Construction method and application method of hybrid prediction model of photovoltaic power station cluster power

Abstract

The invention discloses a construction method and an application method of a hybrid prediction model of photovoltaic power station cluster power. The construction method comprises the following steps: acquiring historical time data corresponding to a photovoltaic power station cluster in # imgabs0 # preset time periods; inputting the # imgabs 1 # historical time data into a hard clustering model and a soft clustering model respectively to obtain # imgabs 2 # hard clustering clusters and # imgabs 3 # soft clustering clusters, and inputting the # imgabs 1 # historical time data into the hard clustering model and the soft clustering model respectively to obtain # imgabs 2 # hard clustering clusters and # imgabs 3 # soft clustering clusters; performing boundary correction on the hard clustering cluster through the soft clustering cluster to obtain a # imgabs4 # weather clustering cluster; establishing an adjacent matrix and node features of an undirected graph corresponding to the cluster; based on the adjacent matrix and the node features in each weather cluster, training the original prediction model of each weather cluster to obtain a corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a time sequence feature extraction sub-model. According to the method, the accuracy of photovoltaic power station cluster power prediction can be effectively improved by establishing the power prediction models corresponding to different weather clusters.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power prediction, and in particular to a hybrid prediction model construction method, application method and system for photovoltaic power station cluster power. Background Art

[0002] With the continuous development of society, human demand for electricity is growing, leading to the intensification of fossil energy consumption. Under the background of the "dual carbon" goal, my country is committed to the development and utilization of renewable energy; among them, photovoltaic power generation technology continues to advance, power generation efficiency is gradually improved, and its share in the power system increases year by year. However, the large-scale access of photovoltaic power stations to the distribution network also brings some challenges; for example, photovoltaic power generation is significantly affected by meteorological factors, the power generation fluctuates greatly, and the power generation of a single photovoltaic power station is small, which puts pressure on the balanced dispatching of the power grid; at the same time, it may also cause reverse flow in the distribution network, thereby affecting the stable operation of the power grid. Therefore, how to accurately predict the short-term power of a photovoltaic power station cluster has become an urgent problem to ensure the stability and security of the power grid. Summary of the invention

[0003] The purpose of the embodiments of the present invention is to provide a hybrid prediction model construction method, application method and system for photovoltaic power station cluster power, which can effectively improve the accuracy of photovoltaic power station cluster power prediction by establishing power prediction models corresponding to different weather clusters.

[0004] The first embodiment of the present invention provides a method for constructing a hybrid prediction model for photovoltaic power station cluster power, comprising: Get the photovoltaic power station cluster in The historical time data corresponding to each preset time period respectively; wherein the historical time data includes: the historical power time series and meteorological data matrix of all photovoltaic power stations in the corresponding preset time period; ; Will The historical time data are input into the hard clustering model and the soft clustering model respectively, and the hard clusters and soft clusters; and modifying the boundaries of the hard clusters by using the soft clusters to obtain weather clusters; among them, ; According to the physical connection relationship between all the photovoltaic power stations, an adjacency matrix of an undirected graph is obtained; Divide each historical time data according to the photovoltaic power station to obtain the node features of the undirected graph; Based on the adjacency matrix and the node features in each weather cluster, the original prediction model of each weather cluster is trained to obtain a corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a temporal feature extraction sub-model.

[0005] Optionally, the training of the original prediction model of each weather cluster based on the adjacency matrix and the node features in each weather cluster to obtain a corresponding power prediction model includes: Inputting the adjacency matrix and the node features in each of the weather clusters into the corresponding graph convolutional neural network for training, respectively, to obtain a spatial feature extraction sub-model for each of the weather clusters, and corresponding spatial feature sample data; The node feature and spatial feature sample data in each of the weather clusters are respectively input into the corresponding gated recurrent unit for training, so as to obtain the temporal feature extraction sub-model of each of the weather clusters; wherein the hyperparameters in the gated recurrent unit are obtained by optimizing the improved artificial bee colony optimization algorithm; The combination of the spatial feature extraction sub-model and the temporal feature extraction sub-model of each weather cluster is used as the corresponding power prediction model.

[0006] Optionally, the hard clustering clusters are obtained by the following steps: Using each of the historical time data as a sample point of the hard clustering model; In the clustering iteration process, cluster assignment is performed based on the Euclidean distance from each sample point to the cluster center, and the clusters, and in each of the clusters, update the corresponding cluster center by using the density peak clustering algorithm; When the iteration stop condition is met, the output The hard clustering clusters.

[0007] Optionally, the soft clustering is used to modify the boundaries of the hard clustering to obtain Weather clusters, including: Assigning a weather label to each of the hard clusters and each of the soft clusters; Obtaining the intersection of the hard clustering cluster and the soft clustering cluster of the same weather label; Determine whether the degree of membership of each first sample point in the intersection is greater than or equal to a preset threshold; If yes, classifying the first sample point into the weather cluster corresponding to the same weather label; If not, reclassifying the first sample point into the weather cluster corresponding to the maximum membership in its own membership vector except for the same weather label; The second sample point in the complement of the intersection with respect to the hard cluster is reclassified as the weather cluster corresponding to the maximum membership in its own membership vector.

[0008] Optionally, obtaining an adjacency matrix of an undirected graph according to the physical connection relationship between all the photovoltaic power stations includes: The elements in the adjacency matrix are determined by the following formula: ; in, is the adjacency matrix Elements in For edge The source node of side The target node.

[0009] Optionally, the node features are specifically: ; in, For the The historical power time series of each PV power station within a preset period; For the Meteorological data matrix of each photovoltaic power station in a preset period of time; For the PV power stations at the predicted time point The previous The power value at each time point; For the PV power stations at the predicted time point The previous Meteorological factors at each time point The value of .

[0010] Optionally, the improved artificial bee colony optimization algorithm guides bees to search for new nectar sources through the following formula: ; in, For the newly generated The location of the honey source of the hyperparameter; For the The current number of bees Hyperparameter values; For the The step factor between the leading bee and the best nectar source corresponding to the hyperparameter; The currently found The global optimal solution corresponding to the hyperparameters; For the The current number of bees Hyperparameter values; For the The hyperparameters correspond to The step size factor of the bees; .

[0011] A second aspect of the present invention provides a hybrid prediction model application method for photovoltaic power station cluster power, including: Acquire the collected data of the photovoltaic power station cluster in the current period; wherein the collected data includes: the power time series and meteorological data matrix of all photovoltaic power stations in the current period; According to the collected data The Euclidean distance between the centers of the weather clusters is used to determine the power prediction model; Dividing the collected data according to photovoltaic power stations to obtain current node features of an undirected graph; The adjacency matrix of the undirected graph and the current node features are input into the power prediction model to obtain the predicted power of the photovoltaic power station cluster; wherein the weather clustering cluster, the adjacency matrix and the power prediction model are obtained by adopting the hybrid prediction model construction method of the photovoltaic power station cluster power described in any one of the first aspects above.

[0012] A third aspect of the present invention provides a system for constructing a hybrid prediction model of photovoltaic power station cluster power, including: The data acquisition module is used to obtain the data of the photovoltaic power station cluster. The historical time data corresponding to each preset time period respectively; wherein the historical time data includes: the historical power time series and meteorological data matrix of all photovoltaic power stations in the corresponding preset time period; ; Weather cluster acquisition module, used to The historical time data are input into the hard clustering model and the soft clustering model respectively, and the hard clusters and soft clusters; and modifying the boundaries of the hard clusters by using the soft clusters to obtain weather clusters; among them, ; An adjacency matrix building module, used to obtain an adjacency matrix of an undirected graph according to the physical connection relationship between all the photovoltaic power stations; A node feature acquisition module, used to divide each historical time data according to photovoltaic power stations, and obtain node features of the undirected graph; A prediction model training module is used to train the original prediction model of each weather cluster based on the adjacency matrix and the node features in each weather cluster to obtain a corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a temporal feature extraction sub-model.

[0013] A fourth aspect of the present invention provides an application system of a hybrid prediction model for photovoltaic power station cluster power, including: The data acquisition module is used to obtain the collected data of the photovoltaic power station cluster in the current period; wherein the collected data includes: the power time series and meteorological data matrix of all photovoltaic power stations in the current period; The prediction model determination module is used to determine the The Euclidean distance between the centers of the weather clusters is used to determine the power prediction model; A node feature determination module, used to divide the collected data according to the photovoltaic power station to obtain the current node features of the undirected graph; A predicted power acquisition module is used to input the adjacency matrix of the undirected graph and the current node features into the power prediction model to obtain the predicted power of the photovoltaic power station cluster; wherein the weather cluster cluster, the adjacency matrix and the power prediction model are obtained by adopting the hybrid prediction model construction method of the photovoltaic power station cluster power described in any one of the first aspects above.

[0014] Compared with the prior art, the embodiment of the present invention provides a hybrid prediction model construction method, application method and system for photovoltaic power station cluster power, the construction method comprises: obtaining the photovoltaic power station cluster in The historical time data corresponding to each preset period; The historical time data are input into the hard clustering model and the soft clustering model respectively, and the hard clusters and soft clusters; and the boundaries of the hard clusters are modified by the soft clusters to obtain weather clustering clusters; according to the physical connection relationship between all photovoltaic power stations, the adjacency matrix of the undirected graph is obtained; each historical time data is divided according to the photovoltaic power station to obtain the node features of the undirected graph; based on the adjacency matrix and the node features in each weather clustering cluster, the original prediction model of each weather clustering cluster is trained to obtain the corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a temporal feature extraction sub-model. The present invention can effectively improve the accuracy of power prediction of photovoltaic power station clusters by establishing power prediction models corresponding to different weather clustering clusters. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of an embodiment of a hybrid prediction model construction method for photovoltaic power station cluster power provided by the present invention; Figure 2 It is a flow chart of another embodiment of the hybrid prediction model construction method of photovoltaic power station cluster power provided by the present invention; Figure 3 It is a flow chart of an embodiment of the improved artificial bee colony algorithm for optimizing hyperparameters in a gated recurrent unit provided by the present invention; Figure 4 It is a flow chart of an embodiment of a hybrid prediction model application method for photovoltaic power station cluster power provided by the present invention; Figure 5 A comparative example diagram of the prediction results under a clear sky provided by the present invention; Figure 6 This is a comparative example diagram of the prediction results under cloudy conditions provided by the present invention; Figure 7 A comparative example diagram of the prediction results of a rainy day provided by the present invention; Figure 8 It is a structural schematic diagram of an embodiment of a hybrid prediction model building system for photovoltaic power station cluster power provided by the present invention; Fig. 9 It is a structural schematic diagram of an embodiment of a hybrid prediction model application system for photovoltaic power station cluster power provided by the present invention. DETAILED DESCRIPTION

[0016] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this technical field without creative work are within the scope of protection of the present invention.

[0017] See also Figure 1 , is a flow chart of an embodiment of a method for constructing a hybrid prediction model for photovoltaic power station cluster power provided by the present invention.

[0018] The first aspect of the present invention provides a method for constructing a hybrid prediction model of photovoltaic power station cluster power, including steps S11 to S15, which are specifically as follows: Step S11: Obtain the photovoltaic power station cluster The historical time data corresponding to each preset time period respectively; wherein the historical time data includes: the historical power time series and meteorological data matrix of all photovoltaic power stations in the corresponding preset time period; ; Step S12: The historical time data are input into the hard clustering model and the soft clustering model respectively, and the hard clusters and soft clusters; and modifying the boundaries of the hard clusters by using the soft clusters to obtain weather clusters; among them, ; Step S13: obtaining an adjacency matrix of an undirected graph according to the physical connection relationship between all the photovoltaic power stations; Step S14: Divide each historical time data according to the photovoltaic power station to obtain the node features of the undirected graph; Step S15: Based on the adjacency matrix and the node features in each weather cluster, the original prediction model of each weather cluster is trained to obtain a corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a temporal feature extraction sub-model.

[0019] It should be noted that the key to establishing a suitable power prediction model to achieve accurate prediction of photovoltaic power station power lies in the in-depth analysis of the prediction object and data. First, the meteorological data of the photovoltaic power station is obtained through the local official meteorological agency. However, since meteorological data is usually released over a wide area, when the cloud changes, it may affect the solar irradiance in the area where the photovoltaic power station is located. Therefore, cluster analysis of weather types can better capture the impact of weather changes, thereby establishing a more accurate prediction model under different weather conditions and improving the prediction accuracy. In addition, considering that there is a significant spatial correlation between the sum of the actual output power of each power station in the photovoltaic power station cluster and the power of each power station, when constructing the prediction model, it is necessary to combine the spatial characteristics and temporal characteristics of each photovoltaic power station in the cluster to comprehensively improve the accuracy and reliability of power prediction.

[0020] In step S11, all historical time data obtained can be obtained through The method identifies and processes abnormal data (missing data or data with large fluctuations). The following is a specific implementation example provided by the present invention: (1) Collect clusters in The historical time data within a preset time period, each historical time data includes the historical power time series and meteorological data matrix of all photovoltaic power stations in the photovoltaic power station cluster within the corresponding preset time period; wherein the meteorological factors in the meteorological data matrix are selected from weather temperature, power, wind speed, total irradiance, direct irradiance, diffuse irradiance, air pressure, humidity and wind direction, and the factors with the greatest influence on the output power of the power station are selected through the Person coefficient.

[0021] (2) Utilization The method is used to detect outliers in historical time data, and the detected outliers are corrected by using the mean value of the previous and next moments.

[0022] In step S12, The optimal number of clusters is the historical power time series of the photovoltaic power station as the clustering input feature, and the evaluation index Gap value is calculated. The larger the Gap value, the better the clustering effect. The clustering effect is better when the Gap value is the largest. The value is used as the number of clusters (usually between 3 and 5).

[0023] In step S14, the sample points of the hard clustering model or the soft clustering model correspond to a historical time data (i.e., the historical power time series and meteorological data matrix of all photovoltaic power stations in the corresponding preset time period). The undirected graph reflects the spatial layout of each photovoltaic power station in the power grid corresponding to the cluster. Therefore, each node in the undirected graph corresponds to a photovoltaic power station, and the node feature is the historical power time series and meteorological data matrix of a photovoltaic power station in a preset time period. In other words, the historical time data in a preset time period is divided according to the photovoltaic power station to obtain the input feature of each node in the undirected graph.

[0024] In step S15, the sample point data in each weather cluster is used to train the corresponding power prediction model. The structure of the power prediction model consists of two main parts. The spatial feature extraction submodel focuses on the spatial correlation between photovoltaic power stations, which can extract the mutual influence of power output between power stations and help capture the spatiotemporal dependency between different power stations in the power grid; the temporal feature extraction submodel focuses on the temporal characteristics of photovoltaic power station power and analyzes its trend and other temporal characteristics.

[0025] See also Figure 2 , is a flow chart of another embodiment of the method for constructing a hybrid prediction model for photovoltaic power station cluster power provided by the present invention.

[0026] The embodiment of the present invention can more accurately predict the short-term power changes of a photovoltaic power station cluster under different weather conditions by establishing power prediction models corresponding to different weather clusters.

[0027] In an optional embodiment, the hard clustering clusters are obtained by the following steps: Using each of the historical time data as a sample point of the hard clustering model; In the clustering iteration process, cluster assignment is performed based on the Euclidean distance from each sample point to the cluster center, and the clusters, and in each of the clusters, update the corresponding cluster center by using the density peak clustering algorithm; When the iteration stop condition is met, the output The hard clustering clusters.

[0028] It should be noted that the hard clustering cluster is to classify the weather types of historical time data by K-Means clustering algorithm (i.e. K-means clustering algorithm); at the same time, the density peaks clustering (DPC) algorithm is used to optimize the clustering center of the K-Means clustering algorithm, corresponding to Figure 2 In other words, when each cluster center is updated, the DPC algorithm is used to update the cluster center.

[0029] Specifically, in the cluster area The historical power time series and meteorological data matrix of a photovoltaic power station in a preset period of time is recorded as As the input feature of clustering; in other words, a historical time data is used as a sample point.

[0030] ;in, Indicates Input characteristics of a power station.

[0031] ;in, Indicates Power value, Indicates Among the meteorological factors data values, Indicates the total number of input features within the preset time period.

[0032] In the first clustering iteration, according to the number of clusters determined in step S12 above, , randomly selected The sample points are taken as cluster centers. Then, the Euclidean distances between the remaining sample points and the cluster centers are calculated and assigned to the cluster with the closest cluster center.

[0033] Next, the DPC algorithm is used to optimize the K-Means cluster center point (that is, when each cluster center is updated, the DPC algorithm is used to update the cluster center) to improve the robustness of clustering for non-uniformly distributed data and reduce the impact of noise data, which can improve the power prediction accuracy to a certain extent. It is worth noting that the embodiment of the present invention clusters the weather types of the sample points, reducing the impact of meteorological factor fluctuations on the power prediction results.

[0034] Specifically, the first step is to calculate the local density value of each point in the cluster , the calculation formula is: ; In the formula, , that is, if Less than ,but , otherwise ; Indicates the interval value between two data; is the cutoff distance.

[0035] Calculate the relative distance from the sample point to the local higher density point , the calculation formula is as follows: ; Select the local density value in the cluster Or relative distance The largest sample point is used as the new cluster center.

[0036] Finally, when the iteration stop condition is met (such as the cluster center no longer changes), the output The hard clustering clusters.

[0037] In an optional embodiment, the soft clustering is used to modify the boundaries of the hard clustering to obtain Weather clusters, including: Assigning a weather label to each of the hard clusters and each of the soft clusters; Obtaining the intersection of the hard clustering cluster and the soft clustering cluster of the same weather label; Determine whether the degree of membership of each first sample point in the intersection is greater than or equal to a preset threshold; If yes, classifying the first sample point into the weather cluster corresponding to the same weather label; If not, reclassifying the first sample point into the weather cluster corresponding to the maximum membership in its own membership vector except for the same weather label; The second sample point in the complement of the intersection with respect to the hard cluster is reclassified as the weather cluster corresponding to the maximum membership in its own membership vector.

[0038] It is worth noting that the soft clustering in the embodiment of the present invention is obtained by the Fuzzy C-Means (FCM) algorithm. The FCM algorithm is used to re-divide the "both this and that" sample points into clusters, that is, to correct the boundaries of the hard clustering clusters, corresponding to Figure 2Optimization II in ; and define each cluster (assign weather labels) according to the official meteorological information, such as sunny, rainy, and cloudy.

[0039] Similarly use As the input feature of clustering division, soft clustering is performed through the FCM algorithm, and it is possible to obtain the sample points contained in each soft cluster (such as cluster 1 (sunny day): [a(0.9), b(0.8), e(0.6)]), as well as the membership vector of each sample point to each soft cluster (that is, the sample point has different membership values ​​to multiple clusters).

[0040] In specific implementation, it is necessary to determine the intersection of hard clusters and soft clusters with the same weather label. For each first sample point in the intersection, calculate its membership value in the soft cluster and compare it with the preset threshold: (1) If the membership degree of the first sample point in the soft cluster is greater than or equal to the preset threshold, the first sample point is considered to belong to the weather label corresponding to the current hard cluster and is classified into the corresponding weather cluster.

[0041] (2) If the membership of the first sample point in the soft cluster is lower than the preset threshold, it cannot be assigned to the current hard cluster. Therefore, it is necessary to reclassify the first sample point and reallocate it to the weather cluster corresponding to the maximum membership in its membership vector except for the current weather label.

[0042] In addition, for the sample points outside the intersection of the hard clustering clusters, that is, the second sample points in the corresponding complement set, they need to be reclassified. According to the membership vector of the second sample point in the soft clustering, they are classified into the weather cluster corresponding to the largest membership value.

[0043] The embodiment of the present invention can effectively correct the classification errors caused by fuzzy boundaries or data uncertainty in hard clustering clusters through the FCM algorithm, improve the consistency of samples within the weather clustering cluster, and provide more reliable data support for the subsequent training of the power prediction model.

[0044] In an optional embodiment, obtaining an adjacency matrix of an undirected graph according to the physical connection relationship between all the photovoltaic power stations includes: The elements in the adjacency matrix are determined by the following formula: ; in, is the adjacency matrix Elements in For edge The source node of side The target node.

[0045] It should be noted that the embodiment of the present invention uses an undirected graph To describe the physical connection relationship between the photovoltaic power stations in the cluster; Represents a set of photovoltaic power stations (a set of nodes), where one node represents one photovoltaic power station; is the edge set connecting each photovoltaic power station; Represents the adjacency matrix.

[0046] Adjacency Matrix Medium Element The calculation formula is as follows: ; In the formula, is the adjacency matrix Elements in For edge The source node of side In other words, if the edge and edge have the same source or target node (indicates that the node and nodes There is a direct physical connection between nodes. and nodes There are connecting edges between them, so .

[0047] In an optional embodiment, the node feature is specifically: ; in, For the The historical power time series of each PV power station within a preset period; For the Meteorological data matrix of each photovoltaic power station in a preset period of time; For the PV power stations at the predicted time point The previous The power value at each time point; For the PV power stations at the predicted time point The previous Meteorological factors at each time point The value of .

[0048] It should be noted that the embodiment of the present invention uses graph convolutional neural networks (CGN) when constructing the spatial feature extraction sub-model. The CGN network is composed of multiple layers of convolutional networks, and these convolutional networks are used to extract the input features of all nodes in the undirected graph. A convolution operation is performed to extract spatial features.

[0049] Specifically, For the The historical power time series of a photovoltaic power station in a preset period of time, that is, ; For the PV power stations at the predicted time point The previous The power value at each time point; For the The meteorological data matrix of a photovoltaic power station in a preset period of time is Dimensional matrix ; For the PV power stations at the predicted time point The previous Meteorological factors at each time point The meteorological factors include total irradiance (horizontal radiation, horizontal diffuse radiation, oblique diffuse radiation, oblique diffuse radiation), ambient temperature and relative humidity.

[0050] Therefore, for the Node (i.e. The node characteristics of a photovoltaic power station are , which is actually expressed as .

[0051] In an optional embodiment, the training of the original prediction model of each weather cluster based on the adjacency matrix and the node features in each weather cluster to obtain the corresponding power prediction model includes: Inputting the adjacency matrix and the node features in each of the weather clusters into the corresponding graph convolutional neural network for training, respectively, to obtain a spatial feature extraction sub-model for each of the weather clusters, and corresponding spatial feature sample data; The node feature and spatial feature sample data in each of the weather clusters are respectively input into the corresponding gated recurrent unit for training, so as to obtain the temporal feature extraction sub-model of each of the weather clusters; wherein the hyperparameters in the gated recurrent unit are obtained by optimizing the improved artificial bee colony optimization algorithm; The combination of the spatial feature extraction sub-model and the temporal feature extraction sub-model of each weather cluster is used as the corresponding power prediction model.

[0052] like Figure 2 As shown, the embodiment of the present invention establishes different power prediction models according to different weather clusters.

[0053] Part 1: When constructing the spatial feature extraction sub-model, the embodiment of the present invention adopts a CGN network for construction. The CGN network extracts the spatial features of each node in the undirected graph through a multi-layer convolutional network, as shown in the following example: (1) Convolutional layer extracts features , the calculation formula is as follows: ; in, Representative The output of the layer; ; express The degree matrix of ; Indicates The weight matrix of the layer; represents a nonlinear activation function, is the identity matrix.

[0054] (2) Based on the features extracted by the convolutional layer, output the spatial features of the training sample data (or spatial feature sample data) , the calculation formula is as follows: ; in, represents the activation function, represents the hyperbolic tangent function.

[0055] Part II: When constructing the time series feature extraction sub-model, the embodiment of the present invention uses a gated recurrent unit (Gated Recurrent Unit, GRU) to extract the time series features of the photovoltaic power stations in the cluster, and optimizes the hyper parameters in the GRU through the modified artificial bee colony (Modified Artificial Bee Colony, MABC) algorithm, such as Figure 2 The "MABC-GRU network" in is a sub-model for temporal feature extraction.

[0056] In other words, the MABC algorithm is used to optimize the hyperparameters in the GRU neural network, and the MABC-GRU network is established to extract the time series characteristics of the photovoltaic power stations in the cluster. (or ) and spatial features (or spatial feature sample data) As the input of the MABC-GRU neural network.

[0057] For example, ;in, For the The spatial characteristic data of each photovoltaic power station within a preset period of time, for dimensional matrix; an input node in the MABC-GRU neural network The data format is The actual output power of the photovoltaic power station cluster is used as the final comparison data set. The model parameters are updated through a reverse iterative algorithm, and the power prediction models under various weather conditions are obtained through continuous cycles.

[0058] See also Figure 3 , is a flowchart of an embodiment of optimizing hyperparameters in a gated recurrent unit (GRU) using an improved artificial bee colony (MABC) algorithm provided by the present invention.

[0059] In an optional embodiment, the improved artificial bee colony optimization algorithm guides bees to search for new nectar sources through the following formula: ; in, For the newly generated The location of the honey source of the hyperparameter; For the The current number of bees Hyperparameter values; For the The step factor between the leading bee and the best nectar source corresponding to the hyperparameter; The currently found The global optimal solution corresponding to the hyperparameters; For the The current number of bees Hyperparameter values; For the The hyperparameters correspond to The step size factor of the bees; .

[0060] It should be noted that the improved artificial bee colony (MABC) algorithm adopted in the embodiment of the present invention is as follows: (1) Initialize the swarm parameters, population size, and maximum number of iterations. The initialization formula is: ; In the formula, , ; and They refer to the dimension of the optimization parameters and the number of populations respectively; are the maximum and minimum values ​​of the optimization parameters respectively; .

[0061] (2) The leading bees search for new nectar sources according to the following improved method, which makes up for the deficiency of the traditional ABC algorithm that is prone to fall into the local optimal solution: ; In the formula, For the newly generated The location of the honey source of hyperparameters (i.e., the solution / hyperparameter of the new search); For the The current number of bees Hyperparameter values; For the The step factor between the leading bee and the best nectar source corresponding to the hyperparameter is used to control the speed at which the leading bee approaches the best nectar source; The currently found The global optimal solution (i.e. the best nectar source) corresponding to the hyperparameters; For the The current number of bees hyperparameter values, and , that is, the nectar source value of other bees in the neighborhood; For the The hyperparameters correspond to The step size factor of the bee is used to control the The search distance between a bee and other bees in its neighborhood.

[0062] (3) After the leading bee finds the nectar source, it passes the information to the follower bee. The follower bee decides whether to continue following the employed bee based on the roulette wheel. Then the follower bee continues to search for a new nectar source near the nectar source according to the formula in step (2) above. If the fitness value is If it is greater than the original nectar source, the location of the nectar source is updated. The probability of a follower bee choosing an employed bee is , the improved calculation formula is as follows: ; ; in: is the fitness A solution; For the The value of an individual for the optimization problem.

[0063] The current optimal nectar source is used as the standard for follower bees to find nectar sources, allowing follower bees to look for nectar sources with high fitness values, reducing the search time and improving the quality of search.

[0064] (3) If a nectar source value passes the exploration limit If the value of the nectar source is still not updated in the next cycle, the scout bee will search for a new nectar source value according to the above step (2).

[0065] (4) When the maximum number of iterations is reached, the loop ends and the optimal hyperparameters of the GRU network are output.

[0066] It is worth mentioning that the optimization of the Artificial Bee Colony (ABC) algorithm makes up for the deficiency of the traditional ABC algorithm that is prone to fall into the local optimal solution. At the same time, it also improves the probability method of follower bees selecting hired bees, reduces the search time, and improves the search quality. Using the improved artificial bee colony algorithm to optimize the hyperparameters in GRU can overcome the randomness problem caused by the artificial setting of hyperparameters in the GRU prediction model, thereby improving the prediction accuracy of the power prediction model.

[0067] See also Figure 4 , is a flow chart of an embodiment of a hybrid prediction model application method for photovoltaic power station cluster power provided by the present invention.

[0068] The second aspect of the present invention provides a hybrid prediction model application method for photovoltaic power station cluster power, including steps S21 to S24, which are as follows: Step S21: acquiring the collected data of the photovoltaic power station cluster in the current period; wherein the collected data includes: the power time series and meteorological data matrix of all photovoltaic power stations in the current period; Step S22: According to the collected data The Euclidean distance between the centers of the weather clusters is used to determine the power prediction model; Step S23: dividing the collected data according to the photovoltaic power station to obtain the current node features of the undirected graph; Step S24: Input the adjacency matrix of the undirected graph and the current node features into the power prediction model to obtain the predicted power of the photovoltaic power station cluster; wherein the weather clustering cluster, the adjacency matrix and the power prediction model are obtained by adopting the hybrid prediction model construction method of the photovoltaic power station cluster power described in any one of the first aspects above.

[0069] It should be noted that the embodiment of the present invention is an ultra-short-term power prediction, the sampling time interval is 15 minutes, and the data collected at 5 time points in the current period (or the preset period) constitute the collected data, that is, the power time series and meteorological data matrix of all photovoltaic power stations in the current period. The cluster power generation at the next time point is predicted by collecting data (data at 5 time points).

[0070] In the specific implementation, based on the sample point data (collected data) at 5 time points before the prediction time , calculate its The center of the weather cluster The Euclidean distance between them is used to determine the weather type at the prediction time and the corresponding power prediction model, which is then brought into the power prediction model for cluster power prediction; , Indicates the time before the prediction time. Meteorological factor value.

[0071] In order to further reflect the technical effects achieved by the hybrid prediction model construction method and application method of photovoltaic power station cluster power provided by the present invention, some specific embodiments are provided below for reference, and the technical effects of the present invention are described: Example 1: Select the meteorological data and power generation of a photovoltaic power station in a certain area from June 1 to September 25, 2021, and calculate the Person correlation coefficient between the historical power generation and each meteorological factor. See Table 1, which is a table of the calculation results of the Person correlation coefficient.

[0072] Table 1. Calculation results of Person correlation coefficient As can be seen from Table 1, total irradiance (horizontal radiation, horizontal diffuse, oblique radiation, oblique diffuse), ambient temperature and relative humidity are meteorological characteristics that affect photovoltaic power generation.

[0073] Example 2: Calculate the clustering evaluation index Gap value to determine the optimal number of clusters See Table 2 for different Gap value table corresponding to the value.

[0074] Table 2. Different Gap value table corresponding to the value As can be seen from Table 2, when the number of clusters When the value is 3, the clustering effect is best.

[0075] Example 3: First, the sample points are hard clustered and assigned weather type labels by using the clustering algorithm (cluster 1) that combines the K-Means and DPC algorithms (i.e., each cluster is defined according to the official meteorological information, namely, sunny, rainy, and cloudy). Then, the sample points are softly clustered and assigned weather type labels by using the FCM clustering algorithm (cluster 2). Next, the intersection of the sunny clusters in cluster 1 and cluster 2 is calculated, and the sunny membership of the first sample point in the intersection is determined. The sample points with sunny membership less than 0.7 are divided into cloudy clusters; similarly, the boundaries of the rainy clusters are corrected. See Table 3, which shows the final clustering results.

[0076] Table 3. Final clustering results Example 4: To verify the superiority of the improved artificial bee colony optimization algorithm, two test functions are selected as shown in Table 4, and the parameters of the MABC optimization algorithm are set as follows: the number of colonies is set to 80, the maximum number of iterations is 100, and the number of runs is set to 10. The following will compare the MABC optimization algorithm in terms of computational accuracy and stability.

[0077] Table 4. Test function table Table 5. Test results comparison table In Table 5, Best represents the minimum error, Worst represents the maximum error, and Mean represents the average error.

[0078] From Table 5, we can see that for the function , the Best of MABC is 0, the optimization result is the same as the optimal value, the Worst of MABC is ten times smaller than that of ABC, and the Mean of MABC is 10 times smaller than that of ABC; for the function , ABC's Best is 100 times that of MABC, which shows that MABC's search quality is much higher than ABC, and MABC's Mean is lower than ABC. In summary, the calculation accuracy of MABC optimization algorithm is higher than that of ABC.

[0079] Example 5: The sampling time of each photovoltaic power station in the photovoltaic power station cluster is 7:20-18:20, with an interval of 15 minutes. The data set is divided into training set and test set in a ratio of 7:3.

[0080] The spatial features are extracted through the graph convolutional neural network, and the hyperparameters of the graph convolutional neural network are set by empirical methods, as shown in Table 6.

[0081] Table 6. Some hyperparameters of graph convolutional neural network The MABC algorithm is used to optimize the hyperparameters of the gated recurrent unit. The results are shown in Table 7.

[0082] Table 7. Some hyperparameters of gated recurrent units The trained power prediction model is used to predict the cluster power of the photovoltaic power station, and the prediction results obtained by the embodiment of the present invention are compared with other prediction models. The comparison results are as follows: Figures 5 to 7 As shown; among them, Figure 5 A comparative example diagram of the prediction results under a clear sky provided by the present invention; Figure 6 This is a comparative example diagram of the prediction results under cloudy days (overcast days) provided by the present invention; Figure 7 This is a comparative example diagram of the prediction results for rainy days provided by the present invention.

[0083] Depend on Figures 5 to 7 It can be seen that under the three weather conditions, the prediction curve corresponding to the power prediction model (hybrid model) of the present invention is closer to the true value curve, especially under the sunny state. In cloudy and rainy days, due to the changeable meteorological conditions, the power generation of each photovoltaic power station fluctuates greatly. Therefore, the power prediction model is less stable than the sunny state, but compared with other prediction models, the prediction curve corresponding to the power prediction model (hybrid model) of the present invention is still closer to the true value curve.

[0084] See also Figure 8 , is a structural schematic diagram of an embodiment of a hybrid prediction model building system for photovoltaic power station cluster power provided by the present invention.

[0085] The third aspect of the present invention provides a system for constructing a hybrid prediction model of a photovoltaic power plant cluster power, which is used to implement the method for constructing a hybrid prediction model of a photovoltaic power plant cluster power described in any embodiment of the first aspect, and the construction system includes: The data acquisition module 11 is used to obtain the data of the photovoltaic power station cluster. The historical time data corresponding to each preset time period respectively; wherein the historical time data includes: the historical power time series and meteorological data matrix of all photovoltaic power stations in the corresponding preset time period; ; The weather cluster acquisition module 12 is used to The historical time data are input into the hard clustering model and the soft clustering model respectively, and the hard clusters and soft clusters; and modifying the boundaries of the hard clusters by using the soft clusters to obtain weather clusters; among them, ; The adjacency matrix establishment module 13 is used to obtain the adjacency matrix of the undirected graph according to the physical connection relationship between all the photovoltaic power stations; A node feature acquisition module 14 is used to divide each historical time data according to the photovoltaic power station and obtain the node features of the undirected graph; The prediction model training module 15 is used to train the original prediction model of each weather cluster based on the adjacency matrix and the node features in each weather cluster to obtain a corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a temporal feature extraction sub-model.

[0086] It should be noted that the hybrid prediction model construction system for photovoltaic power station cluster power provided by the embodiment of the third aspect of the present invention can implement all processes of the hybrid prediction model construction method for photovoltaic power station cluster power described in any embodiment of the first aspect above. The functions of each module in the system and the technical effects achieved are respectively the same as the functions and technical effects achieved by the hybrid prediction model construction method for photovoltaic power station cluster power described in the embodiment of the first aspect above, and will not be repeated here.

[0087] See also Fig. 9 , is a structural schematic diagram of an embodiment of a hybrid prediction model application system for photovoltaic power station cluster power provided by the present invention.

[0088] The fourth aspect of the present invention provides an application system of a hybrid prediction model of photovoltaic power station cluster power, which is used to implement the application method of the hybrid prediction model of photovoltaic power station cluster power described in any embodiment of the second aspect, and the application system includes: The data acquisition module 21 is used to obtain the collected data of the photovoltaic power station cluster in the current period; wherein the collected data includes: the power time series and meteorological data matrix of all photovoltaic power stations in the current period; The prediction model determination module 22 is used to determine the prediction model according to the collected data. The Euclidean distance between the centers of the weather clusters is used to determine the power prediction model; A node feature determination module 23 is used to divide the collected data according to the photovoltaic power station to obtain the current node features of the undirected graph; The predicted power acquisition module 24 is used to input the adjacency matrix of the undirected graph and the current node features into the power prediction model to obtain the predicted power of the photovoltaic power station cluster; wherein the weather clustering cluster, the adjacency matrix and the power prediction model are obtained by adopting the hybrid prediction model construction method of the photovoltaic power station cluster power described in any one of the first aspects above.

[0089] It should be noted that the hybrid prediction model application system for photovoltaic power station cluster power provided in the fourth aspect of the embodiment of the present invention can implement all processes of the hybrid prediction model application method for photovoltaic power station cluster power described in the above-mentioned second aspect embodiment. The functions of each module in the system and the technical effects achieved are respectively the same as the functions and technical effects achieved by the hybrid prediction model application method for photovoltaic power station cluster power described in the above-mentioned second aspect embodiment, and will not be repeated here.

[0090] 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 technical 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 hybrid prediction model construction method for photovoltaic power station cluster power, characterized in that: include: Get the photovoltaic power station cluster in The historical time data corresponding to each preset time period respectively; wherein the historical time data includes: the historical power time series and meteorological data matrix of all photovoltaic power stations in the corresponding preset time period; ; Will The historical time data are input into the hard clustering model and the soft clustering model respectively, and the hard clusters and soft clusters; and modifying the boundaries of the hard clusters by using the soft clusters to obtain weather clusters; among them, ; According to the physical connection relationship between all the photovoltaic power stations, an adjacency matrix of an undirected graph is obtained; Divide each historical time data according to the photovoltaic power station to obtain the node features of the undirected graph; Based on the adjacency matrix and the node features in each weather cluster, the original prediction model of each weather cluster is trained to obtain a corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a temporal feature extraction sub-model.

2. The hybrid prediction model construction method for photovoltaic power station cluster power according to claim 1, characterized in that: The training of the original prediction model of each weather cluster based on the adjacency matrix and the node features in each weather cluster to obtain a corresponding power prediction model includes: Inputting the adjacency matrix and the node features in each of the weather clusters into the corresponding graph convolutional neural network for training, respectively, to obtain a spatial feature extraction sub-model for each of the weather clusters, and corresponding spatial feature sample data; The node feature and spatial feature sample data in each of the weather clusters are respectively input into the corresponding gated recurrent unit for training, so as to obtain the temporal feature extraction sub-model of each of the weather clusters; wherein the hyperparameters in the gated recurrent unit are obtained by optimizing the improved artificial bee colony optimization algorithm; The combination of the spatial feature extraction sub-model and the temporal feature extraction sub-model of each weather cluster is used as the corresponding power prediction model.

3. The hybrid prediction model construction method for photovoltaic power station cluster power according to claim 1, characterized in that: The hard clustering clusters are obtained by the following steps: Using each of the historical time data as a sample point of the hard clustering model; In the clustering iteration process, cluster assignment is performed based on the Euclidean distance from each sample point to the cluster center, and the clusters, and in each of the clusters, update the corresponding cluster center by using the density peak clustering algorithm; When the iteration stop condition is met, the output The hard clustering clusters.

4. The hybrid prediction model construction method for photovoltaic power station cluster power according to claim 1, characterized in that: The soft clustering is used to correct the boundaries of the hard clustering to obtain Weather clusters, including: Assigning a weather label to each of the hard clusters and each of the soft clusters; Obtaining the intersection of the hard clustering cluster and the soft clustering cluster of the same weather label; Determine whether the degree of membership of each first sample point in the intersection is greater than or equal to a preset threshold; If yes, classifying the first sample point into the weather cluster corresponding to the same weather label; If not, reclassifying the first sample point into the weather cluster corresponding to the maximum membership in its own membership vector except for the same weather label; The second sample point in the complement of the intersection with respect to the hard cluster is reclassified as the weather cluster corresponding to the maximum membership in its own membership vector.

5. The hybrid prediction model construction method for photovoltaic power station cluster power according to claim 1, characterized in that: The step of obtaining an adjacency matrix of an undirected graph according to the physical connection relationship between all the photovoltaic power stations comprises: The elements in the adjacency matrix are determined by the following formula: ； in, is the adjacency matrix Elements in For edge The source node of side The target node.

6. The hybrid prediction model construction method for photovoltaic power station cluster power according to claim 1, characterized in that: The node features are specifically: ； in, For the The historical power time series of each PV power station within a preset period; For the Meteorological data matrix of each photovoltaic power station in a preset period of time; For the PV power stations at the predicted time point The previous The power value at each time point; For the PV power stations at the predicted time point The previous Meteorological factors at each time point The value of .

7. The hybrid prediction model construction method for photovoltaic power station cluster power according to claim 2, characterized in that: The improved artificial bee colony optimization algorithm guides bees to search for new nectar sources through the following formula: ； in, For the newly generated The location of the honey source of the hyperparameter; For the The current number of bees Hyperparameter values; For the The step factor between the leading bee and the best nectar source corresponding to the hyperparameter; The currently found The global optimal solution corresponding to the hyperparameters; For the The current number of bees Hyperparameter values; For the The hyperparameters correspond to The step size factor of the bees; .

8. A hybrid prediction model application method for photovoltaic power station cluster power, characterized in that: include: Acquire the collected data of the photovoltaic power station cluster in the current period; wherein the collected data includes: the power time series and meteorological data matrix of all photovoltaic power stations in the current period; According to the collected data The Euclidean distance between the centers of the weather clusters is used to determine the power prediction model; Dividing the collected data according to photovoltaic power stations to obtain current node features of an undirected graph; The adjacency matrix of the undirected graph and the current node features are input into the power prediction model to obtain the predicted power of the photovoltaic power station cluster; wherein the weather clustering cluster, the adjacency matrix and the power prediction model are obtained by adopting the hybrid prediction model construction method of the photovoltaic power station cluster power as described in any one of claims 1 to 7.

9. A system for constructing a hybrid prediction model for photovoltaic power station cluster power, characterized in that: include: The data acquisition module is used to obtain the data of the photovoltaic power station cluster. The historical time data corresponding to each preset time period respectively; wherein the historical time data includes: the historical power time series and meteorological data matrix of all photovoltaic power stations in the corresponding preset time period; ; Weather cluster acquisition module, used to The historical time data are input into the hard clustering model and the soft clustering model respectively, and the hard clusters and soft clusters; and modifying the boundaries of the hard clusters by using the soft clusters to obtain weather clusters; among them, ; An adjacency matrix building module, used to obtain an adjacency matrix of an undirected graph according to the physical connection relationship between all the photovoltaic power stations; A node feature acquisition module, used to divide each historical time data according to photovoltaic power stations, and obtain node features of the undirected graph; A prediction model training module is used to train the original prediction model of each weather cluster based on the adjacency matrix and the node features in each weather cluster to obtain a corresponding power prediction model; wherein the structure of the power prediction model is composed of a spatial feature extraction sub-model and a temporal feature extraction sub-model.

10. An application system of a hybrid prediction model for photovoltaic power station cluster power, characterized in that: include: The data acquisition module is used to obtain the collected data of the photovoltaic power station cluster in the current period; wherein the collected data includes: the power time series and meteorological data matrix of all photovoltaic power stations in the current period; The prediction model determination module is used to determine the The Euclidean distance between the centers of the weather clusters is used to determine the power prediction model; A node feature determination module, used to divide the collected data according to the photovoltaic power station to obtain the current node features of the undirected graph; A predicted power acquisition module is used to input the adjacency matrix of the undirected graph and the current node features into the power prediction model to obtain the predicted power of the photovoltaic power station cluster; wherein the weather cluster cluster, the adjacency matrix and the power prediction model are obtained by adopting the hybrid prediction model construction method of the photovoltaic power station cluster power as described in any one of claims 1 to 7.

Images