A joint optimization clustering distributed photovoltaic power prediction method and system

Abstract

The application discloses a kind of cluster distributed photovoltaic power prediction method and system of joint optimization, it is related to the field of distributed energy management, including: data local call is carried out, timing alignment is carried out, constructs multidimensional feature correlation number table;Extract geographical location information and historical output data to carry out similarity analysis, carry out spectral clustering, output photovoltaic cluster;According to power prediction scale collection meteorological time series data input cluster power prediction model, carry out cluster power prediction;According to cluster level power prediction value carries out grid-connected space-time joint optimization, outputs energy storage charge-discharge scheduling strategy;After carrying out single station fitting conversion, output distributed power prediction value, carry out output deviation identification, locate fault node;Dynamic update energy storage charge-discharge scheduling strategy.Solve the technical problems that existing power prediction has not high precision, it is difficult to realize cross-cluster power optimization scheduling, and grid operation stability is not good, to improve the technical effects of prediction accuracy and grid operation stability.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of distributed energy management, in particular to a joint optimization clustering distributed photovoltaic power prediction method and system. BACKGROUND

[0002] Accurate prediction of distributed photovoltaic power is crucial for stable operation of the power grid and efficient use of energy. At present, the main methods to solve the problem of distributed photovoltaic power prediction are to model and predict based on the historical data of a single photovoltaic power station or unit, or to use a simple clustering method to group photovoltaic power stations for prediction. However, these methods, when faced with the actual situation of small amount of data or incomplete data of distributed photovoltaic power stations, do not fully consider the geographical location difference between photovoltaic power stations, the time alignment of historical output data, and the power complementarity between clusters, resulting in low prediction accuracy and difficulty in realizing cross-cluster power optimization scheduling.

[0003] In the related art, the distributed photovoltaic power prediction has the technical problems of low prediction accuracy, difficulty in realizing cross-cluster power optimization scheduling, and poor stability of power grid operation, especially in the scenario of small amount of data or incomplete data of distributed photovoltaic power stations. SUMMARY

[0004] The present application provides a joint optimization clustering distributed photovoltaic power prediction method and system, which adopts local calling and time alignment preprocessing of multi-source data, constructs a multi-dimensional feature correlation table, extracts information from the table to analyze similarity, performs spectral clustering on photovoltaic power generation units to obtain K photovoltaic clusters, inputs meteorological time series data into a cluster power prediction model, performs cluster power prediction to obtain K cluster-level power prediction values, performs grid-connected space-time joint optimization according to the cluster-level power prediction values, outputs energy storage charging and discharging scheduling strategy to realize cross-cluster power complementarity, fits and converts the distributed power prediction value according to the cluster-level power prediction value, identifies the output deviation to locate the fault node, and dynamically updates the energy storage charging and discharging scheduling strategy. The technical means achieve the technical effects of improving the accuracy of photovoltaic power prediction, realizing cross-cluster power optimization scheduling, and improving the stability of power grid operation.

[0005] The application provides a cluster distributed photovoltaic power prediction method combined with optimization, comprising: calling multiple-source data of a distributed photovoltaic power station locally, and preprocessing the calling results through time sequence alignment to construct a multi-dimensional feature correlation number table; extracting multiple geographical position information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station from the multi-dimensional feature correlation number table for similarity analysis, and performing power generation unit spectrum clustering on the multiple photovoltaic power generation units according to the analysis results, outputting K photovoltaic clusters, K being a natural number greater than or equal to 1; inputting meteorological time sequence data collected according to a power prediction scale into K cluster power prediction models of the K photovoltaic clusters for cluster power prediction, and outputting K cluster-level power prediction values; performing grid-connected space-time joint optimization according to the K cluster-level power prediction values, outputting an energy storage charging and discharging scheduling strategy, and performing cross-cluster power complementation of the K photovoltaic clusters; after single-station fitting conversion is performed according to the K cluster-level power prediction values, distributed power prediction values are output, output deviation of the distributed power prediction values is identified, and a photovoltaic power station fault node is located; and the energy storage charging and discharging scheduling strategy is dynamically updated according to the photovoltaic power station fault node.

[0006] In a possible implementation, the similarity analysis of the multiple geographical position information and the multiple historical output data of the multiple photovoltaic power generation units in the distributed photovoltaic power station from the multi-dimensional feature correlation number table and the power generation unit spectrum clustering of the multiple photovoltaic power generation units according to the analysis results to output K photovoltaic clusters comprise: drawing multiple unit output curves according to the multiple historical output data; outputting an output curve similarity matrix of the multiple photovoltaic power generation units by analyzing the multiple unit output curves through a dynamic time warping algorithm; constructing a spatial correlation matrix according to the multiple geographical position information; and performing power generation unit spectrum clustering by fusing and analyzing the output curve similarity matrix and the spatial correlation matrix, and clustering the multiple photovoltaic power generation units into the K photovoltaic clusters.

[0007] In a possible implementation, the grid-connected space-time joint optimization according to the K cluster-level power prediction values to output an energy storage charging and discharging scheduling strategy and perform cross-cluster power complementation of the K photovoltaic clusters comprise: dynamically loading K real-time grid-connected processing thresholds according to the K cluster-level power prediction values; obtaining K cluster-level power measured values of the K photovoltaic clusters by summing output power monitoring; performing space-time coupling optimization according to the deviation of the K cluster-level power measured values and the K real-time grid-connected processing thresholds to output the energy storage charging and discharging scheduling strategy; and performing cross-cluster power dynamic complementation of the K photovoltaic clusters according to the energy storage charging and discharging scheduling strategy.

[0008] In a possible implementation, the clustering of the plurality of photovoltaic generating units into the K photovoltaic clusters is performed by fusing the output curve similarity matrix and the spatial correlation matrix, including: performing exponential decay mapping on the output curve similarity matrix to generate a first normalized similarity matrix; performing normalization on the spatial correlation matrix according to the matrix scale of the first normalized similarity matrix to obtain a second normalized similarity matrix; superimposing the first normalized similarity matrix and the second normalized similarity matrix based on a preset matrix weight coefficient to obtain a fused similarity matrix; constructing a normalized Laplacian matrix based on the fused similarity matrix; performing eigenvalue decomposition on the normalized Laplacian matrix to select K low-dimensional eigenvectors corresponding to K smallest eigenvalues; mapping the plurality of photovoltaic generating units to a low-dimensional feature space based on the K low-dimensional eigenvectors; and updating the cluster center points in the low-dimensional feature space iteratively to output the K photovoltaic clusters.

[0009] In a possible implementation, the updating of the cluster center points in the low-dimensional feature space iteratively to output the K photovoltaic clusters includes: presetting a center point spacing scale, and initializing K cluster center points in the low-dimensional feature space based on the center point spacing scale; assigning the plurality of photovoltaic generating units in the low-dimensional feature space to the K cluster center points based on the nearest Euclidean distance to obtain K sets of generating unit assignment results; performing random disturbance iteration on the K cluster center points to obtain K first iteration center points; assigning the plurality of photovoltaic generating units in the low-dimensional feature space to the K first iteration center points based on the nearest Euclidean distance to obtain K sets of first iteration assignment results; comparing the distance sum of squares of the K sets of generating unit assignment results and the K sets of first iteration assignment results, and performing iteration retention judgment on the K cluster center points and the K first iteration center points according to the comparison result; and outputting the K photovoltaic clusters when the iteration fluctuation of the distance sum of squares is less than a preset deviation.

[0010] In a possible implementation, the spatio-temporal coupling optimization is performed according to deviations of the K cluster-level power measured values from the K real-time grid processing thresholds, and the energy storage charging and discharging scheduling strategy is output, including: according to deviation directions of the K cluster-level power measured values from the K real-time grid processing thresholds, the K photovoltaic clusters are divided into W excess output clusters and F insufficient output clusters, W is a natural number greater than or equal to 1, and F is a natural number greater than or equal to 1; local energy storage fitting is performed on the W excess output clusters, and W excess output values are output; after F insufficient output values of the F insufficient output clusters are extracted, the W excess output values and the F insufficient output values are processed to obtain a plurality of initial charging and discharging scheduling strategies; and the plurality of initial charging and discharging scheduling strategies are evaluated with the minimum energy storage loss and power transmission cost as the target, and the energy storage charging and discharging scheduling strategy is obtained through screening.

[0011] In a possible implementation, before the K cluster power prediction models of the K photovoltaic clusters are constructed by inputting the meteorological time series data collected according to the power prediction scale to the K cluster power prediction models and performing cluster power prediction and outputting K cluster-level power prediction values, the method further includes: constructing a standard power prediction model by using an LSTM, wherein the input data of the power prediction scale is meteorological time series data and output time series data conforming to the power prediction scale, and the output result is predicted power time series data conforming to the power prediction scale of a preset scale factor; performing time series alignment data calling according to the plurality of historical output data to obtain a plurality of meteorological data records and a plurality of power generation records; taking the plurality of historical output data, the plurality of meteorological data records, and the plurality of power generation records as training data to localize the standard power prediction model to obtain a plurality of initial power prediction sub-models of the plurality of photovoltaic power generation units; performing model parameter federal aggregation on the plurality of initial power prediction sub-models according to the K photovoltaic clusters to obtain K federal power prediction sub-models; replicating the K federal power prediction sub-models according to the number of clusters of the K photovoltaic clusters to obtain K sets of federal power prediction sub-models; the K sets of federal power prediction sub-models are connected in parallel in a group, and a power summing engine is configured at an output end, so as to complete construction of the K cluster power prediction models.

[0012] In a possible implementation, the constructing the spatial correlation matrix according to the plurality of geographic location information comprises: enumerating the plurality of photovoltaic power generation units through a combination algorithm to obtain H groups of photovoltaic power generation units, H being a natural number greater than or equal to 1; combining the plurality of geographic location information according to the H groups of photovoltaic power generation units to obtain H groups of geographic location information; wherein the geographic location information comprises photovoltaic unit longitude and latitude, photovoltaic unit altitude, photovoltaic panel inclination angle and photovoltaic panel azimuth angle; after processing the H groups of geographic location information based on a preset geographic normalization weight, performing inter-group Euclidean distance calculation on the processing result to output H spatial correlation coefficients; constructing a reference correlation matrix according to the plurality of photovoltaic power generation units; filling the H spatial correlation coefficients into the reference correlation matrix with the H groups of photovoltaic power generation units as filling guides, and completing the construction of the spatial correlation matrix.

[0013] In a possible implementation, the multi-source data of the distributed photovoltaic power station is locally called, and the calling result is preprocessed through time sequence alignment to construct a multi-dimensional feature correlation number table, comprising: performing local data calling through a plurality of edge collection terminals arranged at the plurality of photovoltaic power generation units to obtain a plurality of meteorological data records, a plurality of output data records and the plurality of geographic location information; presetting an output sliding window, and performing sliding mean calculation on the plurality of output data records according to the output sliding window to obtain a plurality of down-sampled output data; presetting a meteorological interpolation granularity, and performing linear interpolation processing on the plurality of meteorological data records by using the meteorological interpolation granularity to obtain a plurality of up-sampled meteorological data; after time sequence alignment of the plurality of down-sampled output data and the plurality of up-sampled meteorological data, storing the plurality of down-sampled output data, the plurality of up-sampled meteorological data and the plurality of geographic location information based on the plurality of photovoltaic power generation units, and completing the construction of the multi-dimensional feature correlation number table.

[0014] The application also provides a cluster distributed photovoltaic power prediction system optimized in combination, comprising: a time sequence alignment module, configured to locally call multi-source data of a distributed photovoltaic power station, and to preprocess the calling results by time sequence alignment to construct a multi-dimensional feature correlation number table; a power generation unit spectrum clustering module, configured to extract multiple geographical location information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station from the multi-dimensional feature correlation number table to perform similarity analysis, and to perform power generation unit spectrum clustering on the multiple photovoltaic power generation units according to the analysis results, and to output K photovoltaic clusters, K being a natural number greater than or equal to 1; a cluster power prediction module, configured to input meteorological time sequence data collected according to a power prediction scale into K cluster power prediction models of the K photovoltaic clusters to perform cluster power prediction, and to output K cluster-level power prediction values; an energy storage charging and discharging scheduling strategy output module, configured to perform grid-connected space-time joint optimization according to the K cluster-level power prediction values, and to output an energy storage charging and discharging scheduling strategy to perform cross-cluster power complementation of the K photovoltaic clusters; an output deviation identification module, configured to perform single-station fitting conversion according to the K cluster-level power prediction values to output a distributed power prediction value, and to perform output deviation identification according to the distributed power prediction value to locate a photovoltaic power station fault node; and an energy storage charging and discharging scheduling strategy updating module, configured to dynamically update the energy storage charging and discharging scheduling strategy according to the photovoltaic power station fault node.

[0015] The application provides a cluster distributed photovoltaic power prediction method and system optimized in combination. First, multi-source data of a distributed photovoltaic power station is locally called, and the calling results are preprocessed by time sequence alignment to construct a multi-dimensional feature correlation number table. Then, multiple geographical location information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station are extracted from the multi-dimensional feature correlation number table to perform similarity analysis, and the multiple photovoltaic power generation units are subjected to power generation unit spectrum clustering according to the analysis results, and K photovoltaic clusters are output. Next, meteorological time sequence data collected according to a power prediction scale is input into K cluster power prediction models of the K photovoltaic clusters to perform cluster power prediction, and K cluster-level power prediction values are output. Then, grid-connected space-time joint optimization is performed according to the K cluster-level power prediction values, and an energy storage charging and discharging scheduling strategy is output to perform cross-cluster power complementation of the K photovoltaic clusters. Subsequently, single-station fitting conversion is performed according to the K cluster-level power prediction values to output a distributed power prediction value, and output deviation identification is performed according to the distributed power prediction value to locate a photovoltaic power station fault node. Finally, the energy storage charging and discharging scheduling strategy is dynamically updated according to the photovoltaic power station fault node. The application achieves the technical effects of improving the accuracy of photovoltaic power prediction, realizing cross-cluster power optimization scheduling, and further improving the stability of power grid operation. BRIEF DESCRIPTION OF DRAWINGS

[0016] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings of the embodiments of the present application will be briefly introduced below. The flowcharts are used to illustrate the operations performed by the system according to the embodiments of the present application. It should be understood that the foregoing or the following operations are not necessarily performed in sequence. On the contrary, various steps can be processed in reverse order or simultaneously according to needs. Meanwhile, other operations can be added to these processes, or one or more steps of operations can be removed from these processes.

[0017] Figure 1 A flowchart of a joint optimization clustering distributed photovoltaic power prediction method provided by the embodiments of the present application.

[0018] Figure 2 A structural diagram of a joint optimization clustering distributed photovoltaic power prediction system provided by the embodiments of the present application.

[0019] Legend: timing alignment module 10, power generation unit spectrum clustering module 20, cluster power prediction module 30, energy storage charging and discharging scheduling strategy output module 40, output deviation identification module 50, and energy storage charging and discharging scheduling strategy updating module 60. DETAILED DESCRIPTION

[0020] The above description is only a summary of the technical solutions of the present application. In order to more clearly understand the technical means of the present application, the embodiments of the present application can be implemented according to the content of the specification, and in order to make the above and other purposes, features and advantages of the present application more obvious and easy to understand, the following specific embodiments of the present application are described.

[0021] In order to make the purposes, technical solutions and advantages of the present application more clear, the present application will be further described in detail below with reference to the drawings. The described embodiments should not be regarded as limiting the present application, and all other embodiments obtained by those of ordinary skill in the art without making creative efforts belong to the scope of protection of the present application.

[0022] In the following description, "some embodiments" are referred to, which describe a subset of all possible embodiments, but it can be understood that "some embodiments" can be the same subset or different subsets of all possible embodiments, and can be combined with each other without conflict, and the term "first\second" referred to only distinguishes similar objects, and does not represent a specific order for the objects. The terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or server including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or modules not clearly listed or inherent to these processes, methods, products or devices. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as understood by those skilled in the art to which the present application belongs. The terms used herein are only for the purpose of describing the embodiments of the present application.

[0023] The embodiments of the present application provide a joint optimization clustering distributed photovoltaic power prediction method, as shown in Figure 1 The method comprises the following steps:

[0024] In step S100, multi-source data of the distributed photovoltaic power station is locally called, and the calling results are preprocessed by time sequence alignment to construct a multi-dimensional feature correlation table.

[0025] Specifically, through API interface, database query or file reading, etc., a plurality of types of data such as meteorological data (irradiance, temperature, wind speed, etc.), equipment operation data (inverter state, panel temperature, etc.), historical output data, etc. are obtained from the local system of the distributed photovoltaic power station. Since the data of different data sources have different time stamps or sampling frequencies, a time alignment algorithm (such as linear interpolation, time window aggregation, etc.) is used to unify the data to the same time scale, so as to ensure the consistency and comparability of the data in subsequent analysis. The aligned data is integrated into a multi-dimensional data table (multi-dimensional feature correlation table), wherein each row represents the data of a time point, and each column represents a feature (such as irradiance, temperature, output, etc.). This multi-dimensional feature correlation table is the basis for subsequent feature extraction and similarity analysis. Among them, the distributed photovoltaic power station refers to a power system composed of a plurality of photovoltaic power generation units, which are distributed in different geographical positions and connected to the grid through the grid. Multi-source data refers to data from different data sources.

[0026] In a possible implementation, the distributed photovoltaic power station is called locally for multi-source data, and the calling results are preprocessed by time alignment to construct a multi-dimensional feature correlation table, and step S100 further includes step S110, local data calling is performed by a plurality of edge collection terminals arranged at a plurality of photovoltaic power generation units to obtain a plurality of meteorological data records, a plurality of output data records and the plurality of geographic position information.

[0027] Specifically, an edge collection terminal (a device deployed on the site of a photovoltaic power generation unit for data collection and preliminary processing) is arranged at each photovoltaic power generation unit, and these terminals are responsible for collecting data in real time or periodically. The edge collection terminal calls data from local storage or sensors, including meteorological data (data for describing the state of the atmosphere, which can be data directly affecting the output of photovoltaic power generation, such as irradiance, temperature, wind speed, etc.), output data (actual output power data of the photovoltaic power generation unit, such as inverter output power, etc.) and geographic position information (information describing the position of the photovoltaic power generation unit, including latitude, longitude, altitude, inclination angle of the photovoltaic panel, azimuth angle, etc.), and stores the called data in the form of records.

[0028] Step S120, a sliding window of output is preset, and a plurality of output data records are calculated by sliding mean according to the sliding window of output to obtain a plurality of down-sampled output data.

[0029] Specifically, according to the requirements of data analysis or prediction model, a time window size such as 1 hour, 30 minutes, etc. is set for calculating the sliding mean. In the set time window, the output data is calculated by mean to obtain the down-sampled output data. This method is used to smooth the data fluctuation and reduce the noise influence. The number of data points of the down-sampled output data is reduced, but it can better reflect the overall trend. The down-sampling refers to the process of reducing the sampling frequency or the number of data points. In this application, the original high-frequency output data record is converted into lower-frequency down-sampled output data by calculating the sliding mean of the output data. The purpose of selecting the down-sampled meteorological data is to smooth the data fluctuation, reduce the influence of noise on the data, and make the data better reflect the overall trend; and secondly, to reduce the data volume, reduce the calculation burden of subsequent processing and analysis, and improve the efficiency.

[0030] Step S130, a meteorological interpolation granularity is preset, and linear interpolation processing is performed on a plurality of meteorological data records by using the meteorological interpolation granularity to obtain a plurality of up-sampled meteorological data.

[0031] The upsampling refers to a process of increasing the sampling frequency or the number of data points of data. In this application, the meteorological data records are linearly interpolated using a meteorological interpolation granularity to obtain meteorological data at more time points, i.e., upsampled meteorological data. The upsampled meteorological data is selected to improve the time resolution of the data, so that the meteorological data and the output data after down-sampling can be better matched and aligned in the time scale, providing a more accurate and consistent data basis for subsequent analysis and modeling.

[0032] Specifically, according to the accuracy requirement of the data analysis or prediction model, a time interval such as 5 minutes, 10 minutes, etc. is set for linear interpolation. Within the set time interval, the meteorological data is linearly interpolated, i.e., linearly estimated between known data points, to obtain upsampled meteorological data. The number of data points of the upsampled meteorological data is increased, improving the time resolution of the data. The linear interpolation is a simple and effective data interpolation method, which can estimate unknown data points according to known data points.

[0033] After the time sequence alignment of the plurality of down-sampled output data and the plurality of upsampled meteorological data, the plurality of photovoltaic power generation units are associated with the plurality of down-sampled output data, the plurality of upsampled meteorological data and the plurality of geographic location information to complete the construction of the multi-dimensional feature association number table.

[0034] Specifically, the down-sampled output data and the upsampled meteorological data are aligned according to the time stamp to ensure that the data at each time point is corresponding. The aligned data and the geographic location information are associated and stored according to the photovoltaic power generation units to form a multi-dimensional feature association number table. This multi-dimensional feature association number table is the basis for subsequent feature extraction, similarity analysis and power prediction.

[0035] For example, if each minute is selected as the unified time reference, and if the preset output sliding window is 30 minutes, the average output within each 30 minutes can be calculated, and this average value can be assigned to each minute within the 30 minutes. If the recording interval of the original meteorological data is 15 minutes, the meteorological data can be interpolated to each minute time point using linear interpolation. Through such processing, the down-sampled output data and the upsampled meteorological data can be aligned on the time reference of each minute, providing an accurate and consistent data basis for subsequent analysis and modeling. This implementation method calls local data through the edge collection terminal, and performs sliding mean calculation, linear interpolation processing and time sequence alignment to construct a multi-dimensional feature association number table containing multiple features. This number table provides an accurate and consistent data basis for subsequent feature extraction, similarity analysis and power prediction, which helps to improve the accuracy of prediction and the stability of grid connection.

[0036] Step S200, extracting multiple geographical location information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station from the multi-dimensional feature correlation number table for similarity analysis, and performing power generation unit spectral clustering on the multiple photovoltaic power generation units according to the analysis result, and outputting K photovoltaic clusters, K being a natural number greater than or equal to 1.

[0037] Specifically, geographical location information related to each photovoltaic power generation unit is extracted from the multi-dimensional feature correlation number table, such as longitude, latitude, altitude, etc., which is used for subsequent similarity analysis or cluster division. The output data of each photovoltaic power generation unit at different time points is extracted from the multi-dimensional feature correlation number table to form a historical output sequence, which is the basis for building a power prediction model. By fusing geographical location information and historical output data, the similarity between different photovoltaic power generation units is calculated using similarity measurement methods (such as Euclidean distance, cosine similarity, etc.). Based on the results of similarity analysis, a similarity matrix is constructed, and a spectral clustering algorithm (such as K-means spectral clustering, Normalized Cuts, etc.) is used to divide the photovoltaic power generation units into K clusters. Spectral clustering is a clustering algorithm based on graph theory, which realizes data clustering by constructing a similarity matrix and performing spectral decomposition on the matrix. Spectral clustering algorithm can handle non-spherical data distribution, and has certain robustness to noise.

[0038] In one possible implementation, multiple geographical location information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station are extracted from the multi-dimensional feature correlation number table for similarity analysis, and the multiple photovoltaic power generation units are clustered according to the analysis result, and K photovoltaic clusters are output. Step S200 further includes step S210, drawing multiple unit output curves according to multiple historical output data.

[0039] Specifically, the historical output data of each photovoltaic power generation unit (actual output power data of the photovoltaic power generation unit in the past period of time) is extracted from the multi-dimensional feature correlation number table. For each photovoltaic power generation unit, arrange its historical output data in chronological order. Using time series analysis tools or programming languages (such as matplotlib library of Python), time as the horizontal axis and output as the vertical axis, draw the time-output curve of each unit, i.e. unit output curve, which reflects the change of photovoltaic power generation unit output power with time.

[0040] Step S220, analyzing the output curve similarity matrix of the multiple photovoltaic power generation units by dynamic time warping algorithm.

[0041] Specifically, the dynamic time warping (DTW) algorithm is an algorithm for calculating the similarity between two time series, which can handle the nonlinear changes of time series, and find the best matching path between two time series through dynamic programming, so as to calculate their similarity. The similarity between each pair of photovoltaic power generation unit output curves is calculated using the DTW algorithm. The calculated similarity values are filled into a matrix to form a similarity matrix. The similarity matrix is a two-dimensional array, where each element represents the similarity between two photovoltaic power generation unit output curves.

[0042] Step S230, constructing a spatial correlation matrix according to the plurality of geographic location information.

[0043] Specifically, the spatial correlation between each unit is calculated using the geographic location information of the photovoltaic power generation unit (including latitude, longitude, altitude, photovoltaic panel inclination and azimuth, etc.). The geographic location information reflects the spatial distribution characteristics of the photovoltaic power generation unit, and the spatial correlation describes the correlation strength of the output fluctuation between different units due to geographical proximity. By constructing a spatial correlation matrix, the influence of geographical proximity is represented in numerical form. The spatial correlation matrix is a two-dimensional array, where each element represents the correlation strength of the output fluctuation between two photovoltaic power generation units due to geographical proximity.

[0044] Step S240, performing power unit spectral clustering by fusing and analyzing the output curve similarity matrix and the spatial correlation matrix, and clustering the plurality of photovoltaic power generation units into K photovoltaic clusters.

[0045] Specifically, the output curve similarity matrix and the spatial correlation matrix are weighted and fused to obtain a comprehensive similarity matrix. The spectral clustering algorithm is a clustering algorithm based on graph theory, which uses the comprehensive similarity matrix to construct a graph structure and performs clustering through spectral decomposition of the graph. The spectral clustering algorithm is used to cluster the photovoltaic power generation units to obtain K photovoltaic clusters. The photovoltaic cluster is a collection obtained by grouping photovoltaic power generation units according to similarity. This implementation can accurately capture the output characteristic differences and similarities between photovoltaic power generation units by drawing unit output curves and using the dynamic time warping algorithm to calculate the similarity matrix. At the same time, the spatial correlation matrix is constructed considering the geographic location information, which can reflect the correlation strength of the output fluctuation caused by geographical proximity. Fusing and analyzing these two matrices and performing spectral clustering can obtain more reasonable and effective photovoltaic cluster division results. Such division results are helpful for subsequent cluster-level power prediction and development of energy storage charging and discharging scheduling strategies, thereby improving the overall operation efficiency and grid stability of the distributed photovoltaic power station.

[0046] In a possible implementation, the step S230 of constructing the spatial correlation matrix according to the plurality of geographic location information further includes a step S231 of enumerating a plurality of photovoltaic power generation units by a combination algorithm to obtain H groups of photovoltaic power generation units.

[0047] Specifically, the number of photovoltaic power generation units to be combined (for example, two units per group) is determined, and all possible combinations of photovoltaic power generation units are enumerated by using a combination algorithm to form H groups of combinations, for example, a loop or recursive algorithm is used to enumerate all possible combinations, and H is a natural number greater than or equal to 1.

[0048] In step S232, a plurality of geographic location information is obtained according to H groups of photovoltaic power generation unit combinations.

[0049] The geographic location information includes the longitude and latitude of the photovoltaic unit, the altitude of the photovoltaic unit, the inclination angle of the photovoltaic panel, and the azimuth angle of the photovoltaic panel. Specifically, the geographic location information corresponding to the H groups of photovoltaic power generation units is extracted from a multidimensional feature correlation table or a geographic location information database. The extracted geographic location information is sorted according to the H group combinations to form a set of H group geographic location information. The geographic location information includes the longitude and latitude of the photovoltaic unit, the altitude of the photovoltaic unit, the inclination angle of the photovoltaic panel, and the azimuth angle of the photovoltaic panel, and other information describing the spatial position of the photovoltaic power generation unit.

[0050] In step S233, after processing the H groups of geographic location information based on the preset geographic normalization weight, the inter-group Euclidean distance is calculated for the processing result, and H spatial correlation coefficients are output.

[0051] Specifically, the geographic location information includes multiple dimensions, such as longitude and latitude, altitude, inclination angle of the photovoltaic panel, and azimuth angle. The dimensions have different units and ranges, and in order to make them comparable when calculating the spatial correlation, they need to be normalized. The preset geographic normalization weight is used to adjust the weight of different dimensions to reflect their importance in the spatial correlation. According to the preset geographic normalization weight, the H groups of geographic location information are normalized to eliminate the dimensional differences between different dimensional data. The Euclidean distance formula is used to calculate the distance between each group of geographic location information as a measure of spatial correlation. The calculated distance value is converted into a correlation coefficient, and the smaller the distance, the higher the correlation.

[0052] An example calculation is as follows: assuming that the geographic location information of two groups of photovoltaic power generation units is as shown in Table 1.

[0053] Table 1: Geographic location information

[0054]

[0055] The preset geographical normalization weights are: longitude: 0.3, latitude: 0.3, altitude: 0.2, photovoltaic panel inclination: 0.1, photovoltaic panel azimuth: 0.1. Scale the data of each dimension to the interval [0, 1]. For example, the normalization formula of longitude is: Assuming that the minimum value of longitude in all data is 110.0° and the maximum value is 111.0°, then: normalized longitude 1= = 0.2, normalized longitude 2= = 0.5. According to the preset weights, the weighted sum of the normalized data is obtained. For example: comprehensive normalization value 1 = 0.3 x 0.2 + 0.3 x 0.5 + 0.2 x 0.5 + 0.1 x 0.3 + 0.1 x 0.8 = 0.36, comprehensive normalization value 2 = 0.3 x 0.5 + 0.3 x 0.8 + 0.2 x 0.6 + 0.1 x 0.2 + 0.1 x 0.9 = 0.49. Use the comprehensive normalization value to calculate the Euclidean distance between the two sets of geographical location information: Euclidean distance= = 0.13. Convert the Euclidean distance to a correlation coefficient, the smaller the distance, the higher the correlation. For example, you can use the formula: , assuming that the maximum Euclidean distance is 1.0, then: correlation coefficient = 1- = 0.87. Finally, the processed spatial correlation coefficient is 0.87, indicating that the two sets of photovoltaic power generation units have high correlation in geographical location.

[0056] Step S234, constructing a reference correlation matrix according to the plurality of photovoltaic power generation units.

[0057] Specifically, according to the number of photovoltaic power generation units, a reference correlation matrix is constructed, and the rows and columns of the matrix correspond to different photovoltaic power generation units. Initialize the reference correlation matrix to a zero matrix or initialize it according to certain rules. The reference correlation matrix is a matrix used to store the spatial correlation coefficients between photovoltaic power generation units, serving as a reference for subsequent filling.

[0058] Step S235, filling H spatial correlation coefficients into the reference correlation matrix with the H photovoltaic power generation units as the filling guide, and completing the construction of the spatial correlation matrix.

[0059] Specifically, according to the combination of the H photovoltaic power generation units, the filling positions of the spatial correlation coefficients in the benchmark correlation matrix are determined. The H spatial correlation coefficients are filled into the corresponding positions of the benchmark correlation matrix according to the filling guide. The filling guide is a rule or basis for guiding the filling of the spatial correlation coefficients in the benchmark correlation matrix. The spatial correlation matrix is the benchmark correlation matrix after the spatial correlation coefficients are filled, and is used to represent the spatial correlation between the photovoltaic power generation units. This implementation manner can obtain a matrix accurately reflecting the spatial correlation between the photovoltaic power generation units by combining the enumeration of the photovoltaic power generation units, extracting the geographical position information, calculating the spatial correlation coefficients, and filling them into the benchmark correlation matrix. Such a matrix helps subsequent spectral clustering of the power generation units, clusters the photovoltaic power generation units with similar geographical positions and output characteristics into the same photovoltaic cluster, and thus improves the accuracy and reliability of the cluster power prediction. Meanwhile, the spatial correlation matrix can also provide an important basis for the grid-connected space-time joint optimization, and helps to develop a more reasonable energy storage charging and discharging scheduling strategy, and realize the cross-cluster power complementation of the photovoltaic cluster.

[0060] In a possible implementation manner, the power generation unit spectral clustering is performed by fusing the analysis of the output curve similarity matrix and the spatial correlation matrix, the plurality of photovoltaic power generation units are clustered into the K photovoltaic clusters, and step S240 further includes step S241 of performing exponential decay mapping on the output curve similarity matrix to generate a first normalized similarity matrix.

[0061] Specifically, an exponential decay function is applied to process the elements in the output curve similarity matrix to reduce the influence of large similarity values and enhance the discrimination of small similarity values. For each element in the similarity matrix, an exponential decay function such as exp(-a similarity) is applied, where a is a decay coefficient and similarity is the original similarity value. The processed matrix is normalized so that the element values in the matrix are within a specific range (such as between 0 and 1). The exponential decay mapping is a mathematical transformation method for adjusting the values of the elements in the matrix to change their distribution characteristics. The first normalized similarity matrix is the output curve similarity matrix after the exponential decay mapping and normalization processing.

[0062] Step S242, according to the matrix scale of the first normalized similarity matrix, the spatial correlation matrix is normalized to obtain a second normalized similarity matrix.

[0063] Specifically, the scale of the first normalized similarity matrix (such as the range of element values, etc.) is analyzed. According to the scale of the first normalized similarity matrix, the spatial correlation matrix is normalized correspondingly, so that the two matrices have numerical comparability. That is, according to the numerical range (such as the maximum / minimum value) of the first normalized similarity matrix, the elements in the spatial correlation matrix are linearly scaled or standardized (such as min-max normalization or Z-score standardization) to match the numerical range of the first normalized similarity matrix. The second normalized similarity matrix is the spatial correlation matrix after normalization.

[0064] Step S243, based on the preset matrix weight coefficient, the first normalized similarity matrix and the second normalized similarity matrix are superimposed to obtain a fusion similarity matrix.

[0065] Specifically, the matrix weight coefficient is preset, which is a coefficient for balancing the contribution of the first normalized similarity matrix and the second normalized similarity matrix in the fusion process. The two normalized similarity matrices are weighted and summed according to the weight coefficient to obtain the fusion similarity matrix.

[0066] Step S244, constructing a normalized Laplacian matrix based on the fusion similarity matrix.

[0067] Specifically, first, the degree matrix of the fusion similarity matrix is calculated (i.e., the diagonal matrix, whose diagonal elements are the sum of each row of the fusion similarity matrix), then the Laplacian matrix is calculated (i.e., the degree matrix minus the fusion similarity matrix), and finally the Laplacian matrix is normalized. The normalized Laplacian matrix is a matrix used for spectral clustering.

[0068] Step S245, performing eigenvalue decomposition on the normalized Laplacian matrix to select K low-dimensional eigenvectors corresponding to K smallest eigenvalues.

[0069] Specifically, the normalized Laplacian matrix is decomposed into the form of eigenvalues and eigenvectors. The eigenvectors corresponding to the K smallest eigenvalues are selected as the basis vectors of the low-dimensional feature space, which are used to construct the low-dimensional feature space.

[0070] Step S246, after constructing the low-dimensional feature space based on the K low-dimensional eigenvectors, the plurality of photovoltaic power generation units are mapped to the low-dimensional feature space.

[0071] Specifically, the low-dimensional feature space is a space composed of K low-dimensional feature vectors, which are used to represent the low-dimensional features of the photovoltaic power generation units. The original high-dimensional data (i.e., the output data and geographic location information of the photovoltaic power generation units) is mapped into the low-dimensional feature space. This mapping process is achieved by calculating the projection of the feature vector of each photovoltaic power generation unit in the low-dimensional feature space. Specifically, for each photovoltaic power generation unit, the dot product operation is performed between its high-dimensional feature vector and the K low-dimensional feature vectors, obtaining the coordinate representation of the photovoltaic power generation unit in the low-dimensional feature space. In this way, each photovoltaic power generation unit has a corresponding low-dimensional feature vector in the low-dimensional feature space, thereby realizing the mapping from high-dimensional to low-dimensional. High-dimensional data contains multiple complex and redundant information dimensions, such as output data, geographic location, and other multi-class parameters of photovoltaic power generation units. "Low-dimensional" is the key feature dimension that is most representative and more concise by mathematical methods from high-dimensional data. A small number of core features are used to describe the originally complex high-dimensional data, making the data easier to process and analyze while retaining key information. The low-dimensional feature vector refers to a feature vector with a dimension less than the dimension threshold, which can be set to 5.

[0072] Step S247, updating the cluster center points in the low-dimensional feature space by iteration, and outputting K photovoltaic clusters.

[0073] Specifically, in the low-dimensional feature space, the cluster center points are updated by an iterative algorithm (such as the K-means algorithm), so that the clustering result is more accurate. According to the iteration result, the photovoltaic power generation units are divided into K clusters. This implementation method reflects the similarity between photovoltaic power generation units more comprehensively by fusing two kinds of similarity information (output curve similarity and spatial correlation), thereby improving the accuracy of clustering and providing strong support for subsequent power prediction and grid optimization.

[0074] In one possible implementation, the cluster center points are updated in the low-dimensional feature space by iteration, and the K photovoltaic clusters are output, step S247 further includes step S2471, presetting a center point distance scale, and initializing K cluster center points in the low-dimensional feature space based on the center point distance scale. Specifically, first, a center point distance scale is set, which determines the relative distance or distribution between the initial cluster center points. Then, in the low-dimensional feature space, K points are selected as the initial cluster center points according to the scale randomly or according to a certain strategy (such as uniform distribution). Step S2472, based on the nearest Euclidean distance, a plurality of photovoltaic power generation units in the low-dimensional feature space are assigned to the K cluster center points, obtaining K sets of power generation unit assignment results. Specifically, the Euclidean distance between each photovoltaic power generation unit and the K cluster center points is calculated, and each power generation unit is assigned to the cluster center point closest to it. In this way, each cluster center point will have a group of photovoltaic power generation units closest to it.

[0075] Step S2473, random perturbation iteration is performed on the K cluster center points to obtain K first iteration center points. Specifically, a slight random perturbation is performed on the current K cluster center points (i.e., a small distance is randomly moved near the current position), thereby obtaining new K iteration center points. Step S2474, the multiple photovoltaic power generation units in the low-dimensional feature space are assigned to the K first iteration center points based on the nearest Euclidean distance, thereby obtaining K sets of first iteration assignment results. Specifically, similar to step S2472, but this time the assignment is based on the new K iteration center points. Step S2475, the sum of squared distances of the K sets of power generation unit assignment results and the K sets of first iteration assignment results is calculated, and iteration retention judgment of the K cluster center points and the K first iteration center points is performed according to the calculation result. Specifically, the sum of squared distances between the current assignment result (i.e., the result of step S2472) and the iteration assignment result (i.e., the result of step S2474) is calculated. If the sum of squared distances after iteration is smaller, it means that the iteration improves the clustering effect, and therefore the new iteration center point is retained; otherwise, the original cluster center point is retained. The sum of squared distances refers to the sum of squares of distances of all points to a certain center point, which is used to measure the compactness of clustering.

[0076] Step S2476, when the iteration fluctuation of the sum of squared distances is less than a preset deviation, the K photovoltaic clusters are output. Specifically, steps S2473 to S2475 are repeatedly executed until the iteration fluctuation of the sum of squared distances (i.e., the difference between the sum of squared distances of two consecutive iterations) is less than a preset deviation value (a threshold value for judging whether the iteration converges). At this time, it is considered that the clustering has converged, and the final K photovoltaic clusters are output. This implementation mode updates the cluster center points through random perturbation and iteration, and judges whether the clustering effect is improved by comparing the sum of squared distances, thereby finally obtaining a compact and accurate clustering result. The accurate clustering can improve the accuracy and reliability of power prediction, thereby optimizing the energy storage charging and discharging scheduling strategy, and realizing cross-cluster power complementation of photovoltaic clusters.

[0077] Step S300, input the obtained meteorological time series data according to the power prediction scale into the K cluster power prediction models of the K photovoltaic clusters, perform cluster-type power prediction, and output K cluster-level power prediction values.

[0078] Specifically, according to the characteristics and predicted demand of the photovoltaic cluster, a suitable power prediction model is selected, such as a time series model (ARIMA, Autoregressive Integrated Moving Average (ARIMA), LSTM, Long Short-Term Memory (LSTM), etc.), a machine learning model (random forest, support vector machine, etc.), or a deep learning model. The power prediction model is used to predict the output value of the photovoltaic power generation cluster in the future time period, using historical output data and corresponding meteorological time series data within the cluster as the training set to train the selected model. The training process includes steps such as feature selection and parameter optimization. The trained model is verified using an independent validation set to evaluate its prediction performance (such as accuracy, stability, etc.). According to the time scale of power prediction (such as hour, day, week, etc.), the corresponding meteorological time series data is collected and input into the trained cluster power prediction model for power prediction. Each model outputs a cluster-level power prediction value.

[0079] In one possible implementation, the meteorological time series data obtained according to the power prediction scale is input into K cluster power prediction models of K photovoltaic clusters for cluster power prediction, and before outputting K cluster-level power prediction values, step S300 further includes step S310 of constructing a standard power prediction model using LSTM, wherein the input data of the power prediction scale is meteorological time series data and output time series data conforming to the power prediction scale, and the output result is predicted power time series data conforming to the preset scale factor of the power prediction scale. In this application, the preset scale factor is set to 1 / 12, i.e., the prediction time resolution output by the power prediction model is 12 times finer than the input time scale. For example, if the input is hour-level prediction (60 minutes), the output is 5-minute-level (60 divided by 12) power time series data. Specifically, the long short-term memory network (LSTM) is selected as the basic architecture of the power prediction model, which is good at processing time series data and can capture long-term dependencies in the data. The input and output of the LSTM model are designed. The input data is meteorological time series data and output time series data conforming to the power prediction scale. After preprocessing, these data are input into the LSTM model in chronological order. The output result is predicted power time series data conforming to the preset scale factor of the power prediction scale, i.e., the model can predict the power value in a shorter time period in the future. The power prediction scale refers to the time range of the predicted power, such as hour-level, minute-level, etc. In this application, the preset scale factor can be set to 1 / 12. The meteorological time series data refers to the time-varying meteorological data, such as irradiance, temperature, etc. The output time series data refers to the time-varying photovoltaic power generation unit output data.

[0080] Step S320, according to the multiple historical output data, the timing alignment data is called to obtain multiple weather data records and multiple power generation records.

[0081] Specifically, the time points corresponding to the weather data are extracted from the historical output data, ensuring that the weather data and the power generation data are aligned in time. The stored weather data and power generation data records, which are obtained through historical observation or simulation, are called for model training and verification.

[0082] Step S330, multiple historical output data, multiple weather data records and multiple power generation records are used as training data for the localization of the standard power prediction model, to obtain multiple initial power prediction sub-models of multiple photovoltaic power generation units.

[0083] Specifically, the extracted and called historical output data, weather data records and power generation records are used as the training data set. The parameters of the LSTM model are adjusted through the training process to better fit the training data and improve the prediction accuracy. The localization process includes data preprocessing, model structure adjustment, hyperparameter optimization and other steps. Among them, localization refers to the process of adapting a general model to a specific region or data set to improve the accuracy and applicability of the model.

[0084] Step S340, according to K photovoltaic clusters, model parameter federated aggregation is performed on multiple initial power prediction sub-models to obtain K federated power prediction sub-models.

[0085] Specifically, according to the division of photovoltaic clusters, the initial power prediction sub-models are grouped according to the clusters. The initial power prediction sub-models in each cluster are federated aggregated, which means merging or averaging their model parameters to obtain a federated power prediction sub-model representing the cluster. Federated aggregation can improve the generalization ability of the model because it combines the wisdom of multiple sub-models.

[0086] Step S350, according to the number of clusters of K photovoltaic clusters, K federated power prediction sub-models are replicated to obtain K groups of federated power prediction sub-models.

[0087] Specifically, for each photovoltaic cluster, according to the number of photovoltaic power generation units in the cluster, the corresponding federated power prediction sub-model is replicated, so that each cluster has an independent group of federated power prediction sub-models.

[0088] Step S360, by connecting K groups of federated power prediction sub-models in parallel in the group and configuring a power summing engine at the output end, the construction of K cluster power prediction models is completed.

[0089] Specifically, the federated power prediction sub-model groups of each cluster are connected in parallel, that is, the models are run at the same time, and the outputs of the models are processed. A power addition engine is configured at the output end to add the output power of the parallel models to obtain the predicted power of each cluster. In this way, the predicted power of each cluster can be calculated independently and finally combined into the predicted power of the entire power station. This implementation improves the accuracy of power prediction by combining historical output data and meteorological time series data through LSTM models and federated aggregation technology, providing accurate and reliable power prediction support for subsequent grid connection optimization and energy storage scheduling.

[0090] Step S400, according to the K cluster level power prediction values, the space-time joint optimization of grid connection is carried out, and the energy storage charging and discharging scheduling strategy is output to realize the cross-cluster power complementation of the K photovoltaic clusters.

[0091] Specifically, the power of different clusters has fluctuation characteristics in time and space. An optimization algorithm (such as linear programming, dynamic programming, genetic algorithm, etc.) is used to optimize and schedule the grid-connected power, with the goal of smoothing power fluctuations and improving grid stability. According to the optimization results, the charging and discharging strategy of the energy storage system is formulated. When the power prediction value of a certain cluster is high, the energy storage system can be arranged to charge; when the power prediction value is low, the energy storage system can be arranged to discharge to supplement the power demand. Through the charging and discharging scheduling of the energy storage system, the power complementation between different clusters is realized, further improving the grid performance and economic benefits of the entire distributed photovoltaic power station.

[0092] In one possible implementation, according to the K cluster level power prediction values, the space-time joint optimization of grid connection is carried out, and the energy storage charging and discharging scheduling strategy is output to realize the cross-cluster power complementation of the K photovoltaic clusters, and step S400 further includes step S410, dynamically loading K real-time grid connection processing thresholds according to the K cluster level power prediction values. Specifically, the system adjusts the grid connection processing threshold in real time according to the K cluster level power prediction values (i.e. the predicted power output values of each photovoltaic cluster in the future period obtained by the cluster power prediction model). The grid connection processing threshold is used to judge whether the cluster power meets the requirements of the power grid, and the adjustment of the grid connection processing threshold can be based on the receiving capacity of the power grid, the historical power fluctuation range of the cluster, and the current power load demand and other factors.

[0093] For example, the real-time adjustment of the grid connection processing threshold is as follows: assuming that the power prediction values of 3 photovoltaic clusters are as shown in Table 2.

[0094] Table 2: Cluster level power prediction values

[0095]

[0096] The grid acceptance capacity is no more than 120 kW for each cluster, considering the historical power fluctuation range of the cluster (assuming a fluctuation range of ±20 kW) and the current grid load demand (assuming that the grid wants the cluster power to be stable around the predicted value, but not exceeding the acceptance capacity).

[0097] The grid-connection handling threshold is adjusted as follows: the power prediction value of cluster 1 is 100 kW, and the fluctuation range is ±20 kW, so the adjusted grid-connection handling threshold is [80, 120] kW. The power prediction value of cluster 2 is 150 kW, and the fluctuation range is ±20 kW, but the upper limit of the grid acceptance capacity is 120 kW. Therefore, the lower threshold should be the predicted value minus the fluctuation range, but not exceeding the upper limit of the grid acceptance capacity. The adjusted grid-connection handling threshold should be [120, 120] kW (both the lower and upper limits are the upper limit of the grid acceptance capacity). The power prediction value of cluster 3 is 80 kW, and the fluctuation range is ±20 kW, so the adjusted grid-connection handling threshold is [60, 100] kW.

[0098] Finally, the real-time grid-connection handling thresholds are shown in Table 3.

[0099] Table 3: Real-time grid-connection handling thresholds

[0100]

[0101] In step S420, K cluster-level power actual values of the K photovoltaic clusters are obtained by output power monitoring summation. Specifically, the actual output power of each photovoltaic cluster is monitored in real time by a power meter or a data acquisition system installed at the cluster grid-connection point, and the actual output power of each photovoltaic generating unit is summed up to obtain the actual power output value of each photovoltaic cluster.

[0102] In step S430, according to the deviation of the K cluster-level power actual values and the K real-time grid-connection handling thresholds, a spatio-temporal coupling optimization is performed to output an energy storage charging and discharging scheduling strategy. Specifically, the cluster-level power actual values are compared with the grid-connection handling thresholds to calculate the difference or deviation. Considering the time (such as the power demand change in different time periods) and space (such as the power complementary ability between different clusters) dimensions, the optimal energy storage charging and discharging strategy is determined by an optimization algorithm. According to the optimization result, a specific energy storage charging and discharging scheduling strategy is generated, including charging and discharging time, charging and discharging power, etc.

[0103] At step S440, the cross-cluster power dynamic complementation of the K photovoltaic clusters is performed according to the energy storage charging and discharging scheduling strategy. Specifically, the charging and discharging behaviors of the energy storage system are controlled according to the energy storage charging and discharging scheduling strategy. When the power of a certain photovoltaic cluster is excessive, the excess power is absorbed by the energy storage system; when the power of a certain cluster is insufficient, the stored power is released by the energy storage system to supplement it, so as to realize the power balance among the clusters. This implementation mode can flexibly cope with the fluctuations and changes of the power of the photovoltaic clusters, reduce the impact and instability factors of the power grid, and realize the harmonious coexistence of the photovoltaic clusters and the power grid, through dynamically loading the grid-connected processing threshold, real-time monitoring of the power output, performing the time-space coupling optimization, and performing the cross-cluster power dynamic complementation.

[0104] In a possible implementation, the time-space coupling optimization is performed according to the deviation of the K cluster-level power measured values from the K real-time grid-connected processing thresholds, and the energy storage charging and discharging scheduling strategy is output, and step S430 further includes step S431. According to the deviation direction of the K cluster-level power measured values from the K real-time grid-connected processing thresholds, the K photovoltaic clusters are divided into W power excessive clusters and F power insufficient clusters. Specifically, W is a natural number greater than or equal to 1, F is a natural number greater than or equal to 1, and for each photovoltaic cluster, the cluster-level power measured value is compared with the real-time grid-connected processing threshold. If the measured value is greater than the real-time grid-connected processing threshold, it is determined to be power excessive; if the measured value is less than the real-time grid-connected processing threshold, it is determined to be power insufficient. According to the deviation direction, the K photovoltaic clusters are divided into W power excessive clusters and F power insufficient clusters, and the sum of W and F is equal to K.

[0105] At step S432, the local energy storage fitting is performed on the W power excessive clusters, and W power excessive values are output. Specifically, for each power excessive cluster, the excess power (i.e. the difference between the measured value and the real-time grid-connected processing threshold) is calculated, and the amount of excess power that can be stored is determined. The calculated amount of excess power is output as the power excessive value for subsequent power complementation matching.

[0106] At step S433, after extracting F power insufficient values of the F power insufficient clusters, power complementation matching enumeration is performed according to the W power excessive values and the F power insufficient values, and a plurality of initial charging and discharging scheduling strategies are obtained. Specifically, for each power insufficient cluster, the insufficient power (i.e. the difference between the real-time grid-connected processing threshold and the measured value) is calculated as the power insufficient value. According to the W power excessive values and the F power insufficient values, various possible charging and discharging scheduling schemes are enumerated by a combination algorithm, including the allocation of the excess power of a certain power excessive cluster to one or more power insufficient clusters. All possible charging and discharging scheduling schemes obtained by enumeration are output as initial charging and discharging scheduling strategies.

[0107] At step S434, a plurality of initial charging and discharging scheduling strategies are evaluated with the objective of minimizing the energy storage loss and power transmission cost, and a charging and discharging scheduling strategy of the energy storage is selected. Specifically, the objective of minimizing the energy storage loss and power transmission cost is set. The energy storage loss includes the energy loss in the charging and discharging process, and the power transmission cost includes the cost in the power transmission process. Each initial charging and discharging scheduling strategy is evaluated, and the energy storage loss and power transmission cost thereof are calculated. According to the evaluation result, the charging and discharging scheduling strategy with the minimum energy storage loss and power transmission cost is selected as the final charging and discharging scheduling strategy of the energy storage.

[0108] An example of spatiotemporal coupling optimization is as follows: assuming that there are 3 photovoltaic clusters, and the power measurement values and real-time grid connection processing thresholds are as shown in Table 4.

[0109] Table 4: Power measurement values and real-time grid connection processing thresholds

[0110]

[0111] Cluster 1: measurement value 110kW, grid connection processing threshold [80, 120]kW, in the normal range. Cluster 2: measurement value 130kW, grid connection processing threshold [120, 120]kW, excess output (measurement value > upper limit). Cluster 3: measurement value 50kW, grid connection processing threshold [60, 100]kW, insufficient output (measurement value < lower limit). According to the deviation direction, the 3 photovoltaic clusters are divided into: excess output cluster: cluster 2, insufficient output cluster: cluster 3. Cluster 2: excess output value = measurement value - upper limit threshold = 130kW - 120kW = 10kW, the excess output value of cluster 2 is 10kW. Cluster 3: insufficient output value = lower limit threshold - measurement value = 60kW - 50kW = 10kW (indicating that 10kW needs to be supplemented).

[0112] Enumerate possible charging and discharging scheduling schemes: scheme 1: all 10kW excess power of cluster 2 is allocated to cluster 3. Scheme 2: part of the 10kW excess power of cluster 2 is allocated to cluster 3, and the remaining part is stored in the local energy storage system (for example, 5kW is allocated to cluster 3, and 5kW is stored in the local energy storage system).

[0113] The calculation formula of energy storage loss and power transmission cost: energy storage loss = charging and discharging amount x loss coefficient (assuming that the loss coefficient is 0.1); power transmission cost = power transmission amount x power transmission unit price (assuming that the power transmission unit price is 0.5 yuan / kW).

[0114] Evaluation scheme 1: energy storage loss: 0 (all for power supplement), power transmission cost: 10kWx0.5 yuan / kW=5 yuan, total cost: 5 yuan. Evaluation scheme 2: energy storage loss: 5kWx0.1=0.5kW, power transmission cost: 5kWx0.5 yuan / kW=2.5 yuan, total cost: 0.5kWx0.5 yuan / kW+2.5 yuan=2.75 yuan. The total cost of scheme 2 is 2.75 yuan, which is less than 5 yuan of scheme 1, so scheme 2 is selected as the final energy storage charging and discharging scheduling strategy. The final scheduling strategy is: 5kW of surplus power of cluster 2 is allocated to cluster 3, and the remaining 5kW of cluster 2 is stored in the local energy storage system. This implementation can flexibly cope with the fluctuation and change of the power of the photovoltaic cluster by judging the deviation direction of the cluster power, classifying the power surplus and power shortage clusters, enumerating power complementary matching, and evaluating and screening the optimal energy storage charging and discharging scheduling strategy, improving the utilization rate of renewable energy, optimizing the operation of the power grid, and ensuring the stability and reliability of power supply.

[0115] Step S500, after the single station fitting conversion according to the K cluster level power prediction value is output, the distributed power prediction value is used for output deviation identification, and the fault node of the photovoltaic power station is located.

[0116] Specifically, the cluster level power prediction value is distributed to each photovoltaic power station by using weighted average method or proportional distribution method. If the weighted average method is used, the cluster level power prediction value is distributed to each power station according to the historical power proportion of each photovoltaic power station in the cluster. The weight can be calculated based on historical power generation, geographical location, equipment capacity and other factors. If the proportional distribution method is used, the cluster level power prediction value is distributed to each power station according to the equipment capacity proportion of each photovoltaic power station. Using threshold comparison method, for each photovoltaic power station, the difference (deviation) between its actual power and predicted power is calculated. A reasonable deviation threshold (such as 10% or 20%) is set. If the deviation exceeds the deviation threshold, it is considered that the power station has an abnormality. Combined with geographical location information and historical data, the power station whose deviation exceeds the deviation threshold is further analyzed to determine the fault node, including checking the environmental factors (such as whether there are obstructions, whether it is in the shadow area, etc.) around the fault power station, comparing the historical power data of the power station, and judging whether it is a sudden failure or a long-term problem.

[0117] For example, assume there is a photovoltaic cluster containing 3 photovoltaic power stations (A, B, C), and the cluster-level power prediction value is 300kW. Using the weighted average method, assume that the historical power proportions of A, B, and C power stations are 30%, 40%, and 30% respectively. The allocated power prediction values are: A power station: 300kW x 30% = 90kW; B power station: 300kW x 40% = 120kW; C power station: 300kW x 30% = 90kW. Assume the actual power measurement values are: A power station: 80kW; B power station: 125kW; C power station: 88kW. Calculate the deviation: A power station: |80kW - 90kW| = 10kW; B power station: |125kW - 120kW| = 5kW; C power station: |88kW - 90kW| = 2kW. Assume the deviation threshold is 10kW, and the deviation of A power station is greater than or equal to the deviation threshold, which is identified as an anomaly. Check the geographical location information of A power station, and find that there are newly planted trees around it, which may block part of the solar panels.

[0118] Step S600, dynamically update the energy storage charging and discharging scheduling strategy according to the fault node of the photovoltaic power station.

[0119] Specifically, according to the power deviation of the fault node, adjust the charging and discharging plan of the energy storage system. If the power of the fault power station is insufficient, the energy storage system increases the discharge amount to compensate for the power gap. Through the power meter and the data acquisition system, the actual power of each photovoltaic power station is monitored in real time. According to the real-time data, the charging and discharging plan of the energy storage system is dynamically updated to ensure the stable operation of the power grid.

[0120] For example, assume that A power station is identified as a fault node with insufficient power, with an actual power of 80kW and a predicted power of 90kW, and a deviation of 10kW. The energy storage system increases the discharge amount by 10kW to compensate for the power gap of A power station. The energy storage system dynamically adjusts the discharge amount according to real-time monitoring data to ensure the stable operation of the power grid. An example of dynamic updating of the energy storage charging and discharging scheduling strategy is shown in Table 5.

[0121] Table 5: Example of dynamic updating of energy storage charging and discharging scheduling strategy

[0122]

[0123] The embodiment of the application adopts multi-source data local calling and time sequence alignment preprocessing, constructs a multi-dimensional feature correlation table, extracts information from the table to analyze similarity, obtains K photovoltaic clusters through spectral clustering of the photovoltaic power generation unit, inputs meteorological time sequence data into a cluster power prediction model, obtains K cluster-level power prediction values through cluster power prediction, performs grid space joint optimization according to the cluster-level power prediction values, outputs energy storage charging and discharging scheduling strategies to realize cross-cluster power complementation, obtains distributed power prediction values through single-station fitting and conversion according to the cluster-level power prediction values, identifies power deviation positioning fault nodes, and dynamically updates energy storage charging and discharging scheduling strategies, and the like technical means, and the following technical effects are achieved:

[0124] 1. Through multi-source data local calling and time sequence alignment preprocessing, similarity analysis and spectral clustering are performed in combination with geographic location information and historical output data to construct an accurate cluster power prediction model, thereby solving the problem that distributed photovoltaic power cannot be accurately predicted in the prior art, and the technical effect of improving photovoltaic power prediction accuracy is achieved.

[0125] 2. Grid space joint optimization is performed according to the cluster-level power prediction values, and energy storage charging and discharging scheduling strategies are output to realize cross-cluster power dynamic complementation, thereby solving the problem that cross-cluster power optimization scheduling is difficult to achieve in the prior art, and the technical effect of improving power grid operation stability and economic benefits is achieved.

[0126] 3. Through dynamic monitoring of power output and real-time adjustment of energy storage charging and discharging strategies, power deviation is quickly responded to, power gaps are compensated or excess power is absorbed, thereby solving the problem that power grid operation stability is poor in the prior art, and the technical effect of reducing power grid impact and ensuring stable operation of the power grid is achieved.

[0127] 4. Single-station fitting and conversion and output deviation identification are adopted in combination with geographic location information and historical data analysis to accurately locate power abnormal photovoltaic power stations, thereby solving the problem that fault nodes cannot be effectively identified and located in the prior art, and the technical effect of quickly discovering and processing faults and reducing maintenance costs is achieved.

[0128] In the foregoing, a clustering distributed photovoltaic power prediction method according to an embodiment of the application is described in detail with reference to Figure 1 A clustering distributed photovoltaic power prediction system according to an embodiment of the application is described below with reference to Figure 2

[0129] ​According to an embodiment of the present application, a joint optimization clustering distributed photovoltaic power prediction system is used to solve the technical problems of low prediction accuracy, difficulty in realizing cross-cluster power optimization scheduling, and poor power grid operation stability of existing distributed photovoltaic power prediction. The system achieves the technical effects of improving the accuracy of photovoltaic power prediction, realizing cross-cluster power optimization scheduling, and improving the stability of power grid operation. The joint optimization clustering distributed photovoltaic power prediction system includes a time sequence alignment module 10, a power generation unit spectrum clustering module 20, a cluster power prediction module 30, an energy storage charging and discharging scheduling strategy output module 40, an output deviation identification module 50, and an energy storage charging and discharging scheduling strategy updating module 60.

[0130] The time sequence alignment module 10 is used to locally call multi-source data of a distributed photovoltaic power station and preprocess the calling results through time sequence alignment to construct a multi-dimensional feature correlation number table. The power generation unit spectrum clustering module 20 is used to perform similarity analysis on a plurality of geographical location information and a plurality of historical output data of a plurality of photovoltaic power generation units in the distributed photovoltaic power station extracted from the multi-dimensional feature correlation number table, and perform power generation unit spectrum clustering on the plurality of photovoltaic power generation units according to the analysis results, output K photovoltaic clusters, and K is a natural number greater than or equal to 1. The cluster power prediction module 30 is used to input meteorological time series data collected according to a power prediction scale into K cluster power prediction models of the K photovoltaic clusters to perform cluster power prediction and output K cluster-level power prediction values. The energy storage charging and discharging scheduling strategy output module 40 is used to perform grid-connected space-time joint optimization according to the K cluster-level power prediction values, output an energy storage charging and discharging scheduling strategy, and perform cross-cluster power complementation of the K photovoltaic clusters. The output deviation identification module 50 is used to perform single-station fitting conversion according to the K cluster-level power prediction values to output a distributed power prediction value, perform output deviation identification according to the distributed power prediction value, and locate a photovoltaic power station fault node. The energy storage charging and discharging scheduling strategy updating module 60 is used to dynamically update the energy storage charging and discharging scheduling strategy according to the photovoltaic power station fault node.

[0131] In the following, the specific configuration of the power generation unit spectrum clustering module 20 will be described in detail. As described above, the similarity analysis is performed on the plurality of geographical location information and the plurality of historical output data of the plurality of photovoltaic power generation units in the distributed photovoltaic power station extracted from the multi-dimensional feature correlation number table, and the power generation unit spectrum clustering is performed on the plurality of photovoltaic power generation units according to the analysis result, and K photovoltaic clusters are output. The power generation unit spectrum clustering module 20 can further include: a unit output curve drawing unit for drawing a plurality of unit output curves according to the plurality of historical output data; an output curve similarity matrix output unit for outputting an output curve similarity matrix of the plurality of photovoltaic power generation units by dynamic time warping algorithm analysis; a spatial correlation matrix construction unit for constructing a spatial correlation matrix according to the plurality of geographical location information; and a power generation unit spectrum clustering unit for performing power generation unit spectrum clustering by fusing and analyzing the output curve similarity matrix and the spatial correlation matrix, and clustering the plurality of photovoltaic power generation units into the K photovoltaic clusters.

[0132] In the following, the specific configuration of the energy storage charging and discharging scheduling strategy output module 40 will be described in detail. As described above, the grid-connected space-time joint optimization is performed according to the K cluster-level power prediction values, and the energy storage charging and discharging scheduling strategy is output, and the cross-cluster power complementation of the K photovoltaic clusters is performed. The energy storage charging and discharging scheduling strategy output module 40 can further include: a real-time grid-connected processing threshold loading unit for dynamically loading K real-time grid-connected processing thresholds according to the K cluster-level power prediction values; a cluster-level power actual value acquisition unit for obtaining K cluster-level power actual values of the K photovoltaic clusters by output power monitoring addition; a space-time coupling optimization unit for performing space-time coupling optimization according to the deviation of the K cluster-level power actual values and the K real-time grid-connected processing thresholds, and outputting the energy storage charging and discharging scheduling strategy; and a cross-cluster power dynamic complementation unit for performing cross-cluster power dynamic complementation of the K photovoltaic clusters according to the energy storage charging and discharging scheduling strategy.

[0133] The power generation unit spectrum clustering is performed by fusing the output curve similarity matrix and the spatial correlation matrix, the plurality of photovoltaic power generation units are clustered into the K photovoltaic clusters, and the power generation unit spectrum clustering unit can further include: a first normalized similarity matrix generating sub-unit configured to perform exponential decay mapping on the output curve similarity matrix to generate a first normalized similarity matrix; a second normalized similarity matrix generating sub-unit configured to normalize the spatial correlation matrix according to a matrix scale of the first normalized similarity matrix to obtain a second normalized similarity matrix; a similarity matrix superimposing sub-unit configured to superimpose the first normalized similarity matrix and the second normalized similarity matrix based on a preset matrix weight coefficient to obtain a fused similarity matrix; a normalized Laplacian matrix constructing sub-unit configured to construct a normalized Laplacian matrix based on the fused similarity matrix; an eigenvalue decomposition sub-unit configured to perform eigenvalue decomposition on the normalized Laplacian matrix to select K low-dimensional eigenvectors corresponding to K smallest eigenvalues; a mapping sub-unit configured to map the plurality of photovoltaic power generation units to a low-dimensional feature space after the low-dimensional feature space is constructed based on the K low-dimensional eigenvectors; and a clustering center point iterative updating sub-unit configured to output the K photovoltaic clusters by iteratively updating clustering center points in the low-dimensional feature space.

[0134] The K photovoltaic clusters are output by iteratively updating the clustering center points in the low-dimensional feature space, and the clustering center point iterative updating sub-unit can further include: a clustering center point initialization component configured to preset a center point distance scale and initialize K clustering center points in the low-dimensional feature space based on the center point distance scale; a photovoltaic power generation unit distribution component configured to distribute the plurality of photovoltaic power generation units in the low-dimensional feature space to the K clustering center points based on the nearest Euclidean distance to obtain K sets of power generation unit distribution results; a random disturbance iteration component configured to perform random disturbance iteration on the K clustering center points to obtain K first iteration center points; the photovoltaic power generation unit distribution component is configured to distribute the plurality of photovoltaic power generation units in the low-dimensional feature space to the K first iteration center points based on the nearest Euclidean distance to obtain K sets of first iteration distribution results; a center point iteration retention judgment component configured to compare distance sums of the K sets of power generation unit distribution results and the K sets of first iteration distribution results, and perform iteration retention judgment on the K clustering center points and the K first iteration center points according to a comparison result; and a photovoltaic cluster output component configured to output the K photovoltaic clusters when iteration fluctuations of the distance sums are less than a preset deviation.

[0135] The time-space coupling optimization is performed according to the deviation of the K cluster-level power measured values and the K real-time grid-connected processing thresholds, and the energy storage charging and discharging scheduling strategy is output. The time-space coupling optimization unit can further include: a photovoltaic cluster division subunit configured to divide the K photovoltaic clusters into W excess output clusters and F insufficient output clusters according to the deviation direction of the K cluster-level power measured values and the K real-time grid-connected processing thresholds, W being a natural number greater than or equal to 1, and F being a natural number greater than or equal to 1; a local energy storage fitting subunit configured to fit the W excess output clusters locally to output W excess output values; a power complementary matching enumeration subunit configured to, after extracting F insufficient output values of the F insufficient output clusters, perform power complementary matching enumeration according to the W excess output values and the F insufficient output values to obtain a plurality of initial charging and discharging scheduling strategies; and an energy storage charging and discharging scheduling strategy acquisition subunit configured to evaluate the plurality of initial charging and discharging scheduling strategies with the objective of minimizing energy storage loss and power transmission cost, and screen to obtain the energy storage charging and discharging scheduling strategy.

[0136] Before the cluster-type power prediction module 30 outputs the K cluster-level power prediction values according to the input meteorological time series data collected according to the power prediction scale into the K cluster power prediction models of the K photovoltaic clusters to perform cluster-type power prediction, the cluster-type power prediction module 30 can further include: a standard power prediction model construction unit configured to construct a standard power prediction model using LSTM, wherein the input data of the power prediction scale is meteorological time series data and output time series data conforming to the power prediction scale, and the output result is predicted power time series data conforming to a preset scale factor; a data calling unit configured to perform time series alignment data calling according to the plurality of historical output data to obtain a plurality of meteorological data records and a plurality of power generation records; a standard power prediction model localization unit configured to use the plurality of historical output data, the plurality of meteorological data records, and the plurality of power generation records as training data to localize the standard power prediction model to obtain a plurality of initial power prediction sub-models of the plurality of photovoltaic power generation units; a parameter federated aggregation unit configured to aggregate model parameters of the plurality of initial power prediction sub-models according to the K photovoltaic clusters to obtain K federated power prediction sub-models; a sub-model replication unit configured to replicate the K federated power prediction sub-models according to the number of clusters of the K photovoltaic clusters to obtain K sets of federated power prediction sub-models; and a cluster power prediction model construction unit configured to complete construction of the K cluster power prediction models by connecting the K sets of federated power prediction sub-models in parallel within a group and configuring a power summation engine at an output end.

[0137] The space correlation matrix is constructed according to the plurality of geographic position information, and the space correlation matrix construction unit can further include: a photovoltaic power generation unit combination enumeration subunit configured to enumerate the plurality of photovoltaic power generation units by a combination algorithm to obtain H sets of photovoltaic power generation units, H being a natural number greater than or equal to 1; a geographic position information acquisition subunit configured to combine the plurality of geographic position information according to the H sets of photovoltaic power generation units to obtain H sets of geographic position information, wherein the geographic position information includes photovoltaic unit longitude and latitude, photovoltaic unit altitude, photovoltaic panel inclination angle and photovoltaic panel azimuth angle; an inter-group Euclidean distance calculation subunit configured to, after processing the H sets of geographic position information based on a preset geographic normalization weight, perform inter-group Euclidean distance calculation on the processing result to output H space correlation coefficients; a reference correlation matrix construction subunit configured to construct a reference correlation matrix according to the plurality of photovoltaic power generation units; and a correlation coefficient filling subunit configured to fill the H space correlation coefficients into the reference correlation matrix with the H sets of photovoltaic power generation units as a filling guide, to complete construction of the space correlation matrix.

[0138] In the following, a specific configuration of the time sequence alignment module 10 will be described in detail. As described above, the multi-source data of the distributed photovoltaic power station is locally called, and the multi-dimensional feature correlation number table is constructed by performing time sequence alignment preprocessing on the calling result. The time sequence alignment module 10 can further include: a local data calling unit configured to perform local data calling through a plurality of edge collection terminals arranged at the plurality of photovoltaic power generation units to obtain a plurality of meteorological data records, a plurality of output data records and the plurality of geographic position information; a sliding mean calculation unit configured to preset an output sliding window and perform sliding mean calculation on the plurality of output data records according to the output sliding window to obtain a plurality of down-sampled output data; a linear interpolation processing unit configured to preset a meteorological interpolation granularity and perform linear interpolation processing on the plurality of meteorological data records by using the meteorological interpolation granularity to obtain a plurality of up-sampled meteorological data; and a correlation storage unit configured to, after performing time sequence alignment on the plurality of down-sampled output data and the plurality of up-sampled meteorological data, store the plurality of down-sampled output data, the plurality of up-sampled meteorological data and the plurality of geographic position information based on the plurality of photovoltaic power generation units to complete construction of the multi-dimensional feature correlation number table.

[0139] The joint optimization clustering distributed photovoltaic power prediction system provided in the embodiments of the present application can execute the joint optimization clustering distributed photovoltaic power prediction method provided in any embodiment of the present application, and has the corresponding function modules and beneficial effects of the execution method.

[0140] Although the present application makes various references to certain modules in the system according to the embodiments of the present application, however, any number of different modules can be used and run on the user terminal and / or the server, the various units and modules included are only divided according to the functional logic, but are not limited to the above division, as long as the corresponding functions can be realized; in addition, the specific name of each functional unit is only for the convenience of mutual differentiation, and does not serve to limit the protection scope of the present application.

[0141] The above detailed description does not constitute a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications, combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present application shall be included in the protection scope of the present application. In some cases, the actions or steps described in the present application can be performed in an order different from that in the embodiments and still achieve the desired results. In addition, the processes depicted in the drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multi-task processing and parallel processing are possible or can be advantageous.

Claims

1. A jointly optimized clustering-based distributed photovoltaic power prediction method, characterized in that, The method includes: Multi-source data from distributed photovoltaic power stations are accessed locally, and the access results are preprocessed with time-series alignment to construct a multi-dimensional feature association table. The multidimensional feature association table extracts multiple geographical location information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station for similarity analysis. Based on the analysis results, the multiple photovoltaic power generation units are clustered by power generation unit spectrum to output K photovoltaic clusters, where K is a natural number greater than or equal to 1. Meteorological time-series data collected according to the power prediction scale are input into the K cluster power prediction models of the K photovoltaic clusters to perform cluster-based power prediction and output K cluster-level power prediction values. Based on the K cluster-level power prediction values, grid-connected spatiotemporal joint optimization is performed, and an energy storage charging and discharging scheduling strategy is output to achieve cross-cluster power complementarity among the K photovoltaic clusters; After performing single-station fitting and conversion based on the K cluster-level power prediction values ​​and outputting distributed power prediction values, output deviation identification is performed based on the distributed power prediction values ​​to locate the fault nodes of the photovoltaic power station. The energy storage charging and discharging scheduling strategy is dynamically updated based on the fault nodes of the photovoltaic power station. Before inputting the meteorological time-series data collected according to the power prediction scale into the K cluster power prediction models of the K photovoltaic clusters to perform cluster-based power prediction and output the K cluster-level power prediction values, the method further includes: A standard power prediction model is constructed using LSTM, wherein the input data for the power prediction scale are meteorological time series data and power output time series data that conform to the power prediction scale, and the output result is predicted power time series data that conforms to the power prediction scale with a preset scaling factor. Based on the multiple historical power output data, time-series aligned data calls are performed to obtain multiple meteorological data records and multiple power generation records; The standard power prediction model is localized using the multiple historical power output data, multiple meteorological data records, and multiple power generation records as training data to obtain multiple initial power prediction sub-models for the multiple photovoltaic power generation units. Based on the K photovoltaic clusters, the model parameters of the multiple initial power prediction sub-models are federated and aggregated to obtain K federated power prediction sub-models; Based on the number of the K photovoltaic clusters, the K federated power prediction sub-models are replicated to obtain K sets of federated power prediction sub-models; The K cluster power prediction models are constructed by connecting the K groups of federated power prediction sub-models in parallel within the group and configuring a power summing engine at the output end.

2. The jointly optimized clustered distributed photovoltaic power prediction method as described in claim 1, characterized in that, The multidimensional feature association table is used to extract multiple geographical location information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station for similarity analysis. Based on the analysis results, the multiple photovoltaic power generation units are subjected to power generation unit spectrum clustering, outputting K photovoltaic clusters, including: Based on the aforementioned historical output data, multiple unit output curves were plotted. The power output curves of the multiple units are analyzed using a dynamic time warping algorithm to output a similarity matrix of the power output curves of the multiple photovoltaic power generation units. Construct a spatial correlation matrix based on the multiple geographic location information; By fusing and analyzing the similarity matrix of the output curves and the spatial correlation matrix, the power generation units are clustered into the K photovoltaic clusters.

3. The jointly optimized clustered distributed photovoltaic power prediction method as described in claim 1, characterized in that, Based on the K cluster-level power prediction values, grid-connected spatiotemporal joint optimization is performed to output an energy storage charging and discharging scheduling strategy, and cross-cluster power complementarity is achieved among the K photovoltaic clusters, including: Based on the K cluster-level power prediction values, K real-time grid connection processing thresholds are dynamically loaded; By summing the output power monitoring data, the measured power values ​​of the K photovoltaic clusters are obtained. Based on the deviation between the K cluster-level power measured values ​​and the K real-time grid-connected processing thresholds, spatiotemporal coupling optimization is performed to output the energy storage charging and discharging scheduling strategy. Dynamic power complementarity across the K photovoltaic clusters is performed according to the energy storage charging and discharging scheduling strategy.

4. The jointly optimized clustered distributed photovoltaic power prediction method as described in claim 2, characterized in that, By fusing the similarity matrix of the output curves and the spatial correlation matrix, the power generation units are clustered into K photovoltaic clusters, including: An exponential decay mapping is performed on the power output curve similarity matrix to generate a first normalized similarity matrix; Based on the matrix scale of the first normalized similarity matrix, the spatial correlation matrix is ​​normalized to obtain the second normalized similarity matrix; A fused similarity matrix is ​​obtained by superimposing the first normalized similarity matrix and the second normalized similarity matrix based on the preset matrix weight coefficients; A normalized Laplacian matrix is ​​constructed based on the fused similarity matrix; The normalized Laplacian matrix is ​​subjected to eigenvalue decomposition to select K low-dimensional eigenvectors corresponding to K smallest eigenvalues; After constructing a low-dimensional feature space based on the K low-dimensional feature vectors, the multiple photovoltaic power generation units are mapped to the low-dimensional feature space. The K photovoltaic clusters are output by iteratively updating the cluster center points in the low-dimensional feature space.

5. The jointly optimized clustered distributed photovoltaic power prediction method as described in claim 4, characterized in that, The K photovoltaic clusters are output by iteratively updating the cluster centers in the low-dimensional feature space, including: A preset center point spacing scale is used as a reference to initialize K cluster center points in the low-dimensional feature space. Based on the nearest Euclidean distance, the multiple photovoltaic power generation units in the low-dimensional feature space are assigned to the K cluster centers to obtain the K groups of power generation unit allocation results; Randomly perturb and iterate the K cluster centers to obtain K first iteration centers; Based on the nearest Euclidean distance, the multiple photovoltaic power generation units in the low-dimensional feature space are assigned to the K first iteration center points to obtain K sets of first iteration assignment results; Calculate the sum of squared distances between the K groups of power generation unit allocation results and the K groups of first iteration allocation results, and make iterative retention judgments on the K cluster center points and the K first iteration center points based on the calculation results; When the iterative fluctuation of the sum of squared distances is less than a preset deviation, the K photovoltaic clusters are output.

6. The jointly optimized clustered distributed photovoltaic power prediction method as described in claim 3, characterized in that, Based on the deviations between the K cluster-level measured power values ​​and the K real-time grid-connected processing thresholds, spatiotemporal coupling optimization is performed to output the energy storage charging and discharging scheduling strategy, including: Based on the deviation direction between the K cluster-level measured power values ​​and the K real-time grid connection processing thresholds, the K photovoltaic clusters are divided into W clusters with excess output and F clusters with insufficient output, where W is a natural number greater than or equal to 1 and F is a natural number greater than or equal to 1. Perform local energy storage fitting on the W excess power clusters and output W excess power values; After extracting the F insufficient output values ​​from the F insufficient output clusters, the W excess output values ​​and the F insufficient output values ​​are processed to obtain multiple initial charge and discharge scheduling strategies. With the goal of minimizing energy storage losses and transmission costs, the multiple initial charge and discharge scheduling strategies are evaluated, and the energy storage charge and discharge scheduling strategy is selected.

7. The jointly optimized clustered distributed photovoltaic power prediction method as described in claim 2, characterized in that, A spatial correlation matrix is ​​constructed based on the multiple geographic location information, including: The multiple photovoltaic power generation units are enumerated by a combination algorithm to obtain H groups of photovoltaic power generation units, where H is a natural number greater than or equal to 1; Based on the combination of multiple geographical location information of the H group of photovoltaic power generation units, H group of geographical location information is obtained; wherein, the geographical location information includes the latitude and longitude of the photovoltaic unit, the altitude of the photovoltaic unit, the tilt angle of the photovoltaic panel, and the azimuth angle of the photovoltaic panel; After processing the H groups of geographic location information based on preset geographic normalization weights, the Euclidean distance between the groups is calculated on the processing results, and H spatial correlation coefficients are output. A benchmark correlation matrix is ​​constructed based on the aforementioned multiple photovoltaic power generation units; Using the H groups of photovoltaic power generation units as filling guides, the H spatial correlation coefficients are filled into the benchmark correlation matrix to complete the construction of the spatial correlation matrix.

8. The jointly optimized clustered distributed photovoltaic power prediction method as described in claim 1, characterized in that, Multi-source data from distributed photovoltaic power stations is accessed locally, and the access results are preprocessed with time-series alignment to construct a multi-dimensional feature association table, including: By accessing local data through multiple edge acquisition terminals deployed in the multiple photovoltaic power generation units, multiple meteorological data records, multiple power output data records, and multiple geographical location information are obtained; A preset output sliding window is used, and the sliding average of the multiple output data records is calculated according to the output sliding window to obtain multiple downsampled output data. A preset meteorological interpolation granularity is used, and the multiple meteorological data records are linearly interpolated using the preset meteorological interpolation granularity to obtain multiple upsampled meteorological data. After performing time-series alignment on the multiple downsampled power output data and multiple upsampled meteorological data, the multiple downsampled power output data, multiple upsampled meteorological data and multiple geographical location information are stored in association based on the multiple photovoltaic power generation units, thus completing the construction of the multidimensional feature association table.

9. A jointly optimized clustered distributed photovoltaic power prediction system, characterized in that, The system includes: The time-series alignment module is used to perform local calls to multi-source data from distributed photovoltaic power stations and to construct a multi-dimensional feature association table by performing time-series alignment preprocessing on the call results. The power generation unit spectrum clustering module is used to extract multiple geographical location information and multiple historical output data of multiple photovoltaic power generation units in the distributed photovoltaic power station from the multidimensional feature association table, perform similarity analysis, and perform power generation unit spectrum clustering on the multiple photovoltaic power generation units according to the analysis results, outputting K photovoltaic clusters, where K is a natural number greater than or equal to 1. The cluster power prediction module is used to input meteorological time-series data collected according to the power prediction scale into the K cluster power prediction models of the K photovoltaic clusters, perform cluster power prediction, and output K cluster-level power prediction values. The energy storage charging and discharging scheduling strategy output module is used to perform grid-connected spatiotemporal joint optimization based on the K cluster-level power prediction values, output the energy storage charging and discharging scheduling strategy, and perform cross-cluster power complementarity of the K photovoltaic clusters. The output deviation identification module is used to identify the output deviation and locate the fault node of the photovoltaic power station after performing single-station fitting and conversion based on the K cluster-level power prediction values ​​and outputting distributed power prediction values. An energy storage charging and discharging scheduling strategy update module is used to dynamically update the energy storage charging and discharging scheduling strategy according to the fault nodes of the photovoltaic power station. Before inputting the meteorological time-series data collected according to the power prediction scale into the K cluster power prediction models of the K photovoltaic clusters to perform cluster-based power prediction and output the K cluster-level power prediction values, the system also includes: The standard power prediction model building unit is used to build a standard power prediction model using LSTM. The input data of the power prediction scale are meteorological time series data and power output time series data that conform to the power prediction scale, and the output result is predicted power time series data that conforms to the power prediction scale with a preset scaling factor. Based on the multiple historical power output data, time-series aligned data calls are performed to obtain multiple meteorological data records and multiple power generation records; The standard power prediction model is localized using the multiple historical power output data, multiple meteorological data records, and multiple power generation records as training data to obtain multiple initial power prediction sub-models for the multiple photovoltaic power generation units. Based on the K photovoltaic clusters, the model parameters of the multiple initial power prediction sub-models are federated and aggregated to obtain K federated power prediction sub-models; Based on the number of the K photovoltaic clusters, the K federated power prediction sub-models are replicated to obtain K sets of federated power prediction sub-models; The K cluster power prediction models are constructed by connecting the K groups of federated power prediction sub-models in parallel within the group and configuring a power summing engine at the output end.

Images