A power grid energy storage method considering wind and light uncertainty and timing characteristics

Abstract

The present application relates to the field of power grid energy storage, and discloses a power grid energy storage method considering wind and light uncertainty and timing characteristics, comprising: obtaining wind and light load typical seasonal scenes by SOM clustering method for scene reduction, establishing a multi-time scale energy storage model considering long-term energy storage, and constructing a power grid wind and light storage stochastic programming model considering the economy and low carbon of power grid investment; the present application solves the power balance during the day by constructing a long-short-term multi-time scale hybrid energy storage model suitable for real conditions; and at the same time, the wind and light output uncertainty and the calculation efficiency are considered, typical scenes are generated by using the SOM clustering method, and wind and light load typical seasonal data sets are obtained, so as to ensure the calculation accuracy and improve the calculation efficiency.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of power grid energy storage, and more particularly to a power grid energy storage method considering wind and light uncertainty and timing characteristics. BACKGROUND

[0002] At present, the installed capacity of new energy mainly based on wind and light is increasing year by year, and the output of wind and light is greatly affected by the external environment, showing obvious seasonal characteristics. The main performance is that in summer, the light resource is rich, the photovoltaic output is large, and the wind turbine output is small. In winter, the wind energy resource is rich, the wind turbine output is large, and the photovoltaic output is small. The load demand also has obvious seasonal characteristics. In summer, the demand for air conditioning and other refrigeration loads increases, and in winter, the demand for electric heaters and other heating loads increases. There is a long-term power imbalance between the supply of wind and light on the source side and the demand of users on the load side. The commonly used energy storage at present is energy storage with short-term energy regulation characteristics, such as electrochemical energy storage, which cannot solve the long-term power imbalance between the source and the load. Considering long-term energy storage, it can realize the long-time scale transfer of power, improve the utilization rate of wind and light, meet the load demand, reduce the power supply of thermal power units, and help achieve "carbon peak and carbon neutral". How to use long-term energy storage to solve the long-term power imbalance between the source and the load and ensure the stable operation of the power grid has become one of the research hotspots in recent years

[0003] And the long-term energy storage is in the time scale of weeks, months or even seasons, so the installation and storage cost is much higher than that of short-term energy storage. How to realize the economic operation of the system after considering the long-term energy storage is a difficulty. At the same time, in the multi-time scale hybrid energy storage system composed of long-term energy storage and short-term energy storage, the short-term energy storage realizes the power balance within a day, and the long-term energy storage can also be normally charged and discharged within a day, but how to realize the power balance between days is still a big technical problem, and a long-term and short-term multi-time scale hybrid energy storage model suitable for real conditions needs to be established.

[0004] Therefore, we propose a power grid energy storage method considering wind and light uncertainty and timing characteristics. SUMMARY

[0005] The purpose of the present application is to solve the above problems by providing a power grid energy storage method considering wind and light uncertainty and timing characteristics.

[0006] The above technical purpose of the present application is realized by the following technical scheme: a power grid energy storage method considering wind and light uncertainty and timing characteristics, comprising: obtaining wind, light and load typical seasonal scenes by reducing scenes through a SOM clustering method; considering long-term energy storage, a multi-time scale energy storage model is established; and a power grid wind, light, storage random programming model is constructed by comprehensively considering the economy and low carbon of power grid investment.

[0007] Specifically, the method comprises the following steps:

[0008] S1, a wind and light load seasonal scenario generation method considering uncertainty;

[0009] The method comprises generating a wind and light output scenario considering uncertainty by using an improved generative adversarial network, and on this basis, using a Latin hypercube sampling method and a SOM clustering algorithm to reduce the wind and light load data pairs to obtain typical wind and light load scenarios with seasonal characteristics; S2, a power grid planning method considering multi-time scale energy storage collaboration;

[0010] The method comprises planning a power grid based on typical seasonal scenarios of wind and light loads considering uncertainty, proposing a multi-time scale hybrid energy storage model of short-term energy storage and long-term energy storage collaboration, and on this basis, proposing a power grid planning method considering multi-time scale energy storage collaboration.

[0011] Further, in step S1, the improved generative adversarial network comprises a traditional generative adversarial network including a generator and a discriminator: the generator receives input noise data to generate data, the discriminator judges whether the data generated by the generator is real according to historical data, and after the discriminator judges that the synthesized data is real, the generator and the discriminator will be updated by fixing each other until the discriminator cannot judge whether the data given by the generator is synthesized data, and then the generator and the discriminator stop updating.

[0012] Let historical wind and light output data be x, and its distribution be P data (x), random noise z is described by a Gaussian distribution P z , suppose the generator is G and the discriminator is D, in the training process, noise z ~ P z is sent to the generator, and the output generated wind and light output G(z) ~ P g , at the same time, historical wind and light output is sent to the discriminator, and the discriminator judges the generated data,

[0013] The loss function of the generator is:

[0014] L G =-E z～p(z) [logD(G(z))] (1)

[0015] In the formula, E is the mathematical expectation, D(G(z)) is the judgment result of the discriminator on the generated data,

[0016] The loss function of the discriminator is:

[0017]

[0018] In the formula, D(x) is the judgment result of the discriminator on the real data, and 1-D(G(z)) is the judgment result of the discriminator on the synthesized data,

[0019] The goal of the generator is to generate synthetic data that looks more like real data, and the goal of the discriminator is to identify as much synthetic data as possible, so the objective function of the generative adversarial network is:

[0020]

[0021] The objective function solving process is equivalent to solving P data (x) and P g JS divergence problem:

[0022]

[0023] But P data (x) and P g It is difficult to have an ideal overlap, so the JS divergence is easy to become a constant, which leads to the discriminator being unable to identify the distance between the two distributions, resulting in gradient disappearance and poor training effect. To solve this problem, the traditional generative adversarial network is improved: the wasserstein distance is introduced to optimize the adversarial network, and the wasserstein distance is:

[0024]

[0025] In the formula, P r , P w are two different distributions, ∏(P r , P w ) is the joint distribution of the two distributions (P r , P w ), γ is a possible distribution in the set, x is a real sample, y is a generated sample, (x, y) is generated from γ, ||x-y|| is the distance between the two samples, E (x,y)～γ [||x-y||] is the energy consumption in it,

[0026] The objective function of the improved generative adversarial network is as follows:

[0027]

[0028] In the formula, D∈1-Lipsschitz makes the discriminator conform to the 1-Lipsschitz function, which limits the range of the discriminator, so that P data (x) and P g are not much different,

[0029] The loss function of the generator is:

[0030]

[0031] The loss function of the discriminator is:

[0032]

[0033] Further, the step S1 is combined with the Latin hypercube sampling method and the SOM clustering algorithm to reduce the scene, and the wind and light load data pairs are obtained. The specific steps are as follows:

[0034] Wind and light load seasonal scene reduction

[0035] The improved generative adversarial network generates wind and light data with uncertainty, and 8760 groups of data pairs are formed with the load historical data at the same time of the year. According to the time sequence from small to large, the Latin hypercube sampling method (LHS) is used for sampling, and an initial data set with time sequence characteristics is constructed,

[0036] The specific steps of LHS are as follows:

[0037] Suppose the LHS sampling object size is N, there are k random variables, and the probability density function of the random variable x k is as follows:

[0038] y k =f k (x k ) (9)

[0039]

[0040] 1) Divide the probability space of each random variable x k into N equal intervals, that is: {[0, 1 / N], [1 / N, 2 / N], …, [(N-1) / N, 1]},

[0041] 2) Randomly select a sampling point y ki (i=1,2,…,N)

[0042] 3) Use the inverse transformation function as shown in formula (2) to find the sampling point x ki of each interval, and arrange them in time sequence from small to large to obtain the initial sampling matrix S with k x N order, that is, the wind and light load initial data set of the whole year,

[0043] The obtained wind and light load initial data set of the whole year is normalized, and the self-organizing map (SOM) clustering algorithm is used for scene reduction to obtain 4 groups of clustering results, that is, the wind and light load data set with seasonal characteristics. The four data sets correspond to spring, summer, autumn and winter respectively, and in each data set, the wind and light load data pairs have corresponding seasonal characteristics,

[0044] SOM clustering algorithm automatically classifies data according to its characteristics, the basic principle is to classify two data with small Euclidean distance as a class, and use this way to group data, the specific steps of SOM clustering algorithm are as follows:

[0045] First, set a data set D, the sample in set D is x1, x2, …, x n , that is: D={x1, x2, …, x n}, each sample contains n elements, such as the elements in sample x i are x i1 , x i2 , …, x in , that is: x i =(x i1 , x i2 , …, x in ), similarly, the elements in sample x j are x j1 , x j2 , …, x jn , that is x j =(x j1 , x j2 , …, x jn ),

[0046] Calculate the Euclidean distance between sample x i and sample x j , the Euclidean distance d(x i , x j ) is expressed as follows:

[0047]

[0048] 1) initialization, initialize the SOM network connection weight as a random value, determine the learning initial value, input vector initial value and the number of output neurons, that is, the number of clusters,

[0049] 2) confirm the best matching unit (best matching unit, BMU), calculate and compare the distance between the field nodes by the Euclidean distance formula, find the minimum distance, and confirm the winning neuron,

[0050] 3) weight update adjustment, update the node neuron weight within the BMU radius, use the Gaussian function to shrink the field range of each iteration, and the neighborhood radius is expressed as:

[0051] r(n)=r0exp(-n / a) (12)

[0052] In the formula, r(n) is the field radius of the n th update; r0 is the initial value; a is a constant,

[0053] 4) Repeat step 2) - step 3) until the maximum number of iterations, output results,

[0054] 5) Finally, the number of clusters is obtained by calculating the davies-bouldin index (DBI), and the cluster with the smallest DBI is selected as the final result of clustering, and the calculation method of DBI is as follows:

[0055]

[0056] N i,j =||C i -C j || p (15)

[0057] In the formula: N is the total number of clusters; L i,j The relationship between the i-th cluster and the j-th cluster in the clustering result is described, and the smaller the value is, the better the clustering is; N i,j is the distance between clusters i and j; S i is the Euclidean distance of cluster i, T i is the size of the i-th cluster matrix; Y k is the k-th feature vector in the cluster; C i is the centroid of the i-th cluster; C j is the centroid of the j-th cluster; and p is the dimension of the cluster vector.

[0058] Further, the multi-time scale hybrid energy storage model in step S2 is established as follows:

[0059] The short-term energy storage charges and discharges in a short time scale to maintain power balance in a short time, the time scale is hours, and the model is as follows:

[0060] SOC constraint:

[0061] SOC0=0.5c (16)

[0062] SOC1=SOC 24 (17)

[0063] SOC min ≤SOC≤SOC max (18)

[0064] In the formula, SOC1 is the SOC state of the short-term energy storage at the initial time, c is the capacity of the short-term energy storage, SOC 24 represents the SOC state of the short-term energy storage at the end time, SOC min , SOC max are the maximum and minimum values of the short-term energy storage SOC allowed,

[0065] Equation 16 indicates that the initial time of short-term energy storage is half of the capacity, equation 17 indicates that the initial time and the final time of short-term energy storage in a day are the same, and equation 18 indicates that the short-term energy storage is operated within the allowed range of SOC,

[0066] Power constraint:

[0067] P ESS,min ≤P ESS ≤P ESS,max (19)

[0068] In the formula, P ESS,min , P ESS,max are the minimum and maximum values of the charging and discharging of short-term energy storage, P ESS is the charging and discharging power of short-term energy storage,

[0069] Long-term energy storage is charged and discharged in a long time scale to maintain long-term power balance, and the time scale is season, and the model is as follows:

[0070] SOC constraint:

[0071]

[0072]

[0073] In the formula, is the SOC state of long-term energy storage at the initial time of the first typical day, c ESS is the capacity of long-term energy storage, is the SOC state of long-term energy storage at the final time of the kth typical day, is the SOC state of long-term energy storage at the initial time of the k+1th typical day, is the SOC state of long-term energy storage at the final time of the fourth typical day, is the SOC state of long-term energy storage at time t of the kth typical day, is the SOC state of long-term energy storage at time t-1 of the kth typical day, is the self-discharge rate of long-term energy storage, a ESS is the efficiency of long-term energy storage, and are the charging and discharging amounts of long-term energy storage at time t of the kth typical day, respectively, and Δt is the scheduling duration, SOC ESS,min , SOC ESS,max are the maximum and minimum values of the SOC of long-term energy storage, respectively,

[0074] Formula 20 represents that the long-term energy storage is half of the capacity at the initial time of the first typical day, formula 21 represents that the SOC of the long-term energy storage at the end of the kth typical day is the same as that at the initial time of the k+1th typical day, formula 22 represents that the SOC of the long-term energy storage at the initial time of the first typical day is the same as that at the end of the fourth typical day, formula 23 is a calculation method of the SOC of the long-term energy storage at time t of the kth typical day, and formula 24 represents that the long-term energy storage operates within the allowed range of the SOC,

[0075] Power constraint:

[0076]

[0077] In the formula, is the charge-discharge state of the long-term energy storage, which is 1 or 0, 1 represents that the long-term energy storage is charging and discharging, and 0 represents that the long-term energy storage is not charging and discharging; P ESS,max is the maximum value of the charge-discharge of the long-term energy storage,

[0078] A power grid planning method considering multi-time scale energy storage collaboration

[0079] Based on the multi-time scale hybrid energy storage model obtained above, problem B in the proposed power grid planning model is described as follows,

[0080] Objective function

[0081] The objective function of the optimization problem is to minimize the total cost C of the power grid:

[0082]

[0083] In the formula, C1 is the economic cost of the power grid, and C2 is the green cost of the power grid; p In,RE , p In,Sto are the installation costs per unit capacity of wind and light, and energy storage (including short-term energy storage and long-term energy storage); c RE , c Sto are the installed capacities of wind and light, and energy storage (including short-term energy storage and long-term energy storage); p OM,RE , p OM,Sto are the operation and maintenance costs per unit capacity of wind and light, and energy storage (including short-term energy storage and long-term energy storage); P RE , P Sto are the outputs of wind and light, and energy storage (including short-term energy storage and long-term energy storage); p Buy,MG , p Sel,MG are the purchase and sale prices of electricity in the power grid, and P Buy,MG , P Sel,MG are the purchase and sale quantities of electricity in the power grid; μ is the penalty coefficient of abandoned wind and light; P Waste is the amount of abandoned wind and light:

[0084] P Waste= P RE -P Load -P Sto -P Sel,MG (29)

[0085] wherein P Load is the load power,

[0086] Constraints

[0087] Capacity constraints:

[0088]

[0089] wherein c RE,max is the upper limit of the wind-solar installation capacity, c Sto is the upper limit of the energy storage installation capacity (including short-term energy storage and long-term energy storage),

[0090] Power balance constraints:

[0091] P RE + P Sto -P Sel,MG + P Buy,MG = P Load (31)

[0092] Wind-solar power output constraints:

[0093] P RE,min ≤ P RE ≤ P RE,max (32)

[0094] wherein P RE,min and P RE,max are the minimum value and the maximum value of the wind-solar power output, respectively,

[0095] Large grid power purchase and sale constraints:

[0096]

[0097] wherein P Buy,MG,max and P Buy,MG,min are the upper and lower limits of the power purchase from the large grid, respectively, and P Sel,MG,max and P Sel,MG,min are the upper and lower limits of the power sale to the large grid, respectively,

[0098] In addition to the above constraints, the constraint conditions of the grid planning model also include the above-mentioned equations (9)-(15).

[0099] According to the above scheme,

[0100] The beneficial effects compared with the prior art are as follows:

[0101] The application realizes the daily power balance by constructing a long-short-term multi-time scale hybrid energy storage model suitable for real conditions; and simultaneously taking into account the wind and light output uncertainty and the high efficiency of calculation, using the SOM clustering method to generate typical scenarios, obtaining the wind and light load typical seasonal data set, while ensuring the calculation accuracy and improving the calculation efficiency. DETAILED DESCRIPTION

[0102] The application will be further described in detail below by specific examples, and the following examples are only descriptive, not limiting the protection scope of the application.

[0103] The application comprises: a wind and light load seasonal scenario generation method considering uncertainty and a power grid planning method considering multi-time scale energy storage coordination. The SOM clustering method is used to reduce the scenarios to obtain the wind and light load typical seasonal scenarios, a multi-time scale energy storage model is established considering long-term energy storage, and a power grid wind and light storage stochastic programming model is constructed considering the economy and low carbon of power grid investment.

[0104] 1. A wind and light load seasonal scenario generation method considering uncertainty

[0105] Due to environmental factors, the wind and light output has the characteristics of randomness and volatility. In addition, the wind and light output and the load demand have obvious seasonal characteristics: the wind power plant output is less and the photovoltaic power station output is more in summer, the wind power plant output is more and the photovoltaic power station output is less in winter, and the load demand increases in summer and winter. In power grid planning, the selection of wind and light load seasonal typical scenarios considering uncertainty is very important. Generating wind and light load seasonal typical scenarios containing uncertainty includes two problems, problem A: an improved generative adversarial network is proposed to generate wind and light output scenarios considering uncertainty. On this basis, problem B: a method combining Latin hypercube sampling and SOM clustering algorithm is used to reduce the wind and light load data to obtain wind and light load typical scenarios with seasonal characteristics.

[0106] A. Improved generative adversarial network

[0107] The traditional generative adversarial network includes a generator and a discriminator: the generator receives the input noise data to generate data, and the discriminator judges whether the data generated by the generator is real according to the historical data. When the discriminator judges that the synthesized data is correct, the generator and the discriminator will be updated by fixing each other, until the discriminator cannot judge whether the data given by the generator is synthesized data, then the generator and the discriminator stop updating.

[0108] Let the historical wind and light output data be x, and its distribution be P data (x), and the random noise z adopts a Gaussian distribution P z . Assume that the generator is G and the discriminator is D. In the training process, the noise z ~ P zThe generated wind and light output G(z)~P is sent to the generator, and the generated wind and light output G(z)~P is output g At the same time, the historical wind and light output is sent to the discriminator, and the discriminator judges the generated data.

[0109] The loss function of the generator is:

[0110] L G = -E z～p(z) [logD(G(z))] (1)

[0111] In the formula, E is the mathematical expectation, and D(G(z)) is the judgment result of the discriminator on the generated data.

[0112] The loss function of the discriminator is:

[0113]

[0114] In the formula, D(x) is the judgment result of the discriminator on the real data, and 1-D(G(z)) is the judgment result of the discriminator on the synthesized data.

[0115] The goal of the generator is to generate more synthesized data that looks like real data, and the goal of the discriminator is to identify as many synthesized data as possible. Therefore, the objective function of the generative adversarial network is:

[0116]

[0117] The solving process of the objective function is equivalent to solving the JS divergence problem of P data (x) and P g

[0118]

[0119] However, P data (x) and P g have a very ideal overlap, so the JS divergence is easy to become a constant, which makes the discriminator unable to identify the distance between the two distributions, resulting in gradient disappearance and poor training effect. To solve this problem, the traditional generative adversarial network is improved: the wasserstein distance is introduced to optimize the adversarial network. The wasserstein distance is:

[0120]

[0121] In the formula, P r ,P w are two different distributions, and Π(P r ,P w ) is the joint distribution of the two distributions (P r ,P w ​where X is a set of real samples, Y is a set of generated samples, (x, y) is generated from γ, ||x-y|| is the distance between two samples, and E (x,y)～γ [||x-y||] is the energy consumption in which.

[0122] The improved GAN objective function is as follows:

[0123]

[0124] where D∈1-Lipsschitz makes the discriminator conform to the 1-Lipsschitz function, limits the range of the discriminator, and makes P data (x) and P g are not much different.

[0125] The loss function of the generator is:

[0126]

[0127] The loss function of the discriminator is:

[0128] B. Wind and light load seasonal scenario reduction

[0129] The wind and light data with uncertainty generated by the improved GAN and the load history data at the same time of the year are combined into 8760 data pairs, which are sorted in ascending order according to the time sequence, and are sampled by Latin Hypercube Sampling (LHS) to construct an initial data set with time sequence characteristics.

[0130] The specific steps of LHS are as follows:

[0131] Assuming that the LHS sampling object size is N, there are k random variables, and the random variable x k has the following probability density function:

[0132] y k = f k (x k ) (9)

[0133]

[0134] 4) Divide the probability space of each random variable x k into N equal intervals to form N subintervals, i.e.: {[0, 1 / N], [1 / N, 2 / N], …, [(N-1) / N, 1]}.

[0135] 5) Randomly select a sampling point y ki (i = 1, 2, …, N) in each subinterval

[0136] 6) Using the inverse transform of the function as shown in formula (2) to obtain each interval sampling point x ki , and arranging them in time sequence from small to large to obtain an initial sampling matrix S with sampling value k x N order, that is, the initial annual data set of wind and light load.

[0137] The obtained annual initial data set of wind and light load is normalized, and a self-organizing map (SOM) clustering algorithm is used to reduce scenes to obtain four groups of clustering results, that is, wind and light load data sets with seasonal characteristics. The four data sets correspond to the four seasons of spring, summer, autumn and winter respectively, and in each data set, the wind and light load data pairs have corresponding seasonal characteristics.

[0138] The SOM clustering algorithm automatically classifies data according to their characteristics, and the basic principle is to classify two data with small Euclidean distance as a class, and use this way to group data. The specific steps of the SOM clustering algorithm are as follows:

[0139] First, set a data set D, the samples in set D are x1, x2, …, x n , that is: D = {x1, x2, …, x n}, each sample contains n elements, such as the elements in sample x i are x i1 , x i2 , …, x in , that is: x i = (x i1 , x i2 , …, x in ), and similarly, the elements in sample x j are x j1 , x j2 , …, x jn , that is x j = (x j1 , x j2 , …, x jn ).

[0140] Calculate the Euclidean distance between sample x i and sample x j . The Euclidean distance d(x i , x j ) is represented as follows:

[0141]

[0142] 1) Initialization. Initialize the SOM network connection weight to a random value, determine the learning initial value, input vector initial value and the number of output neurons, that is, the number of clusters.

[0143] 2) Confirm the best matching unit (BMU). Calculate and compare the distance between field nodes by the Euclidean distance formula, find the minimum distance, and confirm the winning neuron.

[0144] 3) Weight update adjustment. Update the node neuron weight within the BMU radius, use the Gaussian function to shrink the field range of each iteration, and the neighborhood radius is expressed as:

[0145] r(n) = r0exp(-n / a) (12)

[0146] In the formula, r(n) is the field radius of the nth update; r0 is the initial value; a is a constant.

[0147] 4) Repeat steps 2) to 3) until the maximum number of iterations, and output the results.

[0148] 5) Finally, the number of clusters is obtained by calculating the Davies-Bouldin index (DBI), and the cluster with the smallest DBI is selected as the final result of clustering. The calculation method of DBI is shown as follows:

[0149]

[0150] N i,j =||C i -C j || p (15)

[0151] In the formula: N is the total number of clusters; Li,j describes the relationship between the ith and jth clusters in the clustering result, the smaller the value, the better the clustering; Ni,j is the distance between clusters i and j; Si is the Euclidean distance of cluster i; Ti is the size of the ith cluster matrix; Yk is the kth feature vector in the cluster; Ci is the centroid of the ith cluster; Cj is the centroid of the jth cluster; p is the dimension of the clustering vector. 2. Power grid planning method considering multi-time scale energy storage collaboration

[0152] Based on the first part, the wind and light load typical seasonal scenarios considering uncertainty are generated to plan the power grid. The proposed power grid planning model includes two problems, problem A: a multi-time scale hybrid energy storage model considering short-term and long-term energy storage collaboration. On this basis, problem B: a power grid planning method considering multi-time scale energy storage collaboration is proposed.

[0153] A. Multi-time scale hybrid energy storage model

[0154] Short-term energy storage charges and discharges in a short time scale to maintain power balance in a short time, the time scale is hours, and its model is as follows:

[0155] 1) SOC constraint:

[0156] SOC0=0.5c (16)

[0157] SOC1=SOC 24 (17)

[0158] SOC min ≤SOC≤SOC max (18)

[0159] In the formula, SOC1 is the SOC state of short-term energy storage at the initial moment, c is the capacity of short-term energy storage, SOC 24 Indicates the SOC state of short-term energy storage at the end of the day. SOC min , SOC max They are the maximum and minimum values ​​allowed for short-term energy storage SOC respectively.

[0160] Formula 16 indicates that the initial time of short-term energy storage is half of the capacity, Formula 17 indicates that the SOC of the short-term energy storage at the initial time and the final time of the day are the same, and Formula 18 indicates that the short-term energy storage operates within the allowable SOC range.

[0161] 2) Power Constraints:

[0162] P ESS,min ≤P ESS ≤P ESS,max (19)

[0163] Where, P ESS,min , P ESS,max are the minimum and maximum values ​​of short-term energy storage charge and discharge, P ESS is the charging and discharging power of short-term energy storage.

[0164] Long-term energy storage charges and discharges over a long time scale to maintain long-term power balance. The time scale is seasonal. Its model is as follows:

[0165] 1) SOC constraint:

[0166]

[0167]

[0168] Where, is the SOC state of long-term energy storage at the initial moment of the first typical day, c ESS is the capacity of long-term energy storage, is the SOC state of long-term energy storage at the end of the kth typical day, is the SOC state of the long-term energy storage at the initial time of the k+1th typical day, is the SOC state of long-term energy storage at the end of the fourth typical day, is the SOC state of long-term energy storage at the kth typical day t, is the SOC state of long-term energy storage at the kth typical day t-1, is the self-discharge rate of long-term energy storage, a ESS is the efficiency of long-term energy storage, and are the charging and discharging amounts of long-term energy storage at the kth typical day t, respectively, and At is the scheduling duration, SOC ESS,min , SOC ESS,max are the maximum and minimum values of the SOC of long-term energy storage, respectively.

[0169] Formula 20 indicates that the long-term energy storage is half of the capacity at the initial moment of the first typical day, formula 21 indicates that the SOC of the long-term energy storage at the end of the kth typical day is the same as that at the initial moment of the k+1th typical day, formula 22 indicates that the SOC of the long-term energy storage at the initial moment of the first typical day is the same as that at the end of the fourth typical day, formula 23 is a calculation method of the SOC of the long-term energy storage at the kth typical day t, and formula 24 indicates that the long-term energy storage is operated within the range of the allowed SOC.

[0170] 2) Power constraint:

[0171]

[0172] In the formula, is the charging and discharging state of long-term energy storage, which is 1 or 0, 1 indicates that the long-term energy storage is charging and discharging, and 0 indicates that the long-term energy storage is not charging and discharging; P ESS,max is the maximum value of the charging and discharging of long-term energy storage.

[0173] B. Power grid planning method considering multi-time scale energy storage coordination

[0174] Based on the multi-time scale hybrid energy storage model obtained above, problem B in the proposed power grid planning model is described as follows.

[0175] 1. Objective function

[0176] The objective function of the optimization problem is to minimize the total cost C of the power grid:

[0177]

[0178] In the formula, C1 is the economic cost of the power grid, and C2 is the green cost of the power grid; p In,RE , p In,Sto are the installation costs per unit capacity of wind and light, and energy storage (including short-term energy storage and long-term energy storage); c RE , c Sto are the installed capacities of wind and light, and energy storage (including short-term energy storage and long-term energy storage).OM,RE , p OM,Sto are respectively the unit operation and maintenance cost of wind and light, energy storage (including short-term energy storage and long-term energy storage); P RE , P Sto are respectively the output of wind and light, energy storage (including short-term energy storage and long-term energy storage); p Buy,MG , p Sel,MG are respectively the purchase and sale electricity price of the power grid, P Buy,MG , P Sel,MG are respectively the purchase and sale electricity quantity of the power grid; μ is the penalty coefficient of abandoned wind and light; P Waste is the abandoned wind and light quantity:

[0179] P Waste = P RE -P Load -P Sto -P Sel,MG (29)

[0180] In the formula, P Load is the load electricity quantity.

[0181] 2. Constraint conditions

[0182] 1) Capacity constraint:

[0183]

[0184] In the formula: c RE,max is the upper limit of wind and light installation capacity, c Sto is the upper limit of energy storage installation capacity (including short-term energy storage and long-term energy storage).

[0185] 2) Power balance constraint:

[0186] P RE + P Sto -P Sel,MG + P Buy,MG = P Load (31)

[0187] 3) Wind and light output constraint:

[0188] P RE,min ≤ P RE ≤ P RE,max (32)

[0189] In the formula: P RE,min , P RE,max are respectively the minimum value and the maximum value of wind and light output.

[0190] 4) Large power grid purchase and sale electricity constraint:

[0191]

[0192] In the formula: PBuy,MG,max , P Buy,MG,min are respectively upper and lower limits of the power purchased from the large power grid, P Sel,MG,max , P Sel,MG,min are respectively upper and lower limits of the power sold to the large power grid.

[0193] In addition to the above constraints, the constraint conditions of the power grid planning model also include the above-mentioned formulas (9)-(15).

[0194] The specific embodiments are only an explanation of the present application, which is not a limitation to the present application, and the person skilled in the art can make modifications to the embodiments without creative contribution according to the needs after reading the specification, but as long as it is within the scope of the claims of the present application, it is protected by the patent law.

Claims

1.A method for grid energy storage considering wind and solar uncertainty and time-of-use characteristics, characterized in that: The wind and light load typical seasonal scene is obtained by reducing the scene through the SOM clustering method, a multi-time scale energy storage model is established considering long-term energy storage, and a grid wind and light storage random programming model is constructed considering the economy and low carbon of grid investment. Specifically, the method comprises the following steps: S1, a wind and light load seasonal scene generation method considering uncertainty; The improved generative adversarial network is used to generate the wind and light output scene considering uncertainty, and then the Latin hypercube sampling method and the SOM clustering algorithm are combined to reduce the scene of the wind and light load data pair, so as to obtain the wind and light load typical scene with seasonal characteristics; S2, a grid planning method considering multi-time scale energy storage coordination; Based on the wind and light load typical seasonal scene considering uncertainty, the grid is planned, a multi-time scale hybrid energy storage model of short-term energy storage and long-term energy storage coordination is established, and then a grid planning method considering multi-time scale energy storage coordination is proposed; The specific steps of reducing the scene of the wind and light load data pair by combining the Latin hypercube sampling method and the SOM clustering algorithm in the step S1 are as follows: Wind and light load seasonal scene reduction The wind and light data with uncertainty generated by the improved generative adversarial network and the historical load data of the same time of the whole year are combined to form 8760 data pairs, which are sorted in time sequence from small to large, and then the Latin hypercube sampling method (Latin Hypercube Sampling, LHS) is used for sampling to construct an initial data set with time sequence characteristics, The specific steps of LHS are as follows: Assume that the LHS sampling object size is N, there are k random variables, and the probability density function of random variable x k is as follows: y k = f k (x k ) (9) 1) Divide the probability space of each random variable x k into N subintervals, That is, {[0, 1 / N], [1 / N, 2 / N], …, [(N-1) / N, 1]}, 2) Randomly select a sample point y within each sub-interval ki , i = 1, 2, …, N 3) Using the inverse transform of the function as shown in formula (2) to find the sampling points x of each interval ki And arrange them in time sequence from small to large, to get the initial sampling matrix S with the sampling value of k x N order, that is, the initial data set of wind and light load throughout the year, The obtained wind and light load annual initial data set is normalized, and the SOM clustering algorithm is used for scene reduction to obtain 4 clustering results, i.e. the wind and light load data set with seasonal characteristics, and the 4 data sets correspond to the four seasons of spring, summer, autumn and winter respectively. In each data set, the wind and light load data pair has corresponding seasonal characteristics, The SOM clustering algorithm automatically classifies data according to their characteristics. The basic principle is to classify two data with small Euclidean distance into one class, and use this way to group data. The specific steps of the SOM clustering algorithm are as follows: First, set a data set collection D, the sample in collection D is x1, x2, …, x n , that is: D = {x1, x2, …, x n}, each sample contains n elements, such as the elements in sample x i x i1 , x i2 , …, x in , that is: x i = (x i1 , x i2 , …, x in ), by analogy, the elements in sample x j x j1 , x j2 , …, x jn , that is x j = (x j1 , x j2 , …, x jn ), Compute the Euclidean distance between the sample x i and the sample x j . The Euclidean distance d(x i , x j ) is given by 1) initialization, initializing the SOM network connection weight as a random value, determining the learning initial value, input vector initial value and the number of output neurons, i.e. the number of clusters, 2) confirm the best matching unit (best matching unit, BMU), calculate and compare the distances between the field nodes by the Euclidean distance formula, find the minimum distance, and confirm the winning neuron, 3) weight update adjustment, update the node neuron weight within the BMU radius, use the Gaussian function to shrink the field range of each iteration, and the neighborhood radius is expressed as: r(n)=r0exp(-n / a) (12) In the formula, r(n) is the field radius of the n th update; r0 is the initial value; a is a constant, 4) repeat step 2) - step 3) until the maximum number of iterations, output the result, 5) finally, the number of clusters is obtained by calculating the davies-bouldin index (DBI), and the clustering result with the minimum DBI is selected as the final clustering result, The calculation method of DBI is as follows: In the step S1, the improved generative adversarial network includes a traditional generative adversarial network including a generator and a discriminator: the generator receives input noise data, and the discriminator judges whether the data generated by the generator is real according to historical data; after the discriminator judges that the synthetic data is correct, the generator and the discriminator will be updated by fixing each other until the discriminator cannot judge whether the data given by the generator is synthetic data, and then the generator and the discriminator stop updating; N i,j =||C i -C j || p (15) where N is the total number of clusters; L i,j The relationship between the i-th cluster and the j-th cluster is described, and the smaller the value is, the better the clustering is; N i,j is the distance between the i-th cluster and the j-th cluster; S i is the Euclidean distance of the i-th cluster, T i is the size of the i-th cluster matrix; Y k is the k-th feature vector in the cluster; C i is the centroid of the i-th cluster; C j is the centroid of the j-th cluster; p is the dimension of the cluster vector. 2.The method of claim 1, wherein, The loss function of the generator is: Let historical wind power data be x, its distribution be P data (x), random noise z adopts Gaussian distribution P z is described, assuming that the generator is G and the discriminator is D, in the training process, the noise z ~ P z is sent to the generator, and the generated wind power G(z) ~ P g , at the same time, the historical wind power is sent to the discriminator, and the discriminator judges the generated data, In the formula, E is the mathematical expectation, D(G(z)) is the judgment result of the discriminator on the generated data, L G = -E z～p(z) [logD(G(z))] (1) The loss function of the discriminator is: In the formula, D(x) is the judgment result of the discriminator on the real data, and 1-D(G(z)) is the judgment result of the discriminator on the synthetic data, The goal of the generator is to generate more synthetic data that looks like real data, and the goal of the discriminator is to identify as many synthetic data as possible, so the objective function of the generative adversarial network is: The objective function of the improved generative adversarial network is as follows: The objective function solving process is equivalent to solving P data (x) and P g JS divergence problem: But P data (x) and P g It is difficult to have an ideal overlap, so the JS divergence is easy to become a constant, which leads to the discriminator being unable to identify the distance between the two distributions, resulting in gradient disappearance, poor training effect. To solve this problem, the traditional generative adversarial network is improved: the wasserstein distance is introduced to optimize the adversarial network, and the wasserstein distance is: where P r , w is two different distributions, ∏(P r , w ) is the set of joint distributions (P r , w ) of the two distributions, γ is one possible distribution in the set, x is a real sample, y is a generated sample, (x, y) is generated from γ, ||x - y|| is the distance between the two samples, E (x,y)～γ [||x - y||] is the energy expenditure in which The loss function of the generator is: where D ∈ 1-Lipsschitz makes the discriminator conform to a 1-Lipsschitz function, limits the range of the discriminator, and makes P data (x) and P g not much different, The loss function of the discriminator is: The specific model establishment method of the multi-time scale hybrid energy storage model in step S2 is as follows: 3.The method of claim 1, wherein, The short-term energy storage charges and discharges in a short time scale to maintain power balance in a short time, and the time scale is hours, and the model is as follows: SOC constraint: SOC0=0.5c (16) Formula (16) indicates that the initial time of the short-term energy storage is half of the capacity, formula (17) indicates that the initial time and the final time of the short-term energy storage in a day are the same, and formula (18) indicates that the short-term energy storage operates within the allowed range of SOC, SOC1 = SOC 24 (17) SOC min ≤ SOC ≤ SOC max (18) In the formula, SOC1 is the SOC state of the short-term energy storage at the initial moment, c is the capacity of the short-term energy storage, SOC 24 represents the SOC state of the short-term energy storage at the final moment, SOC min , SOC max are the maximum and minimum values of the short-term energy storage SOC allowed, respectively, Power constraint: The long-term energy storage charges and discharges in a long time scale to maintain long-term power balance, and the time scale is seasons, and the model is as follows: P ESS,min ≤P ESS ≤P ESS,max (19) In the formula, P ESS,min , P ESS,max are minimum and maximum values of short-term energy storage charging and discharging, respectively, and P ESS is the charging and discharging power of the short-term energy storage. SOC constraint: Formula 20 indicates that the initial time of the long-term energy storage in the first typical day is half of the capacity, formula 21 indicates that the final time of the long-term energy storage in the kth typical day and the initial time of the k+1th typical day are the same, formula 22 indicates that the initial time of the long-term energy storage in the first typical day and the final time of the fourth typical day are the same, formula 23 is a calculation method of the SOC of the long-term energy storage at time t in the kth typical day, and formula 24 indicates that the long-term energy storage operates within the allowed range of SOC, wherein, is the SOC state of the long-term energy storage at the initial time of the first typical day, c ESS is the capacity of the long-term energy storage, is the SOC state of the long-term energy storage at the end time of the kth typical day, is the SOC state of the long-term energy storage at the initial time of the k+1th typical day, is the SOC state of the long-term energy storage at the end time of the fourth typical day, is the SOC state of the long-term energy storage at the t time of the kth typical day, is the SOC state of the long-term energy storage at the t-1 time of the kth typical day, is the self-discharge rate of the long-term energy storage, a ESS is the efficiency of the long-term energy storage, and are the charging and discharging amounts of the long-term energy storage at the t time of the kth typical day, respectively, and At is the scheduling duration, SOC ESS,min , SOC ESS,max are the maximum and minimum values allowed for the long-term energy storage SOC, respectively, Power constraint: A power grid planning method considering multi-time scale energy storage cooperation In the formula, is the charging and discharging state of long-term energy storage, taking 1 or 0, taking 1 indicates that the long-term energy storage is charging and discharging, and taking 0 indicates that the long-term energy storage is not charging and discharging; P ESS,max is the maximum value of the charging and discharging of the long-term energy storage, Based on the multi-time scale hybrid energy storage model obtained above, the problem B in the proposed power grid planning model is described as follows, Objective function The objective function of the optimization problem is to minimize the total cost C of the power grid: Constraint condition In the formula, C1 is the economic cost of the power grid, C2 is the green cost of the power grid; p In,RE , p In,Sto are respectively the unit capacity installation cost of wind and light; c RE , c Sto are respectively the installation capacity of wind and light; p OM,RE , p OM,Sto are respectively the unit capacity operation and maintenance cost of wind and light; P RE , P Sto are respectively the output of wind and light; p Buy,MG , p Sel,MG are respectively the purchase and sale electricity price of the power grid, P Buy,MG , P Sel,MG are respectively the purchase and sale electricity quantity of the power grid; μ is the penalty coefficient of abandoned wind and light; P Waste is the quantity of abandoned wind and light; the energy storage includes short-term energy storage and long-term energy storage P Waste = P RE - P Load - P Sto - P Sel,MG (29) In the formula, P Load is the load power, Capacity constraint: Power balance constraint: where: c RE,max is the wind and solar installation capacity upper limit, c Sto is the energy storage installation capacity upper limit, Wind and light output constraint: P RE +P Sto -P Sel,MG +P Buy,MG = P Load (31) Large power grid purchase and sale power constraint: P RE,min ≤P RE ≤P RE,max (32) wherein: P RE,min , P RE,max are the minimum and maximum values of the wind power output, respectively, In addition to the above constraints, the constraint conditions of the power grid planning model also include formulas (9)-(15) mentioned above. wherein: P Buy,MG,max , P Buy,MG,min are upper and lower limits of the amount of electricity purchased from the large power grid, respectively Sel,MG,max , P Sel,MG,min are upper and lower limits of the amount of electricity sold to the large power grid, respectively ​

Images