Hierarchical harmonic prediction method applied to photovoltaic power under probabilistic framework

Abstract

The invention discloses a hierarchical harmonic prediction method applied to photovoltaic power under a probabilistic framework, and belongs to the technical field of photovoltaic power generation. In order to solve the problem that the current probabilistic hierarchical forecasting method is difficult to ensure the consistency of probability aggregation and is slow in calculation, the method comprises the following steps: collecting photovoltaic power of each hierarchical node of a power system corresponding to the same time resolution and the same forecasting range, performing base forecasting, calculating a base forecasting error of each node, and constructing a base forecasting error matrix; by using a sample rearrangement bottom-to-top aggregation technology, layer-by-layer aggregation is carried out upwards from bottom layer nodes based on the base prediction error matrix, and initial aggregation probabilistic prediction of all the nodes is obtained; constructing a base forecast vector based on the initial aggregation probabilistic forecast and the probabilistic base forecast, and converting the base forecast vector into a probabilistic forecast corrected by a bottom node by using a harmonic matrix; and aggregating the probabilistic forecast corrected by the bottom-layer nodes from bottom to top layer by layer by using a sample rearrangement from bottom to top aggregation technology to obtain the probabilistic forecast of which all the nodes accord with aggregation consistency.

Description

Technical Field

[0001] The present invention belongs to the technical field of photovoltaic power generation, and in particular relates to a hierarchical coordination prediction method applied to photovoltaic power under a probabilistic framework. Background Art

[0002] Photovoltaic power has strong randomness and intermittency, and its grid connection poses a huge threat to the stable operation of the power system. To this end, it is necessary to accurately forecast the photovoltaic power at future moments in order to formulate a reliable power generation plan and maintain the load balance of the power grid. The power system can be regarded as a hierarchical structure, with each node representing the transmission network, distribution network, and photovoltaic station from top to bottom. In order to characterize the uncertainty of the forecast and support risk assessment, each node at each level often uses its own data and forecasting methods to generate a set of probabilistic power forecasts. Although the actual injected photovoltaic power of any parent node in the hierarchical structure is equal to the sum of the photovoltaic power of its subordinate child nodes (aggregate consistency), due to the uncertainty of the forecast itself and the inequality of available information of different participants in the power system, the probabilistic forecasts of nodes at each level usually do not have aggregate consistency, that is, the probability distribution of the parent node prediction variable is not equal to the probability distribution of the sum of the prediction variables of its subordinate child nodes, which seriously hinders the formulation of joint decision-making plans.

[0003] The probabilistic hierarchical harmonic forecasting method is an effective means to achieve probabilistic forecast aggregation consistency. The method can be divided into two steps: the first step is to use an applicable forecasting method to independently generate a probabilistic base forecast of photovoltaic power (abbreviated as base forecast) for each level node; the second step is to use the harmonic technique to correct the base forecast to achieve probabilistic aggregation consistency. The main difference between the proposed probabilistic hierarchical harmonic forecasting methods lies in the different harmonization techniques. Some methods use the harmonization technique to achieve the prediction distribution of each parent node equal to the convolution of the prediction distribution of the subordinate child nodes; some methods aim to correct the mean of the base forecast of all nodes so that it satisfies the deterministic aggregation consistency; and some methods believe that the probabilistic hierarchical forecast is composed of several deterministic hierarchical forecasts (ensemble forecast), and when the forecasts of each group member satisfy the deterministic aggregation consistency, the group forecast also satisfies the probabilistic aggregation consistency. By comparing the three methods, it is not difficult to see that the first method has the most stringent requirements on the consistency of probability aggregation, but the corresponding harmonization technique is also the most complex. It is necessary to obtain the distribution of the sum of random variables through the Copula function, which is slow to calculate for high-dimensional hierarchical structures.

[0004] Therefore, it is of great significance to propose a hierarchical harmonic prediction method that can achieve probabilistic forecasts of nodes at different levels with probability aggregation consistency and computational efficiency. Summary of the invention

[0005] The problem to be solved by the present invention is that the current probabilistic hierarchical forecasting method is difficult to ensure the consistency of probability aggregation and the calculation is slow. A hierarchical harmonic prediction method applied to photovoltaic power under a probabilistic framework is proposed.

[0006] To achieve the above object, the present invention is implemented through the following technical solutions:

[0007] A hierarchical harmonic prediction method for photovoltaic power in a probabilistic framework includes the following steps:

[0008] S1. Collect the photovoltaic power and base forecast of each level node of the power system corresponding to the same time resolution and prediction range, calculate the base forecast error of each node, and construct the base forecast error matrix;

[0009] S2. Using the sample rearrangement bottom-up aggregation technique, the base prediction error matrix obtained in step S1 is aggregated from the bottom nodes upward layer by layer, and finally the initial aggregated probabilistic forecasts of all nodes are obtained;

[0010] S3. constructing a base forecast vector based on the initial aggregated probabilistic forecast and the probabilistic base forecast obtained in step S2, and then converting the base forecast vector into a probabilistic forecast of the underlying node correction using a reconciliation matrix;

[0011] S4. The corrected probabilistic forecasts of the bottom nodes obtained in step S3 are aggregated layer by layer from bottom to top into probabilistic forecasts that meet the aggregation consistency of all nodes using the sample rearrangement and bottom-up aggregation technology.

[0012] Furthermore, the specific implementation method of step S1 includes the following steps:

[0013] S1.1. For each bottom-level node, the ensemble numerical weather forecast is used as the ensemble forecast of the photovoltaic station, and then the quantiles of different quantile levels are calculated to obtain the bottom-level node base forecast;

[0014] S1.2. For the upper nodes, the statistical extrapolation method is used to calculate the quantiles of different quantile levels to obtain the upper node base forecast;

[0015] S1.3. Based on the method of step S1.1 and step S1.2, collect the base forecast of the bottom node, the base forecast of the upper node and the corresponding measured photovoltaic power, and divide the total time into training set, verification set and test set according to 60:20:20. Each sample includes the base forecast and measured value of the nodes at each level at the same time;

[0016] S1.4. Assume that the number of bottom-level nodes is m and the total number of nodes is n, then the probabilistic base prediction mean of node i in the training set is recorded as The measured value of node i in the training set is recorded as y i =(y i,1+h ,yi,2+h ,...,y i,N+h ) T , where N is the total time of the training set, is the mean of the predicted distribution generated by node i at time t for time t+h, y i,t is the measured value of node i at time t, and h is the interval time;

[0017] Then calculate the base prediction error e of node i i (h), calculated as:

[0018]

[0019] Then the base prediction errors of all nodes are combined into the base prediction error matrix E(h), and the expression is:

[0020] E(h)=(e1(h),e2(h),...,e n (h))∈R N×n .

[0021] Furthermore, in step S1, the bottom-level node is a bottom-level photovoltaic station, and the upper-level node includes a middle-level distribution network and an upper-level transmission network.

[0022] Furthermore, the specific implementation method of step S2 includes the following steps:

[0023] S2.1. For node i, set the forecast to be generated at time t for time t+h, l is the length of the historical measured vector, then collect the historical measured vector x of node i i,t =(y i,t-l+1 ,...,y i,t ) T , the quantile level of node i is {τ1,τ2,...,τ Q}, where Q is the number of quantile levels, and the probabilistic base prediction b of node i i,t (h) is:

[0024]

[0025] Set the child nodes of node i to {j,j+1,...,j+n i -1}, where n i is the number of child nodes of node i, then the probabilistic basis predictions of the child nodes of node i are merged into the probabilistic basis prediction matrix B of the child nodes of node i i,t (h), we get:

[0026]

[0027] The initial aggregated probabilistic forecasts of the child nodes of node i are combined into the initial aggregated probabilistic forecast matrix C of the child nodes of node i i,t (h), we get:

[0028]

[0029] Among them, c j,t (h) is the initial aggregated probabilistic forecast of child node j of node i;

[0030] S2.2. Concatenate the historical measured vectors of all nodes in the same sample, and continue to match the concatenated vectors of each training moment in the training set using the nearest neighbor algorithm to obtain the Q closest training moments;

[0031] The measured vectors of each node at time t are concatenated to obtain the concatenated measured vector (x 1,t T ,x 2,t T ,...,x n,t T ) T , use the k-nearest neighbor algorithm to obtain the Q closest training moments closest to the spliced ​​measured vector from the training set, and merge the base prediction deviations of the subordinate nodes of node i at the Q closest training moments into the base prediction error matrix E at time t i,t (h), the expression is:

[0032]

[0033] Among them, e′ j (h) is the base prediction error of child node j at the Q closest training moments;

[0034] S2.3. E i,t Each column vector of (h) calculates the ranking of each element, and the calculation formula is:

[0035]

[0036] Among them, rank(·) is a sorting operation, which calculates the order of each element among all elements in the vector; is the ranked matrix, describing the multivariate dependency structure between child nodes;

[0037] S2.4. Basis The initial aggregated probabilistic forecast matrix is ​​rearranged and a multivariate dependency structure is introduced. The calculation formula is:

[0038]

[0039] in, is the rearranged matrix, perm(c,r) is the rearrangement operation, which rearranges the elements in vector c so that the order of the rearranged elements is the same as that of vector r;

[0040] Z i,t Each row of (h) is regarded as a sample of a multivariate variable consisting of the predictor variables of the child nodes under node i. By summing each row, we can get Q samples of the predictor variables of the parent node. Finally, we calculate the initial aggregate probability distribution of the parent node at the quantile level {τ1,τ2,...,τ Q} and used as the initial aggregated probabilistic forecast c for the node i,t (h);

[0041] S2.5. Based on steps S2.1 to S2.4, aggregate upward from the bottom nodes layer by layer to finally obtain the initial aggregated probabilistic forecasts of all nodes.

[0042] Furthermore, the specific implementation method of step S3 includes the following steps:

[0043] S3.1. Calculate the difference vector between the initial aggregated probabilistic forecast of node i and the probabilistic base forecast. The calculation formula is:

[0044] d i,t (h) = c i,t (h)-b i,t (h)

[0045] Among them, d i,t (h)∈R Q is the difference vector of node i;

[0046] S3.2. Combine the difference vector calculated in step S3.1 with the probabilistic basis prediction of the underlying node into the basis prediction vector D t (h), we get:

[0047] D t (h)=(d 1,t (h),...,d n-m,t (h),b n-m+1,t (h),...,b n,t (h))∈R Q×n ;

[0048] S3.3. Use the harmonic matrix to transform D t (h) is converted into the probabilistic forecast of the bottom node correction, and the expression is:

[0049] L t =PD t (h) T

[0050] Among them, L tT =(q n-m+1,t (h),...,q n,t (h))∈R Q×m is the modified probabilistic forecast of the bottom-level nodes, represents the bottom node k at the quantile level {τ1,τ2,…,τ Q}The corrected quantile forecast; P∈R m×n is the reconciliation matrix to be optimized, which is used to convert the aggregated inconsistency information of a certain quantile level into the modified quantile forecast of all underlying nodes at that quantile level.

[0051] Furthermore, the optimization method of the harmonic matrix in step S3.3 is to construct a small-scale matrix to approximate P by matrix decomposition, and the expression is:

[0052] P=[P1 P2]≈[WVW T P2]

[0053] Where P1∈R m×m It is the matrix composed of the elements on the left side of P, P2∈R m×(n-m) It is the matrix composed of the right part of elements of P;

[0054] Since the number of bottom-level nodes is much larger than the number of upper-level nodes, that is, m>>nm, the number of weight parameters of P1 is much larger than that of P2. m×v With V∈R v×v The product of WVW T Approximate P1, where v <<m, and optimize P by minimizing the quantile loss function, expressed as:

[0055]

[0056] Among them, Loss is the value of the quantile loss function, f(·) is the Huber quantile loss, is the level τ generated by node i at time s for time s+h k quantile of ;

[0057] The Huber quantile loss is expressed as:

[0058]

[0059]

[0060] in, is a function with u as the independent variable, and λ is a value of 2 -8 The constant, τ k is the percentile level.

[0061] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the steps of a hierarchical coordination prediction method applied to photovoltaic power under a probabilistic framework are implemented.

[0062] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements a hierarchical coordination prediction method for photovoltaic power under a probabilistic framework.

[0063] Beneficial effects of the present invention:

[0064] The present invention discloses a hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework, and the harmonic prediction method mainly includes a base prediction vector based on distribution difference, a harmonic matrix based on matrix decomposition technology, and a sample-reordering bottom-up aggregation (SRBU) technology. The base prediction vector based on distribution difference can fully contain the aggregated inconsistency information of the probabilistic base prediction to generate a higher quality corrected probabilistic forecast. The harmonic matrix based on matrix decomposition technology is used to map the base prediction vector into a corrected probabilistic forecast of the bottom node, which uses matrix decomposition technology to approximate a large-scale weight matrix, effectively reducing the number of weight parameters. The SRBU technology can utilize the dependencies between nodes at different levels to aggregate the probabilistic forecasts of the bottom nodes into the probabilistic forecasts of the upper nodes.

[0065] The hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework described in the present invention involves a base forecast vector based on distribution differences that can fully contain the aggregated inconsistency information of the probabilistic base forecast to generate a higher quality corrected probabilistic forecast; the harmonic moment based on matrix decomposition technology uses matrix decomposition technology to approximate a large-scale weight matrix, effectively reducing the number of weight parameters; SRBU combines the nearest neighbor algorithm to accurately characterize the dependency relationship between nodes at different levels, thereby generating a higher quality probabilistic forecast. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 The present invention is a flowchart of a hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework. DETAILED DESCRIPTION

[0067] In order to make the purpose, technical solution and advantages of the present invention more clear, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention, that is, the specific embodiments described are only part of the embodiments of the present invention, rather than all of the specific embodiments. The components of the specific embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations, and the present invention can also have other embodiments.

[0068] Therefore, the following detailed description of the specific embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention claimed for protection, but merely represents the selected specific embodiments of the present invention. Based on the specific embodiments of the present invention, all other specific embodiments obtained by those skilled in the art without making creative work are within the scope of protection of the present invention.

[0069] In order to further understand the content, features and effects of the present invention, the following specific implementation methods are given as examples, and the attached Figure 1 The detailed instructions are as follows:

[0070] Embodiment 1:

[0071] A hierarchical harmonic prediction method for photovoltaic power in a probabilistic framework includes the following steps:

[0072] S1. Collect the photovoltaic power and base forecast of each level node of the power system corresponding to the same time resolution and prediction range, calculate the base forecast error of each node, and construct the base forecast error matrix;

[0073] Furthermore, the specific implementation method of step S1 includes the following steps:

[0074] S1.1. For each bottom-level node, the ensemble numerical weather forecast is used as the ensemble forecast of the photovoltaic station, and then the quantiles of different quantile levels are calculated to obtain the bottom-level node base forecast;

[0075] S1.2. For the upper nodes, the statistical extrapolation method is used to calculate the quantiles of different quantile levels to obtain the upper node base forecast;

[0076] S1.3. Based on the method of step S1.1 and step S1.2, collect the base forecast of the bottom node, the base forecast of the upper node and the corresponding measured photovoltaic power, and divide the total time into training set, verification set and test set according to 60:20:20. Each sample includes the base forecast and measured value of the nodes at each level at the same time;

[0077] S1.4. Assume that the number of bottom-level nodes is m and the total number of nodes is n, then the probabilistic base prediction mean of node i in the training set is recorded as The measured value of node i in the training set is recorded as y i =(y i,1+h ,y i,2+h ,...,y i,N+h ) T , where N is the total time of the training set, is the mean of the predicted distribution generated by node i at time t for time t+h, y i,t is the measured value of node i at time t, and h is the interval time;

[0078] Then calculate the base prediction error e of node i i (h), calculated as:

[0079]

[0080] Then the base prediction errors of all nodes are combined into the base prediction error matrix E(h), and the expression is:

[0081] E(h)=(e1(h),e2(h),...,e n (h))∈R N×n .

[0082] Furthermore, the bottom layer node in step S1 is a bottom layer photovoltaic station, and the upper layer node includes a middle layer distribution network and an upper layer transmission network;

[0083] S2. Using the sample rearrangement bottom-up aggregation technique, the base prediction error matrix obtained in step S1 is aggregated from the bottom nodes upward layer by layer, and finally the initial aggregated probabilistic forecasts of all nodes are obtained;

[0084] Furthermore, the specific implementation method of step S2 includes the following steps:

[0085] S2.1. For node i, set the forecast to be generated at time t for time t+h, l is the length of the historical measured vector, then collect the historical measured vector x of node i i,t =(y i,t-l+1 ,...,y i,t ) T , the quantile level of node i is {τ1,τ2,...,τ Q}, where Q is the total number of quantile levels, and the probabilistic base prediction b of node i i,t (h) is:

[0086]

[0087] Set the child nodes of node i to {j,j+1,...,j+n i -1}, where n iis the number of child nodes of node i, then the probabilistic basis predictions of the child nodes of node i are merged into the probabilistic basis prediction matrix B of the child nodes of node i i,t (h), we get:

[0088]

[0089] The initial aggregated probabilistic forecasts of the child nodes of node i are combined into the initial aggregated probabilistic forecast matrix C of the child nodes of node i i，t (h), we get:

[0090]

[0091] Among them, c j,t (h) is the initial aggregated probabilistic forecast of child node j of node i;

[0092] Furthermore, is the initial aggregation probability distribution of node j at the quantile level {τ1,τ2,...,τ Q}, and the initial aggregated probabilistic forecast of the bottom-level node is equal to its probabilistic base forecast.

[0093] S2.2. Concatenate the historical measured vectors of all nodes in the same sample, and continue to match the concatenated vectors of each training moment in the training set using the nearest neighbor algorithm to obtain the Q closest training moments;

[0094] The measured vectors of each node at time t are concatenated to obtain the concatenated measured vector (x 1,t T ,x 2,t T ,...,x n,t T ) T , use the k-nearest neighbor algorithm to obtain the Q closest training moments closest to the spliced ​​measured vector from the training set, and merge the base prediction deviations of the subordinate nodes of node i at the Q closest training moments into the base prediction error matrix E at time t i,t (h), the expression is:

[0095]

[0096] Among them, e′ j (h) is the base prediction error of child node j at the Q closest training moments;

[0097] S2.3. E i,t Each column vector of (h) calculates the ranking of each element, and the calculation formula is:

[0098]

[0099] Among them, rank(·) is a sorting operation, which calculates the order of each element among all elements in the vector; is the ranked matrix, describing the multivariate dependency structure between child nodes;

[0100] Further, for example, rank((3.4,5.1,2.6,7.3) T )=(2,3,1,4) T , is the ranked matrix, describing the multivariate dependency structure between child nodes. i,t (h) only reflects the marginal distribution of the initial aggregate probability forecast of each child node, and does not include the forecast size dependence between nodes, so it is necessary to Rearrange it;

[0101] S2.4. Basis The initial aggregated probabilistic forecast matrix is ​​rearranged and a multivariate dependency structure is introduced. The calculation formula is:

[0102]

[0103] in, is the rearranged matrix, perm(c,r) is the rearrangement operation, which rearranges the elements in vector c so that the order of the rearranged elements is the same as that of vector r;

[0104] Z i,t Each row of (h) is regarded as a sample of a multivariate variable consisting of the predictor variables of the child nodes under node i. By summing each row, we can get Q samples of the predictor variables of the parent node. Finally, we calculate the initial aggregate probability distribution of the parent node at the quantile level {τ1,τ2,...,τ Q} and used as the initial aggregated probabilistic forecast c for the node i,t (h);

[0105] Further, for example, perm((2.6,3.4,5.1,7.3) T ,(2,3,1,4) T )=(3.4,5.1,2.6,7.3) T ;

[0106] S2.5. Based on steps S2.1 to S2.4, aggregate from the bottom nodes upward layer by layer, and finally obtain the initial aggregated probabilistic forecasts of all nodes;

[0107] S3. constructing a base forecast vector based on the initial aggregated probabilistic forecast and the probabilistic base forecast obtained in step S2, and then converting the base forecast vector into a probabilistic forecast of the underlying node correction using a reconciliation matrix;

[0108] Furthermore, the specific implementation method of step S3 includes the following steps:

[0109] S3.1. Calculate the difference vector between the initial aggregated probabilistic forecast of node i and the probabilistic base forecast. The calculation formula is:

[0110] d i,t (h) = c i,t (h)-b i,t (h)

[0111] Among them, d i,t (h)∈R Q is the difference vector of node i;

[0112] Furthermore, at this time, node i represents the upper node;

[0113] S3.2. Combine the difference vector calculated in step S3.1 with the probabilistic basis prediction of the underlying node into the basis prediction vector D t (h), we get:

[0114] D t (h)=(d 1,t (h),...,d n-m,t (h),b n-m+1,t (h),...,b n,t (h))∈R Q×n ;

[0115] Furthermore, D t Each row of (h) contains not only the information of the probabilistic basis forecast of all nodes at a certain quantile level, but also reflects the aggregation inconsistency of the upper-level nodes;

[0116] S3.3. Use the harmonic matrix to transform D t (h) is converted into the probabilistic forecast of the bottom node correction, and the expression is:

[0117] L t =PD t (h) T

[0118] Among them, L t T =(q n-m+1,t (h),...,q n,t (h))∈R Q×m is the modified probabilistic forecast of the bottom-level nodes, represents the bottom node k at the quantile level {τ1,τ2,…,τ Q}; P∈R m×nis the reconciliation matrix to be optimized, which is used to convert the aggregated inconsistency information of a certain quantile level into the modified quantile forecast of all underlying nodes at that quantile level.

[0119] Furthermore, the optimization method of the harmonic matrix in step S3.3 is to construct a small-scale matrix to approximate P by matrix decomposition, and the expression is:

[0120] P=[P1 P2]≈[WVW T P2]

[0121] Where P1∈R m×m It is the matrix composed of the elements on the left side of P, P2∈R m×(n-m) It is the matrix composed of the right part of elements of P;

[0122] Since the number of bottom-level nodes is much larger than the number of upper-level nodes, that is, m>>nm, the number of weight parameters of P1 is much larger than that of P2. m×v With V∈R v×v The product of WVW T Approximate P1, where v <<m, and optimize P by minimizing the quantile loss function, expressed as:

[0123]

[0124] Among them, Loss is the value of the quantile loss function, f(·) is the Huber quantile loss, is the level τ generated by node i at time s for time s+h k quantile of ;

[0125] The Huber quantile loss is expressed as:

[0126]

[0127]

[0128] in, is a function with u as the independent variable, and λ is a value of 2 -8 The constant, τ k is the percentile level.

[0129] S4. The corrected probabilistic forecasts of the bottom nodes obtained in step S3 are aggregated layer by layer from bottom to top into probabilistic forecasts that meet the aggregation consistency of all nodes using the sample rearrangement and bottom-up aggregation technology.

[0130] The data used in this embodiment comes from the Solar Power Data for Integration studies (SPDIS) developed by the National Renewable Energy Laboratory (NREL). The data set contains simulated solar power generation power for one year at thousands of different locations in the United States, with a time resolution of 5 minutes. According to previous studies, the 5-minute simulated photovoltaic power in 318 areas of California is first averaged as hourly data as the bottom layer (L2) sequence. Subsequently, the k-nearest neighbor algorithm is used to divide these locations into five regions, and the corresponding bottom layer sequences are aggregated to form the first layer (L1) sequence. Finally, the five sequences of the first layer are aggregated to obtain a top layer (L0) sequence representing the overall photovoltaic power generation in California, and then a hierarchical structure is obtained. Among them, the L1 layer sequence is connected to 27, 73, 101, 86 and 31 L2 sequences respectively. For each hierarchical node, the quantile random forest model is used, and the hourly photovoltaic power data of the past 7 days is used as the input of the model to generate a probabilistic base forecast of photovoltaic power every hour of the next day. The first 80% of the dataset in the first half of the year is used to optimize the weight parameters of the quantile random forest and the harmonic matrix, and the last 20% is used to determine the hyperparameters of the quantile random forest and the dimensions of the member matrix in the approximate harmonic matrix. The second half of the dataset is used as a test set to verify the effectiveness of the proposed forecasting method. The forecasting performance is evaluated using the widely used continuous ranked probability score skill score (CRPS skill score):

[0131]

[0132]

[0133] Among them, CRPS ref CRPS for the reference model complete-history persistence ensemble; is the predicted distribution of the i-th sample; z i is the corresponding observed value; 1(·) is 1 when the condition in the brackets is met, and 0 otherwise.

[0134] The proposed method is denoted as NMDHPF. In addition, a variety of comparative models are set to verify the effectiveness of NMDHPF. Among them, BF-BU, BF-OLS, BF-WLS, BF-HLS, and BF-MinT respectively use bottom-up, ordinary least squares, weighted least squares, hierarchical least squares, and minimum trace to perform deterministic reconciliation on the mean of the probabilistic base forecast, and then translate the probabilistic base forecast of each node so that its mean is equal to the result of deterministic reconciliation. SRBU-BU, SRBU-OLS, SRBU-WLS, SRBU-HLS, and SRBU-MinT respectively use bottom-up, ordinary least squares, weighted least squares, hierarchical least squares, and minimum trace to perform deterministic reconciliation on the mean of the initial aggregated probabilistic forecast, and then translate the initial aggregated probabilistic forecast of each node so that its mean is equal to the result of deterministic reconciliation. BF represents the probabilistic base forecast, and SRBU represents the initial aggregated probabilistic forecast. The comparison model NHPF is obtained without matrix decomposition approximation of the harmonic matrix of the proposed method. The effectiveness of the constructed harmonic matrix can be verified by comparing NMDHPF and NHPF. The basis prediction vector of the proposed prediction method is replaced with the basis prediction of all nodes at a certain quantile level, and the matrix decomposition approximation of the harmonic matrix is ​​not performed, and the comparison model BF-NHPF is obtained. The effectiveness of the constructed basis prediction vector can be verified by comparing NHPF and BF-NHPF. The prediction performance of each model at different levels is shown in Table 1:

[0135] Table 1 CRPS skillscore of the model proposed in this embodiment and various hierarchical coordination prediction techniques on the test set

[0136]

[0137]

[0138] It can be seen from Table 1 that, first, through analysis, it can be found that the predictability of the proposed method is better than the comparative model based on deterministic harmony. Second, compared with the probabilistic base forecast BF, the probability prediction performance of the proposed method is significantly improved, which proves the effectiveness of the harmony. Third, compared with the initial aggregated probabilistic forecast SRBU, the forecast quality of the proposed method is higher, which proves the rationality of correcting the base forecast using distribution difference information. Fourth, the CRPS skill score of NHPF is higher than that of BF-NHPF, indicating that the constructed base forecast vector helps to improve the performance of the forecast; finally, compared with NHPF, the performance of the proposed model is higher, proving that the approximate harmonic matrix based on matrix decomposition does not reduce the quality of the forecast while reducing the number of weight parameters. This fully demonstrates the beneficial effects of the invention of this application.

[0139] It should be noted that relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "comprise a ..." do not exclude the existence of other identical elements in the process, method, article or device including the elements.

[0140] Although the present application has been described above with reference to specific embodiments, various modifications may be made thereto and parts thereof may be replaced with equivalents without departing from the scope of the present application. In particular, as long as there is no structural conflict, the various features in the specific embodiments disclosed in the present application may be used in combination with each other in any manner, and the fact that these combinations are not exhaustively described in this specification is only for the sake of omitting space and saving resources. Therefore, the present application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A hierarchical harmonic prediction method for photovoltaic power in a probabilistic framework, characterized in that: The steps include: S1. Collect the photovoltaic power and base forecast of each level node of the power system corresponding to the same time resolution and prediction range, calculate the base forecast error of each node, and construct the base forecast error matrix; S2. Using the sample rearrangement bottom-up aggregation technique, the base prediction error matrix obtained in step S1 is aggregated from the bottom nodes upward layer by layer, and finally the initial aggregated probabilistic forecasts of all nodes are obtained; S3. constructing a base forecast vector based on the initial aggregated probabilistic forecast and the probabilistic base forecast obtained in step S2, and then converting the base forecast vector into a probabilistic forecast of the underlying node correction using a reconciliation matrix; S4. The corrected probabilistic forecasts of the bottom nodes obtained in step S3 are aggregated layer by layer from bottom to top into probabilistic forecasts that meet the aggregation consistency of all nodes using the sample rearrangement and bottom-up aggregation technology.

2. According to claim 1, a hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework is characterized in that: The specific implementation method of step S1 includes the following steps: S1.

1. For each bottom-level node, the ensemble numerical weather forecast is used as the ensemble forecast of the photovoltaic station, and then the quantiles of different quantile levels are calculated to obtain the bottom-level node base forecast; S1.

2. For the upper nodes, the statistical extrapolation method is used to calculate the quantiles of different quantile levels to obtain the upper node base forecast; S1.

3. Based on the method of step S1.1 and step S1.2, collect the base forecast of the bottom node, the base forecast of the upper node and the corresponding measured photovoltaic power, and divide the total time into training set, verification set and test set according to 60:20:

20. Each sample includes the base forecast and measured value of the nodes at each level at the same time; S1.

4. Assume that the number of bottom-level nodes is m and the total number of nodes is n, then the probabilistic base prediction mean of node i in the training set is recorded as The measured value of node i in the training set is recorded as y i =(y i,1+h ,y i,2+h ,...,y i,N+h ) T , where N is the total time of the training set, is the mean of the predicted distribution generated by node i at time t for time t+h, y i,t is the measured value of node i at time t, and h is the interval time; Then calculate the base prediction error e of node i i (h), calculated as: Then the base prediction errors of all nodes are combined into the base prediction error matrix E(h), and the expression is: E(h)=(e1(h),e2(h),...,e n (h))∈R N×n 。 3. The hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework according to claim 2, characterized in that: In step S1, the bottom-level nodes are bottom-level photovoltaic stations, and the upper-level nodes include middle-level distribution networks and upper-level transmission networks.

4. The hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework according to claim 3, characterized in that: The specific implementation method of step S2 includes the following steps: S2.

1. For node i, set the forecast to be generated at time t for time t+h, l is the length of the historical measured vector, then collect the historical measured vector x of node i i,t =(y i,t-l+1 ,...,y i,t ) T , the quantile level of node i is {τ1,τ2,...,τ Q }, where Q is the number of quantile levels, and the probabilistic base prediction b of node i i,t (h) is: Set the child nodes of node i to {j,j+1,...,j+n i -1}, where n i is the number of child nodes of node i, then the probabilistic basis predictions of the child nodes of node i are merged into the probabilistic basis prediction matrix B of the child nodes of node i i,t (h), we get: The initial aggregated probabilistic forecasts of the child nodes of node i are combined into the initial aggregated probabilistic forecast matrix C of the child nodes of node i i,t (h), we get: Among them, c j,t (h) is the initial aggregated probabilistic forecast of child node j of node i; S2.

2. Concatenate the historical measured vectors of all nodes in the same sample, and continue to match the concatenated vectors of each training moment in the training set using the nearest neighbor algorithm to obtain the Q closest training moments; The measured vectors of each node at time t are concatenated to obtain the concatenated measured vector (x 1,t T ,x 2,t T ,...,x n,t T ) T , use the k-nearest neighbor algorithm to obtain the Q closest training moments closest to the spliced ​​measured vector from the training set, and merge the base prediction deviations of the subordinate nodes of node i at the Q closest training moments into the base prediction error matrix E at time t i,t (h), the expression is: Among them, e′ j (h) is the base prediction error of child node j at the Q closest training moments; S2.

3. E i,t Each column vector of (h) calculates the ranking of each element, and the calculation formula is: Among them, rank(·) is a sorting operation, which calculates the order of each element among all elements in the vector; is the ranked matrix, describing the multivariate dependency structure between child nodes; S2.

4. Basis The initial aggregated probabilistic forecast matrix is ​​rearranged and a multivariate dependency structure is introduced. The calculation formula is: in, is the rearranged matrix, perm(c,r) is the rearrangement operation, which rearranges the elements in vector c so that the order of the rearranged elements is the same as that of vector r; Z i,t Each row of (h) is regarded as a sample of a multivariate variable consisting of the predictor variables of the child nodes under node i. By summing each row, we can get Q samples of the predictor variables of the parent node. Finally, we calculate the initial aggregate probability distribution of the parent node at the quantile level {τ1,τ2,...,τ Q } and used as the initial aggregated probabilistic forecast c for the node i,t (h); S2.

5. Based on steps S2.1 to S2.4, aggregate upward from the bottom nodes layer by layer to finally obtain the initial aggregated probabilistic forecasts of all nodes.

5. The hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework according to claim 4, characterized in that: The specific implementation method of step S3 includes the following steps: S3.

1. Calculate the difference vector between the initial aggregated probabilistic forecast of node i and the probabilistic base forecast. The calculation formula is: d i,t (h)=c i,t (h)-b i,t (h) Among them, d i,t (h)∈R Q is the difference vector of node i; S3.

2. Combine the difference vector calculated in step S3.1 with the probabilistic basis prediction of the underlying node into the basis prediction vector D t (h), we get: D t (h)=(d 1,t (h),...,d n-m,t (h),b n-m+1,t (h),...,b n,t (h))∈R Q×n ; S3.

3. Use the harmonic matrix to transform D t (h) is converted into the probabilistic forecast of the bottom node correction, and the expression is: L t =PD t (h) T Among them, L t T =(q n-m+1,t (h),...,q n,t (h))∈R Q×m is the modified probabilistic forecast of the bottom-level nodes, represents the bottom node k at the quantile level {τ1,τ2,…,τ Q }; P∈R m×n is the reconciliation matrix to be optimized, which is used to convert the aggregated inconsistency information of a certain quantile level into the modified quantile forecast of all underlying nodes at that quantile level.

6. The hierarchical harmonic prediction method for photovoltaic power under a probabilistic framework according to claim 5, characterized in that: The optimization method of the harmonic matrix in step S3.3 is to use matrix decomposition to construct a small-scale matrix to approximate P, and the expression is: P=[P1 P2]≈[WVW T P2] Where P1∈R m×m It is the matrix composed of the elements on the left side of P, P2∈R m×(n-m) It is the matrix composed of the right part of elements of P; Since the number of bottom-level nodes is much larger than the number of upper-level nodes, that is, m>>nm, the number of weight parameters of P1 is much larger than that of P2. m×v With V∈R v×v The product of WVW T Approximate P1, where v <<m, and optimize P by minimizing the quantile loss function, expressed as: Among them, Loss is the value of the quantile loss function, f(·) is the Huber quantile loss, is the level τ generated by node i at time s for time s+h k The quantile of y i,s+h is the true value of node i at time s+h; The Huber quantile loss is expressed as: in, is a function with u as the independent variable, and λ is a value of 2 -8 The constant, τ k is the percentile level.

Images