Low-carbon building integrated energy system random scene generation method considering time sequence correlation

Abstract

The invention discloses a low-carbon building comprehensive energy system random scene generation method considering time sequence correlation. The method comprises the following steps: collecting historical data of building renewable energy and load and carrying out combined processing; performing feature analysis and typical season division on historical day scenes; performing scene clustering based on k-means, and calculating a time sequence probability distribution parameter based on maximum likelihood estimation; latin hypercube sampling and Cholesky decomposition are combined, the edge distribution characteristics of all variables are guaranteed through LHS, correlation reconstruction is conducted on a sampling result through Cholesky decomposition, and the time sequence coupling relation between the variables is accurately reflected while the sampling efficiency advantage is kept; and scene set reduction is carried out based on a backward reduction method, and typical scenes are reserved. The method is suitable for processing multi-variable and highly-correlated scene generation requirements such as wind and light load in a building energy system, and a more reliable input basis is provided for subsequent energy system planning and optimal scheduling.

Description

Technical Field

[0001] The present invention relates to the technical field of integrated energy system optimization, and in particular to a random scenario generation method for a low-carbon building integrated energy system considering time series correlation. Background Art

[0002] In the optimization of integrated building energy systems, stochastic scenario generation is a key step in addressing the uncertainty of wind and solar loads. Building loads and renewable energy output exhibit significant seasonal dependencies (e.g., peak cooling loads in summer and dominant heating loads in winter). Existing scenario generation methods often employ a unified, year-round modeling approach, ignoring the inherent differences in energy use between seasons. Consequently, the generated scenarios fail to accurately reflect the energy supply and demand characteristics during extreme weather or transitional seasons.

[0003] While simple and easy to implement, traditional Monte Carlo sampling methods suffer from inherent drawbacks such as slow convergence and low computational efficiency. This is particularly true when generating high-dimensional scenarios, requiring a large number of samples to ensure accuracy. This severely restricts their application in engineering practice. In contrast, Latin Hypercube Sampling (LHS) significantly improves sampling efficiency through a stratified sampling strategy. This method evenly divides the probability space into non-overlapping subintervals and accurately samples within each interval, avoiding sample clustering. With the same sample size, the estimation error is approximately 60% lower than that of Monte Carlo. However, the traditional LHS method still has significant shortcomings when applied to integrated building energy systems: Firstly, it assumes the independence of each variable and fails to preserve the actual spatiotemporal correlations between variables such as photovoltaic output and load demand. Secondly, it lacks specific treatment for the seasonal and time-of-day characteristics unique to building energy.

[0004] These defects make it difficult for the scenario sets generated by existing scenario generation methods to accurately reflect the dynamic characteristics of the actual energy system, and it is difficult to meet the needs of generating multi-variable and strongly correlated scenarios such as wind, solar and load in building energy systems. Summary of the Invention

[0005] The problem to be solved by the present invention is to provide a random scenario generation method for a low-carbon building integrated energy system taking into account time series correlation. It innovatively combines Latin hypercube sampling with Cholesky decomposition, accurately reflects the time series coupling relationship between variables while maintaining the sampling efficiency advantage. It is suitable for processing multi-variable and strongly correlated scenario generation needs such as wind, solar and load in building energy systems, and provides a more reliable input basis for subsequent energy system planning and optimal scheduling.

[0006] The present invention adopts the following technical solution: a method for generating random scenarios of a low-carbon building integrated energy system considering temporal correlation, comprising the following steps:

[0007] Step S1: Collect historical data of renewable energy and load of a building, perform data preprocessing and normalization, and perform joint processing based on time series to obtain historical daily scenarios with source-load time series correlation;

[0008] Step S2: Perform feature analysis on historical daily scenes of the building's source-load time series correlation, divide the scene into typical seasons (summer, winter, and transitional season), and obtain the source-load time series joint data for each season;

[0009] Step S3: Perform k-means classification on the source-load time series joint data of each season to achieve fine classification of data within the same season based on similar characteristics, and select the number of typical scenes for each season based on the elbow method;

[0010] Step S4: Calculate the time series probability distribution parameters of each group in each season based on maximum likelihood estimation (MLE), and establish the source-load probability distribution model at different times in different seasons;

[0011] Step S5: performing scene correlation analysis based on the Spearman correlation coefficient method;

[0012] Step S6: Randomly sample each scene based on the source-load probability distribution, perform stratified sampling using the Latin hypercube sampling method, and adjust the correlation of the sampled data using Cholesky decomposition to generate a scene set with correlation;

[0013] Step S7: Use the backward reduction method to optimize and reduce the scene set generated in step S6, and retain typical scenes.

[0014] Preferably, in step S1, a plurality of source-load time series joint historical scenarios are obtained, and the specific steps include:

[0015] Step S11: Obtain historical data of light radiation and load of a certain building for a period of length m, perform data preprocessing, and fill in missing data based on a linear interpolation method.

[0016] Step S12: using the maximum and minimum value normalization method, normalize the data obtained in step S11 to the interval [0, 1] and perform per-unit value conversion to obtain a medium- and long-term source-load time series with a length of m;

[0017] Step S13: Segment the normalized medium- and long-term historical time series of the source and load by daily scenes, and obtain the daily scene matrix X = [x1, x2, ..., x 24 ] day×24; The hourly data of the source-load variables on the same day are taken as a vector and merged into a multidimensional vector (n*24 dimensions), where n is the number of source-load variables, thereby obtaining the historical daily scenario X=[x1,x2,…,x 24 ] day×24n .

[0018] Preferably, in step S3, the elbow method is used to determine the optimal number of clusters, and the sum of squared Euclidean distances is used to measure the clustering effect. As the number of clusters k increases, an inflection point appears in the change of the sum of squared distances, and the point where the rate of decline suddenly slows down is considered the optimal number of clusters.

[0019] The sum of squares of the Euclidean distance is expressed as SqE, and the calculation formula is as follows:

[0020]

[0021] In the formula, Ci represents the cluster, k represents the number of cluster centers, p represents the samples in a cluster, m i Represents the center of mass.

[0022] Preferably, in step S3, the specific steps of the K-means algorithm are as follows:

[0023] Step S31, iteratively calculate k from 1 to 10, and calculate SqE after each clustering is completed;

[0024] In step S32, SqE gradually decreases, and an inflection point appears during the change process. When the rate of decrease suddenly slows down, it is considered to be the optimal cluster number value.

[0025] Preferably, in step S4, the randomness of the illumination radiation satisfies the Beta distribution, and its probability density function at time t can be expressed as:

[0026]

[0027] Where, is the normalized value of light radiation; Γ() is the Gamma function; α t and β t are the two shape parameters that determine the Beta distribution.

[0028] Theoretically, when there is a large amount of historical load data, the volatility of multi-energy loads approximately follows the traditional normal distribution. However, in practice, due to the existence of maximum and minimum load values, the tail interval outside the upper and lower load boundaries of the traditional normal distribution is truncated to obtain a truncated normal distribution that describes the volatility of multi-energy loads, as shown in the following formula:

[0029]

[0030] Where, and are the load value of the klth energy source at time t and its maximum and minimum values ​​respectively; and are the mean and standard deviation of different loads of the kl-th energy source at time t.

[0031] Preferably, in step S5, the scene correlation analysis is performed based on the Spearman correlation coefficient method. For a certain group, the autocorrelation coefficient r of its internal scene is k,n The calculation method is as follows:

[0032]

[0033] Where a i,k and are the data at time i on day k and the mean of the data on day k; a j,n and are the data at time j on day n and the mean of the data on day n, respectively. For any scenario on day k and day n in this cluster, the closer the days are, the stronger the correlation is.

[0034] In particular, when k = n, the correlation coefficient of the two scenes is 1, so the diagonal elements of the correlation coefficient matrix ρ are 1, and the calculation formula is as follows:

[0035]

[0036] Preferably, in step S6, the Latin hypercube sampling method based on Cholesky decomposition mainly consists of two parts: sampling and sorting. The specific steps are as follows:

[0037] Step S61: stratified sampling of each variable using Latin hypercube sampling technology. The specific steps are as follows:

[0038] Step S611: Each random variable X=[X1, X2, ..., X m ,…,X M ] is used as the input of the probability density function PDF() of group u. Integrating PDF() yields its cumulative distribution function (CDF) F m (x), the calculation formula is as follows:

[0039] Y m =F m (X m ),m=1,2,...,M

[0040] Where M is the sample size.

[0041] Step S612: Divide the [0,1] probability space into N equal-width intervals (N is the number of required scenes), and randomly select a sample point u in each interval. n ~U[(n-1) / N,n / N];

[0042] Step S613: Inverse cumulative probability density function (CDF -1 ) The transformation obtains the sampling value, and the calculation formula is as follows:

[0043]

[0044] Where x m,n is the sample value in each interval.

[0045] Step S614: construct an initial sampling matrix S0, that is, the initial scene set of group u, as shown below:

[0046]

[0047] Each row corresponds to N sample values ​​of a variable.

[0048] Step S62: Implement correlation adjustment of the sampled data through Cholesky decomposition. The specific steps are as follows:

[0049] Step S621: Calculate the source-load time series related historical daily scenario X obtained in step S13 = [x1, x2, ..., x 24 ] day×24n The Spearman rank correlation coefficient matrix ρ ta rget (24n*24n dimension), which is used as the target correlation coefficient matrix;

[0050] Step S622: Generate a sorting matrix L of the same dimension as the initial sampling matrix S0, where each row is composed of random integers from 1 to N, and calculate the autocorrelation coefficient matrix ρ of L L ;

[0051] Step S623: L Perform Cholesky decomposition to eliminate the autocorrelation of elements in L due to random sorting, and then obtain the reconstructed sorting matrix G, as shown below:

[0052] ρ L =DD T

[0053] G=D- 1 L

[0054] Where D is a lower triangular matrix;

[0055] Step S624: Autocorrelation coefficient matrix ρ of historical data target Perform Cholesky decomposition as shown below:

[0056]

[0057] Step S625: D target Multiplying by G can obtain the updated order matrix U, as shown below:

[0058] U=D target G

[0059] Step S624: reorder based on the updated sequence matrix:

[0060] According to the numerical weight of each column in U, the data in each column of the initial sampling matrix S0 (i.e., the sampling values ​​of a certain group at a certain moment) are reordered, and finally a scene sample S with similar correlation to the historical data is obtained.

[0061] Preferably, in step S7, a backward reduction method is used to optimize and reduce the large-scale scene set generated in step S4, significantly reducing the complexity of subsequent optimization calculations while retaining key probability features. The specific steps are as follows:

[0062] Step S71: Use the probability distance (such as Wasserstein distance) to calculate the probability of each pair of scenes (x i , x j ) between the two locations;

[0063] Step S72: Iteratively reduce the scene distances. The specific steps are as follows:

[0064] Initialization: Set the deprecated set Retention set R = {all initial scenes};

[0065] Scenario selection: In the kth iteration, select the scenario l that minimizes the following equation: k Add to deprecation set:

[0066]

[0067] Where p i For scene x i Probability of occurrence;

[0068] Update set: Deprecated set formula: J k =J k-1 ∪{l k};

[0069] Holdout set formula: R k =R k-1 \{lk};

[0070] Termination condition: Stop when only the preset number of scenes are left in the reserved set.

[0071] Step S73, probability merging: for each scenario i in the discarded set J, its probability mass is transferred to the nearest scenario j in the retained set:

[0072]

[0073] Where J(j) is the set of all abandoned scenes with J as the nearest neighbor.

[0074] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0075] 1. The random scenario generation method for the low-carbon building integrated energy system of the present invention innovatively combines Latin hypercube sampling with Cholesky decomposition: LHS is used to ensure the marginal distribution characteristics of each variable, and Cholesky decomposition is used to reconstruct the correlation of the sampling results, thereby accurately reflecting the temporal coupling relationship between variables while maintaining the sampling efficiency advantage.

[0076] 2. The random scenario generation method of the low-carbon building integrated energy system of the present invention is particularly suitable for processing the multi-variable and strongly correlated scenario generation requirements such as wind, solar and load in the building energy system, providing a more reliable input basis for subsequent energy system planning and optimized scheduling. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flow chart of the random scenario generation method of the low-carbon building integrated energy system of the present invention;

[0078] Figure 2 This is a historical data diagram of the source and load of the low-carbon building integrated energy system according to an embodiment of the present invention;

[0079] Figure 3 This is a schematic diagram of determining the optimal number of clusters using the elbow method according to an embodiment of the present invention;

[0080] Figure 4 Generate a result graph for the summer scene of the embodiment of the present invention;

[0081] Figure 5 Generate a result graph for the transition season scenario of an embodiment of the present invention;

[0082] Figure 6 This is a result graph generated for a winter scene according to an embodiment of the present invention. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the application are further elaborated in detail below with reference to the accompanying drawings. The described embodiments are only a part of the embodiments involved in the present invention. All non-innovative embodiments of other researchers in this field on this embodiment fall within the scope of protection of the present invention. At the same time, the step numbers in the embodiments of the present invention are only set for the convenience of explanation and description, and the order between the steps is not limited in any way. The execution order of each step in the embodiment can be adaptively adjusted according to the understanding of those skilled in the art.

[0084] In one embodiment of the present invention, taking a low-carbon building integrated energy system as an example, the uncertain source load factors include photovoltaic, electric load, heating load and cooling load, a random scene generation method for a low-carbon building integrated energy system considering time series correlation is proposed. Figure 1 The specific steps are as follows:

[0085] Step S1: Collect historical data of renewable energy and load of a building, jointly process the photovoltaic and load data according to time series, and obtain several source-load time series joint historical scenarios.

[0086] In this embodiment, the time resolution of the collected historical data is not less than 1 hour, and the data time span includes at least a complete annual cycle; data preprocessing is performed, including missing value filling and outlier correction. Since the units of the historical daily and hourly data of the source load are not unified and the numerical differences are large, in order to eliminate the influence of the data dimension and numerical differences on the convergence of the algorithm, the data is further normalized.

[0087] The specific steps include:

[0088] Step S11: Obtain the historical data of light radiation, electric load, heating load and cooling load of a building every day and every hour in a certain year, such as Figure 2 As shown, the missing data are filled based on the linear interpolation method;

[0089] Step S12: Use the maximum and minimum value normalization method to normalize the data obtained in step S11 to the interval [0,1] for per-unit value conversion; the normalization calculation formula is as follows:

[0090]

[0091] Where S t is the historical data of the scene at time t (including light radiation, electrical load, heating load and cooling load) and its maximum and minimum values; S t * is the normalized scene value.

[0092] Step S13: Segment the normalized medium- and long-term historical time series of the source and load by daily scenes, and obtain the daily scene matrix X = [x1, x2, ..., x 24 ] day×24 ; The hourly data of the source-load variables on the same day are taken as a vector and merged into a multidimensional vector (n*24 dimensions), where n is the number of source-load variables. Thus, the historical daily scenario X=[x1,x2,…,x 24 ] day×24n ;

[0093] Step S2: Analyze the characteristics of the historical data of light radiation, electrical load, heating load, and cooling load of the building, and divide them into typical seasons (summer, winter, and transitional season);

[0094] Specifically, weather changes and building energy use have distinct seasonal characteristics, with cooling loads dominating in summer and heating loads in winter. Based on empirical experience and the building's location, annual data is categorized into transitional periods, for example, March to May and September to November, summer from June to August, and winter from December to February.

[0095] Step S3: Perform k-means classification on the seasonal source-load time series joint data obtained in step S2 above, so as to further refine the classification of data within the same season based on similar characteristics.

[0096] The K-means algorithm is an unsupervised learning algorithm primarily used for data clustering. Determining the K value, or the number of clusters to form, is a key issue. While similarity within a set of clusters increases with the number of clusters, too many clusters reduce the differences between clusters, making it difficult to distinguish between categories. Therefore, it's important to set an appropriate number of clusters.

[0097] Furthermore, the elbow method is used to determine the optimal number of clusters, and the sum of squares of Euclidean distance is used to measure the clustering effect.

[0098] The formula for calculating the sum of squares of Euclidean distance is as follows:

[0099]

[0100] In the formula, Ci represents the cluster, k represents the number of cluster centers, p represents the samples in a cluster, m i Represents the center of mass.

[0101] In this embodiment, the specific steps of the K-means algorithm are as follows:

[0102] Step S31, iteratively calculate k from 1 to 10, and calculate SqE after each clustering is completed;

[0103] In step S32, SqE gradually decreases, and an inflection point appears during the change process. When the rate of decrease suddenly slows down, it is considered to be the optimal cluster number value.

[0104] In this embodiment, Figure 3 As shown, the elbow method determines that the optimal number of clusters is two.

[0105] Step S4: Calculate the time series probability distribution parameters of each group in each season based on maximum likelihood estimation, and establish a source-load probability distribution model at different times in different seasons;

[0106] Photovoltaic output depends on meteorological conditions such as solar radiation, while load is determined by user energy consumption. Probabilistic distribution models fitted based on the statistical patterns of extensive historical data can characterize these uncertainties. Numerous studies have shown that short-term daily variations in solar intensity approximately follow a beta distribution, and variations in multi-energy load can be considered to approximately follow a normal distribution.

[0107] In this embodiment, the randomness of the light radiation satisfies the Beta distribution, α t and β t To determine the two shape parameters of the Beta distribution, the two calculation methods are as follows:

[0108]

[0109] Where μ G,t and δ G,t are the mean and standard deviation of different light radiation at time t respectively.

[0110] Theoretically, when there is a large amount of historical load data, the volatility of multi-energy loads approximately follows the traditional normal distribution. However, in practice, due to the existence of maximum and minimum load values, the tail interval outside the upper and lower load boundaries of the traditional normal distribution is truncated to obtain a truncated normal distribution that describes the volatility of multi-energy loads, as shown in the following formula:

[0111]

[0112] Where, and are the load value of the klth energy source at time t and its maximum and minimum values ​​respectively; and are the mean and standard deviation of different loads of the kl-th energy source at time t.

[0113] Step S5: Perform scene correlation analysis based on the Spearman correlation coefficient method.

[0114] Specifically, for a certain group, the autocorrelation coefficient r of its internal scene k,n The calculation method is as follows:

[0115]

[0116] Where a i,k and are the data at time i on day k and the mean of the data on day k; a j,n and are the data at time j on day n and the mean of the data on day n, respectively. For any scenario on day k and day n in this cluster, the closer the days are, the stronger the correlation is.

[0117] In particular, when k = n, the correlation coefficient of the two scenes is 1, so the diagonal elements of the correlation coefficient matrix ρ are 1, and the calculation formula is as follows:

[0118]

[0119] Step S6: random scene sampling is performed based on the source-load probability distribution;

[0120] Specifically, a set of correlated scenes is generated by combining Latin hypercube sampling with Cholesky decomposition.

[0121] In this embodiment, the Latin hypercube sampling method based on Cholesky decomposition mainly consists of two parts: sampling and sorting. The specific steps are as follows:

[0122] Step S61: stratified sampling of each variable using Latin hypercube sampling technology. The specific steps are as follows:

[0123] For each random variable X=[X1,X2,…,X m ,…,X M ] is used as the input of the probability density function PDF() of group u. Integrating PDF() yields its cumulative distribution function (CDF) F m (x), the calculation formula is as follows:

[0124] Y m =F m (X m ),m=1,2,...,M

[0125] Where M is the sample size.

[0126] Divide the [0,1] probability space into N equal-width intervals (N is the number of required scenes), and randomly select a sample point u in each interval n ~U[(n-1) / N,n / N];

[0127] By inverse cumulative probability density function (CDF -1 ) The transformation obtains the sampling value, and the calculation formula is as follows:

[0128]

[0129] Where x m,n is the sample value in each interval.

[0130] The initial sampling matrix S0, that is, the initial scene set of group u, is constructed as follows:

[0131]

[0132] Each row corresponds to N sample values ​​of a variable.

[0133] Step S62: Implement correlation adjustment of the sampled data through Cholesky decomposition. The specific steps are as follows:

[0134] Calculate the Spearman rank correlation coefficient matrix ρ of historical data target , the calculation formula is as shown in step S3;

[0135] Generate a sorting matrix L of the same dimension as the initial sampling matrix S0, each row of which is composed of random integers from 1 to N, and calculate the autocorrelation coefficient matrix ρ of L L ;

[0136] For the above obtained ρ L Cholesky decomposition is performed to eliminate the autocorrelation of elements in L due to random sorting, and then the reconstructed sorting matrix G is obtained. The formula is shown below.

[0137] ρ L =DD T

[0138] G=D- 1 L

[0139] Where D is a lower triangular matrix.

[0140] Similarly, the autocorrelation coefficient matrix ρ of historical data target Perform Cholesky decomposition as shown below:

[0141]

[0142] D target Multiplying by G can obtain the updated order matrix U, as shown below:

[0143] U=D target G

[0144] Reorder the updated order matrix obtained above. Reorder each column of the initial sampling matrix S0 (i.e., the sampling values ​​of a certain group at a certain moment) according to the numerical weight of each column in U, and finally obtain scene samples S with similar correlation to historical data.

[0145] Step S7: Use backward reduction to reduce the scene set and retain typical scenes.

[0146] Specifically, the backward reduction method is used to optimize and reduce the large-scale scene set generated in step S6, significantly reducing the complexity of subsequent optimization calculations while retaining key probability features. The specific steps are as follows:

[0147] Step S71, calculating the distance between scenes;

[0148] For any two scenes w i and w j , the formula for calculating the sum of squares of the Euclidean distance is:

[0149]

[0150] Where x i.t ,x j,t are any two scenes at a certain time t in the scene sample S obtained in step 4;

[0151] Step S72: Iteratively reduce the distance between scenes based on step S71. The specific steps are as follows:

[0152] Initialization: Set the deprecated set Retention set R = {all initial scenes}

[0153] Scenario selection: In the kth iteration, select the scenario l that minimizes the following equation: k Add to deprecation set:

[0154]

[0155] Where p i For scene x i Probability of occurrence;

[0156] Update set: Deprecated set formula: J k =J k-1 ∪{l k};

[0157] Holdout set formula: R k =R k-1 \{l k};

[0158] Termination condition: Stop when only the preset number of scenes are left in the reserved set.

[0159] Step S73, probability merging: for each scenario i in the discarded set J, its probability mass is transferred to the nearest scenario j in the retained set:

[0160]

[0161] Where J(j) is the set of all abandoned scenes with J as the nearest neighbor.

[0162] In summary, the summer scene generation result obtained by the method of this embodiment is as follows Figure 4 As shown, the transition season scene generation results are as follows Figure 5 As shown, the winter scene generation results are as follows Figure 6 As shown, it can provide a reliable input basis for subsequent energy system planning and optimal scheduling.

[0163] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A random scenario generation method for a low-carbon building integrated energy system considering temporal correlation, characterized in that: The steps include: Step S1: Collect historical data of renewable energy and load of a building, perform data preprocessing and normalization, and perform joint processing based on time series to obtain historical daily scenarios with source-load time series correlation; Step S2: Perform feature analysis on historical daily scenes of the source-load time series correlation of the building, divide the scene into typical seasons, and obtain the source-load time series joint data of each season; Step S3: Perform k-means classification on the source-load time series joint data of each season to achieve fine classification of data within the same season based on similar characteristics, and select the number of typical scenes for each season based on the elbow method; Step S4: Calculate the time series probability distribution parameters of each group in each season based on maximum likelihood estimation, and establish a source-load probability distribution model at different times in different seasons; Step S5: performing scene correlation analysis based on the Spearman correlation coefficient method; Step S6: Randomly sample each scene based on the source-load probability distribution, perform stratified sampling using the Latin hypercube sampling method, and adjust the correlation of the sampled data using Cholesky decomposition to generate a scene set with correlation; Step S7: Use the backward reduction method to optimize and reduce the scene set generated in step S6, and retain typical scenes.

2. The random scenario generation method for a low-carbon building integrated energy system considering time series correlation according to claim 1 is characterized in that: In step S1, a plurality of source-load time series joint historical scenarios are obtained, and the specific steps include: Step S11: Obtain historical data of light radiation and load of a building for a period of m, and perform data preprocessing, including missing value filling and outlier correction; Step S12: using the maximum and minimum value normalization method, normalize the data obtained in step S11 to the interval [0, 1] and perform per-unit value conversion to obtain a medium- and long-term source-load time series with a length of m; Step S13: Segment the normalized medium- and long-term historical time series of the source and load by daily scenes, and obtain the daily scene matrix X = [x1, x2, ..., x 24 ] day×24 ; The hourly data of the source-load variables on the same day are taken as a vector and merged into an n*24-dimensional vector, where n is the number of source-load variables, to obtain the historical daily scenario X=[x1,x2,…,x 24 ] day×24n .

3. The random scenario generation method for a low-carbon building integrated energy system considering time series correlation according to claim 2 is characterized in that: In step S3, the optimal number of clusters is determined according to the elbow method, and the clustering effect is measured using the sum of squares of the Euclidean distances; the sum of squares of the Euclidean distances is calculated using the following formula: Where C i represents the cluster, k represents the number of cluster centers, p represents the samples in a cluster, m i represents the centroid point, SqE represents the sum of squares of Euclidean distances; The clustering calculation is performed iteratively with k ranging from 1 to the preset number of times. SqE is calculated after each clustering is completed. SqE gradually decreases, and an inflection point where the decline rate suddenly slows down during the change process is taken as the optimal cluster number value.

4. The random scenario generation method for a low-carbon building integrated energy system considering time series correlation according to claim 1 is characterized in that: In step 4, the randomness of the illumination radiation satisfies the Beta distribution, and its probability density function at time t is expressed as: Where, is the normalized value of light radiation; Γ() is the Gamma function; α t and β t are the two shape parameters that determine the Beta distribution.

5. The random scenario generation method for a low-carbon building integrated energy system considering time series correlation according to claim 1 is characterized in that: In step 5, the tail interval outside the upper and lower load boundaries of the traditional normal distribution is truncated to obtain a truncated normal distribution that describes the volatility of multi-energy load, as shown in the following formula: Where, and are the load value of the klth energy source at time t and its maximum and minimum values ​​respectively; and are the mean and standard deviation of different loads of the kl-th energy source at time t.

6. The random scenario generation method for a low-carbon building integrated energy system considering temporal correlation according to claim 5 is characterized in that: In step S5, for a certain group, the autocorrelation coefficient r of the internal scene k,n The calculation method is as follows: Where a i,k and are the data at time i on day k and the mean of the data on day k; a j,n and are the data at time j on day n and the mean of the data on day n respectively; For any scenes on the kth day and the nth day in this group, the closer the days are, the stronger the correlation is. When k=n, the correlation coefficient of the two scenes is 1, and the diagonal elements of the correlation coefficient matrix ρ are 1.

7. The random scenario generation method for a low-carbon building integrated energy system considering time series correlation according to claim 3 is characterized in that: In step S6, stratified sampling is performed using the Latin hypercube sampling method, and the specific steps include: Step S611: Each random variable X=[X1, X2, ..., X m ,…,X M ] is used as the input of the probability density function PDF(·) of group u, and the cumulative distribution function F is obtained by integrating PDF(·). m (x), the calculation formula is as follows: Y m =F m (X m ),m=1,2,...,M Where M is the sample size; Step S612: Divide the probability space [0,1] into N equal-width intervals, where N is the number of required scenes; randomly select a sample point u in each interval. n ; Step S613: using the inverse cumulative probability density function F X -1 (x) is transformed to obtain the sampling value, and the calculation formula is as follows: Where x m,n is the sampling value in each interval; Step S614: construct an initial sampling matrix S0 as the initial scene set of group u, as shown below: Each row corresponds to N sample values ​​of a variable.

8. The random scenario generation method for a low-carbon building integrated energy system considering temporal correlation according to claim 7 is characterized in that: In step S6, the correlation adjustment of the sampled data is achieved through Cholesky decomposition, and the specific steps include: Step S621: Calculate the source-load time series related historical daily scenario X obtained in step S13 = [x1, x2, ..., x 24 ] day×24n The Spearman rank correlation coefficient matrix ρ target , with a dimension of 24n*24n, as the target correlation coefficient matrix; Step S622: Generate a sorting matrix L with the same dimensions as the initial sampling matrix S0, where each row is randomly composed of integers from 1 to N, and calculate the autocorrelation coefficient matrix ρ of L L ; Step S623: Get the obtained ρ L Perform Cholesky decomposition to eliminate the autocorrelation of elements in L due to random sorting, and obtain the reconstructed sorting matrix G, as shown below: r L =DD T G=D -1 L Where D is a lower triangular matrix; Step S624: Autocorrelation coefficient matrix ρ of historical data target Perform Cholesky decomposition to obtain matrix D target , as shown below: Step S625: D target Multiply it with G to obtain the updated order matrix U, as shown below: U=D target G Step S626: reorder the data in each column of the initial sampling matrix S0 based on the updated sequence matrix U, and reorder the data in each column according to the numerical weight of each column in U to obtain scene samples S with similar correlation to the historical data.

9. The random scenario generation method for a low-carbon building integrated energy system considering time series correlation according to claim 1 is characterized in that: The backward reduction method in step S7 specifically comprises the following steps: Step S71: Use the probability distance to calculate the probability of each pair of scenes (x i , x j ) between the two locations; Step S72: perform iterative reduction based on the distance between scenes to obtain a discarded set J and a retained set R; Step S73: Probability merging: for each scenario i in the discarded set J, its probability mass is transferred to the nearest scenario j in the retained set: Where J(j) is the set of all abandoned scenes with J as the nearest neighbor.

10. The random scenario generation method for a low-carbon building integrated energy system considering time series correlation according to claim 9 is characterized in that: In step S72, iterative reduction is performed based on the distance between scenes, and the specific steps include: Step S721, initialization: set the abandoned set Retention set R = {all initial scenes}; Step S722, scene selection: In the kth iteration, select the scene l that minimizes the following equation: k Add to deprecation set: Where p i For scene x i The probability of occurrence, d(x i ,x j ) is the scene (x i , x j ) between the two locations; Step S723: Update the set: Deprecation set formula: J k =J k-1 ∪{l k }; Holdout set formula: R k =R k-1 \{l k }; Step S724, termination condition: stop iteration when only a preset number of scenes remain in the retained set.

Images