New energy output medium and long term random scene generation method and device

Abstract

The invention discloses a method and a device for generating a long-term random scene in new energy output, and belongs to the field of power grid operation scheduling. The new energy output medium-and-long-term random scene generation method comprises the steps of performing dimension reduction on a historical day scene to extract significant features of the historical day scene; clustering the salient feature set of the historical day scene by adopting an affinity propagation clustering algorithm, and dividing the generated day scene into a corresponding cluster; then simulating evolution characteristics among different categories of day scenes based on a Markov chain; and generating a medium and long term random scene through Monte Carlo simulation. According to the method, probability distribution, time sequence characteristics and spatial correlation of medium and long-term scenes of new energy output can be accurately fitted, and effective data support can be provided for medium and long-term power transaction and operation mode formulation.

Description

Technical Field

[0001] The present invention belongs to the field of power grid operation and dispatching, and more specifically, relates to a method and device for generating medium- and long-term random scenarios of renewable energy output. Background Art

[0002] Vigorously developing new energy is a key measure for my country to optimize its energy structure and mitigate climate change. The output of distributed new energy stations shows strong uncertainty characteristics such as randomness, volatility, and time-space correlation. Accurately describing the uncertainty of its output is a prerequisite for large-scale integration of new energy power generation, which can effectively promote the efficient consumption of new energy and is also of great significance for maintaining the safe and stable operation of the power system. After the distributed new energy stations are connected to the grid, affected by their uncertainty, the grid operation and dispatch planning problem is transformed from a deterministic problem to an uncertain optimization problem.

[0003] Since the time scales involved in power system scheduling and planning are different, it is necessary to study the generation methods of renewable energy output scenarios at multiple time scales. The key lies in accurately describing the uncertainty of renewable energy output and generating specific scenarios in a controllable manner. For the problem of modeling the uncertainty of renewable energy output, randomly generating a large number of daily scenarios can fully explore whether it can accurately describe the characteristics of renewable energy output, which is of great significance to the study of issues such as power system economic scheduling and unit combination.

[0004] As for the problem of generating specific scenarios in a controllable manner, specifically, combining the day-ahead forecast information to generate the day-ahead renewable energy output scenario can provide data support for the day-ahead dispatch optimization of the power system, and the medium- and long-term scenario is an important basis for formulating power plans and the medium- and long-term operation mode of the power system. However, due to the high variable dimension of medium- and long-term scenario modeling, the variables of the currently commonly used hidden Markov model grow exponentially and have great limitations; the neural network-based method is difficult to train due to its large parameters and significantly fewer training samples than short-term scenarios, and it is easy to cause overfitting. Summary of the invention

[0005] In view of the defects of the related art, the purpose of the present invention is to provide a method and device for generating medium- and long-term random scenarios of new energy output, aiming to solve the problems in the prior art that the dimension of dependent variables is too high, resulting in difficulty in model training, and the medium- and long-term scenario generation problem is highly complex and has low accuracy.

[0006] To achieve the above objectives, in a first aspect, the present invention provides a method for generating medium- and long-term random scenarios of new energy output, comprising:

[0007] S1. Construct an original day scene matrix based on the original day scene samples, perform dimensionality reduction processing, extract its significant features, and obtain a significant feature set;

[0008] S2, clustering is performed based on the significant feature set to obtain a preset number of cluster centers, and the original daily scenes are divided into corresponding cluster centers to form clusters; wherein each cluster represents a state;

[0009] S3, simulating a state transition probability matrix between different states through a Markov chain, and obtaining a cumulative state transition probability matrix between different states corresponding to a preset time according to the state transition probability matrix;

[0010] S4. Based on the cumulative state transition probability matrix, a daily scene state sequence corresponding to a preset time is obtained by Monte Carlo simulation, and medium- and long-term random scenes are obtained by sequentially splicing.

[0011] Optionally, S1 specifically includes:

[0012] S11, constructing an original daily scene matrix according to the original daily scene samples;

[0013] S12, constructing a manifold structure of the original day scene sample in a high-dimensional space based on the graphical representation, and projecting the data points in the original day scene matrix into a low-dimensional space through nonlinear mapping to obtain a low-dimensional projection of the manifold structure;

[0014] S13. Optimize the low-dimensional projection based on stochastic gradient descent to obtain a set of significant features.

[0015] Optionally, S2 specifically includes:

[0016] S21, clustering is performed using a neighbor propagation clustering algorithm, a similarity matrix of a significant feature set corresponding to the original day scene sample is calculated, and attraction and membership are set according to the similarity matrix;

[0017] S22, setting the number of iterations T, initializing the attraction r and the membership a to 0, and iteratively updating the attraction r and the membership a, to determine whether the corresponding original daily scene sample is a cluster center; after T iterations, a preset number of cluster centers are calculated;

[0018] S23, dividing all original daily scene samples into clusters formed by corresponding cluster centers.

[0019] Optionally, S21 specifically includes:

[0020] S211, calculating the similarity matrix of the salient feature set corresponding to the original day scene sample, and normalizing it to the range of [0,1];

[0021] The similarity matrix S is expressed as follows:

[0022]

[0023] Among them, S(Ii , I k ) represents the normalized cosine similarity, which represents sample I k Suitable as sample I i The degree of cluster center, the larger the value, the more appropriate; I i and I k They represent the i-th day scene sample and the k-th day scene sample respectively;

[0024] S212, setting attraction r (I i ,I k ) and membership a(I i ,I k ); wherein the attraction r(I i ,I k ) means that after considering the remaining potential cluster centers, sample I k Suitable as sample I i The degree of cluster center; the membership degree a(I i ,I k ) indicates that when considering sample I k After being used as the cluster center, sample I i Select Sample I k The suitability of the cluster center.

[0025] Optionally, the iterative updating of the attraction r and the membership a to determine whether the corresponding original daily scene sample is a cluster center includes:

[0026] The attraction r(I i ,I k ) is:

[0027]

[0028] The membership degree a(I i ,I k ) is:

[0029]

[0030] By calculating sample I i The sum of attraction and belongingness r(I i ,I k )+a(I i ,I k ), find the sample I that maximizes the sum k ; If i = k, then sample I i is the cluster center, otherwise it means I k is the cluster center and sample I i Belong to I k , and continue to iterate and update.

[0031] Optionally, the calculating of a preset number of cluster centers further includes:

[0032] Construct the clustering effectiveness index DB, and the calculation formula is as follows:

[0033]

[0034] Where: W k , W j C k Cluster and C j The average Euclidean distance between the samples within the cluster and the central sample in the cluster is called the intra-cluster compactness; M Ck,Cj For C k The distance between the cluster and the central sample of the Cj-th cluster is called the inter-class separation;

[0035] The optimal number of clusters K of the clustering algorithm is determined according to the minimum value of the DB indicator.

[0036] Optionally, S3 specifically includes:

[0037] S31. Taking the daily scene state as the state of the Markov chain model, a probabilistic modeling is performed based on the Markov chain for the time series transition characteristics of the daily scene state evolution of new energy, and the daily scene state transition probability P is obtained. r (fz);

[0038] The P r The expression of (fz) is:

[0039] In the formula, x t With x t-1 Respectively represent the cluster number to which the scene belongs on day t and day t-1, n( f,z ) represents the number of times the scene belonging to cluster number f changes to the scene belonging to cluster number z in consecutive days;

[0040] S32, constructing a state transfer matrix for transition probability according to the daily scene state transition probability; wherein the transition probability between each state is a state transfer matrix P r express:

[0041]

[0042] Where P r (fz) represents the state transition probability;

[0043] S33, obtaining a cumulative state transition probability matrix corresponding to a preset time according to the state transition probability matrix;

[0044] The cumulative state transition probability matrix Prcum formula is as follows:

[0045]

[0046] in,

[0047] Optionally, S4 specifically includes:

[0048] S41. Generate a random number w from uniform distribution and compare w with the probability in the cumulative state transition probability matrix. If P rcuml,f-1 <w≤P rcuml,f , then the current state is determined to be the output state corresponding to the f-th scenario cluster;

[0049] S42, generate a random number u, if P rcum(f,z-1) <u≤P rcum(f,z) , then the next state is determined to be the output state corresponding to the z-th scene cluster;

[0050] S43, judging whether the generated output state sequence reaches the specified sequence length; if not, returning to S42; if so, extracting the daily scenes in the corresponding cluster one by one according to the generated output state sequence, and splicing them to form a new energy medium- and long-term scene.

[0051] In a second aspect, the present invention further provides a device for generating medium- and long-term random scenarios of new energy output, comprising:

[0052] The feature extraction module is used to construct the original day scene matrix according to the original day scene samples, perform dimensionality reduction processing, extract its significant features, and obtain a significant feature set;

[0053] A clustering module, used for clustering based on the significant feature set to obtain a preset number of cluster centers, and dividing the original daily scenes into corresponding cluster centers to form clusters; wherein each cluster represents a state;

[0054] A state transfer module is used to simulate the state transfer probability matrix between different daily scene states through a Markov chain, and obtain the cumulative state transfer probability matrix between different daily scene states corresponding to a preset time according to the state transfer probability matrix;

[0055] The prediction module is used to obtain the daily scene state sequence corresponding to the preset time through Monte Carlo simulation based on the cumulative state transition probability matrix, and to obtain the medium- and long-term random scenes in sequence.

[0056] Compared with the prior art, the above technical solution conceived by the present invention can achieve the following beneficial effects:

[0057] 1. The embodiment of the present invention provides a method for generating medium- and long-term random scenarios of renewable energy output, which reduces the dimension of historical daily scenarios to extract their significant features, clusters the significant feature sets of historical daily scenarios using the neighbor propagation clustering algorithm, and divides the generated daily scenarios into corresponding clustering clusters, and then simulates the evolution characteristics between different categories of daily scenarios based on the Markov chain, and finally generates medium- and long-term random scenarios through Monte Carlo simulation. By extracting features through dimensionality reduction, the accuracy and stability of clustering results are improved; the MCMC method can effectively simulate the state evolution law of random variables, wherein the Markov chain can perform probabilistic modeling for the time series transfer characteristics of random variables, which can be better applied to the state evolution law of renewable energy output daily scenarios; this method can accurately fit the probability distribution, time series characteristics and spatial correlation of renewable energy output medium- and long-term scenarios, reduce the complexity of scenario prediction, improve the accuracy of scenario prediction, and provide effective data support for medium- and long-term power trading and operation mode formulation.

[0058] 2. The embodiment of the present invention provides a method for generating medium- and long-term random scenarios of new energy output, which adopts nonlinear dimensionality reduction, retains more global structures, and makes feature extraction more accurate. It adopts the neighbor propagation clustering algorithm, which does not require the number of clusters to be specified in advance, is insensitive to the initial value of the model, and the clustering results are more stable and the square error of the clustering results is smaller. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 It is a flow chart of a method for generating medium- and long-term random scenarios of new energy output provided by an embodiment of the present invention;

[0060] Figure 2 The historical daily scene clustering cluster based on significant features provided by the embodiment of the present invention;

[0061] Figure 3 A probability distribution diagram of the monthly generated scenario and the historical monthly scenario provided by the embodiment of the present invention;

[0062] Figure 4 A time series curve of the monthly generated scenario and the historical monthly scenario provided by the embodiment of the present invention;

[0063] Figure 5 The annual scenario probability histogram provided by the embodiment of the present invention. DETAILED DESCRIPTION

[0064] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0065] The contents involved in the above embodiment are described below in conjunction with a preferred embodiment.

[0066] Embodiment 1

[0067] like Figure 1 As shown, the present invention provides a method for generating medium- and long-term random scenarios of new energy output, comprising:

[0068] S1. Construct an original day scene matrix based on the original day scene samples, perform dimensionality reduction processing, extract its significant features, and obtain a significant feature set;

[0069] S2, clustering is performed based on the significant feature set to obtain a preset number of cluster centers, and the original daily scenes are divided into corresponding cluster centers to form clusters; wherein each cluster represents a state;

[0070] S3, simulating a state transition probability matrix between different states through a Markov chain, and obtaining a cumulative state transition probability matrix between different states corresponding to a preset time according to the state transition probability matrix;

[0071] S4. Based on the cumulative state transition probability matrix, a daily scene state sequence corresponding to a preset time is obtained by Monte Carlo simulation, and medium- and long-term random scenes are obtained by sequentially splicing.

[0072] In order to study the evolution law of the daily scenario of renewable energy output, it is necessary to first study the daily scenario clustering method. The dimension of the daily scenario matrix of renewable energy output is N*T, where N is the number of stations and T is the number of power sampling points in a day. The dimension of the daily scenario matrix of renewable energy output increases significantly with the increase in the number of stations and the shortening of the power point sampling interval. The output sequences of wind farms and photovoltaic power stations that are geographically close have a strong correlation, so the daily scenario matrix shows multicollinearity, which will lead to unstable clustering results.

[0073] In order to achieve effective clustering of daily scenes, its redundant features can be ignored, and the focus is on the significant features of the high-dimensional daily scene matrix. The linear dimensionality reduction method is simple and efficient, and performs well when processing linearly related data, but it does not work well for complex data. Nonlinear dimensionality reduction obtains the low-dimensional manifold of the high-dimensional matrix through complex nonlinear transformations, which can not only maintain the global structure of the original scene matrix, but also characterize its local structure, so it can handle complex nonlinear relationships. Since the daily scene of new energy output has significant nonlinearity, a nonlinear dimensionality reduction method is selected to extract the significant features of the original matrix. In this application, UMAP is selected to reduce the dimensionality of the original daily scene matrix. UMAP is essentially a manifold based on a graph layout algorithm to represent high-dimensional data in a low-dimensional space. It constructs a graphical representation of the original data in a high-dimensional space and projects it to a low-dimensional space to obtain a significant feature set. Among them, the significant features usually include: clustering characteristics of daily scene data (such as high or low average output), changing trends (such as seasonal trends or periodic fluctuations), representative data points (such as peak and valley values), etc.

[0074] Markov chain can be used to perform probabilistic modeling for the time series transfer characteristics of random variables, so the state evolution law of the new energy output daily scenario can be approximately characterized based on the Markov chain. The Markov chain model assumes that the next state of the random variable depends only on the previous state, and the cluster category to which the new energy output daily scenario belongs is used as the state of the Markov chain model, and each cluster corresponds to a state.

[0075] The MCMC method is based on the cumulative state transition probability matrix and determines the state jump of random variables through Monte Carlo sampling. Specifically, based on the Markov chain simulation of the evolution characteristics of the daily scene state, the Monte Carlo method is used to simulate the daily scene state sequence of any length based on the cumulative state transition probability matrix between daily scene clusters.

[0076] Optionally, S1 specifically includes:

[0077] S11, constructing an original daily scene matrix according to the original daily scene samples;

[0078] S12, constructing a manifold structure of the original day scene sample in a high-dimensional space based on the graphical representation, and projecting the data points in the original day scene matrix into a low-dimensional space through nonlinear mapping to obtain a low-dimensional projection of the manifold structure;

[0079] S13. Optimize the low-dimensional projection based on stochastic gradient descent to obtain a set of significant features.

[0080] The optimized low-dimensional projection keeps its similarity with the global graph in high-dimensional space.

[0081] Among them, S11 specifically includes: obtaining the output data of each day from the historical data; taking the daily output data recorded by hour as an example, the output data can be represented as a time series vector, and the output data of each hour is an element of this vector; for each day's scene, the time series vector of the day is used as a row in the matrix, and each row represents a day's scene. Multiple daily scenes are combined to obtain the original daily scene matrix for subsequent mathematical processing.

[0082] Optionally, S12 specifically includes:

[0083] S121, find neighboring points to build a local graph, and obtain a global graph through joint analysis;

[0084] S122: searching for a corresponding set of original daily scene samples according to the global graph, so as to represent a low-dimensional projection of the global graph in a low-dimensional space.

[0085] The UMAP method is used to reduce the dimensionality of the original day scene, which includes two steps: first, the manifold structure of the original day scene in high-dimensional space is constructed based on the graphical representation, and then the low-dimensional projection of the manifold is obtained through nonlinear mapping. In order to use the graph to represent the manifold structure of the original data in high-dimensional space, the local KNN (knearest neighbors) graph is first constructed based on the Nearest-Neighbor-Descent algorithm to find neighboring points. In this process, the number of neighboring points needs to be determined. The selection of the number of neighboring points determines whether UMAP can effectively balance the global and local structures of the original data. When the number of neighboring points is large, UMAP has a more complete representation of the global structure, but lacks attention to the local structure of the original data. Smaller neighbor values ​​make UMAP pay more attention to the local structure of high-dimensional data. After the KNN graph is constructed, the global graph is obtained through joint analysis. The global graph can be used as an approximate representation of the manifold structure of the original data in high-dimensional space. Then, the global graph is approximated by a limited set of data samples in the low-dimensional space, the Euclidean distance is used to measure the correlation between local data, and the layout of the low-dimensional space is optimized based on stochastic gradient descent to maintain its similarity with the global graph in the high-dimensional space as much as possible. In the subsequent simulation analysis, UMAP is used to reduce the dimensionality of the daily scene.

[0086] Optionally, S2 specifically includes:

[0087] S21. Use the neighbor propagation clustering algorithm to perform clustering, calculate the similarity matrix of the significant feature set corresponding to the original day scene sample, and set the attraction and membership according to the similarity matrix.

[0088] The neighbor propagation clustering algorithm does not need to specify the number of clusters in advance, is insensitive to the initial value of the model, and has more stable clustering results. It selects cluster centers based on the information propagation of neighboring points. At the beginning of clustering, all daily scene samples are regarded as potential cluster centers, and samples iteratively compete for cluster centers.

[0089] Among them, S21 specifically includes:

[0090] S211, calculating the similarity matrix of the salient feature set corresponding to the original day scene sample, and normalizing it to the range of [0,1];

[0091] The similarity matrix S is expressed as follows:

[0092]

[0093] Among them, S(I i , I k ) represents the normalized cosine similarity, which represents sample I k Suitable as sample I i The degree of cluster center, the larger the value, the more appropriate; I i and I k They represent the i-th day scene sample and the k-th day scene sample respectively;

[0094] The cosine similarity is the off-diagonal element S(I i ,I k ), determine the reference degree as the diagonal element S(I i, I i ), the reference degree S(I i, I i ) represents sample I i Whether it can become a cluster center. The number of clusters is positively correlated with the initial reference value. Usually, the reference value is set to the median or mean of the similarity matrix at the beginning.

[0095] S212, setting attraction r (I i ,I k ) and membership a(I i ,I k ); wherein the attraction r(I i ,I k ) means that after considering the remaining potential cluster centers, sample I k Suitable as sample I i The degree of cluster center; the membership degree a(I i ,I k ) indicates that when considering sample I k After being used as the cluster center, sample I i Select Sample I kThe suitability of the cluster center.

[0096] S22, setting the number of iterations T, initializing the attraction r and the membership a to 0, and iteratively updating the attraction r and the membership a to determine whether the corresponding original daily scene sample is a cluster center; after T iterations, a preset number of cluster centers are calculated.

[0097] Optionally, the iterative updating of the attraction r and the membership a to determine whether the corresponding original daily scene sample is a cluster center includes:

[0098] The attraction r(I i ,I k ) is:

[0099]

[0100] The membership degree a(I i ,I k ) is:

[0101]

[0102] By calculating sample I i The sum of attraction and belongingness r(I i ,I k )+a(I i ,I k ), find the sample I that maximizes the sum k ; If i = k, then sample I i is the cluster center, otherwise it means I k is the cluster center and sample I i Belong to I k , and continue to iterate and update.

[0103] In order to avoid numerical oscillation, the damping coefficient λ is often introduced to attenuate the update of membership and attraction:

[0104] r t+1 (i,k)=λ·r t (i,k)+(1-λ)r t+1 (i,k)

[0105] a t+1 (i,k)=λ·a t (i,k)+(1-λ)a t+1 (i,k)

[0106] Each iteration will update the attraction and membership until the preset number of iterations is reached or the cluster center no longer changes, then the iteration terminates.

[0107] Optionally, the calculating of a preset number of cluster centers further includes:

[0108] Construct the clustering effectiveness index DB, and the calculation formula is as follows:

[0109]

[0110] Where: W k , W j C k Cluster and C j The average Euclidean distance between the samples within the cluster and the central sample in the cluster is called the intra-cluster compactness; M Ck,Cj For C k The distance between the cluster and the central sample of the Cj-th cluster is called the inter-class separation;

[0111] By comparing the DB index under different numbers of clusters, the optimal number of clusters K of the clustering algorithm is determined according to the minimum value of the DB index.

[0112] Among them, the lower the DB index is, the more accurately the redundant daily scenes with the same fluctuation characteristics, statistical characteristics and spatiotemporal correlation can be divided into the same cluster during clustering.

[0113] S23, dividing all original daily scene samples into clusters formed by corresponding cluster centers.

[0114] Optionally, S3 specifically includes:

[0115] S31. Taking the daily scene state as the state of the Markov chain model, a probabilistic modeling is performed based on the Markov chain for the time series transition characteristics of the daily scene state evolution of new energy, and the daily scene state transition probability P is obtained. r (fz);

[0116] The P r The expression of (fz) is:

[0117] In the formula, xt and xt-1 represent the cluster numbers to which the scene belongs on day t and day t-1, respectively, and n(f,z) represents the number of times the scene belonging to cluster number f changes to the scene belonging to cluster number z in consecutive days;

[0118] Among them, the Markov chain model assumes that the next state of the random variable depends only on the previous state. The cluster category to which the new energy output daily scenario belongs is used as the state of the Markov chain model, and each cluster corresponds to a state.

[0119] S32, constructing a state transfer matrix for transition probability according to the daily scene state transition probability; wherein the transition probability between each state is a state transfer matrix P rexpress:

[0120]

[0121] Where P r (fz) represents the state transition probability;

[0122] Among them, each daily scene cluster represents a state. Therefore, for K daily scene clusters, the transition probability between states is expressed by the state transition matrix P r express.

[0123] P r (fz) represents the state transition probability, and its formula is as follows:

[0124]

[0125] In the formula, x t With x t-1 represent the cluster numbers to which the scenes belong on days t and t-1 respectively, and n(f,z) represents the number of times the scene belonging to cluster number f changes to the scene belonging to cluster number z in consecutive days.

[0126] S33, obtaining a cumulative state transition probability matrix corresponding to a preset time according to the state transition probability matrix;

[0127] The cumulative state transition probability matrix Prcum formula is as follows:

[0128]

[0129] in,

[0130] Optionally, S4 specifically includes:

[0131] S41. Generate a random number w from uniform distribution and compare w with the probability in the cumulative state transition probability matrix. If P rcuml,f-1 <w≤P rcuml,f , then the current state is determined to be the output state corresponding to the f-th scenario cluster;

[0132] S42, generate a random number u, if P rcum(f,z-1) <u≤P rcum(f,z) , then the next state is determined to be the output state corresponding to the z-th scene cluster;

[0133] S43, judging whether the generated output state sequence reaches the specified sequence length; if not, returning to S42; if so, extracting the daily scenes in the corresponding cluster one by one according to the generated output state sequence, and splicing them to form a new energy medium- and long-term scene.

[0134] In this embodiment, the preset time is one month. In order to avoid the phenomenon of ramping, that is, the output at 0 o'clock next day is too different from the output at 23 o'clock today, it is required that the daily scene extracted from the z-th generation scene cluster meets the requirement that the difference between its output at 0 o'clock and the output at 23 o'clock of the previous day scene does not exceed the extreme value of the first-order difference of the historical sequence. If no daily scene meets the requirement in the traversed z scene clusters, a daily scene with a smaller output difference is selected, and the output at 23 o'clock of the previous day scene and the output at 0 o'clock of the current day scene are median filtered.

[0135] In a specific embodiment, the 2020 historical annual measured new energy output scenarios of three wind farms and four photovoltaic power stations geographically adjacent to each other in a province of my country are used as a data set to verify the effectiveness of the proposed method. The output data sampling interval is 1 hour, and the data is normalized to the [-1,1] interval.

[0136] Step 1: Daily scene clustering

[0137] First, a large number of new energy random daily scenes are generated based on the improved diffusion model, and then the significant features of the daily scenes are extracted based on UMAP. Then, the significant feature set of historical daily scenes is clustered using the nearest neighbor propagation clustering algorithm, and the generated daily scenes are divided into corresponding clusters. According to the above ideas, the nearest neighbor propagation algorithm is used for clustering based on the significant features of the daily scenes, and the UMAP method is used to reduce the original daily scenes to 3 dimensions. The corresponding clustering results are shown in Figure 2 The 24-hour power curves of the three wind farms on the same day are connected into 72 power sampling points as the wind power joint day scenario, and the 24-hour power curves of the four photovoltaic power stations on the same day are connected into 96 power sampling points as the photovoltaic joint day scenario.

[0138] Depend on Figure 2 It can be seen that the optimal clustering cluster based on significant feature clustering is 8 categories, and there is no imbalance in the category samples in its clustering results, and the number of samples in each category is relatively balanced. The DB index in the two cases was further calculated. The DB index based on significant feature clustering is 1.65, which effectively avoids the problem of imbalance in category samples, and is therefore of great significance for subsequent analysis and research. Compared with the optimal clustering K-means, the DB index of the optimal clustering K-means is 1.902, which is higher than the nearest neighbor propagation clustering algorithm. Therefore, clustering based on the nearest neighbor propagation clustering algorithm is more effective, and the same type of redundant daily scenes can be accurately covered in the same clustering cluster.

[0139] Step 2: Monthly scenario generation

[0140] Taking the monthly scenario as an example, the above method is used to generate the scenario. The generation effect is evaluated by comparing the fitting of the probability distribution, time series correlation and spatial correlation of the generated monthly output scenario. The probability histograms of the monthly generated scenario and the historical monthly scenario are shown respectively ( Figure 3) and timing curves ( Figure 4 ).

[0141] It can be seen from the result graph that when there are obvious differences in the time series curves of the generated monthly scenarios and the historical monthly scenarios, for three wind farms and four photovoltaic power stations, the probability distribution of the generated monthly scenarios is relatively consistent with the historical monthly scenarios, which shows that the proposed method can effectively learn the probability distribution statistical characteristics and spatiotemporal correlation characteristics of the historical monthly output scenarios. Figure 5 The probability histograms of annual scenarios for multiple new energy stations are displayed, which intuitively shows that the probability distribution of the annual scenarios based on nearest neighbor propagation clustering-MCMC (daily scenario generation is based on the improved diffusion model) is closest to the historical annual scenarios.

[0142] In this scheme, the neighbor propagation clustering algorithm is used to cluster the significant features of the historical output daily scenes, and the cluster number is used to represent the daily scene state. The state evolution characteristics of the daily scene are described based on the Markov chain, and finally the medium- and long-term random scenes are generated through Monte Carlo simulation. The problem that the model is difficult to train due to the high dimension of the dependent variable, the medium- and long-term scene generation problem is complex and the accuracy is low is solved. The probability distribution, time series characteristics and spatial correlation of the medium- and long-term scenarios of new energy output are accurately fitted, which reduces the complexity of scenario prediction and improves the accuracy of scenario prediction, which can provide effective data support for the formulation of medium- and long-term power trading and operation mode.

[0143] Embodiment 2

[0144] The present invention also provides a device for generating medium- and long-term random scenarios of new energy output, comprising:

[0145] The feature extraction module is used to construct the original day scene matrix according to the original day scene samples, perform dimensionality reduction processing, extract its significant features, and obtain a significant feature set;

[0146] A clustering module, used for clustering based on the significant feature set to obtain a preset number of cluster centers, and dividing the original daily scenes into corresponding cluster centers to form clusters; wherein each cluster represents a state;

[0147] A state transfer module is used to simulate the state transfer probability matrix between different daily scene states through a Markov chain, and obtain the cumulative state transfer probability matrix between different daily scene states corresponding to a preset time according to the state transfer probability matrix;

[0148] The prediction module is used to obtain the daily scene state sequence corresponding to the preset time through Monte Carlo simulation based on the cumulative state transition probability matrix, and to obtain the medium- and long-term random scenes in sequence.

[0149] A medium- and long-term random scenario generation device for new energy output provided by an embodiment of the present invention can execute a medium- and long-term random scenario generation method for new energy output provided by any embodiment of the present invention, and has functional modules and beneficial effects corresponding to the execution method.

[0150] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for generating medium- and long-term random scenarios of renewable energy output, characterized in that: include: S1. Construct an original day scene matrix based on the original day scene samples, perform dimensionality reduction processing, extract its significant features, and obtain a significant feature set; S2, clustering is performed based on the significant feature set to obtain a preset number of cluster centers, and the original daily scenes are divided into corresponding cluster centers to form clusters; wherein each cluster represents a state; S3, simulating a state transition probability matrix between different states through a Markov chain, and obtaining a cumulative state transition probability matrix between different states corresponding to a preset time according to the state transition probability matrix; S4. Based on the cumulative state transition probability matrix, a daily scene state sequence corresponding to a preset time is obtained by Monte Carlo simulation, and medium- and long-term random scenes are obtained by sequentially splicing.

2. The method according to claim 1, characterized in that S1 specifically includes: S11, constructing an original daily scene matrix according to the original daily scene samples; S12, constructing a manifold structure of the original day scene sample in a high-dimensional space based on the graphical representation, and projecting the data points in the original day scene matrix into a low-dimensional space through nonlinear mapping to obtain a low-dimensional projection of the manifold structure; S13. Optimize the low-dimensional projection based on stochastic gradient descent to obtain a set of significant features.

3. The method according to claim 1, characterized in that S2 specifically includes: S21, clustering is performed using a neighbor propagation clustering algorithm, a similarity matrix of a significant feature set corresponding to the original day scene sample is calculated, and attraction and membership are set according to the similarity matrix; S22, setting the number of iterations T, initializing the attraction r and the membership a to 0, and iteratively updating the attraction r and the membership a, to determine whether the corresponding original daily scene sample is a cluster center; after T iterations, a preset number of cluster centers are calculated; S23, dividing all original daily scene samples into clusters formed by corresponding cluster centers.

4. The method according to claim 3, characterized in that S21 specifically includes: S211, calculating the similarity matrix of the salient feature set corresponding to the original day scene sample, and normalizing it to the range of [0,1]; The similarity matrix S is expressed as follows: Among them, S(I i , I k ) represents the normalized cosine similarity, which represents sample I k Suitable as sample I i The degree of cluster center, the larger the value, the more appropriate; I i and I k They represent the i-th day scene sample and the k-th day scene sample respectively; S212, setting attraction r (I i ,I k ) and membership a(I i ,I k ); wherein the attraction r(I i ,I k ) means that after considering the remaining potential cluster centers, sample I k Suitable as sample I i The degree of cluster center; the membership degree a(I i ,I k ) indicates that when considering sample I k After being used as the cluster center, sample I i Select Sample I k The suitability of the cluster center.

5. The method according to claim 4, characterized in that The iterative updating of the attraction r and the membership a to determine whether the corresponding original daily scene sample is a cluster center includes: The attraction r(I i ,I k ) is: The membership degree a(I i ,I k ) is: By calculating sample I i The sum of attraction and belongingness r(I i ,I k )+a(I i ,I k ), find the sample I that maximizes the sum k ; If i = k, then sample I i is the cluster center, otherwise it means I k is the cluster center and sample I i Belong to I k , and continue to iterate and update.

6. The method according to claim 3, characterized in that The calculation to obtain a preset number of cluster centers also includes: Construct the clustering effectiveness index DB, and the calculation formula is as follows: Where: W k , W j C k Cluster and C j The average Euclidean distance between the samples within the cluster and the central sample in the cluster is called the intra-cluster compactness; M Ck,Cj For C k The distance between the cluster and the central sample of the Cj-th cluster is called the inter-class separation; The optimal number of clusters K of the clustering algorithm is determined according to the minimum value of the DB indicator.

7. The method according to claim 1, characterized in that S3 specifically includes: S31. Taking the daily scene state as the state of the Markov chain model, a probabilistic modeling is performed based on the Markov chain for the time series transition characteristics of the daily scene state evolution of new energy, and the daily scene state transition probability P is obtained. r (fz); The P r The expression of (fz) is: In the formula, x t With x t-1 Respectively represent the cluster number to which the scene belongs on day t and day t-1, n( f,z ) represents the number of times the scene belonging to cluster number f changes to the scene belonging to cluster number z in consecutive days; S32, constructing a state transfer matrix for transition probability according to the daily scene state transition probability; wherein the transition probability between each state is a state transfer matrix P r express: Where P r (fz) represents the state transition probability; S33, obtaining a cumulative state transition probability matrix corresponding to a preset time according to the state transition probability matrix; The cumulative state transition probability matrix Prcum formula is as follows: in, 8. The method according to claim 1, characterized in that S4 specifically includes: S41. Generate a random number w from uniform distribution and compare w with the probability in the cumulative state transition probability matrix. If P rcuml,f-1 <w≤P rcuml,f , then the current state is determined to be the output state corresponding to the f-th scenario cluster; S42, generate a random number u, if P rcum(f,z-1) <u≤P rcum(f,z) , then the next state is determined to be the output state corresponding to the z-th scene cluster; S43, judging whether the generated output state sequence reaches the specified sequence length; if not, returning to S42; if so, extracting the daily scenes in the corresponding cluster one by one according to the generated output state sequence, and splicing them to form a medium- and long-term scenario of new energy.

9. A device for generating medium- and long-term random scenarios of new energy output, characterized in that: include: The feature extraction module is used to construct the original day scene matrix according to the original day scene samples, perform dimensionality reduction processing, extract its significant features, and obtain a significant feature set; A clustering module, used for clustering based on the significant feature set to obtain a preset number of cluster centers, and dividing the original daily scenes into corresponding cluster centers to form clusters; wherein each cluster represents a state; A state transfer module is used to simulate the state transfer probability matrix between different daily scene states through a Markov chain, and obtain the cumulative state transfer probability matrix between different daily scene states corresponding to a preset time according to the state transfer probability matrix; The prediction module is used to obtain the daily scene state sequence corresponding to the preset time through Monte Carlo simulation based on the cumulative state transition probability matrix, and to obtain the medium- and long-term random scenes in sequence.

Images