Distributed photovoltaic access area-oriented net load prediction method

Abstract

The invention discloses a distributed photovoltaic access area-oriented net load prediction method, which relates to the technical field of photovoltaic power generation and is used for carrying out net load prediction by constructing a graph structure data form based on geographic coordinate data, meteorological characteristics and a user behavior mechanism and combining a graph neural network and a Transform architecture suitable for time sequence characteristic extraction. According to the method, a design of combining a graph neural network with a Transform architecture is adopted, and the graph neural network aggregates node neighborhood information through a message passing mechanism and captures local spatial dependence, such as correlation of photovoltaic output of adjacent regions; the Transformer architecture globally models a space-time sequence by means of a self-attention mechanism, automatically learns long-term dependency, effectively solves the problem of long-range dependency deficiency of a traditional time sequence model, and through the synergistic effect of the two, the model has local space perception and global space-time reasoning capabilities at the same time, the net load prediction precision is improved, and the prediction efficiency is improved. And key technical support is provided for efficient utilization of distributed energy and safe operation of a power grid.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation technology, and more specifically, to a net load forecasting method for distributed photovoltaic (PV) grid connection areas. Background Technology

[0002] Driven by energy transition, distributed photovoltaic (PV) power has been deployed on a large scale in rural areas of provinces, becoming an important way to popularize clean energy. However, these areas have low load density, scattered geographical distribution, and micro-topographical differences that can easily cause significant fluctuations in local meteorological conditions, making it difficult to match PV output with user load in time and space. At the same time, net load data is affected by the coupling between PV and load, exhibiting strong nonlinearity and randomness, which increases the complexity of prediction.

[0003] Traditional modeling relies on the physical topology of the power grid to define node connections, which cannot capture local spatial relationships under the coupling of "geography-meteorology-users", such as the correlation of photovoltaic output in adjacent areas, and it is difficult to aggregate the neighborhood information of nodes. At the same time, traditional time series models lack global spatiotemporal modeling capabilities and cannot effectively learn the long-term dependence and short-term fluctuations of net load. In addition, a single model cannot simultaneously possess local spatial perception and global spatiotemporal reasoning capabilities, resulting in insufficient accuracy of net load prediction. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, the present invention provides a net load forecasting method for distributed photovoltaic (PV) grid connection areas to address the problems mentioned in the background art.

[0005] This application provides a net load forecasting method for distributed photovoltaic (PV) grid integration areas, applied to the power grid dispatching system of such areas. The method includes:

[0006] Collect geographic coordinate data of distributed photovoltaic installations and user load nodes, load curve data of user nodes for at least one year, and meteorological data corresponding to the time axis of the load curve data;

[0007] Based on geographic coordinate data, meteorological data, corresponding meteorological characteristics, and user behavior mechanisms, a graph-structured data format is constructed. ,in Contains a set of nodes Adjacency matrix of multi-relation edges Node set Each node Associated with geographic latitude and longitude coordinates Distributed photovoltaic installed capacity and time-varying user behavior trend feature vector and static features ;

[0008] Using graph neural networks to process graph structured data. Spatial correlation fusion is performed on the node features to extract the node set. internal nodes The spatial relationship characteristics are used to obtain time series data of user nodes that are fused with spatial correlations; the input of the graph neural network includes time indexes. The full-graph node feature matrix below and normalized adjacency matrix ,in It is a composite adjacency matrix. For degree matrix, These are the matrix weights;

[0009] A deep learning model using the Transformer architecture is constructed, and the Transformer architecture is used to extract load curve data of user nodes and distributed photovoltaic installed capacity. The corresponding photovoltaic power output data time series mode, and the deep learning model takes the fused spatially correlated user node time series data as input, and outputs node Projected future net load.

[0010] In some embodiments of this application, the distributed photovoltaic installed capacity Used to indicate the first Whether the node has a photovoltaic system installed, and the capacity of the distributed photovoltaic system. Set binary indicator To further clarify the nodes Whether to install a photovoltaic system, if the node Without a photovoltaic system installed, the distributed photovoltaic installed capacity is... =0; Time-varying user behavior trend feature vector Including historical net load Photovoltaic output provided by observation or forecasting agencies Local weather conditions and time- and date-related indicators, including temperature. and radiation Time and date related indicators include one-hot encoding of hours or holidays.

[0011] In some embodiments of this application, graph structure data is constructed. At that time, graph structure data Weight of the middle edge Based on geographical proximity Meteorological similarity Historical net load correlation Photovoltaic or similar capacity relationship Similarity to user attributes or land use Calculated;

[0012] Node set The connection structure between internal nodes is determined by the full graph node feature matrix. Obtained by calculating the distances between different nodes With X t-1 The absolute value of the cosine similarity of the difference is used to determine the actual dynamic relationship between nodes;

[0013] Simultaneously set threshold Used to filter closely related nodes to reduce data redundancy in the algorithm, when the absolute value of cosine similarity is greater than a threshold. When the connection is active, the two nodes remain connected; otherwise, they are disconnected.

[0014] In some embodiments of this application, geographical proximity The weight formula is ,in It is a node and nodes The distance between them, σ is the standard deviation, if Exceeding the preset maximum distance ,but .

[0015] In some embodiments of this application, historical net load correlation weight Determined by the similarity of historical net load sequences, the calculation formula is as follows: ,in Represents a node and nodes The correlation coefficient of the historical net load series, Corresponding nodes , The historical net load sequence.

[0016] In some embodiments of this application, a similarity measurement method based on climate factors is used to measure meteorological similarity. The weighting formula is ,in Represents a node and nodes A measure of difference on the meteorological feature vector R. Corresponding nodes , The meteorological feature vector.

[0017] In some embodiments of this application, the distributed photovoltaic installed capacity Constructing photovoltaic or capacity similarity relationships based on capacity factors When, its weighting formula is ,in To adjust the node , Distributed photovoltaic installed capacity , The parameter that determines the contribution of differences to similarity.

[0018] In some embodiments of this application, user attributes and land use similarity relationships are discussed. The weight formula is ,in Represents a node , Distance measure between corresponding user attributes or land use patterns Parameters for adjusting the degree of influence of differences in user attributes;

[0019] Geographical proximity Meteorological similarity Historical net load correlation Photovoltaic or similar capacity relationship Similarity to user attributes or land use The graph structure data is determined by fusion using a weighted method. The weight of each edge in the equation.

[0020] In some embodiments of this application, the Transformer architecture is based on a multi-head attention mechanism, preserves temporal sequence information through positional encoding, and extracts global and local temporal features in parallel using an encoder-decoder architecture, as shown in the following formula:

[0021]

[0022] in, , , , The weight parameter matrix, It is a high-dimensional power sequence The resulting query matrix, key matrix, and value matrix are obtained through transformation. Indicates the length of the input sequence. This represents the output of the self-attention mechanism. This represents the dimension of the key matrix. The number of heads representing multi-head attention. This is the aggregated time-series feature sequence.

[0023] In some embodiments of this application, the batch gradient descent algorithm is used to train the deep learning model, and the mean squared error is used as the basic loss function. The formula for the basic loss function is as follows: ,in The node predicted by the model Future net load value, It is a node Future net load true value For the sample size, To predict the step size.

[0024] Compared with the prior art, the beneficial effects of the present invention are:

[0025] 1. This invention adopts a design that combines graph neural networks and Transformer architecture. The graph neural network aggregates the neighborhood information of nodes through the message passing mechanism to capture local spatial dependencies, such as the correlation of photovoltaic power output in adjacent areas. The Transformer architecture uses the self-attention mechanism to globally model the spatiotemporal sequence and automatically learn long-term dependencies, effectively solving the problem of missing long-range dependencies in traditional time series models. The two work together to enable the model to have both local spatial perception and global spatiotemporal reasoning capabilities, thereby improving the accuracy of net load prediction.

[0026] 2. When constructing graph structure data, the connection structure between nodes is dynamically determined by the absolute value of the cosine similarity of the difference between the feature matrices of all nodes in the graph. A threshold is set to filter closely related nodes. This dynamic modeling method can automatically adjust the node connection weights according to real-time meteorological data, avoid prediction deviations caused by spatiotemporal changes in static graph structures, and enhance the generalization ability of the model.

[0027] 3. Traditional methods rely on the physical topology of the power grid to define node connections. However, this invention constructs a graph structure by integrating geographical and meteorological mechanisms with user behavior characteristics. It is not limited to physical connections. By dynamically constructing graph connection relationships based on meteorological similarities such as solar radiation and cloud cover, and using date and time encoding to capture similarities in user electricity consumption patterns such as agricultural irrigation cycles and residential electricity consumption habits, it enables geographically non-adjacent photovoltaic nodes with similar meteorological conditions to form strong connections. This makes the implicit associations between nodes explicit, more realistically reflects the spatiotemporal synergy of photovoltaic output, and improves prediction accuracy.

[0028] 4. This invention uniformly processes photovoltaic-load coupled nodes and pure load nodes, and achieves joint prediction of the two types of nodes by sharing graph structure and model parameters. This avoids the accumulation of errors from separate model training, and can also explore the interaction between different types of nodes. The prediction results can provide the provincial power grid with a regional and node-specific spatiotemporal distribution map of net load, directly supporting the accurate formulation of distributed photovoltaic consumption, energy storage configuration and demand response strategies.

[0029] 5. Considering the low coverage of monitoring equipment and inconsistent data quality in rural areas with distributed photovoltaic access, this invention guides the construction of a graph structure based on prior knowledge of geographical and meteorological mechanisms, reducing reliance on high-density real-time data. At the same time, the Transformer's self-attention mechanism can automatically filter key spatiotemporal features, suppress noise data interference, enhance the robustness of the model in data-scarce scenarios, and better adapt to the actual operating conditions of rural power grids. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0031] Figure 1 A flowchart provided for this invention. Detailed Implementation

[0032] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0034] It should be noted that net load forecasting for distributed photovoltaic (PV) grid integration areas is a technical support for the grid dispatching system to achieve optimized dispatching and safe operation of distributed energy resources. Its core is to solve the grid supply-demand matching problem caused by the volatility of distributed PV output and the randomness of user load by predicting the future net load of user nodes within the area. This application uses a rural distributed PV grid integration area within a province as its application scenario, applying the net load forecasting method for distributed PV grid integration areas to the grid dispatching system of this area. This assists in realizing distributed PV integration, energy storage configuration, and demand response strategy formulation, thereby improving grid operation stability and energy utilization efficiency.

[0035] The specific implementation of this application is illustrated below through examples.

[0036] See appendix Figure 1 A net load forecasting method for distributed photovoltaic (PV) grid integration areas, applied to the grid dispatching system of such areas, includes:

[0037] S1: Collect geographic coordinate data of distributed photovoltaic installations and user load nodes, load curve data of user nodes for at least one year, and meteorological data corresponding to the time axis of the load curve data.

[0038] Specifically, the geographic coordinate data is obtained from the power supply companies of each county and city. Database acquisition accurately marks the latitude and longitude locations of distributed photovoltaic installation nodes and user load nodes in the distribution network topology; load curve data of user nodes are acquired through the distribution automation system and smart meter acquisition system, with a time resolution of 15 minutes to ensure coverage of at least one full calendar year to include seasonal load variation characteristics; meteorological data integrates data from ground observation stations of the China Meteorological Administration and satellite inversion fusion data, including parameters such as temperature, humidity, rainfall, and irradiance. The regional meteorological data is mapped to the location of each user node using an inverse distance weighting method. At the same time, all collected data are time-aligned, missing value filled, and outlier corrected to ensure that the time axis of the load curve data and the meteorological data are completely matched.

[0039] S2: Based on geographic coordinate data, meteorological data, corresponding meteorological characteristics, and user behavior mechanisms, construct a graph-structured data form. Formula, in which Contains a set of nodes Adjacency matrix of multi-relation edges Node set Each node The connection includes geographical latitude and longitude. Standard, Distributed photovoltaic installed capacity Quantitative and time-varying user behavior trend feature vectors and static features .

[0040] Specifically, nodes Geographic latitude and longitude coordinates Directly using collected geographic coordinate data; distributed photovoltaic installed capacity Data obtained through distributed photovoltaic grid connection filing, if nodes If no photovoltaic system is installed, Simultaneously set binary indicator , Represents a node Install a photovoltaic system. Indicates not installed; static characteristics Fixed attributes, including user type, agricultural electricity consumption ratio, and industrial and commercial electricity consumption ratio, are extracted from the user profile system of the electricity consumption behavior analysis platform; time-varying user behavior trend feature vectors are also included. Includes historical net load Photovoltaic output provided by observation or forecasting agencies Local weather conditions and time- and date-related indicators, where local weather conditions are for each node. Real-time corresponding temperature and radiation Time and date related indicators use one-hot encoding for hours and holidays.

[0041] S3: Using graph neural networks for graph structured data. Spatial correlation fusion is performed on the node features to extract the node set. internal nodes The spatial relationship characteristics are used to obtain time series data of user nodes that are fused with spatial correlations; the input of the graph neural network includes time indexes. The full-graph node feature matrix below and normalized adjacency matrix ,in It is a composite adjacency matrix. For degree matrix, These are the matrix weights.

[0042] Specifically, the feature matrix of nodes in the entire graph middle, For a set of nodes The total number of nodes, Time-varying user behavior trend feature vector Dimensions; Normalized adjacency matrix Through formula calculate, for An identity matrix of order 1. For the comprehensive adjacency matrix degree matrix diagonal elements ); Comprehensive Adjacency Matrix Through formula calculate, For a set of relations, Let be the matrix weights of each relation adjacency matrix and satisfy... In this embodiment The initial value was set to 0.2, and optimized through model training; the graph neural network aggregated each node through a message passing mechanism. The neighborhood characteristics of the nodes Spatial relationship features are generated by fusing self-features and neighborhood features, combined with a time index. The time-series information is used to form user node time-series data that integrates spatial relationships.

[0043] S4: Construct a deep learning model using the Transformer architecture, and use the Transformer architecture to extract load curve data of user nodes and distributed photovoltaic installed capacity. The corresponding photovoltaic power output data time series mode, and the deep learning model takes the fused spatially correlated user node time series data as input, and outputs node Projected future net load.

[0044] Specifically, the Transformer architecture includes an encoder and a decoder. The encoder consists of multiple layers of encoding units, each containing a multi-head attention layer and a feedforward neural network layer. The decoder consists of multiple layers of decoding units, each containing a masked multi-head attention layer, a multi-head attention layer, and a feedforward neural network layer. It integrates spatially correlated user node time-series data by node. The input is split, the encoder extracts historical temporal features, and the decoder generates nodes. Net load forecast for the next 24 hours ( (Corresponding to 15-minute resolution).

[0045] Furthermore, the installed capacity of distributed photovoltaic power... Used to indicate the first Whether the node has a photovoltaic system installed, and the capacity of the distributed photovoltaic system. Set binary indicator To further clarify the nodes Whether to install a photovoltaic system, if the node Without a photovoltaic system installed, the distributed photovoltaic installed capacity is... Time-varying user behavior trend feature vector Including historical net load Photovoltaic output provided by observation or forecasting agencies Local weather conditions and time- and date-related indicators, including temperature. and radiation Time and date related indicators include one-hot encoding of hours or holidays.

[0046] Specifically, binary pointers and One-to-one correspondence, hour , hour Historical net load For nodes The actual net load of the previous time step; photovoltaic output Real-time scenarios use weather forecast values, while offline scenarios use observed values. Hourly one-hot encoding is a 24-dimensional vector. In holiday one-hot encoding, holidays are set to 1 and weekdays to 0. The two types of encoding together constitute the time feature.

[0047] Furthermore, construct graph-structured data. At that time, graph structure data Weight of the middle edge Based on geographical proximity Meteorological similarity Historical net load correlation Photovoltaic or similar capacity relationship Similarity to user attributes or land use Calculated; Node set The connection structure between internal nodes is determined by the full graph node feature matrix. Obtained by calculating the distances between different nodes and The absolute value of the cosine similarity of the difference is used to determine the actual dynamic relationship between nodes; a threshold is also set. Used to filter closely related nodes to reduce data redundancy in the algorithm, when the absolute value of cosine similarity is greater than a threshold. When the connection is active, the two nodes remain connected; otherwise, they are disconnected.

[0048] Specifically, threshold Set to 0.6, when The nodes remain connected; when a summer afternoon thunderstorm causes a sudden drop in the irradiance of some nodes, the absolute value of the cosine similarity between these nodes and other nodes drops below 0.3, and the graph structure automatically disconnects them to reduce abnormal interference.

[0049] Furthermore, geographical proximity The weight formula is ,in It is a node and nodes The distance between them It is the standard deviation, if Exceeding the preset maximum distance ,but .

[0050] Specifically, Calculate straight-line distance (unit: km) using latitude and longitude. (Provincial scenario) km; for example, a node ( , With nodes ( )of km, ;like km> ,but .

[0051] Furthermore, historical net load correlation The weight is determined by the similarity of the historical net load sequence, and the calculation formula is: ,in Represents a node and nodes The correlation coefficient of the historical net load series, , Corresponding nodes , The historical net load sequence.

[0052] Specifically, , Using load curve data from the past 90 days, The Pearson correlation coefficient (values ​​[-1, 1]); for example, node , All are agricultural users. ,but If node For industrial and commercial users, ,but .

[0053] Furthermore, a similarity metric based on climate factors is used to measure meteorological similarity. The weighting formula is ,in Represents a node and nodes In meteorological feature vectors Difference measurement on , Corresponding nodes , The meteorological feature vector was obtained by using the Pearson correlation coefficient.

[0054] Specifically, meteorological feature vectors (Temperature, humidity, rainfall, irradiance). This is the average daily data over the past 30 days; if ,but ;like ,but .

[0055] Furthermore, the installed capacity of distributed photovoltaic power... Constructing photovoltaic or capacity similarity relationships based on capacity factors When, its weighting formula is ,in To adjust the node , Distributed photovoltaic installed capacity , The parameter that determines the contribution of differences to similarity.

[0056] Specifically, The settings are based on the regional average photovoltaic installed capacity (in this example, the average is 50kW, therefore...). );like =60kW =55kW, then ;like =10kW, then .

[0057] Furthermore, user attributes and land use similarity relationships The weight formula is ,in Represents a node , Distance measure between corresponding user attributes or land use patterns Parameters for adjusting the degree of influence of differences in user attributes; geographical proximity. Meteorological similarity Historical net load correlation Photovoltaic or similar capacity relationship Similarity to user attributes or land use The graph structure data is determined by fusion using a weighted method. The weight of each edge in the equation.

[0058] Specifically, Static feature vectors , Euclidean distance =[Agricultural electricity consumption ratio, industrial and commercial electricity consumption ratio, residential electricity consumption ratio]). ;For example , , , During fusion, the weights of each relation are first normalized to the [0,1] interval, and then... The final edge weights are obtained by weighted summation.

[0059] Furthermore, the Transformer architecture uses a multi-head attention mechanism at its core, preserves temporal sequence information through positional encoding, and extracts global and local temporal features in parallel using an encoder-decoder architecture, as shown in the following formula:

[0060]

[0061] in, , , , The weight parameter matrix, , , These are the query matrix, key matrix, and value matrix obtained from high-dimensional power sequence transformations. This represents the time-series data of user nodes that are fused spatially. Indicates the length of the input sequence. This represents the output of the self-attention mechanism. This represents the dimension of the key matrix. The number of heads representing multi-head attention. This is the aggregated time-series feature sequence.

[0062] Specifically, the position encoding uses sine and cosine encoding. , (Transformer feature dimension); , Bullish attention will , , After calculating in 8 groups, the parts are spliced ​​together, and then... Mapped output.

[0063] Furthermore, the batch gradient descent algorithm is used to train the deep learning model, with mean squared error as the basic loss function. The formula for the basic loss function is as follows: ,in The node predicted by the model Future net load value, It is a node Future net load true value For the sample size, To predict the step size.

[0064] Specifically, the training dataset consists of data from the past two years (split into a 7:3 ratio for training and validation), and batch processing is performed by time windows (batch size = 32); the optimizer is Adam (initial learning rate = 0.001), and the validation set... The learning rate is halved if it doesn't decrease for 5 consecutive epochs; the early stopping strategy is used for the validation set. Training stops if the value increases for 10 consecutive epochs; in this example, S=10000, H=96 (predicting 24 hours), and the model objective is to minimize... .

[0065] To verify the effectiveness of this method, the GNN+Transformer model is compared with LSTM, CNN, and Transformer models. The test dataset consists of net load data from rural distributed photovoltaic grid connection areas within a province. The evaluation metrics are shown in the table below:

[0066] Model MAE / MW RMSE / MW <![CDATA[R 2 ]]> LSTM 1.81 2.32 0.909 CNN 1.72 2.34 0.909 Transformer 1.64 2.31 0.908 GNN+Transformer (this method) 1.26 2.20 0.907

[0067] As shown in the table above, the MAE and RMSE of this method are lower than those of the other three models. Specifically, the MAE is reduced by 29.3% compared to LSTM, 26.7% compared to CNN, and 23.2% compared to Transformer, indicating that the predicted values ​​of this method have smaller deviations from the true values ​​and higher prediction accuracy. The R² is close to that of other models, indicating that this method has a good fitting ability to the net load change pattern and better overall prediction performance.

[0068] In summary, the net load forecasting method for distributed photovoltaic (PV) grid access areas provided in this application constructs a dynamic graph structure that integrates geographical, meteorological, and user behavior features. By combining graph neural networks and the Transformer architecture, it achieves deep fusion of spatiotemporal features, overcoming the limitations of traditional models that rely on physical topology and single-time-series modeling, thus improving the accuracy of net load forecasting. At the same time, through dynamic node connections and multi-relationship weight fusion, it enhances the model's adaptability to complex operating conditions in distributed PV grid access areas, providing accurate net load forecasting data for the power grid dispatching system and supporting the optimized dispatching of distributed energy resources and the safe operation of the power grid.

[0069] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A net load forecasting method for distributed photovoltaic (PV) grid connection areas, characterized in that, A power grid dispatching system applied to a distributed photovoltaic (PV) grid integration area, the method comprising: Collect geographic coordinate data of distributed photovoltaic installations and user load nodes, load curve data of user nodes for at least one year, and meteorological data corresponding to the time axis of the load curve data; Based on the geographic coordinate data, the meteorological characteristics corresponding to the meteorological data, and user behavior mechanisms, a graph-structured data format is constructed. ,in Contains a set of nodes Adjacency matrix of multi-relation edges The set of nodes Each node Associated with geographic latitude and longitude coordinates Distributed photovoltaic installed capacity and time-varying user behavior trend feature vector and static features ; Using a graph neural network to process the graph structure data form Spatial association fusion is performed on the node features to extract the node set. internal nodes The spatial relationship characteristics are used to obtain time series data of user nodes with fused spatial associations; the input of the graph neural network includes a time index. The full-graph node feature matrix below and normalized adjacency matrix ,in It is a composite adjacency matrix. For degree matrix, These are the matrix weights; A deep learning model using the Transformer architecture is constructed, and the Transformer architecture is used to extract the load curve data of the user nodes and the distributed photovoltaic installed capacity. The corresponding photovoltaic output data time series mode, and the deep learning model takes the user node time series data associated with the fused space as input, and outputs the node. Projected future net load.

2. The net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 1, characterized in that, The distributed photovoltaic installed capacity Used to indicate the first Whether the node has a photovoltaic system installed, and regarding the distributed photovoltaic installed capacity. Set binary indicator To further clarify the nodes Whether to install a photovoltaic system, if the node If no photovoltaic system is installed, then the distributed photovoltaic installed capacity is... =0; Time-varying user behavior trend feature vector Including historical net load Photovoltaic output provided by observation or forecasting agencies Local meteorological conditions and time- and date-related indicators, wherein the local meteorological conditions include temperature and radiation The time and date related indicators include one-hot encoding of hours or holidays.

3. The net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 1, characterized in that, Construct the graph structure data At that time, graph structure data Weight of the middle edge Based on geographical proximity Meteorological similarity Historical net load correlation Photovoltaic or similar capacity relationship Similarity to user attributes or land use Calculated; The set of nodes The connection structure between internal nodes is achieved through the full graph node feature matrix. Obtained by calculating the distances between different nodes and X t-1 The absolute value of the cosine similarity of the difference is used to determine the actual dynamic relationship between nodes; Simultaneously set threshold Used to filter closely related nodes to reduce data redundancy in the algorithm, when the absolute value of the cosine similarity is greater than a threshold. When the connection is active, the two nodes remain connected; otherwise, they are disconnected.

4. The net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 3, characterized in that, The geographical proximity relationship The weight formula is ,in It is the node and nodes The distance between them σ It is the standard deviation, if Exceeding the preset maximum distance ,but .

5. A net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 3, characterized in that, The historical net load correlation weight Determined by the similarity of historical net load sequences, the calculation formula is as follows: ,in Represents the node and nodes The correlation coefficient of the historical net load series, Corresponding nodes , The historical net load sequence.

6. The net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 3, characterized in that, The meteorological similarity was measured using a climate factor-based similarity metric. The weighting formula is ,in Represents the node and nodes In meteorological feature vectors R Difference measurement on Corresponding nodes , The meteorological feature vector.

7. A net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 3, characterized in that, The distributed photovoltaic installed capacity The capacity factor is used to construct the photovoltaic or capacity similarity relationship. When, its weighting formula is ,in To adjust the nodes , Distributed photovoltaic installed capacity , The parameter that determines the contribution of differences to similarity.

8. A net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 3, characterized in that, The user attributes and land use similarity The weight formula is ,in Represents the node , Distance measure between corresponding user attributes or land use patterns Parameters for adjusting the degree of influence of differences in user attributes; Geographical proximity Meteorological similarity Historical net load correlation Photovoltaic or similar capacity relationship Similarity to user attributes or land use The graph structure data is determined by fusion using a weighted method. The weight of each edge in the equation.

9. A net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 1, characterized in that, The Transformer architecture is based on a multi-head attention mechanism. It preserves temporal sequence information through positional encoding and uses an encoder-decoder architecture to extract global and local temporal features in parallel, as shown in the following formula: in, The weight parameter matrix, It is a high-dimensional power sequence The resulting query matrix, key matrix, and value matrix are obtained through transformation. Indicates the length of the input sequence. This represents the output of the self-attention mechanism. This represents the dimension of the key matrix. The number of heads representing multi-head attention. This is the aggregated time-series feature sequence.

10. A net load forecasting method for distributed photovoltaic (PV) grid connection areas according to claim 1, characterized in that, The deep learning model is trained using batch gradient descent, with mean squared error as the base loss function, the formula of which is: ,in The node predicted by the model Future net load value, It is a node Future net load true value For the sample size, To predict the step size.

Images