A method for analyzing the combined power characteristics of high-proportion distributed photovoltaic power plants

Abstract

This invention discloses a method for analyzing the joint power characteristics of high-proportion distributed photovoltaic (PV) power plants, belonging to the field of PV power analysis technology. The method includes: collecting historical operating data, historical meteorological data, and grid topology data of each distributed PV power plant within a target area; constructing a spatiotemporal graph structure; constructing a spatiotemporal graph neural network model to obtain a joint power characteristic twin model; inputting simulated input data into the joint power characteristic twin model to perform scenario extrapolation of joint power, and outputting the time series, fluctuation characteristics, and spatial distribution characteristics of joint power in the target area. This invention solves the technical problems in existing PV power characteristic analysis, such as the difficulty in scenario extrapolation and the lack of effective solutions to support scheduling decisions, leading to low analysis accuracy and difficulty in scheduling optimization. It achieves scenario extrapolation of PV joint power, resulting in accurate analysis of the joint power characteristics of PV power plants and optimized scheduling decisions.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power analysis technology, and specifically to a method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power plant. Background Technology

[0002] As the penetration rate of distributed photovoltaic power stations in the power system continues to increase, their output is significantly affected by factors such as meteorological conditions and grid topology, exhibiting significant fluctuations and uncertainties. Traditional power analysis methods often rely on a single data dimension or simplified models, making it difficult to accurately characterize the electrical, geographical, and meteorological correlations between multiple power stations. This results in large deviations in joint power time series predictions, insufficient analysis of fluctuation characteristics, and incomplete characterization of spatial distribution features. Furthermore, they lack constraints based on physical laws and the ability to adapt to different scenarios, thus failing to provide accurate and reliable power characteristic data support for grid dispatch.

[0003] Existing technologies have limitations in analyzing the joint power characteristics of photovoltaic power plants, making it difficult to perform scenario simulations and lacking effective solutions to support scheduling decisions. This results in low analysis accuracy and difficulties in scheduling optimization. Summary of the Invention

[0004] This application provides a method for analyzing the joint power characteristics of high-proportion distributed photovoltaic power plants, which addresses the technical problems in existing photovoltaic power plant joint power characteristic analysis, such as difficulty in scenario extrapolation and lack of effective solutions to support scheduling decisions, resulting in low analysis accuracy and difficulty in scheduling optimization.

[0005] In view of the above problems, this application provides a method for analyzing the combined power characteristics of high-proportion distributed photovoltaic power plants.

[0006] This application provides a method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station, the method comprising:

[0007] Historical operational data, historical meteorological data, and power grid topology data of each distributed photovoltaic power station within the target area are collected. Based on the power grid topology data, a spatiotemporal graph structure is constructed with each distributed photovoltaic power station as a node and the spatial relationship between stations as edges. Based on the spatiotemporal graph structure, a spatiotemporal graph neural network model is constructed, and physical constraints are introduced. The model is trained using the historical operational data and the historical meteorological data to obtain a joint power characteristic twin model. Simulated input data is input into the joint power characteristic twin model to perform scenario simulation of joint power, and the joint power time series, fluctuation characteristics, and spatial distribution characteristics of the target area are output.

[0008] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0009] Historical operational data, historical meteorological data, and grid topology data of each distributed photovoltaic (PV) power station within the target area are collected. A spatiotemporal graph structure is constructed using each PV power station as a node and the spatial relationships between stations as edges. A spatiotemporal graph neural network model is built, and physical constraints are introduced. The model is trained using the historical operational data and historical meteorological data to obtain a joint power characteristic twin model. Simulated input data is input into the joint power characteristic twin model to perform joint power scenario simulation, outputting the joint power time series, fluctuation characteristics, and spatial distribution characteristics of the target area. This achieves PV joint power scenario simulation, enabling precise analysis of the joint power characteristics of PV power stations and optimization of scheduling decisions. Attached Figure Description

[0010] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0011] Figure 1 This is a schematic flowchart of a method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station, provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of the process for constructing a spatiotemporal graph structure in a method for analyzing the joint power characteristics of a high-proportion distributed photovoltaic power station provided in an embodiment of this application. Detailed Implementation

[0013] This application provides a method for analyzing the joint power characteristics of high-proportion distributed photovoltaic power plants, which addresses the technical problems in existing photovoltaic power plant joint power characteristic analysis, such as difficulty in scenario extrapolation and lack of effective solutions to support scheduling decisions, resulting in low analysis accuracy and difficulty in scheduling optimization.

[0014] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0015] Examples, such as Figure 1 As shown, this application provides a method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station, the method comprising:

[0016] Step S100: Collect historical operating data, historical meteorological data, and power grid topology data of each distributed photovoltaic power station in the target area.

[0017] Specifically, firstly, the scope and objects of data collection must be clearly defined, focusing on the target area to be analyzed and covering all distributed photovoltaic power stations within that area to ensure comprehensive data collection. For historical operational data, the focus should be on collecting actual operational parameters of each distributed photovoltaic power station at different time periods, such as daily, monthly, and seasonal periods. This includes, but is not limited to, real-time output power, cumulative power generation, inverter operating status (e.g., operating efficiency, start-up and shutdown records, fault alarm information), and photovoltaic module operating temperature. These data must accurately correspond to specific time points to reflect the temporal changes in the power station's operational status. For historical meteorological data, on the one hand, overall meteorological records of the target area should be collected, covering macro-meteorological parameters such as solar intensity, solar irradiance per unit time, ambient temperature, relative humidity, wind speed, wind direction, and cloud cover. On the other hand, for each distributed photovoltaic power station, supplementary micro-meteorological data at the station level should be collected to ensure accurate matching of meteorological data with the power station location and consistency with historical operational data in terms of time dimension, in order to establish the correlation between meteorological conditions and power station power output. In acquiring power grid topology data, the main focus is on obtaining structural information about the power grid within the target area. This includes the location of each distributed photovoltaic power station's connection point to the grid, the distribution and capacity of substations, the type, length, and impedance parameters of transmission lines, the division of power supply areas to which each power station belongs, and electrical connection paths between stations, such as whether there are direct electrical pathways. This data will provide fundamental support for subsequently constructing spatial associations and electrical relationships between power stations. Through the systematic acquisition and processing of these three types of data, a complete raw dataset is formed.

[0018] Step S200: Based on the power grid topology data, construct a spatiotemporal graph structure with each distributed photovoltaic power station as a node and the spatial relationship between the stations as edges.

[0019] Specifically, feature vectors are constructed for each distributed photovoltaic power station within the target area. Each feature vector covers at least the core attribute parameters of the corresponding power station, such as installed capacity, historical average unit capacity power, and geographical coordinates. The feature vectors of all power stations are then stacked in a preset order to form an initial feature matrix representing the basic information of the nodes in the representation graph structure, thus clarifying the inherent attributes of each node. Next, based on the power grid topology data and the spatial relationships between stations, including electrical distance, geographical distance, and meteorological correlation, the effective neighborhoods of each node are screened and the edge weights are calculated: For any node, the K closest nodes are first selected as candidate neighborhoods based on geographical distance. Then, by querying the power grid topology data, nodes without direct electrical paths to the node or located in different power supply areas are removed, forming a neighborhood set that retains only effective associations. Subsequently, within the effective neighborhoods, three types of weights are calculated for each pair of nodes: 1. Electrical distance weight, which is inversely proportional to the electrical coupling impedance magnitude of the two nodes in the power grid topology; 2. Geographical distance weight, which is inversely proportional to the actual geographical distance between the two nodes; 3. Meteorological correlation weight, which is the Pearson correlation coefficient of the historical power sequences of the two nodes. After normalizing the three types of weights, they are weighted and summed according to a preset ratio to obtain the association strength between nodes. The weight of the "edge" in the graph structure is defined by this association strength, generating a weighted adjacency matrix that represents the spatial association between nodes. Finally, by combining the initial feature matrix and the weighted adjacency matrix, a spatiotemporal graph structure that combines node attributes, spatial correlations between nodes, and temporal characteristics is constructed. This structure can accurately reflect the inherent characteristics of each distributed photovoltaic power station and their electrical, geographical, and meteorological correlations.

[0020] Step S300: Based on the spatiotemporal graph structure, construct a spatiotemporal graph neural network model, and introduce physical constraint terms to train the model using the historical operating data and the historical meteorological data to obtain a joint power characteristic twin model.

[0021] Specifically, based on the constructed spatiotemporal graph structure, a spatiotemporal graph neural network model is built. This model can capture the spatial correlation between nodes of each distributed photovoltaic power station through graph convolutional layers, and mine the temporal variation law of power output through temporal convolutional layers or recurrent layers, thereby achieving collaborative extraction of spatiotemporal dimension features. Subsequently, physical constraints that conform to the operating law of photovoltaic power stations are introduced, including at least one of the following: an energy conservation constraint, which constrains the deviation range between the total irradiance energy of the constrained area and the theoretical maximum power generation potential, to prevent the model output from deviating from the law of energy conservation; a spatial correlation constraint, which constrains the power correlation between stations in the model output to show a physical attenuation trend with distance, conforming to the actual geographical and electrical correlation characteristics; and a fluctuation limit constraint, which constrains the power change rate of the model output to not exceed the limit value based on the physical performance of the inverter, ensuring that the output meets the operating capacity of the equipment. Next, two training scenario libraries were constructed by combining historical operational data and historical meteorological data: the first training scenario library takes different meteorological scenarios, such as combinations of solar intensity and temperature changes, as input, and outputs the joint power time series, fluctuation characteristics, and spatial distribution characteristics of the target area under the corresponding scenario; the second training scenario library takes power grid structure change parameters, such as line impedance adjustment, changes in substation access status, and fixed meteorological scenarios, as input, and outputs the joint power characteristics under the corresponding conditions. Then, physical constraints were incorporated into the model training loss function, and constraint biases were included in the loss calculation through a weighted approach to improve the model's adherence to physical laws. The first twin channel was obtained by training the model using the first training scenario library, focusing on the impact of meteorological changes on joint power; the second twin channel was obtained by training the model using the second training scenario library, focusing on the impact of power grid structure changes on joint power. Finally, the two twin channels were connected and integrated to form a joint power characteristic twin model that can simultaneously adapt to changes in meteorology and power grid structure.

[0022] Step S400: Input the simulated input data into the joint power characteristic twin model, perform scenario simulation of joint power, and output the joint power time series, fluctuation characteristics and spatial distribution characteristics of the target region.

[0023] Specifically, the input simulated data is categorized to determine whether it belongs to meteorological scenario data (such as simulations of different combinations of solar radiation intensity and temperature) or power grid structure change data (such as simulations of line impedance adjustment and power station access status changes). It must also be paired with fixed meteorological parameters to determine the matching twin channel in the joint power characteristic twin model. If it is meteorological scenario data, the first twin channel that focuses on responding to meteorological changes is triggered; if it is power grid structure change data, the second twin channel that focuses on responding to power grid structure adjustments is triggered.

[0024] Subsequently, the simulated input data is preprocessed according to the format required by the model and then input into the matched twin channel for joint power scenario simulation: the spatiotemporal graph neural network in the channel will dynamically calculate the power output of each distributed photovoltaic power station in the target area based on the spatiotemporal correlation rules and physical constraint rules learned in the previous training, combined with the characteristics of the simulated input data, while taking into account the impact of electrical, geographical and meteorological correlations between stations on the joint power.

[0025] Ultimately, the model outputs three core results: first, joint power time series, which presents the changes in total joint power in the target area at different times under the simulated scenario in the form of time series curves; second, fluctuation characteristics, which include statistical data such as power fluctuation amplitude and fluctuation frequency, reflecting the stability of joint power; and third, spatial distribution characteristics, which show the differences in power output and spatial correlation of each distributed photovoltaic power station under the simulated scenario in the form of heat maps or data tables, providing accurate data support for photovoltaic power scheduling, grid planning and risk prediction in the target area.

[0026] In one possible implementation, step S200 further includes:

[0027] Step S210: The spatial relationship includes electrical distance, geographical distance and meteorological correlation.

[0028] Specifically, spatial relationships, including electrical distance, geographical distance, and meteorological correlation, are the core dimensions for quantifying the correlation strength between nodes of distributed photovoltaic power stations when constructing the spatiotemporal graph structure. Electrical distance, calculated based on grid topology data, reflects the degree of electrical coupling between two power stations in the grid system, and its correlation weight is inversely proportional to the magnitude of the electrical coupling impedance between the two power stations. Geographical distance is determined by the geographical coordinates of the power stations, representing the physical spatial interval between the two power stations, and its correlation weight is inversely proportional to the actual geographical distance, and is a key basis for selecting candidate neighborhoods of nodes. Meteorological correlation is quantified by analyzing the Pearson correlation coefficient of the historical power sequences of the two power stations, reflecting the synchronicity of the influence of meteorological conditions on the two power stations. The higher the correlation coefficient, the more obvious the coordinated change characteristics of power output. The three factors complement each other from three levels: grid electrical characteristics, physical spatial location, and the synergy of meteorological influence, providing a comprehensive basis for subsequently defining the edge weights in the spatiotemporal graph structure and constructing a weighted adjacency matrix.

[0029] In one possible implementation, such as Figure 2 As shown, step S200 further includes:

[0030] Step S220: Construct each feature vector for each distributed photovoltaic power station. Each feature vector includes at least the installed capacity, historical average power per unit capacity, and geographical coordinates of the corresponding power station.

[0031] Step S230: Based on the power grid topology data, with each distributed photovoltaic power station as a node, stack the feature vectors to form the initial feature matrix of the node.

[0032] Step S240: Based on the spatial relationship, calculate the association strength between each pair of nodes, define the weight of the edges, perform a weighted linear combination, generate a weighted adjacency matrix representing the spatial relationship between nodes, and establish the spatiotemporal graph structure.

[0033] Specifically, the feature vector is constructed by identifying all distributed photovoltaic (PV) power stations within the target area, ensuring comprehensive coverage of every station participating in the joint power characteristic analysis. Next, core attribute parameters are extracted for each power station to construct its feature vector. The installed capacity parameter directly uses the power station's designed or registered rated installed capacity to intuitively reflect its maximum power generation potential. The historical average power per unit capacity parameter is calculated by retrieving historical operating data, such as actual output power records from the past year or three quarters, to determine the average output power per kilowatt within the statistical period, reflecting the power station's long-term power generation efficiency. Geographic coordinate parameters are obtained through GPS positioning or grid archive queries, accurately recording the longitude and latitude of the power station, laying the foundation for subsequent calculations of geographical distances between stations and analysis of spatial relationships. Finally, these three types of parameters are integrated into a vector form in a preset order, forming a unique feature vector for each distributed PV power station. This ensures that the vector comprehensively and accurately represents the key inherent attributes of the corresponding power station, providing standardized basic data for constructing the initial feature matrix of the nodes.

[0034] Based on the power grid topology data, the connection location and relationship of each distributed photovoltaic power station in the target area within the power grid system are clearly defined. This serves as the basis for determining the node identity of each power station in the spatiotemporal diagram structure, ensuring that the node division matches the actual power grid structure. Next, the feature vectors constructed for each distributed photovoltaic power station are retrieved. Each vector contains core parameters such as installed capacity, historical average power per unit capacity, and geographical coordinates. According to the preset node sorting rules, all feature vectors are arranged in an ordered manner based on the connection order of the power stations in the power grid topology and the geographical order from east to west / from north to south. Finally, the sorted feature vectors are stacked and integrated in a row or column manner. If each feature vector is m-dimensional, such as the 3 core parameters corresponding to 3 dimensions, and there are n distributed photovoltaic power stations in the target area, the stacked vectors form an n x m or m x n matrix. This matrix is ​​the initial feature matrix of the nodes. Each row (or each column) in the matrix corresponds to the complete feature information of a power station node, clearly presenting the inherent attributes of all nodes.

[0035] Based on spatial relationships including electrical distance, geographical distance, and meteorological correlation, this method first filters the effective neighborhood of each node in a distributed photovoltaic power station. The K closest nodes are selected as candidate neighbors based on geographical distance. Then, nodes without direct electrical paths or located in different power supply areas are removed using grid topology data to ensure that only node pairs with actual correlation are calculated. Next, three types of correlation weights are calculated for each pair of nodes within the effective neighborhood: the electrical distance weight is inversely proportional to the electrical coupling impedance magnitude between the two nodes; the geographical distance weight is inversely proportional to the actual geographical distance between the two nodes; and the meteorological correlation weight is calculated by taking two values. The Pearson correlation coefficient of the node's historical power sequence is normalized for the three types of weights and then summed according to a preset ratio to obtain the correlation strength between nodes. This correlation strength is used to define the weight of the edge between two nodes. Subsequently, the edge weights of all node pairs are integrated through a weighted linear combination to generate an n-row n-column weighted adjacency matrix, where n is the total number of nodes. Each element in the matrix corresponds to the correlation weight of a pair of nodes, which intuitively reflects the tightness of the spatial correlation between nodes. Finally, by combining the initial node feature matrix and the weighted adjacency matrix generated in this step, a spatiotemporal graph structure that combines the inherent attributes of nodes with the spatial correlation information between nodes is constructed.

[0036] In one possible implementation, step S240 further includes:

[0037] Step S241: Calculate the electrical distance weight between each pair of nodes, wherein the electrical distance weight is inversely proportional to the electrical coupling impedance magnitude of the two nodes in the power grid topology.

[0038] Step S242: Calculate the geographical distance weight between each pair of nodes, wherein the geographical distance weight is inversely proportional to the geographical distance between the two nodes.

[0039] Step S243: Calculate the meteorological correlation weight between each pair of nodes, where the meteorological correlation weight is the Pearson correlation coefficient of the historical power sequences of the two nodes.

[0040] Step S244: After normalizing the electrical distance weight, geographical distance weight, and meteorological correlation weight, sum them by weight to obtain the correlation strength between each pair of nodes.

[0041] Step S245: Define the weights of the edges between each pair of nodes based on the association strength, and construct the weighted adjacency matrix.

[0042] Specifically, the calculation targets all distributed photovoltaic power station nodes within the target area, i.e., pairs of each power station, ensuring coverage of all potentially electrically connected node pairs. Next, based on previously collected grid topology data, the electrical connection parameters of each node pair (power station) in the grid system are extracted, focusing on key data such as the impedance of the transmission line between the two nodes and the equivalent impedance of the transformer. The electrical coupling impedance magnitude between the two nodes is calculated using circuit theory, i.e., the absolute value of the impedance, reflecting the degree of obstruction in the electrical connection. Subsequently, a calculation model is constructed based on the rule that the electrical distance weight is inversely proportional to the electrical coupling impedance magnitude, using the form: electrical distance weight = 1 / (electrical coupling impedance magnitude + ε), where ε is a very small positive number to avoid a zero denominator. A smaller electrical coupling impedance magnitude between two nodes means less energy loss and interference during transmission between them, a tighter electrical connection, and a larger corresponding electrical distance weight. Finally, the electrical distance weight between all pairs of nodes is calculated using this model.

[0043] The calculation targets all pairs of nodes in the distributed photovoltaic power station within the target area, ensuring coverage of all node pairs whose spatial relationships need to be analyzed. Next, the geographical coordinates (latitude and longitude) within the feature vectors constructed for each node are retrieved. Using the Euclidean distance formula selected based on the size of the region, the actual straight-line distance between each pair of nodes in physical space is calculated, yielding a quantified geographical distance value between the two nodes. Subsequently, a calculation model is constructed based on the rule that geographical distance weight is inversely proportional to geographical distance. For example, a geographical distance weight formula of 1 / (geographical distance + ε) is used, where ε is a very small positive number to avoid abnormal weights due to geographical distance approaching zero. This ensures that the closer the geographical distance between two nodes, the more similar their influence from the same regional microclimate, such as local cloud cover, small-scale wind speed changes, and terrain conditions, resulting in a stronger fundamental correlation in power output characteristics and a larger corresponding geographical distance weight. Finally, this model completes the calculation of the geographical distance weights between all pairs of nodes, providing a quantitative basis at the physical space level for subsequent fusion of multi-dimensional weights and determination of the comprehensive correlation strength between nodes.

[0044] The calculation targets pairwise combinations of all distributed photovoltaic power station nodes within the target area, ensuring coverage of all node pairs requiring meteorological correlation analysis. Next, historical operational data for each node is retrieved, and power output records corresponding to the same time period are extracted to form a historical power sequence for each node. Power value sequences are recorded at the hourly or minute level to ensure complete alignment of the power sequences of the two nodes in the time dimension. Subsequently, the Pearson correlation coefficient formula is used to perform correlation analysis on the historical power sequences of each pair of nodes. By calculating the ratio of the product of the covariance and standard deviation of the two sequences, a correlation coefficient ranging from [-1, 1] is obtained. The closer the coefficient is to 1, the more consistent the trends of power output changes of the two nodes with meteorological conditions such as solar intensity, temperature, and cloud cover, and the stronger the synchronicity under the influence of the same meteorological factor. Finally, the calculated Pearson correlation coefficient is directly used as the meteorological correlation weight for the pair of nodes, thereby quantifying the power synergy characteristics of the two nodes due to meteorological conditions.

[0045] For the three types of weights calculated—electrical distance weight, geographical distance weight, and meteorological correlation weight—normalization was performed. The min-max normalization method was used to map the value range of each weight to the interval [0, 1]. Specifically, for all node logarithmic values ​​in each weight type, the formula (original value - minimum value) / (maximum value - minimum value) was used to eliminate the influence of differences in units and numerical ranges between different weight dimensions, ensuring the comparability and superposition of the three weight types. Next, based on the actual operating characteristics and analytical needs of the photovoltaic power stations in the target area, the proportion coefficients of the three weight types in the comprehensive correlation strength were preset. For example, the electrical distance weighting coefficient is 0.4, the geographical distance weighting coefficient is 0.3, and the meteorological correlation weighting coefficient is 0.3, with the sum of the coefficients being 1. This coefficient can be dynamically adjusted based on the focus of power grid dispatching or the effect of historical data analysis. Subsequently, for each pair of nodes, the normalized three types of weights are multiplied by their corresponding coefficients, and the product results are summed to obtain the correlation strength of the pair of nodes, comprehensively integrating the correlation information of electrical coupling, physical space, and meteorological driving. Finally, the correlation strength between all pairs of nodes is calculated through the above process, providing a comprehensive quantitative basis for defining edge weights and constructing a weighted adjacency matrix.

[0046] The calculated pairwise correlation strength between nodes is used as the weight value of the corresponding edge in the spatiotemporal graph structure. This weight value comprehensively reflects the degree of correlation between the two nodes in the three dimensions of electricity, geography, and meteorology. Next, according to the preset sorting of distributed photovoltaic power station nodes in the target area, such as according to the power station number or the access order in the power grid topology, the correlation strength of all node pairs is matrix-arranged according to the order of the corresponding nodes in the row and column. If there are n nodes in the target area, an n x n matrix is ​​constructed. The element value of the i-th row and j-th column in the matrix is ​​the correlation strength between the i-th node and the j-th node, i.e., the edge weight. The diagonal elements and the correlation between a node and itself can be set to 0 or 1, depending on the model requirements. Finally, the matrix formed by the above arrangement is the weighted adjacency matrix. This matrix completely and accurately represents the spatial correlation relationship between all nodes. Together with the initial node feature matrix, it constitutes the core data of the spatiotemporal graph structure, providing structured graph data input for the subsequent training of the spatiotemporal graph neural network model.

[0047] In one possible implementation, step S240 further includes:

[0048] For any given node, select the K nearest nodes as a candidate neighborhood set based on geographical distance.

[0049] The candidate neighborhood set is traversed, and nodes that have no direct electrical path to any of the nodes or are located in different power supply areas are eliminated by querying the power grid topology data, thus forming an effective neighborhood set.

[0050] Within the effective neighborhood set, calculate the electrical distance weight, geographical distance weight, and meteorological correlation weight between any node and other nodes in the set.

[0051] Specifically, the system retrieves the geographical coordinates (latitude and longitude data) of the target node and all other nodes within the target area. Using an Euclidean distance formula adapted to the regional range, it calculates the straight-line distance between the target node and each other node, obtaining a quantitative result of the geographical distance between the target node and all other nodes. Next, these geographical distances are sorted in ascending order of value, and the K nodes with the smallest distances are selected from the sorted results. K is a neighborhood size parameter preset according to actual analysis needs, such as K=5 or K=10, which can be dynamically adjusted according to the power plant density and power grid structure complexity of the target area. Finally, these K closest nodes are integrated to form a candidate neighborhood set of the target node, initially identifying the potential neighboring nodes most closely related to the target node in terms of physical space.

[0052] For any given distributed photovoltaic power station node (target node) and its candidate neighborhood set, each node in the set is traversed one by one, and each node to be verified and the target node are treated as a pair of nodes to be checked. Next, the previously collected power grid topology data is retrieved, and the connection relationship of this pair of nodes in the power grid system is queried. On the one hand, it is checked whether there is a direct electrical path between the two, such as direct connection through transmission lines or access to the power grid through the same distribution branch, to exclude cases where there is no direct line connection or where the two belong to different branches of the power grid. On the other hand, it is confirmed whether the two are in the same power supply area, that is, whether they are powered by the same substation or the same distribution transformer, to exclude cases where they belong to different power supply management units. Finally, nodes that do not meet the condition of "having a direct electrical path and belonging to the same power supply area" are removed from the candidate neighborhood set, and the remaining nodes are integrated to form the effective neighborhood set of the target node, ensuring that the nodes in the set and the target node have both geographical proximity and a correlation basis at the actual power grid operation level.

[0053] For any distributed photovoltaic power station node (target node) and its effective neighborhood set, other nodes in the set are selected one by one to form node pairs with the target node. Three types of weights are calculated. First, the electrical coupling impedance magnitude of the node pair is extracted based on the power grid topology data. According to the rule that the electrical distance weight is inversely proportional to the impedance magnitude, the electrical distance weight is calculated using the formula 1 / (impedance magnitude + ε), where ε is a very small positive number. Second, the geographical coordinates of the two nodes are retrieved, and the actual geographical distance is calculated using the Euclidean distance formula. According to the rule that the geographical distance weight is inversely proportional to the geographical distance, the geographical distance weight is calculated using 1 / (geographical distance + ε), where ε is a very small positive number. Finally, the historical power sequences aligned with the time dimension of the two nodes are retrieved. The correlation coefficient with a value range of [-1, 1] is obtained by calculating the ratio of the product of covariance and standard deviation using the Pearson correlation coefficient formula. This coefficient is directly used as the meteorological correlation weight. Finally, the three types of weights are calculated for the target node and all other nodes in the effective neighborhood set, providing accurate local weight data for subsequent integration of correlation strength and construction of a weighted adjacency matrix.

[0054] In one possible implementation, step S300 further includes:

[0055] Step S310: The physical constraint term includes at least one of the following: energy conservation constraint term, spatial correlation constraint term, and fluctuation limit constraint term.

[0056] Among them, the energy conservation constraint term is used to constrain the deviation range between the total regional irradiance and the theoretical maximum power generation potential; the spatial correlation constraint term is used to constrain the inter-site power correlation and distance of the model output to conform to the physical attenuation law; and the fluctuation limit constraint term is used to constrain the power change rate of the model output to not exceed the limit value based on the physical performance of the inverter.

[0057] Specifically, physical constraints are crucial for ensuring that the output of the joint power characteristic twin model conforms to the actual operating laws of photovoltaic power plants. These constraints include energy conservation constraints, spatial correlation constraints, and fluctuation limit constraints; at least one of these must be selected. Their specific definitions and functions are as follows: The energy conservation constraint is based on the total irradiance of the target area. It calculates the theoretical maximum power generation potential of the area by combining parameters such as the installed capacity and photoelectric conversion efficiency of each distributed photovoltaic power plant. By setting a deviation threshold, such as ±5% of the theoretical value, it constrains the deviation range between the model's output of the joint total power and the theoretical maximum power generation potential, preventing unreasonable outputs that deviate from the law of energy conservation due to data noise or overfitting. The spatial correlation constraint is based on the principle that the closer the distance between power plants, the stronger the synchronicity of power output affected by meteorological and geographical factors, and the higher the correlation... The higher the physical law, the more constrained the power correlation between different stations in the model output must conform to this decay trend through a preset correlation decay function, such as exponential decay with increasing distance. This prevents results that violate actual physical logic, such as high power correlation between distant stations. The fluctuation limit constraint is based on the inverter's physical performance parameters, such as maximum ramp rate and maximum ramp rate, to determine the power change rate limit. This constraint ensures that the power change rate at any given moment in the model output, i.e., the ratio of the power difference between adjacent time steps to the time step, does not exceed this limit. This ensures that the power fluctuation in the model output conforms to the actual operating capacity of the equipment and avoids extreme fluctuations that exceed the inverter's adjustment range. These three types of constraints can be used individually or in combination according to the needs of the model training scenario, providing physical-level guarantees for the rationality of the model output.

[0058] In one possible implementation, step S300 further includes:

[0059] Step S320: Determine the training loss function of the spatiotemporal graph neural network model.

[0060] Step S330: Combine the physical constraint term with the training loss function, and use the historical operating data and the historical meteorological data to train the model to obtain the joint power characteristic twin model.

[0061] Specifically, based on the model's core objective—accurately predicting the temporal, fluctuation, and spatial distribution characteristics of the combined power in the target region—the loss function must revolve around quantifying the deviation between the model's predicted and actual values, and must be adapted to the continuous regression characteristics of the power data. Next, a basic loss function type is selected, with the mean squared error (MSE) loss function being preferred. Its calculation logic is as follows: extract the actual combined power data from historical operating data, such as the total combined power of the region at each time step and the power output values ​​of each power station, as labels; compare this with the predicted power data output by the model; calculate the squared difference between each set of predicted and actual values; and then take the average of all squared differences to highlight larger deviations. The impact of the difference on the loss enhances the model's fitting accuracy to key power nodes. Simultaneously, auxiliary loss terms can be added according to the training scenario requirements. For example, a mean absolute error (MAE) loss can be added for predicting fluctuation characteristics, calculating the average absolute difference between the predicted and actual fluctuation amplitudes; or a cosine similarity loss can be added for spatial distribution characteristics, quantifying the similarity deviation between the predicted power plant spatial distribution and the actual distribution, avoiding fitting bias caused by a single loss function. Finally, the basic loss function and auxiliary loss terms are weighted according to preset weights, such as MSE accounting for 0.7 and MAE accounting for 0.3, to form a complete training loss function, ensuring it fully covers the prediction dimensions that the model needs to optimize.

[0062] The defined physical constraints are transformed into quantifiable constraint loss terms. For example, the energy conservation constraint is calculated by taking the squared deviation between the model's output regional total power and the theoretical maximum power generation potential; the spatial correlation constraint is calculated by taking the deviation between the predicted site power correlation and the distance decay law; and the fluctuation limit constraint is calculated by taking the absolute value of the power change rate exceeding the limit. Next, weight coefficients are assigned to each constraint loss term, and a weighted sum is performed with the defined basic training loss function to construct a comprehensive loss function incorporating physical rules. Subsequently, historical meteorological data, such as time-series data on sunshine duration, temperature, and wind speed, and power grid topology characteristics are used as model inputs, along with historical operational data... The actual power output, including the power time series of each power station and the total regional power, is used as labels. The data is divided into training and validation sets in chronological order. The parameters of the spatiotemporal graph neural network model are iteratively optimized through the backpropagation algorithm. During the training process, the comprehensive loss function value and the degree of satisfaction of each physical constraint term are monitored in real time. When the prediction error of the model on the validation set continues to decrease and the deviation of the physical constraint is stable within the preset threshold, such as energy conservation deviation <5% and fluctuation over-limit number <1%, the iteration is stopped. Finally, the model parameters at this time are saved to obtain a joint power characteristic twin model that can accurately fit historical data and strictly follow physical laws, providing a reliable tool for subsequent joint analysis and prediction of regional photovoltaic power.

[0063] In one possible implementation, step S320 further includes:

[0064] Step S321: Construct a first training scenario library and a second training scenario library using the historical meteorological data and the historical operational data.

[0065] In the first training scenario library, the input of each training scenario is a meteorological scenario, and the output is the joint power time series, fluctuation characteristics and spatial distribution characteristics of the target area. In the second training scenario library, the input of each training scenario is the power grid structure change parameters and a fixed meteorological scenario, and the output is the joint power time series, fluctuation characteristics and spatial distribution characteristics of the target area.

[0066] Step S322: Combine the physical constraint term with the training loss function, train the spatiotemporal graph neural network model with the first training scene library to obtain the first twin channel, and train the spatiotemporal graph neural network model with the second training scene library to obtain the second twin channel.

[0067] Step S323: Connect the first twin channel and the second twin channel to generate the joint power characteristic twin model.

[0068] Specifically, the historical meteorological data used to build the scenario library includes time-series data such as solar radiation intensity, temperature, wind speed, and cloud cover at different times, and covers typical weather types such as sunny, cloudy, partly cloudy, and rainy days. The historical operational data includes real-time power output, total regional combined power, and inverter operating status of each distributed photovoltaic power station in the target area. The two types of data are fully aligned in the time dimension, which can accurately match the correspondence between meteorological conditions and power output. Next, based on the core objective of separating the impact of meteorological variables and power grid structure variables on combined power, two types of training scenario libraries are constructed. The first training scenario library focuses on the impact of meteorological factors on combined power. The input of each training scenario is a meteorological scenario, which is meteorological time series data extracted from historical meteorological data within a continuous time window, such as a sunshine-temperature-wind speed sequence with a time step of 1 hour and a continuous 24-hour time window, and weather type labels are also added. The output is the time series, fluctuation characteristics, and spatial distribution characteristics of the combined power of the target area extracted from historical operating data within the time window corresponding to the meteorological scenario. The time series of combined power is the total power value hourly over 24 hours, the fluctuation characteristics include the maximum increase, decrease, number of fluctuations, and standard deviation of fluctuation amplitude within 24 hours, and the spatial distribution characteristics are the proportion of the hourly power of each power station to the total power of the area. This realizes the scenario mapping between meteorological input and power multi-dimensional output. The second training scenario library focuses on the impact of power grid structure changes on combined power. To eliminate interference from meteorological variables, the inputs for each training scenario are set as power grid structure change parameters and a fixed meteorological scenario. The power grid structure change parameters are extracted from power grid operation and maintenance records, including parameters such as power station access or decommissioning status, transmission line impedance adjustment, and transformer capacity changes. The fixed meteorological scenario selects a standard scenario with high stability from historical meteorological data, such as a clear weather time series with no fluctuations for 7 consecutive days, to ensure consistent meteorological conditions. The output is also the time series, fluctuation characteristics, and spatial distribution characteristics of the combined power of the target area extracted from historical operation data under the corresponding power grid structure and fixed meteorological scenario. This achieves scenario mapping of power grid structure change and fixed meteorological combined input and multi-dimensional power output, ultimately through the construction of two types of scenario libraries.

[0069] The defined physical constraints are fused with the defined training loss function. At least one of the energy conservation constraints, spatial correlation constraints, and fluctuation limit constraints is transformed into a quantifiable constraint loss. Each constraint loss is then assigned a corresponding weight coefficient, and the weighted sum with the basic training loss function forms a comprehensive loss function that balances prediction accuracy and physical laws. Next, for the first training scenario library, meteorological scenarios from each training scenario are used as input, and the corresponding target region's joint power temporal series, fluctuation characteristics, and spatial distribution characteristics are used as output labels. The spatiotemporal graph neural network model is iteratively trained using the comprehensive loss function. Through backpropagation, the model parameters are continuously optimized, allowing the model to gradually learn the mapping relationship between meteorological factors and the multi-dimensional characteristics of joint power. Ultimately, a first twin channel focused on fitting the influence of meteorological conditions on joint power is obtained. Meanwhile, for the second training scenario library, the power grid structure change parameters and fixed meteorological scenarios of each training scenario are used as inputs, and the joint power time series, fluctuation characteristics and spatial distribution characteristics of the corresponding target area are used as output labels. The same comprehensive loss function is used to train another spatiotemporal graph neural network model. The focus is on optimizing the model's ability to fit the relationship between power grid structure changes and joint power response, resulting in a second twin channel that focuses on capturing the impact of power grid structure changes on joint power. This enables the two twin channels to have the ability to accurately model different influencing factors.

[0070] Based on the output characteristics of the first and second twin channels, a channel fusion mechanism is designed. This mechanism employs an attention weight allocation strategy, dynamically adjusting the fusion ratio of the two channel outputs by analyzing the input data type. For example, in scenarios dominated by meteorological data, the weight of the first channel is increased; in scenarios dominated by power grid structure changes, the weight of the second channel is increased. Next, a feature splicing layer and a collaborative optimization layer are set in the model architecture. The feature splicing layer splices the intermediate feature vectors of the joint power time series, fluctuation characteristics, and spatial distribution characteristics of the two channel outputs dimensionally, forming a comprehensive feature containing dual influencing factors. The collaborative optimization layer uses a multilayer perceptron to perform nonlinear transformation on the spliced ​​features, further eliminating redundant information between the two channel outputs and strengthening complementarity. This ensures that the fusion result retains both the dynamic power response characteristics driven by meteorology and the power distribution adjustment patterns brought about by power grid structure changes. Finally, the output layer decodes the collaboratively optimized features to generate joint power time series, fluctuation characteristics, and spatial distribution characteristics that simultaneously satisfy the dual influences of meteorology and power grid structure, ultimately forming a complete joint power characteristic twin model. This model can adaptively utilize the advantages of the two channels according to the actual input scenario, achieving accurate simulation and prediction of joint power characteristics under complex operating conditions.

[0071] In one possible implementation, step S400 further includes:

[0072] Step S410: Identify the data category of the simulated input data and determine the triggered matching twin channel.

[0073] Step S420: Input the simulated input data into the matched twin channel to perform scenario simulation and output the joint power timing, fluctuation characteristics and spatial distribution characteristics of the target region.

[0074] Specifically, the simulated input data undergoes feature analysis to extract the core parameter types and clarify the scenario attributes reflected by the data. If the simulated input data only contains meteorological parameters, such as the predicted solar intensity sequence for future periods, temperature change trends, wind speed and duration distribution, and cloud cover dynamics, and does not involve any information on power grid structure changes, then its data category is identified as meteorological-driven data. In this case, the determination should trigger the first twin channel focusing on the mapping relationship between meteorological factors and joint power. If the simulated input data contains parameters related to power grid structure changes, such as new power station access schemes, line impedance adjustment values, transformer capacity change parameters, and status indicators of some power stations being taken out of operation, and also includes fixed meteorological scenario parameters, such as setting them as typical sunny weather or standardized meteorological conditions for the same period in history, then its data category is identified as power grid structure-driven data. In this case, the determination should trigger the second twin channel focusing on the relationship between power grid structure changes and joint power response. If the simulated input data includes both meteorological parameters and power grid structure change parameters, the model will use its preset scenario priority rules. For example, based on application requirements, the first twin channel will be triggered primarily by meteorological parameters, with power grid structure parameters serving as auxiliary correction items; or the second twin channel will be triggered primarily in the power grid transformation scenario, with meteorological parameters maintaining a fixed baseline value. Ultimately, a unique matching twin channel will be determined to ensure that subsequent scenario simulations can call the most suitable model channel for calculation.

[0075] Based on the identified matching twin channels, the simulated input data undergoes targeted preprocessing. If the matching is the first twin channel focusing on meteorological-power mapping, the simulated meteorological data, such as future sunshine duration and temperature time series, needs to be time-series aligned and numerically standardized according to the format standards used during channel training to ensure that the data dimensions and time steps are consistent with the channel input requirements. If the matching is the second twin channel focusing on grid structure-power mapping, the simulated grid structure change parameters, such as power plant access status and line impedance values, need to be formatted and integrated with fixed meteorological scenario parameters to supplement grid topology correlation features to adapt to the channel input logic. Next, the preprocessed simulated input data is input into the matching twin channel. The spatiotemporal graph neural network model within the channel will call the mapping rules learned during the training phase, i.e., the power response rules under meteorological drive or the power adjustment rules under grid structure changes, to perform scenario extrapolation. During the extrapolation process, the model will combine built-in physical constraints to verify the intermediate output results in real time. For example, the energy conservation constraint term corrects the deviation between the total regional power and the irradiance energy, and the fluctuation limit constraint term limits the power change rate from exceeding the inverter performance, ensuring that the extrapolation process conforms to physical laws. Finally, after decoding by the model output layer, the joint power time series of the target area is generated, such as the hourly total power change curve for the next 24 hours, including peak power, valley power and their occurrence time, fluctuation characteristics, such as the maximum power fluctuation amplitude of adjacent time steps, fluctuation frequency statistics, and over-limit risk assessment and spatial distribution characteristics, such as the power proportion of each distributed photovoltaic power station at different time periods, and the spatial clustering of high power output stations. The complete output scenario simulation results support the application needs of grid dispatch, operation and maintenance decision-making, etc.

[0076] In one possible implementation, step S300 further includes:

[0077] Step S340: Using the aforementioned joint power characteristic twin model, generate multiple sets of regional joint power simulation data under different historical meteorological conditions.

[0078] Step S350: Unsupervised temporal clustering is used to analyze the joint power simulation data of the multiple groups of regions and automatically identify the power curve categories with different morphological characteristics.

[0079] Step S360: For each identified power curve category, calculate the center trajectory as a representative power curve, and statistically analyze the fluctuation range, occurrence probability, and main associated meteorological characteristics to establish each typical scenario.

[0080] Step S370: Store each typical scenario and map it with historical power dispatching decisions to establish a typical scenario library for random optimization dispatching.

[0081] Specifically, diverse historical meteorological conditions are selected from collected historical meteorological data. These conditions must cover typical weather types commonly seen in the target area, such as sunny, cloudy, partly cloudy, and rainy days, while also incorporating potential special weather conditions such as sandstorms and haze, ensuring coverage of meteorological scenarios with varying irradiance, temperature, wind speed, and cloud cover characteristics. Next, each selected historical meteorological condition is preprocessed according to the input format requirements of the joint power characteristic twin model. The time step of the meteorological data is standardized to ensure that the temporal dimensions of parameters such as irradiance, temperature, and wind speed are consistent with the input standards used during model training. Data values ​​are also standardized to ensure that the parameter ranges are compatible with the model's computational logic. Then, the preprocessed historical meteorological conditions are input into the trained joint power characteristic twin model. The model uses the mapping rules between meteorological factors and joint power learned during training, combined with built-in physical constraints, to verify the extrapolation process, ensuring that the output results conform to physical laws such as energy conservation and power fluctuation limits. Finally, for each type of historical meteorological condition, multiple sets of regional joint power simulation data are generated through multiple model simulations. Each set of data contains complete power time series information, ultimately forming a regional joint power simulation dataset that covers different historical meteorological conditions, is sufficient in quantity, and can reflect the power change characteristics under various meteorological scenarios.

[0082] Preprocessing was performed on the generated joint regional power simulation data. The power time series corresponding to each data set was converted into a standardized vector format, unifying the time series length and time step of all data. This eliminated analytical interference caused by differences in duration and numerical magnitude between different data sets, ensuring that all power curves to be analyzed were in the same data dimension. Next, an unsupervised clustering algorithm adapted to the characteristics of the time series data was selected. The morphological characteristics of the power curves were used as the core clustering basis. By calculating the similarity between different power time series, including distance measurement of key features such as the dynamic change trend of the power curves, the time position of peak and trough values, and the rate of power increase and decrease, power curves with similar morphological characteristics were grouped into the same cluster. The entire clustering process did not require manual labeling of categories in advance. The algorithm automatically determined the appropriate number of clusters based on the morphological patterns inherent in the data itself. For example, power curves with obvious morphological differences, such as "rapid power increase in the morning, reaching a high value at noon and then stabilizing", "generally low power with small fluctuations throughout the day", and "power sudden rise and fall in multiple periods with changes in cloud cover", were clearly distinguished into different categories. Finally, the clustering results are validated using clustering effectiveness evaluation metrics to ensure that the internal morphological consistency of each power curve category is high and the morphological differences between different categories are significant. Ultimately, power curve categories with different morphological characteristics are automatically identified.

[0083] For each automatically identified power curve category, the power time series corresponding to all regional joint power simulation data under that category is extracted. Using the time step as a benchmark, the mean of all power values ​​at each time node is calculated. These means are then connected chronologically to form the central trajectory of that category. This central trajectory is determined as the representative power curve for that category, intuitively reflecting the core trend of power changes within that category. Next, the fluctuation range of the power curve for that category is statistically analyzed. By calculating the deviation of all power values ​​at each time node from their corresponding mean, the power fluctuation intervals at each time node are determined. Integrating the fluctuation intervals of all time nodes yields the power fluctuation range over the entire time series period, clearly presenting the stability characteristics of the power output for that category. Finally, the probability of occurrence of that power curve category is calculated. By statistically analyzing the proportion of power simulation data sets included in that category to the total number of regional joint power simulation data sets, the probability of occurrence of that category under historical meteorological conditions is obtained, reflecting its frequency of occurrence. Simultaneously, historical meteorological conditions corresponding to the generation of power simulation data for this category were traced back, and meteorological features highly correlated with the power characteristics of this category were analyzed and extracted. For example, high irradiance and low wind speed are often associated with power curves of "stable high power output at noon," while cloudy skies and intermittent irradiance are often associated with power curves of "multi-period power fluctuations." Finally, representative power curves, fluctuation ranges, occurrence probabilities, and main associated meteorological features of each power curve category were integrated to form typical scenarios containing the core information of this category, thus completing the construction of each typical scenario.

[0084] Each constructed typical scenario undergoes standardized format processing. Core information such as representative power curves, fluctuation ranges, occurrence probabilities, and key associated meteorological characteristics are organized according to a pre-defined data structure, such as structured data tables or Extensible Markup Language (EXPLAIN), ensuring clear classification and consistent fields for easy storage and retrieval. Next, all processed typical scenarios are batch-stored in a designated database using a hierarchical storage strategy, categorized and archived according to key associated meteorological characteristics, such as sunny, cloudy, or partly cloudy fluctuations, or probability levels, improving scenario retrieval efficiency. Subsequently, the historical power dispatch decision database is retrieved to extract historical dispatch cases matching the characteristics of each typical scenario. These cases must include power scenario descriptions and dispatch strategies for the corresponding time periods, such as energy storage charging and discharging adjustments, conventional power output optimization, grid-connected line power allocation, and dispatch effect evaluation data. By comparing the power and meteorological characteristics of typical scenarios with the scenario conditions of historical dispatch cases, a mapping relationship is established between the two. For example, the typical scenario of "stable high-power output with a high probability of occurrence" is associated with the historical dispatch decision of "prioritizing full grid connection and reducing energy storage investment," and the typical scenario of "frequent and large power fluctuations" is associated with the historical dispatch decision of "strengthening energy storage frequency regulation and dynamically adjusting the reserve capacity of conventional power sources." Finally, all typical scenarios and their matching historical power dispatch decisions are integrated to form a complete dataset containing scenario information, mapping decisions, and decision effects. This dataset is used to construct a typical scenario library for stochastic optimization dispatch. This scenario library can provide historical experience reference and data support for the formulation of dispatch strategies under different power scenarios in subsequent grid stochastic optimization dispatch.

[0085] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0086] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0087] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station, characterized in that, include: Collect historical operating data, historical meteorological data, and power grid topology data for each distributed photovoltaic power station within the target area; Based on the aforementioned power grid topology data, a spatiotemporal graph structure is constructed, using each distributed photovoltaic power station as a node and the spatial relationships between stations as edges, including: Each feature vector is constructed for each of the distributed photovoltaic power stations. Each feature vector includes at least the installed capacity, historical average power per unit capacity, and geographical coordinates of the corresponding power station. Based on the power grid topology data, with each distributed photovoltaic power station as a node, the feature vectors are stacked to form the initial feature matrix of the node; Based on the spatial relationship, the association strength between each pair of nodes is calculated, the weight of the edge is defined, a weighted linear combination is performed, a weighted adjacency matrix representing the spatial relationship between nodes is generated, and the spatiotemporal graph structure is established. Based on the aforementioned spatiotemporal graph structure, a spatiotemporal graph neural network model is constructed, and physical constraints are introduced. The model is trained using the historical operational data and historical meteorological data to obtain a joint power characteristic twin model, including: Determine the training loss function of the spatiotemporal graph neural network model; By combining the physical constraint term with the training loss function and using the historical operational data and historical meteorological data for model training, the joint power characteristic twin model is obtained, including: The first training scenario library and the second training scenario library are constructed using the historical meteorological data and the historical operational data. In the first training scenario library, the input of each training scenario is a meteorological scenario, and the output is the joint power time series, fluctuation characteristics and spatial distribution characteristics of the target area. In the second training scenario library, the input of each training scenario is the power grid structure change parameters and a fixed meteorological scenario, and the output is the joint power time series, fluctuation characteristics and spatial distribution characteristics of the target area. The physical constraint term is combined with the training loss function, and the spatiotemporal graph neural network model is trained with the first training scene library to obtain the first twin channel. The spatiotemporal graph neural network model is trained with the second training scene library to obtain the second twin channel. Connect the first twin channel and the second twin channel to generate the joint power characteristic twin model; The simulated input data is input into the joint power characteristic twin model to perform scenario simulation of joint power, and output the joint power time series, fluctuation characteristics and spatial distribution characteristics of the target region. The physical constraints include at least one of the following: energy conservation constraints, spatial correlation constraints, and fluctuation limit constraints. Among them, the energy conservation constraint term is used to constrain the deviation range between the total regional irradiance and the theoretical maximum power generation potential; the spatial correlation constraint term is used to constrain the inter-site power correlation and distance of the model output to conform to the physical attenuation law; and the fluctuation limit constraint term is used to constrain the power change rate of the model output to not exceed the limit value based on the physical performance of the inverter.

2. The method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station as described in claim 1, characterized in that, The spatial relationships include electrical distance, geographical distance, and meteorological correlation.

3. The method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station as described in claim 1, characterized in that, Based on the spatial relationship, the strength of the association between each pair of nodes is calculated, the weights of the edges are defined, and a weighted linear combination is performed, including: Calculate the electrical distance weights between each pair of nodes, where the electrical distance weights are inversely proportional to the magnitudes of the electrical coupling impedances between the two nodes in the power grid topology; Calculate the geographical distance weight between each pair of nodes, where the geographical distance weight is inversely proportional to the geographical distance between the two nodes; Calculate the meteorological correlation weight between each pair of nodes, where the meteorological correlation weight is the Pearson correlation coefficient of the historical power sequences of the two nodes; After normalizing the electrical distance weight, geographical distance weight, and meteorological correlation weight, the weighted sum is obtained to obtain the pairwise correlation strength between nodes. The weighted adjacency matrix is ​​constructed by defining the weights of the edges between each pair of nodes based on the association strength.

4. The method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station as described in claim 3, characterized in that, Based on the spatial relationship, the association strength between each pair of nodes is calculated, and the weight of the edges is defined. This also includes: For any given node, select the K nearest nodes as a candidate neighborhood set based on geographical distance; Traverse the candidate neighborhood set, and by querying the power grid topology data, remove nodes that have no direct electrical path to any of the nodes or are located in different power supply areas to form an effective neighborhood set; Within the effective neighborhood set, calculate the electrical distance weight, geographical distance weight, and meteorological correlation weight between any node and other nodes in the set.

5. The method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station as described in claim 1, characterized in that, The simulated input data is fed into the joint power characteristic twin model to perform scenario extrapolation of the joint power, and outputs the joint power time series, fluctuation characteristics, and spatial distribution characteristics of the target region, including: Identify the data category of the simulated input data and determine the triggered matching twin channel; The simulated input data is input into the matching twin channel for scenario simulation, and the joint power timing, fluctuation characteristics and spatial distribution characteristics of the target region are output.

6. The method for analyzing the combined power characteristics of a high-proportion distributed photovoltaic power station as described in claim 1, characterized in that, After obtaining the twin model of the joint power characteristics, it also includes: Using the aforementioned joint power characteristic twin model, multiple sets of regional joint power simulation data under different historical meteorological conditions are generated; Unsupervised temporal clustering is used to analyze the joint power simulation data of the multiple regions and automatically identify the power curve categories with different morphological characteristics. For each identified power curve category, the center trajectory is calculated as a representative power curve, and the fluctuation range, occurrence probability, and main associated meteorological characteristics are statistically analyzed to establish each typical scenario. Each typical scenario is stored and mapped to historical power dispatching decisions to establish a typical scenario library for random optimization dispatching.

Images