Distributed photovoltaic resource output curve deduction method, system, device and medium

Abstract

A distributed photovoltaic resource output curve deduction method, system, device and medium, the method comprises: constructing a weighted undirected space-time graph to represent the relationship between distributed photovoltaic resources; solving the time characteristics of each node in the weighted undirected space-time graph; extracting the space-time characteristics of each node in the weighted undirected space-time graph, capturing the output mode characteristics of each node in the space-time characteristics, splicing the time characteristics and the output mode characteristics of each node and inputting them into a multi-layer fully connected network, using the fully connected network to output the output prediction value of the distributed photovoltaic resources in the corresponding area, and drawing the distributed photovoltaic resource output deduction curve. The present application comprehensively utilizes multi-source data, extracts time characteristics combined with meteorological and illumination information of the target region, extracts spatial characteristics of the target region photovoltaic module based on graph network driven deep learning, and learns appropriate mapping relationship according to long-term and short-term weather conditions, and accurately deduces the future output curve of the distributed photovoltaic resource.

Description

Technical Field

[0001] This invention belongs to the field of power grid multi-source data analysis and processing technology, specifically involving a method, system, equipment and medium for deducing the output curve of distributed photovoltaic resources. Background Technology

[0002] As regions adjust their electricity generation ratios, the output of traditional energy sources such as hydropower and thermal power is continuously decreasing. Therefore, accurately predicting the output capacity of photovoltaic (PV) systems is crucial. Rationally coordinating the output of PV systems with other power sources can ensure the safe and stable operation of urban power grids, providing stable and reliable power support for production, daily life, and high-quality economic and social development, thereby achieving the goal of grid-load-storage integration. However, the power output of a PV system is a complex, non-linear process, making effective prediction of PV output a highly challenging task. Because PV power generation is affected by multiple factors, changes in weather, sunlight, or PV module parameters can cause significant random fluctuations in the output of PV modules, affecting the overall power output of the network. Furthermore, PV modules in similar or adjacent areas often receive similar disturbances from weather and sunlight, exhibiting a perceptible local effect.

[0003] Currently, domestic and international solutions for photovoltaic power prediction can be broadly categorized into three types: physical methods, which derive power output patterns based on models of solar radiation and photovoltaic modules; statistical methods, which predict power output based on statistical patterns between the input and output factors of prediction models; and machine learning methods, which deduce the power output of photovoltaic modules by optimizing objective functions based on various models. While these power prediction solutions appear feasible, they still have shortcomings in practical applications. A major shortcoming is that these solutions focus too much on the model performance for short-term, small-scale deployments, neglecting the predictive performance of long-term power output curves during large-scale deployments.

[0004] Furthermore, the aforementioned solutions also have a series of shortcomings in model training. The aforementioned physical methods rely on detailed information about the photovoltaic modules and the environment, have extremely high requirements for observational data, and cannot model photovoltaic output under extreme weather conditions; the aforementioned statistical methods rely on a large amount of historical data and have poor robustness; the prior knowledge learned by the aforementioned machine learning methods is often obtained through training on long-term business data under existing models, and cannot reasonably evaluate the output curve for short-term fine-grained indicators, especially when encountering sudden situations. Summary of the Invention

[0005] The purpose of this invention is to address the problems in the prior art by providing a method, system, device, and medium for extrapolating the output curve of distributed photovoltaic resources. This method can accurately and reasonably extrapolate the future output curve of distributed photovoltaic resources under extreme weather conditions and unforeseen circumstances, thus meeting the predictive performance requirements of long-term output curves for large-scale deployments.

[0006] To achieve the above objectives, the present invention provides the following technical solution: Firstly, a method for deriving the output curve of distributed photovoltaic resources is provided, including: Input the geographical location of each distributed photovoltaic resource and the historical photovoltaic output data under different environmental conditions at different times, and construct a weighted undirected spatiotemporal graph to represent the relationship between distributed photovoltaic resources; Solve for the temporal characteristics of each node in the weighted undirected spatiotemporal graph. The temporal characteristics are intermediate prediction data of distributed photovoltaic resource output under environmental conditions at different times. The spatiotemporal features of each node in the weighted undirected spatiotemporal graph are extracted. The spatiotemporal features are spatial features that include time features, and the spatial features are obtained from the weighted undirected spatiotemporal graph. The output mode features of each node in the spatiotemporal features are captured. The time features and output mode features of each node are concatenated and input into a multi-layer fully connected network. The target output prediction value of the distributed photovoltaic resources in the corresponding region is output using the fully connected network, and the output projection curve of the distributed photovoltaic resources is plotted.

[0007] As a preferred embodiment, the step of constructing a weighted undirected spatiotemporal graph to characterize the relationships between distributed photovoltaic resources includes: Time nodes An undirected graph of an electrical grid is defined as follows: ,in, For a set of nodes, Let be the set of edges. for The feature values ​​of all nodes at time t; Calculate the next node at the current time. With nodes Historical photovoltaic power output data and Correlation coefficient between ; Define the adjacency matrix Indicates photovoltaic resource nodes With nodes The interconnection relationship between them, the adjacency matrix elements in The value is determined as follows:

[0008] In the formula, This is the average of the output correlation coefficients for all node pairs. For a moment Next node With nodes Historical photovoltaic power output data and The Euclidean distance between them; The above formula establishes the connection relationships on the undirected graph of the power grid, forming a weighted undirected spatiotemporal graph representing the correlation degree of different photovoltaic nodes.

[0009] As a preferred embodiment, the steps for solving the temporal characteristics of each node in the weighted undirected spatiotemporal graph include: Establish an LSTM network model. The LSTM network model is a Long Short Time Memory network model. For the node data at each time step... Update the cell state using the forget gate. :

[0010]

[0011] In the formula, This represents the forget gate, used to control the cell state at the previous time step. For the current cell state The extent of the impact; This represents the cell state update value. right The effect is caused by the input gate To take control; Update input gate by pressing formula With cell state update value :

[0012]

[0013] The output gates in the LSTM network model are defined by the following formula. Update and calculate the output of the output unit. :

[0014]

[0015] definition , , , , This represents historical weather data, including temperature, humidity, wind speed, and wind direction. , , This represents historical solar irradiance, vertical irradiance, and solar radiation factor data, among which... For time step; The LSTM network model described is a two-layer model. At any given time, an LSTM network model is used to perform regression prediction on the historical output data of the photovoltaic power plant, with the objective being to output... Real-time photovoltaic output data, before use Photovoltaic output data, weather data, and solar irradiance data for specific time segments, for Periodically fit the photovoltaic output data at any given time; By combining the outputs of a two-layer LSTM network model, a mapping function is fitted based on historical photovoltaic power output data. and the corresponding parameter set Output the time features of each node. The expression is as follows:

[0016] in, and These correspond to the two-layer structure of the selected LSTM network model.

[0017] As a preferred embodiment, the step of solving the temporal features of each node in the weighted undirected spatiotemporal graph further includes: learning from the source domain. A Source-LSTM model, Each Source-LSTM model is pre-trained using a sufficient amount of data from the corresponding source domain; the base-LSTM model in the target domain generates the basic output. ,Depend on Each Source-LSTM model generates source output. ; The model's output is processed through an attention network. and Perform weighting; The attention network described is designed based on a multilayer perceptron, as shown in the following equation:

[0018] in, For input data, For learnable parameters, This represents predictions made using an LSTM network model. The dimension is The output of the multilayer perceptron is converted into model weights by a soft attention mechanism, and the attention network generates multiple non-zero weights according to the following formula. :

[0019] In the formula, ; Based on the attention network, the output of the complete model is... Defined as a Source-LSTM model Output and Base-LSTM model output The weighted sum is expressed as follows:

[0020] Output The temporal characteristics of each photovoltaic node were obtained.

[0021] Furthermore, in the target domain, the model's parameters are updated according to two functions: One is the parameters of the Base-LSTM model, output from the target domain. and the true value of the target domain Calculated loss function Update; another is the parameters in the attention network, which are based on the overall output of the model. and the true value of the target domain Calculated loss function renew; In both the Base-LSTM and attention networks, there are two independent backpropagation mechanisms. The loss function for both is calculated using the mean absolute error, as shown in the following expression:

[0022] In the formula, It is the number of photovoltaic nodes. The set of predicted output values ​​for all nodes. The set of actual values ​​of the effort contributed by all nodes. For the first Predicted output values ​​for each node For the first The actual value of the output of each node. It is the set of all learnable parameters.

[0023] As a preferred embodiment, the step of extracting the spatiotemporal features of each node in the weighted undirected spatiotemporal graph includes: establishing a spatiotemporal graph autoencoder (ST-GAE) and processing the input weighted undirected spatiotemporal graph sequence. Encode and reconstruct data patterns in a weighted undirected spatiotemporal graph; learn deep encoding mapping relationships. Weighted undirected spacetime graph sequence Mapped to a series of hidden space vectors ,exist In the middle, the first Columns represent weighted undirected spacetime graphs The Middle The spatiotemporal features of each node are analyzed; two decoder modules are designed to learn the deep decoding mapping relationship, including feature-level encoding. Encoding at the edge level .

[0024] As a preferred embodiment, the steps of capturing the power output mode features of each node in the spatiotemporal features, concatenating the temporal features and power output mode features of each node and inputting them into a multi-layer fully connected network, and using the fully connected network to output the target power output prediction value of the distributed photovoltaic resources in the corresponding region include: For up to a specific time point time window The mean vector of the corresponding hidden layer space vector is calculated as the encoded feature using the following formula: , No. List For the corresponding number The feature vector encoded by each node; for all , The sparse correlation vector is calculated using the following formula. Dictionary: Capture weighted undirected spatiotemporal graph sequences Input data Components of each node , The corresponding salient spatiotemporal patterns, dictionary and sparse correlation vectors It can be applied in the following ways:

[0025] In the formula, During model training, sparse correlation vectors are randomly initialized. and dictionary And backpropagation optimization is performed based on the error between the model's predicted output value and the actual value; based on the target hidden layer space variables. , ,for Nodes Solving for the dictionary and A sparse pattern Then through mapping Get the current node Output mode characteristics ; The time characteristics obtained from the solution With output mode characteristics Splicing as a complete feature of photovoltaic nodes Input a multi-layer fully connected network and use the ReLU activation function to output the predicted output value of the distributed photovoltaic resources in the corresponding region.

[0026] Secondly, a distributed photovoltaic resource output curve derivation system is provided, including: The spatiotemporal graph construction module is used to input the geographical location of each distributed photovoltaic resource and the historical photovoltaic output data under the corresponding environmental conditions at different times, and to construct a weighted undirected spatiotemporal graph to represent the relationship between distributed photovoltaic resources; The node time feature solving module is used to solve the time features of each node in the weighted undirected spatiotemporal graph. The time features are intermediate prediction data of distributed photovoltaic resource output under environmental conditions at different times. The curve plotting module is used to extract the spatiotemporal features of each node in the weighted undirected spatiotemporal graph. The spatiotemporal features are spatial features that include time features, which are obtained from the weighted undirected spatiotemporal graph. The module captures the output mode features of each node in the spatiotemporal features, concatenates the time features and output mode features of each node and inputs them into a multi-layer fully connected network. The fully connected network is used to output the target output prediction value of the distributed photovoltaic resources in the corresponding region and plot the output projection curve of the distributed photovoltaic resources.

[0027] As a preferred embodiment, the spatiotemporal graph construction module will use time nodes... An undirected graph of an electrical grid is defined as follows: ,in, For a set of nodes, Let be the set of edges. for The feature values ​​of all nodes at time t; Calculate the next node at the current time. With nodes Historical photovoltaic power output data and Correlation coefficient between ; Define the adjacency matrix Indicates photovoltaic resource nodes With nodes The interconnection relationship between them, the adjacency matrix elements in The value is determined as follows:

[0028] In the formula, This is the average of the output correlation coefficients for all node pairs. For a moment Next node With nodes Historical photovoltaic power output data and The Euclidean distance between them; The above formula establishes the connection relationships on the undirected graph of the power grid, forming a weighted undirected spatiotemporal graph representing the correlation degree of different photovoltaic nodes.

[0029] As a preferred embodiment, the node time feature solving module establishes an LSTM network model. The LSTM network model is a Long Short Time Memory network model, which calculates the node data at each time step. Update the cell state using the forget gate. :

[0030]

[0031] In the formula, This represents the forget gate, used to control the cell state at the previous time step. For the current cell state The extent of the impact; This represents the cell state update value. right The effect is caused by the input gate To take control; Update input gate by pressing formula With cell state update value :

[0032]

[0033] The output gates in the LSTM network model are defined by the following formula. Update and calculate the output of the output unit. :

[0034]

[0035] definition , , , , This represents historical weather data, including temperature, humidity, wind speed, and wind direction. , , This represents historical solar irradiance, vertical irradiance, and solar radiation factor data, among which... For time step; The LSTM network model described is a two-layer model. At any given time, an LSTM network model is used to perform regression prediction on the historical output data of the photovoltaic power plant, with the objective being to output... Real-time photovoltaic output data, before use Photovoltaic output data, weather data, and solar irradiance data for specific time segments, for Periodically fit the photovoltaic output data at any given time; By combining the outputs of a two-layer LSTM network model, a mapping function is fitted based on historical photovoltaic power output data. and the corresponding parameter set Output the time features of each node. The expression is as follows:

[0036] in, and These correspond to the two-layer structure of the selected LSTM network model.

[0037] As a preferred embodiment, the node time feature solving module also learns from the source domain. A Source-LSTM model, Each Source-LSTM model is pre-trained using a sufficient amount of data from the corresponding source domain; the base-LSTM model in the target domain generates the basic output. ,Depend on Each Source-LSTM model generates source output. ; The model's output is processed through an attention network. and Perform weighting; The attention network described is designed based on a multilayer perceptron, as shown in the following equation:

[0038] in, For input data, For learnable parameters, This represents predictions made using an LSTM network model. The dimension is The output of the multilayer perceptron is converted into model weights by a soft attention mechanism, and the attention network generates multiple non-zero weights according to the following formula. :

[0039] In the formula, ; Based on the attention network, the output of the complete model is... Defined as a Source-LSTM model Output and Base-LSTM model output The weighted sum is expressed as follows:

[0040] Output The temporal characteristics of each photovoltaic node were obtained. In the target domain, the model's parameters are updated according to two functions: One is the parameters of the Base-LSTM model, output from the target domain. and the true value of the target domain Calculated loss function Update; another is the parameters in the attention network, which are based on the overall output of the model. and the true value of the target domain Calculated loss function renew; In both the Base-LSTM and attention networks, there are two independent backpropagation mechanisms. The loss function for both is calculated using the mean absolute error, as shown in the following expression:

[0041] In the formula, It is the number of photovoltaic nodes. The set of predicted output values ​​for all nodes. The set of actual values ​​of the effort contributed by all nodes. For the first Predicted output values ​​for each node For the first The actual value of the output of each node. It is the set of all learnable parameters.

[0042] As a preferred embodiment, the step of the curve plotting module extracting the spatiotemporal features of each node in the weighted undirected spatiotemporal graph includes: establishing a spatiotemporal graph autoencoder (ST-GAE) and processing the input weighted undirected spatiotemporal graph sequence. Encode and reconstruct data patterns in a weighted undirected spatiotemporal graph; learn deep encoding mapping relationships. Weighted undirected spacetime graph sequence Mapped to a series of hidden space vectors ,exist In the middle, the first Columns represent weighted undirected spacetime graphs The Middle The spatiotemporal features of each node are analyzed; two decoder modules are designed to learn the deep decoding mapping relationship, including feature-level encoding. Encoding at the edge level .

[0043] As a preferred embodiment, the curve plotting module captures the power output mode features of each node in the spatiotemporal characteristics, concatenates the time features and power output mode features of each node and inputs them into a multi-layer fully connected network, and uses the fully connected network to output the target power output prediction value of the distributed photovoltaic resources in the corresponding region. The steps include: For up to a specific time point time window The mean vector of the corresponding hidden layer space vector is calculated as the encoded feature using the following formula: , No. List For the corresponding number The feature vector encoded by each node; for all , The sparse correlation vector is calculated using the following formula. Dictionary: Capture weighted undirected spatiotemporal graph sequences Input data Components of each node , The corresponding salient spatiotemporal patterns, dictionary and sparse correlation vectors It can be applied in the following ways:

[0044] In the formula, During model training, sparse correlation vectors are randomly initialized. and dictionary And backpropagation optimization is performed based on the error between the model's predicted output value and the actual value; based on the target hidden layer space variables. , ,for Nodes Solving for the dictionary and A sparse pattern Then through mapping Get the current node Output mode characteristics ; The time characteristics obtained from the solution With output mode characteristics Splicing as a complete feature of photovoltaic nodes Input a multi-layer fully connected network and use the ReLU activation function to output the predicted output value of the distributed photovoltaic resources in the corresponding region.

[0045] Thirdly, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the distributed photovoltaic resource output curve derivation method.

[0046] Fourthly, a computer-readable storage medium is provided, the computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the distributed photovoltaic resource output curve derivation method.

[0047] Compared with the prior art, the first aspect of the present invention has at least the following beneficial effects: This invention uses a weighted undirected spatiotemporal graph to characterize the relationships between distributed photovoltaic (PV) resources, effectively constructing these relationships. It infers from unstructured data, learning both temporal and untimely characteristics, and thus characterizing the dependence of different nodes on grid output both locally and globally. Simultaneously, by extracting the spatiotemporal features of each node in the weighted undirected spatiotemporal graph, it captures the output pattern features of each node within these features. The temporal and output pattern features of each node are concatenated and input into a multi-layer fully connected network as the complete feature of the PV node. This fully connected network then outputs the target output prediction value of the distributed PV resources in the corresponding region, thereby plotting the distributed PV resource output projection curve. This invention comprehensively utilizes data from various sources, combining meteorological and solar illumination information of the target region for temporal feature extraction. Based on a graph network-driven deep learning algorithm, it extracts spatial features of PV modules in the target region. It does not rely on extremely high-quality observation data or large amounts of historical data, and can accurately and reasonably predict the output of distributed PV resources under extreme weather conditions and emergencies, meeting the prediction performance requirements of long-term output curves for large-scale deployment of distributed PV resources.

[0048] It is understood that the beneficial effects of the second to fourth aspects mentioned above can be found in the relevant descriptions in the first aspect mentioned above, and will not be repeated here. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 Flowchart of the method for extrapolating the output curve of distributed photovoltaic resources according to an embodiment of the present invention; Figure 2A schematic diagram of the temporal feature extraction model based on LSTM network in an embodiment of the present invention; Figure 3 A schematic diagram of the attention transfer learning model structure based on LSTM in an embodiment of the present invention; Figure 4 A schematic diagram of the ST-GAE (Spatiotemporal Graph Autoencoder) model structure according to an embodiment of the present invention; Figure 5 A schematic diagram of a multi-layer fully connected network structure according to an embodiment of the present invention. Detailed Implementation

[0051] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0052] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0053] Example 1 Please see Figure 1 An embodiment of the present invention provides a method for extrapolating the output curve of distributed photovoltaic resources, comprising the following steps: S1: Construction of the output simulation system diagram model; First, a graph structure is constructed based on historical data of photovoltaic nodes, by building time nodes. undirected graph To represent time The situation of all nodes at that time, where, For the set of all photovoltaic nodes, Let be the set of edges, and Used to represent time nodes The feature values ​​of all nodes are considered. Since the output of photovoltaic nodes is greatly affected by the environment, nodes that are geographically close have a high correlation. This invention considers nodes with a correlation coefficient greater than the average level as having a relationship, and then constructs edges for them, ultimately forming a weighted undirected spatiotemporal graph.

[0054] S2: A time feature extraction scheme based on LSTM (Long Short-Term Memory) network, such as... Figure 2 As shown; To extract the temporal dependency features of the input data, an LSTM model is used to model and persist the temporal dependencies. First, update the unit state based on the content of the forget gate. Forgotten Gate Used to control the cell state of the previous time step. To what extent does it affect the current cell state? Input gate Control unit status update value right The impact on the output gates in LSTM Update and calculate the output of the output unit. Designing a two-layer LSTM model for... Regression prediction is performed using historical output data of photovoltaic power plants to output... Real-time photovoltaic power output data.

[0055] S3: Attention transfer learning model based on LSTM (Long Short-Term Memory network), such as Figure 3 As shown; LSTM models suffer from insufficient available business data in real-world deployment scenarios, leading to inaccurate temporal features learned by the model. To address this, this invention proposes an attention transfer learning model based on LSTM. Since the source domain contains sufficient training data, transfer learning draws from the source domain. The Source-LSTM model models rich temporal features, which can better assist in prediction tasks; while the Base-LSTM network in the target domain has low prediction accuracy due to limited data in real training scenarios. Therefore, the source domain model is transferred to the target domain through an attention network that learns to assign weights, which helps to complete the extraction of temporal features in the target domain when data is limited.

[0056] S4: A spatiotemporal feature extraction scheme based on ST-GAE (Spatiotemporal Graph Autoencoder), such as... Figure 4 As shown; Based on the aforementioned weighted undirected spatiotemporal graph, a spatiotemporal graph autoencoder (ST-GAE) is first constructed to encode and reconstruct the input spatiotemporal graph sequence, learning deep encoding mapping relationships. Thus Mapped into a series of hidden space vectors Next, two decoder modules are designed to learn the deep decoding mapping relationship. This encoder-decoder architecture promotes the learning of latent spatiotemporal features in the ST-GAE learning graph.

[0057] S5: Deduction of future photovoltaic system output curves based on spatiotemporal graph dictionary learning; Based on the aforementioned ST-GAE, the time window The mean of the hidden layer space vector is calculated. ,in, For the corresponding number The features are encoded by each node. Then, a sparse correlation vector is calculated. dictionary This is used to capture significant spatiotemporal patterns in the input spatiotemporal graph sequence. The target hidden layer spatial variables are then solved. Afterwards, for Nodes Optimize to obtain dictionary and its A sparse pattern Subsequently, through mapping We can obtain the current time node Output mode feature sequence .

[0058] Based on the aforementioned attention transfer model of the Long Short-Term Memory network, as well as the spatiotemporal graph autoencoder and dictionary learning model, fine-grained spatiotemporal feature information of photovoltaic nodes was obtained, and finally, the power output mode features were concatenated. With time characteristics Input a multi-layer fully connected network, such as Figure 5 As shown, this allows for the projection of the photovoltaic system's output curve at future moments.

[0059] This invention is based on multi-source sensor data from actual deployments. The proposed scheme can comprehensively utilize data from various sources, combine meteorological and solar illumination information of the target area to extract temporal features, extract spatial features of photovoltaic modules in the target area based on graph network-driven deep learning algorithms, and learn appropriate mapping relationships based on long-term and short-term weather conditions, thereby predicting the future output curve of county-level distributed photovoltaic resources. The three main parts of the method of this invention are described in detail below.

[0060] Part 1: Correlation Analysis and Modeling of Photovoltaic Systems Based on Spatiotemporal Graph Model; First, to analyze and model the correlations among distributed photovoltaic (PV) resources in a county-level scenario, this invention constructs relationships among these resources based on a spatiotemporal graph model. It infers from unstructured data, learning its temporal and untemporal characteristics, and thus characterizing the dependence of different nodes on grid output both locally and globally. The proposed scheme characterizes the relationships between PV nodes based on a series of spatiotemporal graphs, effectively constructing relationships between PV resources.

[0061] Specifically, the time nodes An undirected graph of an electrical grid is defined as follows: ,in, For a set of nodes, Then it is the set of edges. Then it is The characteristic values ​​of all nodes at any given time. The output of photovoltaic nodes is greatly affected by the environment. Photovoltaic units geographically close to each other may share similar solar radiation, cloud cover measurements, wind speed, and other weather conditions. To further characterize the output correlation between photovoltaic resource nodes, when the nodes... With nodes When there is a high degree of environmental correlation between nodes, With nodes The nodes are connected. To characterize this correlation, methods such as cosine similarity can be used to calculate the node at the current time. With nodes Historical photovoltaic power output data and Correlation coefficient between .

[0062] In addition, define the adjacency matrix. Indicates photovoltaic resource nodes With nodes Interconnection relationships between them. Adjacency matrix. elements in The value is determined as follows:

[0063] in, This is the average of the output correlation coefficients for all node pairs. For a moment Next node With nodes Historical photovoltaic power output data and The Euclidean distance between them. This invention proposes that when the correlation coefficient of the historical output data of two photovoltaic power plants is higher than the average level, the graph at the corresponding time... There are correlations between them. Based on the above formula, the connection relationships on the graph can be constructed, ultimately forming a weighted undirected spatiotemporal graph representing the correlations between different photovoltaic nodes.

[0064] Part 2: Temporal Feature Extraction and Attention Transfer Learning Based on LSTM (Long Short-Term Memory) Networks; 2.1 Temporal Feature Extraction Model Based on LSTM Network; Based on the aforementioned weighted undirected spatiotemporal graph, this invention performs multi-dimensional mining of the spatiotemporal characteristics between nodes to solve for the temporal features of the nodes in the graph. Specifically, an LSTM (Long Short-Term Memory) network model is used to process the aforementioned weighted undirected spatiotemporal graph to solve for the temporal features of the nodes. The proposed model handles the node data at each time step. Update the cell state using the forget gate. :

[0065]

[0066] in This represents the forget gate, used to control the cell state at the previous time step. To what extent does it affect the current cell state? . This represents the cell state update value, which is related to... The effect is caused by the input gate To take control.

[0067] Next, update the input gate using the following scheme. With cell state update value :

[0068]

[0069] Finally, the output gates in the LSTM Update and calculate the output of the output unit. This serves as the spatiotemporal characteristic of the node.

[0070]

[0071]

[0072] Considering the strong correlation between the state of photovoltaic modules and past information, this invention uses an LSTM network to extract the dependencies in time-series data, thereby enabling the extraction of temporal features in distributed photovoltaic scenarios. Definition , , , , This represents historical weather data, including temperature, humidity, wind speed, and wind direction. , , This represents historical solar irradiance, vertical irradiance, and solar radiation factor data, among which... The time step is defined as [missing information]. To characterize the output curve information of time-series data at different granularities, this invention constructs a two-layer LSTM model. First, in [missing information]... At any given time, LSTM is used to perform regression prediction on the historical output data of a photovoltaic power station, with the objective being to output... Real-time photovoltaic output data. Subsequently, before use... A time segment, namely The photovoltaic power output data, weather data, and sunshine data are used for... The photovoltaic power output data at any given time is periodically fitted. Then, by combining the outputs of two LSTM modules, a mapping function is fitted based on historical photovoltaic power output data. and the corresponding parameter set Output the time features of each node. :

[0073] in, and These correspond to the two-layer structure of the selected LSTM.

[0074] In particular, considering that the proposed solution suffers from insufficient available business data in actual deployment scenarios, making it difficult to train the LSTM model, this invention further proposes an attention transfer learning model based on LSTM.

[0075] 2.2 Attention Transfer Learning Model Based on LSTM Inspired by the applications of attention and adaptive transfer learning in deep learning, this invention proposes an attention transfer learning framework based on an LSTM model for transferring knowledge from an LSTM model trained in the source domain. This framework learns rich spatiotemporal features and dependencies in the source domain using existing large-scale datasets, and trains on data with limited features in the target domain. Finally, the two are combined through an attention network, and parameters are updated using a loss function to complete the transfer learning task.

[0076] For the source domain part in transfer learning, learning from the source domain A Source-LSTM model is proposed. Considering that LSTM models trained on source domain data can extract rich temporal features and perform well in regression prediction tasks, a transfer learning framework for Source-LSTM is designed to transfer rich temporal feature knowledge from the source domain to the target domain.

[0077] Regarding the target domain, since the data in the target domain all come from real-world scenarios, the Base-LSTM model trained with limited data in the target domain often has low prediction accuracy and needs to be combined with the rich temporal features learned from the source domain. Each Source-LSTM model is pre-trained using a sufficient amount of data from the corresponding source domain. The Base-LSTM model generates the basic output. , Individual source model generates source output .

[0078] This invention designs an attention network to monitor the model's output. and Weights are then processed. The attention network is designed based on a multi-layer perceptron (MLP) and can be represented as follows:

[0079] in, For input data, For learnable parameters, This indicates that predictions are made using an LSTM model. The dimension is The output of the MLP is converted into model weights by a soft attention mechanism; the attention network can generate multiple non-zero weights. :

[0080] in .

[0081] Based on the attention network, the output of the complete model is... Defined as a Source-LSTM model Output and Base-LSTM model output Weighted sum:

[0082] This invention designs different methods to update the parameters in transfer learning and attention networks. The source-LSTM models in the source domain are pre-trained using source domain data, and their parameters remain unchanged throughout the training of the transfer learning models. In the target domain, the model parameters are updated according to two functions: one is the parameters of the Base-LSTM, output from the target domain. and the true value of the target domain Calculated loss function Update; one is the parameters in the attention network, which are based on the overall output of the model. and the true value of the target domain Calculated loss function renew.

[0083] In both the Base-LSTM and attention networks, there are two independent backpropagation propagations. Both loss functions are calculated using the mean absolute error.

[0084] In the formula, It is the number of photovoltaic nodes. The set of predicted output values ​​for all nodes. The set of actual values ​​of the effort contributed by all nodes. For the first Predicted output values ​​for each node For the first The actual value of the output of each node. It is the set of all learnable parameters.

[0085] Finally, the transfer model outputs The temporal characteristics of each photovoltaic node were obtained.

[0086] Part Three: Spatiotemporal Feature Extraction Scheme and Future Photovoltaic System Output Curve Deduction; Based on the weighted undirected spatiotemporal graph of photovoltaic nodes, this invention uses ST-GAE (Spatiotemporal Graph Autoencoder) to extract the spatiotemporal features between nodes in the weighted undirected spatiotemporal graph, and further uses a fully connected network to extrapolate the output curve of the future photovoltaic system.

[0087] 3.1 Spatiotemporal feature extraction scheme based on ST-GAE (Spatiotemporal Graph Autoencoder) network model; First, a spatiotemporal graph autoencoder (ST-GAE) is established to process the input spatiotemporal graph sequence. Encoding and reconstruction are performed to capture meaningful data patterns in the spatiotemporal graph. During this process, deep encoding mapping relationships are learned. Thus Mapped to a series of hidden space vectors . In the middle, the first Column Representation Diagram The Middle The spatiotemporal features of each node are analyzed. Based on this, two decoder modules are designed to learn the deep decoding mapping relationship, including feature-level encoding. Encoding at the edge level These two encoding-decoding architectures facilitate ST-GAE learning. Medium to highly nonlinear spatiotemporal latent features.

[0088] 3.2 Derivation of future photovoltaic system output curves based on spatiotemporal graph dictionary learning; Next, this invention performs in-depth analysis of the spatiotemporal graph. For a specific time point... time window The mean vector of the corresponding hidden layer space vector is calculated as the encoded feature. Its first List For the corresponding number The feature vector encoded by each node. For all Calculate the sparse correlation vector for it. dictionary to capture The first input data Components of each node The corresponding salient spatiotemporal patterns. (Referring to the aforementioned dictionary) and sparse correlation vectors Apply it using the following method:

[0089] In the formula, The fitted value for the corresponding vector. During model training, sparse correlation vectors are randomly initialized. and dictionary And backpropagation optimization is performed based on the error between the model's predicted and actual output values. Based on the target hidden layer spatial variables... , can be Nodes Solving for the dictionary and A sparse pattern Subsequently, through mapping We can obtain the current time node Output mode characteristics .

[0090] Finally, the time characteristics obtained from the above solution are... With output mode characteristics Splicing as a complete feature of photovoltaic nodes Input a multi-layer fully connected network and use the ReLU activation function to output the predicted output value of the distributed photovoltaic resources in the corresponding region.

[0091] Example 2 Another embodiment of the present invention also provides a distributed photovoltaic resource output curve extrapolation system, comprising: The spatiotemporal graph construction module is used to input the geographical location of each distributed photovoltaic resource and the historical photovoltaic output data under the corresponding environmental conditions at different times, and to construct a weighted undirected spatiotemporal graph to represent the relationship between distributed photovoltaic resources; The node time feature solving module is used to solve the time features of each node in the weighted undirected spatiotemporal graph. The time features are intermediate prediction data of distributed photovoltaic resource output under environmental conditions at different times. The curve plotting module is used to extract the spatiotemporal features of each node in the weighted undirected spatiotemporal graph. The spatiotemporal features are spatial features that include time features, which are obtained from the weighted undirected spatiotemporal graph. The module captures the output mode features of each node in the spatiotemporal features, concatenates the time features and output mode features of each node and inputs them into a multi-layer fully connected network. The fully connected network is used to output the target output prediction value of the distributed photovoltaic resources in the corresponding region and plot the output projection curve of the distributed photovoltaic resources.

[0092] In one possible implementation, the spatiotemporal graph construction module will use time nodes. An undirected graph of an electrical grid is defined as follows: ,in, For a set of nodes, Let be the set of edges. for The feature values ​​of all nodes at time t; Calculate the next node at the current time. With nodes Historical photovoltaic power output data and Correlation coefficient between ; Define the adjacency matrix Indicates photovoltaic resource nodes With nodes The interconnection relationship between them, the adjacency matrix elements in The value is determined as follows:

[0093] In the formula, This is the average of the output correlation coefficients for all node pairs. For a moment Next node With nodes Historical photovoltaic power output data and The Euclidean distance between them; The above formula establishes the connection relationships on the undirected graph of the power grid, forming a weighted undirected spatiotemporal graph representing the correlation degree of different photovoltaic nodes.

[0094] In one possible implementation, the node time feature solving module establishes an LSTM network model, and for the node data at each time step... Update the cell state using the forget gate. :

[0095]

[0096] In the formula, This represents the forget gate, used to control the cell state at the previous time step. For the current cell state The extent of the impact; This represents the cell state update value. right The effect is caused by the input gate To take control; Update input gate by pressing formula With cell state update value :

[0097]

[0098] The output gates in the LSTM network model are defined by the following formula. Update and calculate the output of the output unit. :

[0099]

[0100] definition , , , , This represents historical weather data, including temperature, humidity, wind speed, and wind direction. , , This represents historical solar irradiance, vertical irradiance, and solar radiation factor data, among which... For time step; The LSTM network model described is a two-layer model. At any given time, an LSTM network model is used to perform regression prediction on the historical output data of the photovoltaic power plant, with the objective being to output... Real-time photovoltaic output data, before use Photovoltaic output data, weather data, and solar irradiance data for specific time segments, for Periodically fit the photovoltaic output data at any given time; By combining the outputs of a two-layer LSTM network model, a mapping function is fitted based on historical photovoltaic power output data. and the corresponding parameter set Output the time features of each node. The expression is as follows:

[0101] in, and These correspond to the two-layer structure of the selected LSTM network model.

[0102] In one possible implementation, the node temporal feature solving module also learns from the source domain. A Source-LSTM model, Each Source-LSTM model is pre-trained using a sufficient amount of data from the corresponding source domain; the base-LSTM model in the target domain generates the basic output. ,Depend on Each Source-LSTM model generates source output. ; The model's output is processed through an attention network. and Perform weighting; The attention network described is designed based on a multilayer perceptron, as shown in the following equation:

[0103] in, For input data, For learnable parameters, This represents predictions made using an LSTM network model. The dimension is The output of the multilayer perceptron is converted into model weights by a soft attention mechanism, and the attention network generates multiple non-zero weights according to the following formula. :

[0104] In the formula, ; Based on the attention network, the output of the complete model is... Defined as a Source-LSTM model Output and Base-LSTM model output The weighted sum is expressed as follows:

[0105] Output The temporal characteristics of each photovoltaic node were obtained. In the target domain, the model's parameters are updated according to two functions: One is the parameters of the Base-LSTM model, output from the target domain. and the true value of the target domain Calculated loss function Update; another is the parameters in the attention network, which are based on the overall output of the model. and the true value of the target domain Calculated loss function renew; In both the Base-LSTM and attention networks, there are two independent backpropagation mechanisms. The loss function for both is calculated using the mean absolute error, as shown in the following expression:

[0106] In the formula, It is the number of photovoltaic nodes. The set of predicted output values ​​for all nodes. The set of actual values ​​of the effort contributed by all nodes. For the first Predicted output values ​​for each node For the first The actual value of the output of each node. It is the set of all learnable parameters.

[0107] In one possible implementation, the step of the curve plotting module extracting the spatiotemporal features of each node in the weighted undirected spatiotemporal graph includes: establishing a spatiotemporal graph autoencoder (ST-GAE) and processing the input weighted undirected spatiotemporal graph sequence. Encode and reconstruct data patterns in a weighted undirected spatiotemporal graph; learn deep encoding mapping relationships. Weighted undirected spacetime graph sequence Mapped to a series of hidden space vectors ,exist In the middle, the first Columns represent weighted undirected spacetime graphs The Middle The spatiotemporal features of each node are analyzed; two decoder modules are designed to learn the deep decoding mapping relationship, including feature-level encoding. Encoding at the edge level .

[0108] In one possible implementation, the steps of the curve plotting module capturing the power output mode features of each node in the spatiotemporal features, concatenating the time features and power output mode features of each node and inputting them into a multi-layer fully connected network, and using the fully connected network to output the target power output prediction value of the distributed photovoltaic resources in the corresponding region include: For up to a specific time point time window The mean vector of the corresponding hidden layer space vector is calculated as the encoded feature using the following formula: , No. List For the corresponding number The feature vector encoded by each node; for all , The sparse correlation vector is calculated using the following formula. Dictionary: Capture weighted undirected spatiotemporal graph sequences Input data Components of each node , The corresponding salient spatiotemporal patterns, dictionary and sparse correlation vectors It can be applied in the following ways:

[0109] In the formula, During model training, sparse correlation vectors are randomly initialized. and dictionary And backpropagation optimization is performed based on the error between the model's predicted output value and the actual value; based on the target hidden layer space variables. , ,for Nodes Solving for the dictionary and A sparse pattern Then through mapping Get the current node Output mode characteristics ; The time characteristics obtained from the solution With output mode characteristics Splicing as a complete feature of photovoltaic nodes Input a multi-layer fully connected network and use the ReLU activation function to output the predicted output value of the distributed photovoltaic resources in the corresponding region.

[0110] Example 3 Another embodiment of the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the distributed photovoltaic resource output curve derivation method.

[0111] Example 4 Another embodiment of the present invention provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the steps of the distributed photovoltaic resource output curve derivation method.

[0112] The computer program includes computer program code, which can be in the form of source code, object code, executable file, or some intermediate form. The computer-readable storage medium can include any entity or device capable of carrying the computer program code, a medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory, a random access memory, an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, the computer-readable medium does not include electrical carrier signals and telecommunication signals. For ease of explanation, the above content only shows the parts related to the embodiments of the present invention; for specific technical details not disclosed, please refer to the method section of the embodiments of the present invention. This computer-readable storage medium is non-transitory and can be stored in storage devices formed by various electronic devices, enabling the execution process described in the method of the embodiments of the present invention.

[0113] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0114] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0115] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0116] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for deriving the output curve of distributed photovoltaic resources, characterized in that, include: Input the geographical location of each distributed photovoltaic resource and the historical photovoltaic output data under different environmental conditions at different times, and construct a weighted undirected spatiotemporal graph to represent the relationship between distributed photovoltaic resources; Solve for the temporal characteristics of each node in the weighted undirected spatiotemporal graph. The temporal characteristics are intermediate prediction data of distributed photovoltaic resource output under environmental conditions at different times. The spatiotemporal features of each node in the weighted undirected spatiotemporal graph are extracted. The spatiotemporal features are spatial features that include time features, and the spatial features are obtained from the weighted undirected spatiotemporal graph. The output mode features of each node in the spatiotemporal features are captured. The time features and output mode features of each node are concatenated and input into a multi-layer fully connected network. The target output prediction value of the distributed photovoltaic resources in the corresponding region is output using the fully connected network, and the output projection curve of the distributed photovoltaic resources is plotted. Time nodes An undirected graph of an electrical grid on a network is defined as follows: ,in, For the set of all photovoltaic resource nodes, Let be the set of edges. for The feature values ​​of all nodes at time t; The steps for solving the temporal characteristics of each node in the weighted undirected spacetime graph include: Establish an LSTM network model. The LSTM network model is a Long Short Time Memory network model. For the node data at each time step... Update the cell state using the forget gate. : In the formula, This represents the forget gate, used to control the cell state at the previous time step. For the current cell state The degree of impact; This represents the cell state update value. right The effect is caused by the input gate To take control; Update input gate by pressing formula With cell state update value : The output gates in the LSTM network model are defined by the following formula. Update and calculate the output of the output unit. : definition , , , , This represents historical weather data, including temperature, humidity, wind speed, and wind direction. , , This represents historical solar irradiance, vertical irradiance, and solar radiation factor data, among which... For time step; The LSTM network model described is a two-layer model. At any given time, an LSTM network model is used to perform regression prediction on the historical output data of the photovoltaic power plant, with the objective being to output... Real-time photovoltaic output data, before use Photovoltaic output data, weather data, and solar irradiance data for specific time segments, for Periodically fit the photovoltaic output data at any given time; By combining the outputs of a two-layer LSTM network model, a mapping function is fitted based on historical photovoltaic power output data. and the corresponding parameter set Output the time features of each node. The expression is as follows: in, and These correspond to the two-layer structure of the selected LSTM network model; The steps for solving the temporal features of each node in the weighted undirected spacetime graph also include: learning from the source domain. A Source-LSTM model, Each Source-LSTM model is pre-trained using a sufficient amount of data from the corresponding source domain; the base-LSTM model in the target domain generates the basic output. ,Depend on Each Source-LSTM model generates source output. ; The model's output is processed through an attention network. and Perform weighting; The attention network described is designed based on a multilayer perceptron, as shown in the following equation: in, For input data, For learnable parameters, This represents predictions made using an LSTM network model. The dimension is The output of the multilayer perceptron is converted into model weights by a soft attention mechanism, and the attention network generates multiple non-zero weights according to the following formula. : In the formula, ; Based on the attention network, the output of the complete model is... Defined as a Source-LSTM model Output and Base-LSTM model output The weighted sum is expressed as follows: Output The temporal characteristics of each photovoltaic node were obtained.

2. The method for extrapolating the output curve of distributed photovoltaic resources according to claim 1, characterized in that, The steps for constructing a weighted undirected spatiotemporal graph to characterize the relationships between distributed photovoltaic resources include: Calculate the next node at the current time. With nodes Historical photovoltaic power output data and Correlation coefficient between ; Define the adjacency matrix Indicates photovoltaic resource nodes With nodes The interconnection relationship between them, the adjacency matrix elements in The value is determined as follows: In the formula, This is the average of the output correlation coefficients for all node pairs. For a moment Next node With nodes Historical photovoltaic power output data and The Euclidean distance between them; The above formula establishes the connection relationships on the undirected graph of the power grid, forming a weighted undirected spatiotemporal graph representing the correlation degree of different photovoltaic nodes.

3. The method for extrapolating the output curve of distributed photovoltaic resources according to claim 1, characterized in that, In the target domain, the model's parameters are updated according to two functions: One is the parameters of the Base-LSTM model, output from the target domain. and the true value of the target domain Calculated loss function Update; another is the parameters in the attention network, which are based on the overall output of the model. and the true value of the target domain Calculated loss function renew; In both the Base-LSTM and attention networks, there are two independent backpropagation mechanisms. The loss function for both is calculated using the mean absolute error, as shown in the following expression: In the formula, It is the number of photovoltaic nodes. The set of predicted output values ​​for all nodes. The set of actual values ​​of the effort contributed by all nodes. For the first Predicted output values ​​for each node For the first The actual value of the output of each node. It is the set of all learnable parameters.

4. The method for extrapolating the output curve of distributed photovoltaic resources according to claim 1, characterized in that, The steps for extracting the spatiotemporal features of each node in the weighted undirected spatiotemporal graph include: establishing a spatiotemporal graph autoencoder (ST-GAE) and processing the input weighted undirected spatiotemporal graph sequence. Encode and reconstruct data patterns in a weighted undirected spatiotemporal graph; learn deep encoding mapping relationships. , For the target hidden layer space variables, The number of nodes will be used to weight the undirected spacetime graph sequence. Mapped to a series of hidden space vectors ,exist In the middle, the first Columns represent weighted undirected spacetime graphs The Middle The spatiotemporal features of each node are analyzed; two decoder modules are designed to learn the deep decoding mapping relationship, including feature-level encoding. Encoding at the edge level .

5. The method for extrapolating the output curve of distributed photovoltaic resources according to claim 4, characterized in that, The steps of capturing the output mode features of each node in the spatiotemporal features, concatenating the temporal features and output mode features of each node and inputting them into a multi-layer fully connected network, and using the fully connected network to output the target output prediction value of the distributed photovoltaic resources in the corresponding region include: For up to a specific time point time window The mean vector of the corresponding hidden layer space vector is calculated as the encoded feature using the following formula: , No. List For the corresponding number The feature vector encoded by each node; for all , The sparse correlation vector is calculated using the following formula. Dictionary: Capture weighted undirected spatiotemporal graph sequences Input data Components of each node , The corresponding salient spatiotemporal patterns, dictionary and sparse correlation vectors It can be applied in the following ways: In the formula, During model training, sparse correlation vectors are randomly initialized. and dictionary And backpropagation optimization is performed based on the error between the model's predicted output value and the actual value; based on the target hidden layer space variables. , ,for Nodes Solving for the dictionary and A sparse pattern Then through mapping Get the current node Output mode characteristics ; The time characteristics obtained from the solution Characteristics of power output mode Splicing as a complete feature of photovoltaic nodes Input a multi-layer fully connected network and use the ReLU activation function to output the predicted output value of the distributed photovoltaic resources in the corresponding region.

6. A distributed photovoltaic resource output curve derivation system, characterized in that, include: The spatiotemporal graph construction module is used to input the geographical location of each distributed photovoltaic resource and the historical photovoltaic output data under the corresponding environmental conditions at different times, and to construct a weighted undirected spatiotemporal graph to represent the relationship between distributed photovoltaic resources; The node time feature solving module is used to solve the time features of each node in the weighted undirected spatiotemporal graph. The time features are intermediate prediction data of distributed photovoltaic resource output under environmental conditions at different times. The curve plotting module is used to extract the spatiotemporal features of each node in the weighted undirected spatiotemporal graph. The spatiotemporal features are spatial features that include time features, and the spatial features are obtained from the weighted undirected spatiotemporal graph. The module captures the output mode features of each node in the spatiotemporal features, concatenates the time features and output mode features of each node and inputs them into a multi-layer fully connected network. The fully connected network is used to output the target output prediction value of the distributed photovoltaic resources in the corresponding region and plot the output projection curve of the distributed photovoltaic resources. The spatiotemporal graph construction module will use time nodes An undirected graph of an electrical grid on a network is defined as follows: ,in, For the set of all photovoltaic resource nodes, Let be the set of edges. for The feature values ​​of all nodes at time t; The node time feature solving module establishes an LSTM network model, which is a Long Short Time Memory network model, for the node data at each time step. Update the cell state using the forget gate. : In the formula, This represents the forget gate, used to control the cell state at the previous time step. For the current cell state The degree of impact; This represents the cell state update value. right The effect is caused by the input gate To take control; Update input gate by pressing formula With cell state update value : The output gates in the LSTM network model are defined by the following formula. Update and calculate the output of the output unit. : definition , , , , This represents historical weather data, including temperature, humidity, wind speed, and wind direction. , , This represents historical solar irradiance, vertical irradiance, and solar radiation factor data, among which... For time step; The LSTM network model described is a two-layer model. At any given time, an LSTM network model is used to perform regression prediction on the historical output data of the photovoltaic power plant, with the objective being to output... Real-time photovoltaic output data, before use Photovoltaic output data, weather data, and solar irradiance data for specific time segments, for Periodically fit the photovoltaic output data at any given time; By combining the outputs of a two-layer LSTM network model, a mapping function is fitted based on historical photovoltaic power output data. and the corresponding parameter set Output the time features of each node. The expression is as follows: in, and These correspond to the two-layer structure of the selected LSTM network model; The node time feature solving module also learns from the source domain. A Source-LSTM model, Each Source-LSTM model is pre-trained using a sufficient amount of data from the corresponding source domain; the base-LSTM model in the target domain generates the basic output. ,Depend on Each Source-LSTM model generates source output. ; The model's output is processed through an attention network. and Perform weighting; The attention network described is designed based on a multilayer perceptron, as shown in the following equation: in, For input data, For learnable parameters, This represents predictions made using an LSTM network model. The dimension is The output of the multilayer perceptron is converted into model weights by a soft attention mechanism, and the attention network generates multiple non-zero weights according to the following formula. : In the formula, ; Based on the attention network, the output of the complete model is... Defined as a Source-LSTM model Output and Base-LSTM model output The weighted sum is expressed as follows: Output The temporal characteristics of each photovoltaic node were obtained.

7. The distributed photovoltaic resource output curve extrapolation system according to claim 6, characterized in that, The spatiotemporal graph construction module calculates the nodes at the current time. With nodes Historical photovoltaic power output data and Correlation coefficient between ; Define the adjacency matrix Indicates photovoltaic resource nodes With nodes The interconnection relationship between them, the adjacency matrix elements in The value is determined as follows: In the formula, This is the average of the output correlation coefficients for all node pairs. For a moment Next node With nodes Historical photovoltaic power output data and The Euclidean distance between them; The above formula establishes the connection relationships on the undirected graph of the power grid, forming a weighted undirected spatiotemporal graph representing the correlation degree of different photovoltaic nodes.

8. The distributed photovoltaic resource output curve extrapolation system according to claim 6, characterized in that, In the target domain, the nodal time feature solving module updates the model parameters according to two functions: One is the parameters of the Base-LSTM model, output from the target domain. and the true value of the target domain Calculated loss function Update; another is the parameters in the attention network, which are based on the overall output of the model. and the true value of the target domain Calculated loss function renew; In both the Base-LSTM and attention networks, there are two independent backpropagation mechanisms. The loss function for both is calculated using the mean absolute error, as shown in the following expression: In the formula, It is the number of photovoltaic nodes. The set of predicted output values ​​for all nodes. The set of actual values ​​of the effort contributed by all nodes. For the first Predicted output values ​​for each node For the first The actual value of the output of each node. It is the set of all learnable parameters.

9. The distributed photovoltaic resource output curve extrapolation system according to claim 6, characterized in that, The curve plotting module extracts the spatiotemporal features of each node in the weighted undirected spatiotemporal graph through the following steps: establishing a spatiotemporal graph autoencoder (ST-GAE) and processing the input weighted undirected spatiotemporal graph sequence. Encode and reconstruct data patterns in a weighted undirected spatiotemporal graph; learn deep encoding mapping relationships. , For the target hidden layer space variables, The number of nodes will be used to weight the undirected spacetime graph sequence. Mapped to a series of hidden space vectors ,exist In the middle, the first Columns represent weighted undirected spacetime graphs The Middle The spatiotemporal features of each node are analyzed; two decoder modules are designed to learn the deep decoding mapping relationship, including feature-level encoding. Encoding at the edge level .

10. The distributed photovoltaic resource output curve extrapolation system according to claim 9, characterized in that, The curve plotting module captures the output mode features of each node in the spatiotemporal features, concatenates the time features and output mode features of each node and inputs them into a multi-layer fully connected network, and uses the fully connected network to output the target output prediction value of the distributed photovoltaic resources in the corresponding region. The steps include: For up to a specific time point time window The mean vector of the corresponding hidden layer space vector is calculated as the encoded feature using the following formula: , No. List For the corresponding number The feature vector encoded by each node; for all , The sparse correlation vector is calculated using the following formula. Dictionary: Capture weighted undirected spatiotemporal graph sequences Input data Components of each node , The corresponding salient spatiotemporal patterns, dictionary and sparse correlation vectors It can be applied in the following ways: In the formula, During model training, sparse correlation vectors are randomly initialized. and dictionary And backpropagation optimization is performed based on the error between the model's predicted output value and the actual value; based on the target hidden layer space variables. , ,for Nodes Solving for the dictionary and A sparse pattern Then through mapping Get the current node Output mode characteristics ; The time characteristics obtained from the solution Characteristics of power output mode Splicing as a complete feature of photovoltaic nodes Input a multi-layer fully connected network and use the ReLU activation function to output the predicted output value of the distributed photovoltaic resources in the corresponding region.

11. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the steps of the distributed photovoltaic resource output curve derivation method as described in any one of claims 1 to 5.

12. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of the distributed photovoltaic resource output curve derivation method as described in any one of claims 1 to 5.

Images