , is the standard deviation of y (N) ; and n is the number of data points used for calculation in one period.

[0066] The value of Pearson correlation coefficient is between [-1, 1], and the positive and negative values respectively represent that the two variables belong to positive correlation and negative correlation, and the larger the value is, the stronger the correlation is. The relationship between the specific value range and the correlation intensity is shown in Table 1.

[0067] Table 1 Relationship between correlation coefficient and correlation intensity

[0068]

[0069]

[0070] The influencing factors with moderate correlation and above are selected as the feature vectors of the distributed power and the load.

[0071] Suppose that the distributed power and the load have p and q influencing factors with moderate correlation and above, respectively, so the feature vectors can be respectively represented as: G = [g1, g2, …, g p ] and L = [l1, l2, …, l q ].

[0072] S3: Rough clustering of data of each power node based on Canopy algorithm

[0073] First, the historical data of each power node are respectively subjected to Canopy rough clustering, and the specific steps are as follows:

[0074] (1) Input the set List composed of original data, and set data distance thresholds T1 and T2, and T1 > T2.

[0075] (2) Randomly select a data point P from the set List, take the data point P as the first data center Canopy, and delete it from the set List.

[0076] (3) Take data point Q from the set List, calculate the distance between data point Q and all the data centers Canopy that have been generated; if the distance between data point Q and a data center is less than the data distance threshold T2, then add data point Q to the data center Canopy and delete it from the set List, that is, consider that the distance between data point Q and the data center Canopy is close enough, and data point Q cannot be the center of other data center Canopy; if the distance between data point Q and all the data centers Canopy is greater than the data distance threshold T1, then take data point Q as a new data center Canopy and delete it from the set List; if the distance between data point Q and a data center Canopy is between the data distance thresholds T2 and T1, then add data point Q to the data center Canopy but do not delete it from the set List, and continue to participate in subsequent calculations.

[0077] (4) Repeat step (3) for other points in List until List is empty

[0078] When step (3) is performed, the distance between two data points is a weighted Euclidean distance with a correlation coefficient as a weight. Take the power node as an example, and the weight of each influencing factor is as follows:

[0079]

[0080] In the formula, w m represents the weight of the mth influencing factor, |c m | and |c n | represent the absolute value of the correlation coefficient of the mth and nth influencing factors, and p represents the number of influencing factors.

[0081] The weighted Euclidean distance between two data points P and Q is represented as:

[0082]

[0083] In the formula, w h represents the weight of the hth influencing factor, g hP and g hQ represent the value of the hth influencing factor of point P and point Q, and p represents the number of influencing factors.

[0084] The number of coarse clustering centers of each power node kp is calculated, where the most frequently occurring kp in the coarse clustering number is taken as the optimal clustering number of the typical operating scenario of the power, and kp is taken as the input parameter of the next step of the Gaussian mixture model.

[0085] The coarse clustering method of load typical scenario is similar to power supply, taking the most frequent kl in coarse clustering number as the best clustering number of load typical operation scenario, which is the input parameter of next step Gaussian mixture model.

[0086] S4: Typical operation scenario extraction based on Gaussian mixture model

[0087] Substitute the best clustering number kp of power supply typical operation scenario into Gaussian mixture model, firstly initialize the mean value, covariance matrix and mixing coefficient of kp Gaussian distribution, mean value μ1, μ2, …, μ kp Initialize randomly between (0, 1), covariance matrix Σ1, Σ2, …, Σ kp Initialize as p-dimensional unit positive definite matrix, mixing coefficient Π1, Π2, …Π kp Initialize as 1 / kp.

[0088] Then use the method of expectation maximization to calculate E-step and M-step in turn, to determine the mean value, covariance matrix and mixing coefficient of Gaussian mixture model.

[0089] E-step: for each data point D ai , calculate its probability belonging to distribution C1, C2, …, C kp .

[0090]

[0091] In the formula, D ai belongs to a distribution C, Π c represents the mixing coefficient of the distribution, N(D ai ; μ c , Σ c ) represents the probability density function of the distribution, which is represented by the following formula:

[0092]

[0093] M-step: using the probability estimated by E-step, calculate the Gaussian mixture model parameters of new iteration

[0094]

[0095] In the formula, N represents the total number of data points.

[0096]

[0097] In the formula, μ c Apply the updated value of this round.

[0098]

[0099] Repeat the calculation of E-step and M-step until convergence, and the parameters of the Gaussian mixture model are found, where the mean of each Gaussian distribution μ can be regarded as the center of the distribution, which can be used as the typical operating scenario of the power supply of the power distribution network. The extraction method of the load typical scenario is similar to that of the power supply.

[0100] After obtaining the typical operating scenarios of the power supply (assuming m0) and the load (assuming n0), respectively, the typical scenarios of the power supply and the load are combined two by two, and a total of m0×n0 power distribution network typical operating scenarios can be obtained, which basically cover the source-load matching scenarios that may occur.

[0101] In summary, the traditional typical operating scenario generation method relies on human experience, which is not only labor-intensive but also subjective. The present application is a data-driven model that can improve the typicality of the generated scenarios and reduce the workload. The present application uses the Canopy algorithm for coarse clustering, which enhances the application effect of the Gaussian mixture model. In addition, the present application is a generative model, which can be used to generate new sample points when a large amount of scenario data is needed for calculation and verification.

Claims

1. A method for generating typical operation scenarios of a power distribution network based on Gaussian mixture model, characterized in that, The method comprises the following steps: S1: normalizing different types of data in the power distribution network operation scene original data in the power distribution network node, and converting text data in the original data set into discrete data for quantization; S2: calculating the annual correlation coefficient between the output of the distributed power supply and the load and different influence factors by using the Pearson correlation coefficient, and selecting the influence factors meeting the set requirements as the feature vectors of the distributed power supply and the load respectively; S3: using the Canopy algorithm to perform coarse clustering on the historical data of the typical operation scenes of the power supply nodes and the load, and determining the number of cluster centers; S4: performing clustering based on the Gaussian mixture model to obtain the typical operation scenes of the power supply and the load; The S3 specifically comprises: S3.1: inputting a set List composed of power distribution network operation scene original data, and setting data distance thresholds T1 and T2, and T1>T2; S3.2: randomly selecting a data point P from the set List, taking the data point P as a first data center Canopy, and deleting the data point P from the set List; S3.3: taking a data point Q from the set List, calculating the distance of the data point Q to all the generated data centers Canopy; if the distance of the data point Q to a certain data center is less than the data distance threshold T2, the data point Q is added to the data center Canopy, and the data point Q is deleted from the set List; if the distance of the data point Q to all the data centers Canopy is greater than the data distance threshold T1, the data point Q is taken as a new data center Canopy, and the data point Q is deleted from the set List; if the distance of the data point Q to a certain data center Canopy is between the data distance thresholds T2 and T1, the data point Q is added to the data center Canopy, but the data point Q is not deleted from the set List, and continues to participate in subsequent calculation; S3.4: repeating the operation of S3.3 for other points in the set List until the set List is empty; S3.5: calculating the number of coarse clustering centers kp of each power supply node, taking the most frequently occurring kp in the coarse clustering number as the best clustering number of the typical operation scene of the power supply, and taking the kp as an input parameter of the next Gaussian mixture model; S3.6: calculating the number of coarse clustering centers kl of each load typical operation scene, taking the most frequently occurring kl in the coarse clustering number as the best clustering number of the load typical operation scene, and taking the kl as an input parameter of the next Gaussian mixture model; The S4 specifically comprises: S4.1: Substitute the best clustering number kp of power supply typical operation scenario into the Gaussian mixture model; firstly, initialize the mean, covariance matrix and mixing coefficient of kp Gaussian distribution; the mean μ1, μ2, …, μ kp Initialize randomly between (0, 1), the covariance matrix Σ1, Σ2, …, Σ kp Initialize as a p-dimensional unit positive definite matrix, the mixing coefficient Π1, Π2, …Π kp Initialize as 1 / kp; S4.2: using the expectation maximization method to calculate the E-step and the M-step in sequence to determine the mean, the covariance matrix and the mixing coefficient of the Gaussian mixture model; E-step: For each data point D ai , compute its probability of belonging to the distributions C1, C2,... kp , C ； wherein point D ai the probability of belonging to a distribution C, Π c denotes the mixing coefficient of the distribution C, denotes the probability density function of the distribution C, is represented by the following equation: ； wherein denotes the covariance matrix of the distribution C; denotes the mean of the distribution C; M-step: using the probability estimated by the E-step to calculate the Gaussian mixture model parameters in a new round of iteration; The mean of the distribution C is: ； In the formula, N represents the total number of data points; The covariance matrix of the distribution C is: ； where the mean μ of the distribution C c Apply the updated value for this round; The mixing coefficient of the distribution C is: ； S4.3: Repeat the calculation of E-step and M-step until convergence, then get the parameters of the Gaussian mixture model; wherein the mean of each Gaussian distribution μ is regarded as the center of the distribution, as the typical operating scenario of the power supply of the power distribution network; the extraction method of the typical scenario of the load is the same; S4.4: After obtaining m0 typical operating scenarios of the power supply and n0 typical operating scenarios of the load respectively, the typical scenarios of the power supply and the load are combined two by two to obtain a total of m0*n0 typical operating scenarios of the power distribution network.

2. The method for generating typical operation scenarios of power distribution network based on Gaussian mixture model according to claim 1, characterized in that, The annual correlation coefficient between the output of the distributed power supply and the load and different influencing factors is calculated by the following formula: ； wherein, and are the measured values of the two related variables at the jth time point in the Nth cycle, respectively; and are the average values of the two variables, respectively; x represents the influencing factors, including light radiation, temperature, relative humidity and wind speed; y represents the distributed power output and load; represents the covariance of x (N) and y (N) ; is the standard deviation of x (N) , is the standard deviation of y (N) ; n is the number of data points used for calculation in one cycle.

3. The method for generating typical operation scenarios of power distribution network based on Gaussian mixture model according to claim 1, characterized in that, The distance between the two data points in S3.3 adopts a weighted Euclidean distance with the correlation coefficient as the weight, and the weight of each influencing factor is: ； In the formula, w m represents the weight of the mth influence factor, |c m |and |c n | represents the absolute value of the correlation coefficient of the mth and nth influence factors, and p represents the number of influence factors. The weighted Euclidean distance between the two data points P and Q is expressed as: ； where w h represents the weight of the hth influencing factor, g hP and g hQ respectively represent the value of the hth influencing factor of the P point and the Q point.

Images