A wind-solar load multi-stage scene tree generation method based on source load uncertainty feature extraction

Abstract

The application discloses a wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction, comprising the following steps: extracting wind and light load scene features based on a constructed stacked sparse autoencoder model to obtain a wind and light load scene feature set; clustering the wind and light load scene feature set based on a density peak value improved affinity propagation algorithm to obtain a wind and light load typical daily curve as a root node of a scene tree; generating a wind and light load scene tree year by year based on considering different growth modes of the load according to the root node of the scene tree; and reducing the scene tree by a scene tree reduction method of Sinkhorn distance to reduce the size of the scene tree. The wind and light load multi-stage scene tree generated by the application can reflect the timing information of new energy output and load, represent the randomness and volatility thereof, and reflect the uncertainty of long-term growth of new energy output and load, and is of great significance to the planning and construction of a power grid.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of power grid planning and scenario generation, and particularly relates to a wind-solar-load multi-stage scenario tree generation method based on source-load uncertainty feature extraction. BACKGROUND

[0002] New energy and load and the like introduce uncertainty factors to the power grid, increasing the difficulty of system planning and operation decision-making. At present, the method of scenario tree generation is often used to describe the uncertainty of new energy output and load in a long time scale. In the traditional method, only a single scenario is considered when defining the root node of the scenario tree, and the scenario often only contains limited data values, which cannot represent the uncertainty information, which is not conducive to the subsequent planning and construction of the power grid.

[0003] Therefore, a new technical solution is needed to solve this problem. SUMMARY

[0004] The purpose of the application is to overcome the deficiencies in the prior art, and provide a wind-solar-load multi-stage scenario tree generation method based on source-load uncertainty feature extraction. The generated wind-solar-load multi-stage scenario tree can reflect the time sequence information of new energy output and load, represent its randomness and volatility, and can reflect the uncertainty of long-term growth of new energy output and load, which is of great significance to the planning and construction of the power grid.

[0005] The technical scheme is that in order to achieve the above purpose, the application provides a wind-solar-load multi-stage scenario tree generation method based on source-load uncertainty feature extraction, comprising the following steps:

[0006] S1: Extracting wind-solar-load scenario features based on the constructed stacked sparse autoencoder model to obtain a wind-solar-load scenario feature set;

[0007] S2: Clustering the wind-solar-load scenario feature set based on the density peak improved affinity propagation algorithm to obtain a wind-solar-load typical daily curve as the root node of the scenario tree;

[0008] S3: Generating a wind-solar-load scenario tree year by year based on the root node of the scenario tree and considering different growth modes of the load;

[0009] S4: Reducing the scenario tree by using a scenario tree reduction method based on Sinkhorn distance to reduce the size of the scenario tree.

[0010] Further, the extraction method of wind-solar-load scenario features in step S1 is specifically:

[0011] A1: Collecting annual wind-solar-load historical data, P WT , P PV and P LThe wind power, photovoltaic and load unit data sets are respectively as follows:

[0012]

[0013]

[0014]

[0015] In the formula: n um is the total number of days in a year; P WT (i), P PV (i) and P L (i) are respectively the data sets of the unit composition of wind power, photovoltaic and load of each hour on the i-th day of the year; P WT，j , P PV，j and P L,j are respectively the unit values of wind power, photovoltaic and load at the k-th hour on the i-th day of the year;

[0016] s f is a full-time sequence scene set composed of P WT , P PV and P L , as shown below:

[0017]

[0018] In the formula: s f has n um samples, each sample containing a wind-light-load scene; the i-th sample in s f is m is the sample dimension in s f .

[0019] A2: using the constructed stacked sparse autoencoder SSAE model to perform feature extraction on s f to obtain the wind-light-load scene feature extraction result of the scene set.

[0020] The SSAE model in the step A2 is composed of M SAEs, and the SSAE uses the hidden layer of the previous SAE as the input of the input layer of the next SAE. The training process of the SAE includes encoding and decoding. In the encoding process, the scene set data features are extracted, as shown in formula (5):

[0021]

[0022] In the formula: is the output of the hidden layer; W a is the encoding weight matrix; b a is the encoding bias; σ h = 1 / (1+e -x) is the activation function;

[0023] During the decoding process, the data is reconstructed, and the extracted scene features are output as shown in Equation (6):

[0024]

[0025] In the formula: For output layer output; W b To decode the weight matrix; b b For decoding bias; σ y For activation function, the same as σ h ;

[0026] The decoded output data and the input scene set data constitute an error function, as shown in equation (7):

[0027]

[0028] In the formula: J SAE ρ is the error function; h is the set sparsity coefficient; ρ is the error function; ρ is the sparsity coefficient; h is the sparsity coefficient. j (s f,i ) is the input s f,i The activation value of neuron j in the hidden layer; Let s be the average activation level of the j-th neuron in the hidden layer. f The input is fed into the SSAE model, and each SAE in the SSAE model is trained according to equation (5-7) until all SAEs have been trained, resulting in s f Feature set y M The i-th feature sample is

[0029] In this invention, the stacked sparse autoencoder (SSAE) can acquire lower-dimensional and deeper features of wind, solar and load scenes through the stacking of SAEs.

[0030] Furthermore, in step S2, s f Feature set Y M Clustering is performed using clustering metrics to generate a typical time-series set of wind, solar, and load scenarios s. c As the root node of the multi-stage scene tree for wind, light, and lotus, the specific steps include the following:

[0031] B1: Similarity Calculation: and Y M Calculate the similarity between the i'-th and j'-th samples to form a similarity matrix.

[0032]

[0033] B2: Density peak value calculation: Calculate the density peak value γ between sample i' and other samples from local density and relative distance i' ;

[0034] B3: Update similarity matrix: Update the similarity matrix with the calculated density peak value γ i' as the diagonal elements in the similarity matrix

[0035] s i'i' = -γ i' (9)

[0036] B4: Calculate the belonging degree matrix and the attraction degree matrix

[0037]

[0038]

[0039] where: a i'k' , s i'k' are the belonging degree and similarity value of the i'th sample to other samples except the j'th sample; r i”j' is the attraction degree value obtained from other samples;

[0040] B5: Introduce damping factor: Prevent oscillation in the iteration process

[0041]

[0042]

[0043] where: d is the iteration number;

[0044] B6: Determine the clustering center: Determine whether it is a clustering center according to formula (14) If formula (14) is greater than 0, then the corresponding full-time scene s f,i can be used as the clustering center

[0045]

[0046] B7: Determine whether to terminate iteration: When the following conditions are met: 1) The set iteration number is reached; 2) The clustering center no longer changes, the iteration can be terminated; if not, return to step B5.

[0047] The above method can obtain several light and shadow scene sets containing clustering centers c1, c2,..., c K ​​K is the number of cluster centers. The cluster center is the typical day scene of the whole year wind load, which is the root node of the scene tree. In addition, the number of each scene set obtained by statistical clustering is counted, and the frequency is calculated as the probability of the scene set. After clustering, three commonly used clustering indicators, contour coefficient, CH index and DB index, are selected to analyze the clustering results and test the effectiveness of clustering.

[0048] Further, the step S3 is specifically:

[0049] C1: Considering the uncertainty of load growth, set three different load growth schemes of high, medium and low;

[0050] C2: According to the K typical day scenes of the root node of the scene tree, assume that the load grows year by year and shows high, medium and low speed growth modes with different probabilities, and gradually generate a scene tree. Each point of the scene tree represents a 72-dimensional wind load scene of a typical day, and the probability of each scene is determined by the probability of each scene of the root node and the probability of different load growth. The generated wind load scene tree has K possible typical day wind load scenes in the first year, 3K possible typical day wind load scenes in the second year, and so on. The size of the scene tree grows exponentially with the number of stages, and the scene tree needs to be reduced to reduce its size.

[0051] Further, the scene tree reduction method by Sinkhorn distance in the step S4 reduces the scene tree, which specifically includes the following steps:

[0052] D1: Each node of the original scene tree is recorded as T m,τ , each node represents a 72-dimensional wind load scene, wherein τ is the time step of the original scene tree year, and m is the number of each node under τ; the scene probability corresponding to any node in T m,τ is recorded as p m,τ , and the sum of the scene probabilities of each column is 1; divide the planning stage by n year years, denoted as t, and the root node of the scene tree is recorded as "stage 1". The next stage numbers are 2, 3, etc. The scene probability sum of each stage is also 1; each stage has multiple wind load scene sets to be reduced, and an optimal subset is found to make the new probability of each scene in the subset as close as possible to the original scene set probability, so as to reduce the size of the scene tree; at any stage t, the number of wind load scene sets to be reduced is equal to the number of nodes in the previous stage, and the wind load scene set to be reduced in each stage is denoted as , t is the stage number, t = 1, 2, …, T; w = 1, 2, …, n t-1 , which represents the index of the scene set to be reduced corresponding to each node of the previous stage, n t-1The number of nodes in the previous stage, n1 is the number of typical daily light load scenarios generated for the first stage by default; the number of leaf nodes generated by each node in the previous stage is The number of leaf nodes in the next stage is

[0053] D2: When t = 2, the number of nodes in the previous stage is n1, then the scene set to be reduced has n1, each scene set to be reduced is composed of scenes in the column belonging to the second stage in the original scene tree, and the row number is the number corresponding to each node in the previous stage, and the super set is reduced to by using the SD (Sinkhorn distance, SD) method, thereby obtaining the scenes and the corresponding probability in each cluster; when t ≥ 3, each scene set to be reduced is composed of scenes in the column belonging to the tth stage in the original scene tree, and the row number is the number corresponding to each point in each cluster in the previous stage, thereby outputting the scene tree of the leaf node and the corresponding probability stage by stage to form a multi-stage scene tree.

[0054] Further, the operation of the SD method in the step D2 specifically includes the following steps:

[0055] E1: Introducing the regularization coefficient γ and the entropy H = -∑ i,j π i,j [log(π i,j )-1] into the objective function to improve the calculation efficiency of finding the optimal light load scenario set; the model is shown in equation (15)

[0056]

[0057] In the formula: π i,j is an element in the optimal transport plan matrix Π, representing the transition probability of an element (x i ~ P(x)) in the original scenario set to a reduced scenario set (y j ~ Q(y)); the transport matrix Π is an M × N decision matrix, where M and N are the sizes of the scenario set X to which x i belongs and the scenario set Y to which y i belongs; c i,j is the cost of the transition element x i to y i , which is usually calculated using the 2-norm; a i and b j are the probabilities of each scenario in the scenario set X and Y, respectively; the optimal objective value is denoted as S D ;

[0058] E2: To solve equation (15), Sinkhorn-knopp algorithm is used to introduce Lagrange multipliers to transform the constrained optimization problem of equation (15) into an unconstrained problem.

[0059]

[0060] where α M×1 and β N×1 are the Lagrange multipliers corresponding to the equality constraints; Π is the matrix [π i,j ];

[0061] E3: Take the derivative of equation (16) to get

[0062]

[0063] Solve π i,j to get the expression of the optimal transportation plan

[0064]

[0065] where π i,j is the element in the optimal transportation matrix Π; exp(-α i / γ) and exp(-β j / γ) are the diagonal elements of U and V respectively; exp(-c i,j / γ) is the element in the positive definite kernel cost matrix M;

[0066] E4: Then calculate vectors u and v, which are constructed from the diagonal elements of U and V respectively, and their corresponding iterative algorithm is as follows:

[0067]

[0068] where g is the iteration number, when g = 1, all elements in v (g) are 1; the error related to vectors u and v is given by

[0069]

[0070] E5: Continue to calculate vectors u and v iteratively until the error terms ε a and ε b are below the specified error threshold At this time, the u and v vectors are used as the main diagonal elements to construct matrices U and V, and the elements in the optimal transportation plan matrix are calculated by equation (18), thus obtaining the calculation formula of S D

[0071] S D =∑c i,j π i,j +γ∑π​i,j [log(π i,j )-1] (21)

[0072] Further, the step S4 of generating the optimal subset of wind and light load scenarios includes the original wind and light load scenario set I as input, and the optimal subset S containing r' scenarios is generated by reduction * The specific calculation process is as follows:

[0073] F1: the wind and light load scenario set to be reduced is I, the size of the scenario set after reduction is r', the error threshold is β2, the initial iteration number is d' = 1;

[0074] F2: define the wind and light load subset S (d) When d' = 1, S (d') is r' scenarios randomly selected from the superset I, and when the iteration number d ≥ 2, S (d'-1) ;

[0075] F3: calculate the Sinkhorn distance S (d') between I and S D (d') and the optimal transport plan matrix Π (d') using formulas (18-20);

[0076] F4: by finding the column index corresponding to the maximum value of each row of the transport matrix Π (d') , the cluster in which each scenario in the scenario set I is located can be determined, and thus the clusters C1, C2,..., C (d') are formed around S r' ;

[0077] F5: calculate the relative error, when d' = 1, continue step F3, and when d' ≥ 2, calculate the relative error

[0078] F6: if the relative error is greater than the error threshold, calculate the transport distance for each wind and light load scenario in the cluster as the intra-cluster center, and obtain the minimum intra-cluster transport distance as the new intra-cluster center, so as to obtain the updated subset S (d') for the next iteration;

[0079] F7: when the relative error is less than the error threshold, the final reduced subset S * is obtained: S (d') , and the corresponding clusters C1, C2,... C r' ;

[0080] F8: set the appearance probability of all scenarios in I, and thus calculate the probability of the optimal subset S * :

[0081]

[0082] F9: output final output optimal subset S * comprise r' wind and load scenes, each cluster C1, C2, … C r' and the corresponding probability vector.

[0083] Further, the process distance (PD) is introduced in step S4 to measure the approximation degree between the generated wind and load scene tree after reduction and the original wind and load scene tree, so as to evaluate the superiority of the proposed SD-based scene tree reduction method. Based on the introduction of the process distance, the similarity between the original wind and load scene tree and the reduced scene tree is measured, which is defined by formula (22)

[0084]

[0085] In the formula: N T and N' T are all node sets of two scene trees respectively; P1 and P2 are probability sets of two scene trees respectively; symbols f→j and e→i are the predecessors of intermediate nodes f and e, which are leaf nodes j and i respectively; p i , p' j , p e and p' f are probabilities of reaching nodes i, j, e and f respectively; formula (22) is a linear programming problem, which can be solved by moving from one stage to another in a decomposed manner.

[0086] The present application provides a wind and load multi-stage scene tree generation method based on source and load uncertainty feature extraction. The generated wind and load multi-stage scene tree can reflect the uncertainty of medium and long term wind and load output and load growth. In order to improve the clustering efficiency of wind and load scenes, a wind and load scene feature extraction method based on stacked sparse autoencoder and a clustering method based on density peak improved affinity propagation are proposed to obtain typical wind and load curves as root nodes of the scene tree. In order to avoid the size of the generated scene tree being too large, a scene tree reduction method based on Sinkhorn distance is proposed to reduce the size of the scene tree.

[0087] The wind and load multi-stage scene tree generation method based on complex feature extraction and Sinkhorn distance provided by the present application can effectively reflect the timing information of new energy output and load, characterize the randomness and volatility of new energy output and load, and reflect the uncertainty of new energy output and load in a long time scale. It can assist in making various decisions in system planning and operation, thereby reducing the possible safety and economic risk, and further providing support for system operation regulation, source and load dynamic planning, improving new energy consumption capacity, and reducing the influence of multi-dimensional uncertainty.

[0088] Beneficial effects: compared with the prior art, the present application improves the traditional single scene only considered in the definition of the scene tree root node, the scene often only contains limited data values, and cannot represent uncertain information. The generated wind and light multi-stage scene tree can effectively reflect the time sequence information of new energy output and load, represent the randomness and volatility of new energy output and load, and reflect the uncertainty of new energy output and load in a long time scale, further providing support for system operation regulation, source and load dynamic planning, improving new energy consumption capacity, reducing the influence of multi-dimensional uncertainty, etc. BRIEF DESCRIPTION OF DRAWINGS

[0089] Figure 1 is a flowchart of the method of the present application;

[0090] Figure 2 is a structural diagram of the stacked sparse autoencoder model in the present application;

[0091] Figure 3 is a schematic diagram of generating a wind and light load scene tree year by year in the present application;

[0092] Figure 4 is a schematic diagram of generating a simple 3-stage scene tree in the present application;

[0093] Figure 5 is a 3-stage wind and light load scene tree result diagram generated by the method of the present application;

[0094] Figure 6 is a wind and light load scene result diagram represented by each point of the 3-stage scene tree generated by the method of the present application. DETAILED DESCRIPTION

[0095] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that these embodiments are only used to illustrate the present application and not to limit the scope of the present application. After reading the present application, those skilled in the art can make various equivalent modifications of the present application, which all fall within the scope defined by the claims attached hereto.

[0096] The present application provides a wind and light load multi-stage scene tree generation method based on source and load uncertainty feature extraction, as shown in Figure 1 , which includes the following steps:

[0097] S1: Extract wind and light load scene features based on the constructed stacked sparse autoencoder model to obtain a wind and light load scene feature set, specifically:

[0098] A1: Collect annual wind and light load historical data, P WT , P PV and P L are wind, photovoltaic and load unit value data sets, respectively, as follows:

[0099]

[0100]

[0101]

[0102] Where: n um P represents the total number of days in a year. WT (i), P PV (i) and P L (i) represents the data set consisting of the per-unit values ​​of wind power, solar power, and load for each hour on the i-th day of the year; P WT,j P PV,j and P L,j These are the per-unit values ​​for wind power, solar power, and load, respectively, for the i-th day and k-th hour of the year;

[0103] s f For P WT P PV and P L The complete time-series scene set composed of the three is shown below:

[0104]

[0105] In the formula: s f There is n um Each sample contains a landscape scene; let s be the sample size. f The i-th sample is m is s f Medium sample dimension;

[0106] A2: Utilize the constructed stacked sparse autoencoder (SSAE) model to analyze s f Feature extraction is performed to obtain the scene feature extraction results for the scene set.

[0107] like Figure 2 As shown, the SSAE model consists of M SAEs. The SSAE uses the hidden layer of the previous SAE as the input of the input layer of the next SAE. The training process of the SAE includes encoding and decoding. During the encoding process, scene set data features are extracted, as shown in Equation (5):

[0108]

[0109] In the formula: The output of the hidden layer; W a b is the encoding weight matrix; a For encoding bias; σ h =1 / (1+e) -x ) is the activation function;

[0110] Reconstruct data in the decoding process, output extraction scene features, as shown in formula (6):

[0111]

[0112] In the formula: Output of the output layer; W b Decoding weight matrix; b b Decoding bias; sigma y Activation function, same as sigma h ;

[0113] The output data after decoding and the input scene set data constitute an error function, as shown in formula (7):

[0114]

[0115] In the formula: J SAE Error function; rho is the set sparsity coefficient; h j (s f,i ) is the activation value of hidden layer neuron j when the input is s f,i ; Average activation amount of the jth neuron in the hidden layer. Input s f to the SSAE model, and each SAE in the SSAE model is trained according to formulas (5-7) until all SAEs are trained to obtain the feature set y f of s M The ith feature sample is

[0116] In the application, the SSAE based on the stacked sparse autoencoder can obtain lower dimension and deeper features of the wind and light load scene through the stacking of SAE.

[0117] S2: The wind and light load scene feature set is clustered based on the density peak improved affinity propagation algorithm, and the wind and light load typical daily curve is obtained as the root node of the scene tree;

[0118] The feature set Y f of s M is used as a clustering index for clustering, and the generated wind and light load typical time sequence scene set s c is used as the root node of the wind and light load multi-stage scene tree, which specifically includes the following steps:

[0119] B1: Similarity calculation: And The ith and jth samples in Y M are calculated, and a similarity matrix is constructed

[0120]

[0121] B2: Density peak value calculation: Calculate the density peak value γ between sample i' and other samples from local density and relative distance i' ;

[0122] B3: Update similarity matrix: Update the density peak value γ calculated as the preference value in the diagonal element of the similarity matrix i'

[0123] s i'i' = -γ i' (9)

[0124] B4: Calculate the belonging degree matrix and the attraction degree matrix

[0125]

[0126]

[0127] where: a i'k' , s i'k' is the belonging degree and similarity value of the i'th sample to other samples except the j'th sample; r i”j' is the attraction degree value obtained from other samples;

[0128] B5: Introduce damping factor: Prevent oscillation in the iteration process

[0129]

[0130]

[0131] where: d is the iteration number;

[0132] B6: Determine the clustering center: Determine whether it is a clustering center according to formula (14) If formula (14) is greater than 0, then The corresponding full-time sequence s f,i can be used as a clustering center

[0133]

[0134] B7: Determine whether to terminate iteration: When the following conditions are met: 1) The set iteration number is reached; 2) The clustering center no longer changes, the iteration can be terminated; if not, return to step B5.

[0135] The above method can obtain a number of light and shadow scene sets containing clustering centers c1, c2, …, c K ​​​K is the number of cluster centers. The cluster center is the typical day scene of the whole year wind and light load, which is the root node of the scene tree. In addition, the number of each scene set obtained by statistical clustering is counted, and the frequency is calculated as the probability of the scene set. After clustering, three commonly used clustering indexes, contour coefficient, CH index and DB index, are selected to analyze the clustering results and test the effectiveness of clustering.

[0136] S3: According to the root node of the scene tree, the wind and light load scene tree is generated year by year based on considering different growth modes of load, as shown in Figure 3 , specifically:

[0137] C1: Considering the uncertainty of load growth, three different load growth schemes, high, medium and low, are set;

[0138] C2: According to the K typical day scenes of the root node of the scene tree, it is assumed that the load grows year by year and behaves as high, medium and low speed growth modes with different probabilities, and a scene tree is gradually generated by generating scenes. Each point of the scene tree represents a 72-dimensional wind and light load scene of a typical day, and the probability of each scene is determined by the probability of each scene of the root node and the probability of different load growth. The first year of the generated wind and light load scene tree has K possible typical day wind and light load scenes, the second year has 3K possible typical day wind and light load scenes, and so on. The size of the scene tree grows exponentially with the number of stages, and the scene tree needs to be reduced to reduce its size.

[0139] S4: The scene tree is reduced by the scene tree reduction method of Sinkhorn distance to reduce the size of the scene tree;

[0140] As shown in Figure 4 , a simple 3-stage scene tree generation is shown in Figure 4 , the scene tree is reduced by the scene tree reduction method of Sinkhorn distance, which includes the following steps:

[0141] D1: Each node of the original scene tree is denoted as T m,τ , each node represents a 72-dimensional wind and light load scene, where τ is the time step of the original scene tree, and m is the number of nodes under τ; T m,τ The scene probability corresponding to any node in T m,τ is denoted as p yearThe planning stage is divided by year as a unit, denoted as t, the root node of the scenario tree is denoted as "stage 1", and the subsequent stage numbers are 2, 3, etc. The scenario probability and of each stage are also 1; each stage has multiple wind and light load scenario sets to be reduced, and an optimal subset is sought, so that the new probability of each scenario in the subset can be as close as possible to the original scenario set probability, so as to reduce the size of the scenario tree; at any stage t, the number of wind and light load scenarios to be reduced is equal to the number of nodes of the previous stage, and the wind and light load scenario set to be reduced at each stage is denoted as , t is the stage number, t = 1, 2,..., T; w = 1, 2,..., n t-1 , n t-1 indicates the index of the scenario set to be reduced corresponding to each node of the previous stage, n i,j is the number of nodes of the previous stage, and by default n1 is the number of typical daily wind and light load scenarios generated in the first stage; let the number of leaf nodes generated by each node of the previous stage be , then the number of leaf nodes of the next stage is

[0142] D2: when t = 2, the number of nodes of the previous stage is n1, then the scenario set to be reduced has n1, each scenario set to be reduced is composed of the scenarios in the column belonging to the second stage in the original scenario tree, and the row number is the number corresponding to each node of the previous stage. The super set is reduced to by using the SD (Sinkhorn distance, SD) method, thereby obtaining the scenarios and the corresponding probability in each cluster; when t ≥ 3, each scenario set to be reduced is composed of the scenarios in the column belonging to the tth stage in the original scenario tree, and the row number is the number corresponding to each point in the cluster of the previous stage, thereby outputting the scenario tree of the leaf node and the corresponding probability stage by stage.

[0143] The operation of the SD method specifically includes the following steps:

[0144] E1: Introduce the regularization coefficient γ and the entropy H = -∑ i,j π i,j [log(π i,j )-1] into the objective function to improve the calculation efficiency of finding the optimal wind and light load scenario set; the model is shown in formula (15)

[0145]

[0146] In the formula: π i,j is an element in the optimal transportation plan matrix Π, which represents an element (x i~P(x)) to the reduced scene set (y j The transition probability of ~Q(y)); the transportation matrix Π is an M×N decision matrix, where M and N are x i The scene set X and y belong to i Size of the scene set Y; c i,j For the transfer element x i to y i The cost is typically calculated using the 2-norm; a i and b j Let S be the probability of each scene in scene sets X and Y, respectively; and let S be the optimal objective value. D ;

[0147] E2: To solve equation (15), the Sinkhorn-knopp algorithm is used, and Lagrange multipliers are introduced to transform the constrained optimization problem of equation (15) into an unconstrained problem.

[0148]

[0149] In the formula: α M×1 and β N×1 These are the Lagrange multiplier vectors corresponding to the equality constraints; Π is the matrix [π i,j ];

[0150] E3: Differentiating equation (16), we get

[0151]

[0152] Solve for π i,j The expression for obtaining the optimal transportation plan.

[0153]

[0154] Where: π i,j It is an element in the optimal transport matrix Π; exp(-α) i / γ) and exp(-β) j / γ) are the diagonal elements of U and V in the diagonal matrix, respectively; exp(-c i,j / γ) are elements in the positive definite kernel cost matrix M;

[0155] E4: Next, calculate vectors u and v, constructed from the diagonal elements of U and V respectively. The corresponding iterative algorithm is as follows:

[0156]

[0157] In the formula: g is the iteration number; when g = 1, v (g) All elements in the equation are 1; the error associated with vectors u and v is given by the following equation.

[0158]

[0159] E5: The vectors u and v continue to iterate until the error term ε a and ε b is below the specified error threshold The u and v vectors at this point form the matrices U and V with the main diagonal elements, and the elements of the optimal transport plan matrix are calculated by equation (18), and thus S D is obtained

[0160] S D =∑c i,j π i,j +γ∑π i,j [log(π i,j )-1] (21)

[0161] Based on this optimal wind and light load scene set (the original wind and light load scene set I is input, and an optimal subset S * containing r' scenes is generated by reducing it), the algorithm flow is as follows:

[0162] F1: The wind and light load scene set to be reduced is I, the scene set size after reduction is r', the error threshold is β2, the initial iteration number is d' = 1;

[0163] F2: Define the wind and light load subset, when d' = 1, S (d) is r' scenes randomly selected from the super set I, when the iteration number d ≥ 2, S (d') = S (d'-1) ;

[0164] F3: Calculate the Sinkhorn distance S (d') between I and S D (d') and the optimal transport plan matrix Π (d') using equations (18-20);

[0165] F4: By finding the column index corresponding to the maximum value of each row of the transport matrix Π (d') , the cluster in which each scene in the scene set I is located can be determined, and thus the clusters C1, C2, …, C r' are formed around S (d') ;

[0166] F5: Calculate the relative error, when d' = 1, continue step F3, when d' ≥ 2, calculate the relative error

[0167] F6: If the relative error is greater than the error threshold, traverse each wind and load scene in the cluster as the cluster center to calculate the transport distance, and get the minimum intra-cluster transport distance as the new intra-cluster center, so as to get the updated subset S (d') The next iteration is performed;

[0168] F7: When the relative error is less than the error threshold, the final reduced subset S is obtained * :=S (d') , and the corresponding clusters C1, C2,...C r' ;

[0169] F8: Set the appearance probability of all scenes in I, and calculate the optimal subset S * probability:

[0170]

[0171] F9: Output the final optimal subset S * contains r' wind and load scenes, each cluster C1, C2,...C r' and the corresponding probability vector.

[0172] The process distance (PD) is introduced in step S4 to measure the approximation between the wind and load scene tree generated after reduction and the original wind and load scene tree. Based on the introduction of the process distance, the similarity between the original wind and load scene tree and the reduced scene tree is measured, which is defined by formula (22)

[0173]

[0174] In the formula: N T and N' T are all node sets of two scene trees; P1 and P2 are probability sets of two scene trees; symbols f→j and e→i are the predecessors of intermediate nodes f and e, which are leaf nodes j and i; p i , p' j , p e and p' f are the probabilities of reaching nodes i, j, e and f; formula (22) is a linear programming problem, which can be solved by moving from one stage to another in a decomposed manner.

[0175] Based on the above wind and load multi-stage scene tree generation method oriented to source and load uncertainty feature extraction, simulation analysis is performed in this embodiment, as follows:

[0176] The embodiment adopts 8760h wind power, photovoltaic power and load unit value in a certain place in 2023 to form a wind-solar-load full-time sequence scene, and uses the clustering result obtained by the SSAE and improved AP clustering algorithm. In order to further compare the advantages and disadvantages of the clustering effect, three clustering validity test indexes of the original AP clustering, the improved AP clustering, the traditional k-means clustering, the PCA+improved AP clustering and the SSAE+improved AP clustering based on the SSAE+improved AP clustering are calculated respectively, and the results are shown in Table 1.

[0177] Table 1 Comparison of clustering results

[0178]

[0179] The original AP clustering algorithm obtains 23 clustering cluster numbers, which are relatively large. The optimal clustering cluster number obtained by the improved AP clustering algorithm is 3, which greatly reduces the clustering cluster number. In the three clustering indexes, the larger the silhouette coefficient, the larger the CH index and the smaller the DB index, the more compact the cluster itself, the more dispersed the cluster, and the better the clustering effect. It is found from Table 2 that the clustering effect after feature extraction is better than that before feature extraction, and the clustering effect after SSAE feature extraction is better than that after PCA, which indicates that compared with PCA, the method proposed in the application has a more effective ability to extract scene features, and improves the clustering quality of the improved AP algorithm.

[0180] According to the scene tree root node generated above, it is assumed that the load grows year by year, and the load grows at a high speed (4.5%), a medium speed (4%) and a low speed (2%) with probabilities of 0.22, 0.55 and 0.23, respectively, and a 5-year wind-solar-load scene tree is gradually generated, and each point of the scene tree represents a 72-dimensional wind-solar-load scene of a typical day. The large-scale wind-solar-load scene tree generated above is reduced according to the method proposed in the application, and a 3-stage wind-solar-load scene tree is obtained, and the scene number and probability are shown in Table 2, and the scene represented by each node is shown in Table 3. Figure 5 Figure 6 In addition, in order to evaluate the superiority of the Sinkhorn-based scene tree reduction method proposed in the application, a linear programming method based on the traditional optimal transport theory is used as a comparison method, and the two methods are evaluated from the calculation accuracy and calculation efficiency. The process distance and total calculation time calculated by the two methods are shown in Table 2. As shown in Table 2, the method of the application has certain improvement in calculation accuracy and calculation efficiency.

[0181] Table 2 Process distance and time calculated by two methods

[0182]

[0183] ​The above examples only illustrate the technical idea of the present application, and cannot be used to limit the protection scope of the present application. Any modification made according to the technical idea of the present application on the basis of the technical scheme falls within the protection scope of the present application. The technologies not involved in the present application can be realized by using the prior art.

Claims

1. A method for generating a wind-solar load multi-stage scenario tree based on source load uncertainty feature extraction, characterized in that, Comprise the following steps: S1: based on the constructed stack sparse autoencoder model extraction wind and light load scene characteristics, get wind and light load scene feature set; S2: based on density peak improved near neighbor propagation algorithm for wind and light load scene feature set clustering, obtain wind and light load typical daily curve, as the root node of the scene tree; S3: according to the root node of the scene tree, based on considering the different growth mode of load, generate wind and light load scene tree year by year; S4: through the scene tree reduction method of Sinkhorn distance to reduce the scene tree, to reduce the size of the scene tree; In step S2, the feature set Y f is clustered with the clustering index s M , and the generated typical time sequence scene set s c of wind and light loads is taken as the root node of the wind and light load multi-stage scene tree, specifically including the following steps: B1: Similarity Calculation: and Y M Calculate the similarity between the i'-th and j'-th samples to form a similarity matrix. B2: Density peak value calculation: Calculate the density peak value γ between sample i' and other samples from local density and relative distance i' ; B3: Update the similarity matrix: update the density peak value γ i' is the preference value, as the diagonal element in the similarity matrix; s i'i' = -γ i' (9) B4: Compute the Affinity Matrix and the Attraction Matrix of each element in the matrix wherein: a i'k' , s i'k' is the affinity value of the i'th sample to other samples except the j'th sample; r i”j' is the attraction value obtained from other samples; B5: introduce damping factor: prevent the occurrence of shock in the iteration process In the formula: d is the number of iterations; B6: judging cluster center: judging according to formula (14) whether it is a cluster center, if formula (14) is greater than 0, then corresponding all-time sequence scene s f,i as a cluster center; B7: determine whether to terminate iteration: when meeting: 1) reach the set number of iterations; 2) the clustering center no longer changes, the iteration can be terminated; if not, return to step B5; Step S3 is specifically: C1: considering the uncertainty of load growth, set three different load growth schemes of high, medium and low; C2: according to the K typical daily scenes of the root node of the scene tree, assume that the load grows year by year, and shows high, medium and low speed growth mode with different probability, and gradually generates a scene tree, each point of the scene tree represents a 72-dimensional wind and light load scene of a typical day, and the probability of each scene is determined by the probability of each scene of the root node and the probability of different load growth.

2. The wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction according to claim 1, characterized in that, The extraction method of wind and light load scene characteristics in step S1 is specifically: A1: Collecting the annual wind and light load history data, P WT , P PV and P L are the wind power, photovoltaic and load unit data sets, respectively, as follows: where n um is the total number of days in a year; P WT is the i-th day of the year; P PV is the i-th day of the year; P L is the i-th day of the year; P WT,j is the i-th day of the year; P PV,j is the i-th day of the year; P L,j is the i-th day of the year; P s f For P WT , P PV and P L the full timing scenario set consisting of the three, as follows: In the formula: s f There is n um Each sample contains a landscape scene; let s be the sample size. f The i-th sample is m is s f Medium sample dimension; A2: using the constructed stacked sparse auto-encoder SSAE model to extract features of s f to obtain the landscape scene feature extraction result of the scene set.

3. The wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction according to claim 2, characterized in that, The SSAE model in step A2 is composed of M SAEs, the hidden layer of the previous SAE is used as the input of the input layer of the next SAE, and the training process of SAE includes encoding and decoding. In the encoding process, the scene set data characteristics are extracted, as shown in formula (5): where: is the output of the hidden layer; W a is the encoding weight matrix; b a is the encoding bias; σ h = 1 / (1 + e -x ) is the activation function. In the decoding process, the data is reconstructed, and the extracted scene characteristics are output, as shown in formula (6): where: is the output layer output; W b is the decoding weight matrix; b b is the decoding bias; σ y is the activation function, same as σ h ; The output data after decoding and the input scene set data constitute the error function, as shown in formula (7): In the formula: J SAE is an error function; p is a set sparsity coefficient; h j (s f,i ) is the activation value of the hidden layer neuron j when the input is s f,i ; is the average activation amount of the jth neuron in the hidden layer; s f is input into the SSAE model, and the SAE in the SSAE model is trained one by one according to formulas (5)-(7) until all the SAEs are trained, and the feature set y f of s M is obtained, and the ith feature sample is 4. The wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction according to claim 3, characterized in that, In step S4, the scene tree is reduced by the scene tree reduction method of Sinkhorn distance, which specifically includes the following steps: D1: Each node of the original scenario tree is denoted as T m,τ , where τ is the time step of the original scenario tree, and m is the number of each node at τ; T m,τ The scenario probability of any node in T m,τ is denoted as p year , and the sum of the scenario probabilities of each column is 1; the planning phase is divided into n t-1 years, denoted as t, and the root node of the scenario tree is denoted as "Phase 1", and the sum of the scenario probabilities of each phase is also 1; each phase has multiple wind power load scenario sets to be reduced, and an optimal subset is sought so that the new probabilities of each scenario in the subset can be as close as possible to the original scenario set probabilities, so as to reduce the size of the scenario tree; at any stage t, the number of wind power load scenario sets to be reduced is equal to the number of nodes of the previous stage, and the wind power load scenario set to be reduced at each stage is denoted as , t is the stage number, t = 1, 2, …, T; w = 1, 2, …, n t-1 indicates the index of the scenario set to be reduced corresponding to each node of the previous stage, n t-1 is the number of nodes of the previous stage, and by default n1 is the number of typical daily wind power load scenarios generated in the first section; let the number of leaf nodes generated by each node of the previous stage be , then the number of leaf nodes of the next stage is D2: when t=2, the number of nodes in the previous stage is n1, then the scene set to be reduced has n1, each scene set to be reduced is composed of the scenes in the column belonging to the second stage in the original scene tree, the row number is the corresponding number of each node in the previous stage, and the superset is reduced to Thus, the scenes in each cluster and the corresponding probability are obtained; when t≥3, each scene set to be reduced is composed of the scenes in the column belonging to the tth stage in the original scene tree, the row number is the corresponding number of each point in each cluster in the previous stage, thus the scene tree of the leaf node is output stage by stage, as well as the corresponding probability to form a multi-stage scene tree.

5. The wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction according to claim 4, characterized in that, The operation of SD method in step D2 specifically includes the following steps: E1: Introduce the regularization coefficient γ and the entropy H = -∑ i,j π i,j [log(π i,j )-1] into the objective function to improve the computational efficiency of finding the optimal wind-solar load scenario set; the model is shown in equation (15) where: π i,j is an element in the optimal transportation plan matrix Π, representing the transition probability from an element (x i ) in the original scenario set to an element (y j ) in the reduced scenario set; the optimal transportation plan matrix Π is an M x N decision matrix, where M and N are the sizes of the x i and y i scenario sets, respectively; c i,j is the cost of transitioning from an element x i to an element y i ; a i and b j are the probabilities of each scenario in the scenario sets X and Y, respectively; and the optimal objective value is denoted as S D . E2: for solving formula (15), Sinkhorn-knopp algorithm is adopted, Lagrange multiplier is introduced, and the constrained optimization problem of formula (15) is converted into an unconstrained problem; where: a M×1 and β N×1 are the Lagrange multiplier vectors corresponding to the equality constraints; and Π is the optimal transportation plan matrix [π i,j ]. E3: derive formula (16) to get Solving π i,j , obtaining an expression for the optimal transportation plan where exp(-a i / γ) and exp(-β j / γ) are the diagonal elements of U and V respectively; and exp(-c i,j / γ) are the elements of the positive definite kernel cost matrix M. E4: then calculate the vectors u and v, which are constructed by the diagonal elements of U and V respectively, and the corresponding iterative algorithm is as follows: where: g is the number of iterations, when g = 1, v (g) All elements in the middle are 1; the error associated with the vectors u and v is given by E5: The vectors u and v continue to be iteratively computed until the error term ε a and ε b is below a specified error threshold The u and v vectors at this point form matrices U and V with the main diagonal elements, and the elements of the optimal transportation plan matrix S are computed from equation (18) D The formula for the computation of S D =∑c i,j π i,j +γ∑τ i,j [log(π i,j )-1] (21).

6. The wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction according to claim 5, characterized in that, The optimal wind light load scene set generation in step S4 includes the original wind light load scene set I as input, and reduces it to generate an optimal subset S containing r' scenes * The specific calculation process is as follows: F1: the wind and light load scene set to be reduced is I, the size of the reduced scene set is r', the error threshold is β2, the initial iteration number is d' = 1; F2: define a subset of wind and light, S when d' = 1 (d) are r' scenes randomly selected from the super set I when the iteration number d ≥ 2, S (d') = S (d'-1) ; F3: Compute I and S using equations (18)-(20) (d') D (d') and the optimal transport plan matrix P;​ F4: Determine the cluster each scene in the scene set I belongs to by finding the column index corresponding to the maximum value in each row of the optimal transport plan matrix Π, thus forming clusters C1, C2,..., C (d') around S r' ; F5: Calculate relative error, continue with step F3 when d' = 1, calculate relative error when d' > 2 F6: If the relative error is greater than the error threshold, traverse each scenic field in the cluster as the cluster center to calculate the transport distance, and obtain the minimum intra-cluster transport distance as the new cluster center, thereby obtaining the updated subset S (d') Proceed to the next iteration; F7: obtaining a final reduced subset S when the relative error is less than the error threshold * : = S (d') and the corresponding clusters C1, C2,..., C r' ; F8: Set the probabilities of occurrence of all scenes in Set I, thereby computing the optimal subset S * the probability that F9: output final output optimal subset S * comprising r' wind and light load scenarios, each cluster C1, C2,..., C r' and the corresponding probability vector.

7. The wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction according to claim 1, characterized in that, In step S4, the process distance is introduced to measure the approximation degree between the reduced wind and light load scene tree and the original wind and light load scene tree.

8. The wind and light load multi-stage scene tree generation method based on source load uncertainty feature extraction according to claim 7, characterized in that, In step S4, the similarity between the original wind and light load scene tree and the reduced scene tree is measured based on the introduction of process distance, which is defined by formula (22) where N T and N' T are the sets of all nodes of the two scene trees respectively; P1 and P2 are the probability sets of the two scene trees respectively; the symbols f→j and e→i are the predecessors of the intermediate nodes f and e are the leaf nodes j and i respectively; p i , p' j , p e and p' f are the probabilities of reaching the nodes i, j, e and f respectively; the equation (22) is a linear programming problem which is solved by moving from one stage to another in a decomposed way.

Images