A regional power grid line loss prediction method integrating spatiotemporal characteristics
By generating representative samples and constructing structure diagrams, combining the dual-column GRU model of the relationship diagram convolution network and attention mechanism, the problem of not comprehensive consideration of structural and timing characteristics in the existing power grid line loss prediction is solved, and high accuracy is achieved regional grid line loss prediction.
Patent Information
- Application Number
- CN202411383807.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-09-30
- Publication Date
- 2024-12-06
- Estimated Expiration
- 2044-09-30
AI Technical Summary
The existing grid line loss prediction methods ignore the comprehensive consideration of grid structure and timing characteristics, resulting in insufficient prediction accuracy. Especially in the complex topology of the power grid and the diversity of timing data, it is difficult to achieve high-accurate line loss prediction.
The regional grid line loss prediction method is adopted to generate representative samples through sub-mode and greedy methods, construct regional grid structure diagrams, extract structural features using relational graph convolution networks, and double-column GRU model based on attention mechanism Extract timing features and combine feedforward neural network to predict line loss.
The accuracy and efficiency of regional power grid line loss prediction is improved, the problems of structural feature loss and incomplete feature selection in traditional methods are overcome, and comprehensive consideration of factors affecting line loss and high accuracy prediction are achieved.
Smart Images

Figure CN118940915B_ABST
Abstract
Description
Technical Field
[0001] The invention discloses a regional power grid line loss prediction method integrating spatiotemporal features, which involves power grid sample generation, power grid structure information extraction and power grid time series feature extraction, and belongs to the field of spatiotemporal series prediction. Background Art
[0002] In recent years, the accelerated advancement of smart grid construction, the explosive growth of grid operation data, and the rapid development of big data and artificial intelligence technologies have provided a solid data foundation and technical support for applications such as grid line loss prediction, load forecasting, and resource allocation optimization. Line loss is the energy loss caused by the transmission of electric energy through transmission lines. The line loss rate is the percentage of line loss to the electric energy supplied by the entire power network. Line loss and line loss rate are important technical and economic indicators for measuring the management level of power supply companies. Accurate prediction and efficient optimization of line loss are important aspects of power system management. By estimating line loss and fluctuation range in the future through line loss prediction, technical support and decision-making basis can be provided for power supply companies to reduce costs and increase efficiency.
[0003] The known data-driven methods for predicting power line loss can be mainly divided into two categories: traditional machine learning models and time series prediction models. Traditional machine learning models include support vector machines, decision tree models, random forests, and neural networks. For example, Chen Shibing et al. (Prediction of line loss rate of 10kV and below power grids based on improved gradient boosting decision tree, "System Simulation Technology", 2023) proposed a prediction method for line loss rate of 10 kilovolt (kV) and below power grids based on improved gradient boosting decision tree (Gradient Boosting Decision Tree, GBDT). The feature importance and Spearman Correlation coefficient of GBDT are used to jointly extract features, and the density-based spatial clustering algorithm is used to divide the data to solve the problem of discrete distribution of line loss caused by differences in power grid structure. Zhang Yitao et al. (10kV Power Grid Line Loss Prediction Based on Grey Relation Analysis and Improved Neural Network, Power System Technology, 2019) used Grey Relation Analysis to select electrical indicators related to line loss, determined the optimal neural network structure through cross-validation and trial-and-error, and used an adaptive genetic algorithm to improve the back propagation neural network, accelerate network convergence and avoid falling into local optimality, thereby improving the accuracy of line loss prediction results. The above method considers limited electrical and non-electrical features, but ignores some hidden features, lacks comprehensive consideration of different related factors affecting line loss, and cannot guarantee high accuracy of line loss prediction.
[0004] In addition, line loss changes dynamically over time and is affected by the time series data of current, voltage, temperature, etc. in the power grid, which also changes dynamically over time. They are collectively referred to as the time series characteristics of the power grid. It is of great significance to establish a spatiotemporal series prediction method to improve the accuracy and effectiveness of power grid line loss prediction. For example, Yu Wei et al. (A method for predicting line loss in a substation based on feature selection and GRU network, patent<CN202211387524.1> , 2022) Taking the average percentage error as the benchmark, multi-dimensional electrical characteristic indicators are selected to establish a line loss prediction model based on the gated recurrent unit (GRU), and the line loss prediction results of different substations are obtained. Huang Yangjue et al. (Line loss prediction method, device, equipment and medium based on LSTM network, <Patent CN202210324444.5>, 2022) use the long short-term memory network (Long Short-Term Memory, LSTM) to build a line loss prediction model, so that the model maintains the long sequence characteristics of the data and reflects the correlation between the input data. Ma Pengcheng et al. (Research on line loss prediction calculation strategy of bidirectional long short-term memory under the power Internet of Things, "Electrical Materials", 2023) proposed a line loss prediction method based on the bidirectional long short-term memory network (Bidirectional LSTM) under the power Internet of Things. On the basis of LSTM, the ability of mutual iterative training is added to improve the prediction accuracy, solving the problem of difficulty in operation and regulation of the power Internet of Things under renewable energy generation and flexible load grid connection. The above methods focus on the impact of time series on line losses, but do not adequately consider the spatial topological structure between key power grid equipment. Summary of the invention
[0005] In order to overcome the shortcomings of the above-mentioned prior art, and in view of the problems that direct measurement of line loss in the power grid during power transmission and power supply is difficult, involves complex factors, and has low accuracy, the present invention proposes a regional power grid line loss prediction method that integrates time and space characteristics, comprehensively considers the impact of the structure and timing characteristics of the regional power grid on the transmission loss, and realizes accurate prediction of regional power grid line loss, providing scientific basis and technical support for optimizing the power grid structure, reducing power grid line loss, energy conservation and environmental protection, and reducing costs and increasing efficiency.
[0006] The present invention first uses sub-modularity and greedy methods to generate representative regional power grid samples, characterize different types of power grids, thereby obtaining different types of regional power grid structure information and reducing the training overhead of the power grid transmission loss model. Then, considering the regional power consumption and substation attributes, the substation and distribution line are used as nodes, and the physical connection relationship between them is used as the edge to construct a regional power grid structure diagram, and the relational graph convolutional network (R-GCN) is used to aggregate the regional power grid structure features. Finally, a double-column GRU model based on the attention mechanism is constructed to extract the time series features of the substation and distribution line respectively, and splice and fuse them with the regional power grid structure features to obtain the spatiotemporal features of the regional power grid, and then use two feedforward neural networks (FNN) to predict the line loss of the substation and distribution line respectively, effectively realizing the fine-grained prediction of the line loss of the regional power grid.
[0007] The present invention is implemented by the following technical solutions: A method for predicting regional power grid line loss by integrating temporal and spatial characteristics, comprising the following steps:
[0008] Step 1, regional power grid sample generation: According to the transmission characteristics of electricity from the power generation end to the power consumption end, a metapath construction method is designed, and the Metapath2vec model is used to obtain the embedded representation of the regional power grid. A measurement function of the regional power grid difference is designed, with the goal of maximizing the difference between different regional power grids. A submodule function is designed and representative samples are generated based on a greedy strategy;
[0009] Step 2, Extraction of regional power grid structure features: Taking the substations and distribution lines as nodes, and the spatial connectivity between them as the edges between nodes, construct a regional power grid structure diagram, use R-GCN to extract regional power grid structure features, establish a READOUT function based on hierarchical aggregation, and represent the extracted structural features as low-dimensional vectors;
[0010] Step 3, regional power grid time series feature extraction: construct a double-column GRU model based on the attention mechanism, process the time series data of the substation and distribution line respectively, calculate the influence of different attributes and their previous moments on the line loss at the current moment, and obtain the time series feature vectors of each substation and distribution line;
[0011] Step 4, line loss prediction: Each node in the regional power grid structure diagram obtains the time series characteristics through Step 3, and then integrates it with the overall structural characteristics of the regional power grid obtained in Step 2 to obtain the spatiotemporal feature vector of each node. Finally, two FNNs are used to accurately predict the line losses of the two types of nodes, the substation and the distribution line, in the power grid structure diagram.
[0012] The step 1 of generating the regional power grid sample specifically includes the following operations:
[0013] Step 1.1 Data normalization:
[0014] The original power data has different dimensions and value ranges. First, the power data is normalized and converted to a value between 0 and 1 using Min-Max Scaling to speed up the model convergence and improve the model's stability and generalization ability. and minimum value , for each value in the sequence Perform normalization:
[0015] (1)
[0016] in, is the original value in the feature sequence, is the normalized value;
[0017] Step 1.2 Calculation of regional power grid differences:
[0018] Taking into account the characteristics of regional power grids with different network structures, the present invention designs a regional power grid difference calculation method based on cosine similarity, aiming to quantify the differences between different regional power grids.
[0019] In order to fully describe the structure of regional power grids and effectively measure the differences between them, we first construct a power transmission network diagram. During the power transmission process, the power is transmitted from the power generation end to the power consumption end through six types of equipment, namely, high-voltage transmission line, main step-down transformer, high-voltage distribution line, secondary step-down transformer, low-voltage distribution line, and power consumption end, which are respectively recorded as , , , , and , these categories of devices are regarded as nodes in the graph, then M={ } represents a set of node types, and then construct edges according to the physical connection relationship between device nodes to obtain the power transmission network diagram ,in, is a node set, is the set of edges;
[0020] In order to fully explore and preserve the structural characteristics of the regional power grid, so as to accurately and efficiently calculate the difference of the regional power grid, the above-constructed power transmission network graph is embedded and represented. Taking into account the diversity of node categories in the power transmission network graph, the meta-path-based graph representation learning method Metapath2vec is extended, and the Skip-gram model is used to maximize the co-occurrence probability of nodes and their context nodes, and generate a low-dimensional vector representation of the nodes in the power transmission network graph; the present invention defines the power grid meta-path and performs a random walk (RandomWalk) to obtain the distribution sample of power grid equipment, so as to obtain the Skip-gram model training Vector representation of midpoints;
[0021] Specifically, the present invention first follows the regional power grid power transmission standards and specifications, and designs the following four types of regional power grid element paths according to the influence of each node on the power grid difference during the transmission process: :
[0022] :High-voltage transmission line—main step-down transformer—high-voltage distribution line, describing the impact of voltage conversion within the power grid and the layout structure of the transmission line on the power grid differentiation;
[0023] :Main step-down transformer - high voltage distribution line - secondary step-down transformer, describing the role of transformers of different levels in the power grid and the impact of the voltage level inside the power grid on the difference of the power grid;
[0024] :High voltage distribution line—secondary step-down transformer—low voltage distribution line, describing the impact of the layout structure of different levels of distribution lines within the power grid on the grid differentiation;
[0025] :Secondary step-down transformer - low voltage distribution line - power consumption end, describing the impact of the access of end users within the power grid and the layout structure of distribution lines on the grid differentiation;
[0026] Then, follow the meta path right Nodes in Perform random walks to obtain the random walk sequence of power grid equipment nodes :
[0027] (2)
[0028] in, for The maximum number of visits to each node in ;
[0029] Then, based on the generated random walk sequence, the corresponding negative samples are obtained by randomly selecting non-element path nodes, and the Skip-gram model based on the negative sample generation strategy is trained to obtain the embedded representation of each power grid device:
[0030] (3)
[0031] in, is the size of the context window, is the dimension of the node vector, Embedding representation matrix for all devices in the power grid;
[0032] Since the four meta-paths contained in each regional power grid have different effects on node representation, the present invention proposes a regional power grid difference calculation method based on the meta-path of power grid equipment, which converts the representation vector of each node into Functions are aggregated to uniformly represent the node types under the regional power grid meta-path:
[0033] (4)
[0034] in, is the meta-path of the regional power grid, and ; Metapath Included class node, and ; for The set of nodes contained in ; Represents a meta path The grid nodes below;
[0035] Finally, in the meta path The cosine mean of the node type representation vectors in any two regional power grids 𝒜 and ℬ is calculated to measure the metapath The difference between 𝒜 and ℬ in the lower region is:
[0036] (5)
[0037] in, Represents the regional power grid element path The set of node types included. for The specific types included in express Regional Grid Metapath Down A vector representation of type , and , for Regional Grid Metapath Down A vector representation of type , and , regional power grid 𝒜 and In the meta path The greater the difference in characteristics, The larger the value of is, the smaller it is;
[0038] Step 1.3 Regional power grid sample selection:
[0039] The selection of regional power grid samples is an optimization problem that aims to maximize the differences between different regional power grids. When the objective function satisfies submodularity, the results obtained by using the greedy strategy to solve this problem can guarantee an approximate optimal solution. Based on the above considerations, let For any subset of the set of grids in a given region, calculate The difference between any two regional power grids in the , and the sum of all regional differences is obtained The overall difference of all regional power grids in:
[0040] (6)
[0041] in, Represents a regional power grid collection Any two different regional power grids;
[0042] make is a subset of the set of power grids in a given region, and , if it does not exist Make , then it is called is the subset with the maximum difference, that is The difference between the regional power grids in is the largest; Formula 1-6 satisfies the submodularity theory, As the objective function, the greedy method is used to efficiently solve the optimal subset ;
[0043] During the generation of regional power grid samples, one is selected each time so that Regional grid with the largest gain , so that The specific sample generation steps include:
[0044] ① Initialization , used to store the selected regional power grid;
[0045] ② Randomly select an initial regional power grid , add it to middle;
[0046] ③For each unselected regional grid , calculate and add it After the gain ;
[0047] ④Select Regional grid with the largest gain , add it to the collection middle;
[0048] ⑤ Repeat steps ③ and ④ until the selected regional grids reach 55% to 65% of the total number of regional grids to form , and finally obtain the regional power grid sample set with the largest difference .
[0049] The regional power grid structure feature extraction in Step 2 specifically includes the following operations:
[0050] Step 2.1 Construction of regional power grid structure diagram:
[0051] In order to effectively extract the structural characteristics of the regional power grid, the present invention divides the substations in the regional power grid into and distribution line assembly Considered as a core node, is the number of stations, is the number of distribution lines, and the regional power grid structure diagram is constructed based on the physical connection relationship between nodes ,in, is a node set, is the edge set, is a collection of node types, is a set of edge types;
[0052] In order to more comprehensively describe the characteristics of the distribution area and distribution lines, the area characteristic matrix is introduced and the distribution line characteristic matrix , respectively save the initial line loss characteristic information related to the substation and distribution line, is the number of characteristics of the nodes in the area, is the characteristic number of the distribution line node; and Integrate it into the regional power grid structure diagram and use a fully connected neural network to and Convert to a feature matrix with the same dimensions :
[0053] (7)
[0054] Among them, X is or , is the Sigmoid activation function, is the weight matrix, is the bias vector;
[0055] Will and Enter formula (7) respectively, and the two feature dimensions are converted to The matrix is then concatenated to obtain a new feature matrix. ;
[0056] Step 2.2 Regional power grid structure diagram feature fusion:
[0057] The regional power grid structure graph contains multiple types of nodes and edges, which is a complex graph structure with rich relationship information. The present invention extends R-GCN, which can fully represent the relationship information, to capture the node characteristics and topological structure in the regional power grid structure graph. layer, Through the aggregation node The neighbor information of Feature representation of the layer :
[0058] (8)
[0059] in, Representation Node On the edge The neighbor set corresponding to type, For Node In the The hidden state of the layer, and The neural network The weight matrix of the layer;
[0060] In the regional power grid structure diagram, in order to overcome the problem of hierarchical information loss when representing the power grid structure characteristics, the present invention designs a READOUT function based on hierarchical aggregation, retaining the hierarchical structure and dependency relationship of different node types in the regional power grid, thereby more accurately reflecting the structural characteristics of the entire regional power grid; the first-level READOUT function aggregates the characteristic information of the substation node to the distribution line node connected to it:
[0061] (9)
[0062] in, Indicates the area node The characteristic vector of Represents a distribution line node The neighbor set of express The updated vector;
[0063] The characteristics of the substation nodes are aggregated to the distribution line nodes to enrich the information of the distribution line nodes so that they contain the characteristic information of themselves and the connected substation nodes; the second-layer READOUT function aggregates all the distribution line nodes processed by the first-layer READOUT function to obtain the overall structural characteristics of the regional power grid. :
[0064] (10).
[0065] The regional power grid time series feature extraction in Step 3 specifically includes the following operations:
[0066] Step 3.1 Attribute attention calculation:
[0067] Considering the different effects of different attributes of the substation and distribution line at any time on the line loss at the current time, an attention mechanism is designed to calculate the influence weights of different attributes:
[0068] (11)
[0069] in, is the query matrix of the substation and distribution line attributes, is the key matrix of the substation and distribution line attributes, is the value matrix of the substation and distribution line loss;
[0070] The influence weights of different attributes on line loss at different times are calculated by formula (11), and the attribute representation vectors are weighted summed and spliced to the original substation and distribution line representation vectors as the updated feature vector representation of the substation and distribution line at the current time. ;
[0071] Step 3.2 Temporal attention calculation:
[0072] Based on the current time obtained in Step 3.1, the updated feature vector representation of the substation and distribution line , further considering the characteristic vectors of the substation and distribution line at the previous moment , expand the two-column GRU model to update the representation of the current station area and distribution line:
[0073] (12)
[0074] in, Indicates element-wise multiplication, update gate Determines the characteristics of the previous moment Features of the current moment the extent of the impact;
[0075] Based on formula (12), the current time of the regional power grid and distribution line is expressed as , and then calculate Always The attention weight at each moment is used to obtain the influence of historical features on current features:
[0076] (13)
[0077] in, is the query matrix of the substation and distribution line at different preceding times, is the key matrix of the substation and distribution line at different preceding moments, for The transposed matrix of is the value matrix of the substation and distribution line at the current moment, for The dimension of the vector in
[0078] The dual-column GRU model based on the attention mechanism processes the substation nodes and distribution line nodes respectively, and extracts the time series features of all nodes through Step 3.1 and Step 3.2.
[0079] The line loss prediction of Step 4 specifically includes the following operations:
[0080] The regional power grid structure diagram The time series characteristics of each node in the grid are combined with the overall structural characteristics of the regional power grid to obtain the spatiotemporal characteristic vector of each node. , whose length is the sum of the length of the time series feature vector and the length of the overall structure feature of the regional power grid. The fused vector is used as the input of the FNN, and the output is the line loss of the node calculated by formula (14):
[0081] (14)
[0082] in, is the weight matrix from the input layer to the hidden layer, each row corresponds to the weight vector of a hidden layer neuron, is the bias vector of the hidden layer, is the Sigmoid activation function, is the weight matrix from the hidden layer to the output layer, is the bias vector of the output layer, is the line loss prediction value;
[0083] There are obvious differences in line losses between substations and distribution lines, and power data may show different distribution characteristics in different regions. In order to improve the accuracy of line loss prediction, two FNNs are used to construct substation and distribution line loss functions respectively, so as to predict the line losses of substation nodes and distribution line nodes; Enhance the model's robustness to outliers and introduce loss weights In order to balance the impact of the two, the following loss function is designed:
[0084] (15)
[0085] in, and are the predicted line loss values of the substation and distribution line, and are the true values of line losses in the substation and distribution lines in the training data, is the preset threshold, and , to ensure the stability of the loss function value, and thus ensure the effectiveness of the model training results.
[0086] Compared with the prior art, the present invention has the following advantages and positive effects:
[0087] 1. To address the problem of slow model training due to the large number of regional power grids, the present invention generates regional power grid samples based on the greedy method, and uses a small number of representative regional power grids to characterize different types of power grids as much as possible, thereby reducing the cost of model training and improving model training efficiency and prediction accuracy.
[0088] 2. In view of the complex topological structure of the regional power grid, the present invention constructs a regional power grid structure diagram with substations and transmission lines as nodes, adopts the R-GCN model to extract power grid structure information, and designs a READOUT function based on hierarchical aggregation to effectively aggregate node features, thus overcoming the shortcomings of traditional methods inaccurate line loss prediction due to the loss of power grid structure.
[0089] 3. In view of the characteristics of power data with strong time series, multiple attributes and complex correlation, the present invention proposes a double-column GRU model based on the attention mechanism, which alleviates the problem of incomplete feature selection of the traditional method affecting the line loss, comprehensively considers the different factors affecting the line loss and their correlation, effectively captures and comprehensively considers the impact of the structural characteristics and time series characteristics of the regional power grid on the line loss, and improves the accuracy of the regional power grid line loss prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0090] Figure 1 It is a flow chart of specific steps of the present invention;
[0091] Figure 2 A schematic diagram of a partial power grid structure in an embodiment;
[0092] Figure 3 for The power transmission network diagram;
[0093] Figure 4 Generate flow charts for a sample of regional power grids;
[0094] Figure 5 for The local topological structure diagram of
[0095] Figure 6 Schematic diagram of the GRU model based on the attention mechanism. DETAILED DESCRIPTION
[0096] The specific implementation of the present invention is described below in conjunction with the accompanying drawings so that those skilled in the art can better understand the present invention. It should be noted that in the following description, when the detailed description of known functions and designs may dilute the main content of the present invention, these descriptions will be omitted here.
[0097] Example 1
[0098] From the 100 regional power grids provided by a power grid company, 60 representative areas are selected as sample training models, and line loss prediction is performed on the other 40 areas. Figure 2 A partial grid structure of an embodiment is shown.
[0099] Step 1, regional power grid sample generation: According to the transmission characteristics of electricity from the power generation end to the power consumption end, a metapath construction method is designed, and the Metapath2vec model is used to obtain the embedded representation of the regional power grid. A measurement function of the regional power grid difference is designed, with the goal of maximizing the difference between different regional power grids. A submodule function is designed and representative samples are generated based on a greedy strategy;
[0100] Step 1.1 Data normalization:
[0101] The original power data has different dimensions and value ranges. First, the power data is normalized and converted to a value between 0 and 1 using Min-Max Scaling to speed up the model convergence and improve the model's stability and generalization ability. and minimum value , for each value in the sequence Perform normalization:
[0102] (1)
[0103] in, is the original value in the feature sequence, is the normalized value;
[0104] According to Step 1.1, in the regional power grid There is a set of nodes in the area and distribution line node collection Each node has two attribute characteristics: static and dynamic. The characteristic sequence of static attributes is constant, such as parameters such as the radius of the substation and the length of the transmission line; the characteristic sequence of dynamic attributes changes over time, such as parameters such as weather conditions and load capacity.
[0105] For the area nodes A static property sequence of
[0106] , first find The minimum value in and maximum value , the normalized data is obtained by formula (1) . Area Node With dynamic attribute sequence , where 61 is the number of time steps, split it into 61 vectors of length 40, normalize each vector as above, and deform it back to the original shape (61×8×5). The other nodes are normalized in the same way.
[0107] Step 1.2 Calculation of regional power grid differences:
[0108] Taking into account the characteristics of regional power grids with different network structures, the present invention designs a regional power grid difference calculation method based on cosine similarity, aiming to quantify the differences between different regional power grids.
[0109] In order to fully describe the structure of regional power grids and effectively measure the differences between them, we first construct a power transmission network diagram. During the power transmission process, the power is transmitted from the power generation end to the power consumption end through six types of equipment, namely, high-voltage transmission line, main step-down transformer, high-voltage distribution line, secondary step-down transformer, low-voltage distribution line, and power consumption end, which are respectively recorded as , , , , and , these categories of devices are regarded as nodes in the graph, then M={ } represents a set of node types, and then construct edges according to the physical connection relationship between device nodes to obtain the power transmission network diagram ,in, is a node set, is the set of edges;
[0110] In order to fully explore and preserve the structural characteristics of the regional power grid, so as to accurately and efficiently calculate the difference of the regional power grid, the above-constructed power transmission network graph is embedded and represented. Taking into account the diversity of node categories in the power transmission network graph, the meta-path-based graph representation learning method Metapath2vec is extended, and the Skip-gram model is used to maximize the co-occurrence probability of nodes and their context nodes, and generate a low-dimensional vector representation of the nodes in the power transmission network graph; the present invention defines the power grid meta-path and performs a random walk (RandomWalk) to obtain the distribution sample of power grid equipment, so as to obtain the Skip-gram model training Vector representation of midpoints;
[0111] Specifically, the present invention first follows the regional power grid power transmission standards and specifications, and designs the following four types of regional power grid element paths according to the influence of each node on the power grid difference during the transmission process: :
[0112] :High-voltage transmission line—main step-down transformer—high-voltage distribution line, describing the impact of voltage conversion within the power grid and the layout structure of the transmission line on the power grid differentiation;
[0113] :Main step-down transformer - high voltage distribution line - secondary step-down transformer, describing the role of transformers of different levels in the power grid and the impact of the voltage level inside the power grid on the difference of the power grid;
[0114] :High voltage distribution line—secondary step-down transformer—low voltage distribution line, describing the impact of the layout structure of different levels of distribution lines within the power grid on the grid differentiation;
[0115] :Secondary step-down transformer - low voltage distribution line - power consumption end, describing the impact of the access of end users within the power grid and the layout structure of distribution lines on the grid differentiation;
[0116] Then, follow the grid element path right Nodes in Perform random walks to obtain the random walk sequence of power grid equipment nodes :
[0117] (2)
[0118] in, for The maximum number of visits to each node in ;
[0119] Then, based on the generated random walk sequence, the corresponding negative samples are obtained by randomly selecting non-element path nodes, and the Skip-gram model based on the negative sample generation strategy is trained to obtain the embedded representation of each power grid device:
[0120] (3)
[0121] in, is the size of the context window, is the dimension of the node vector, Embedding representation matrix for all devices in the power grid;
[0122] Since the four meta-paths contained in each regional power grid have different effects on node representation, the present invention proposes a regional power grid difference calculation method based on the meta-path of power grid equipment, which converts the representation vector of each node into Functions are aggregated to uniformly represent the node types under the regional power grid meta-path:
[0123] (4)
[0124] in, is the meta-path of the regional power grid, and ; Metapath Included class node, and ; for The set of nodes contained in ; Represents a meta path The grid nodes below;
[0125] Finally, in the meta path The cosine mean of the node type representation vectors in any two regional power grids 𝒜 and ℬ is calculated to measure the metapath The difference between 𝒜 and ℬ in the lower region is:
[0126] (5)
[0127] in, is the regional power grid element path The set of node types included. for The specific types included in express Regional Grid Metapath Down A vector representation of type , and , for Regional Grid Metapath Down A vector representation of type , and , regional power grid 𝒜 and In the meta path The greater the difference in characteristics, The larger the value of is, the smaller it is;
[0128] According to Step 1.2, first find the high-voltage transmission lines, main step-down transformers, high-voltage distribution lines, secondary step-down transformers, low-voltage distribution lines and power consumption ends of each regional power grid, and the physical connection relationship between them, and construct the power transmission network diagram , regional power grid The constructed power transmission diagram is shown in the attached figure. Figure 3 As shown. Next, a large number of walk sequences are obtained under different meta-paths and different walk starting points through the random walk strategy. The sequence set obtained by the random walk strategy is used as a positive sample, and at the same time, 7 nodes that will not appear in the walk strategy are randomly selected as negative samples. Some walk strategies and positive and negative samples are shown in Table 1:
[0129] Table 1 Regional power grid Positive and negative samples
[0130]
[0131] Then, the Skip-gram model is used to get node embeddings, and training is performed using positive and negative samples. For each positive sample, the center node is set as the input, and the context node and the negative node are set as the output. For example, the edge node , and The new positive sample is obtained by replacing the central node of the positive sample to ensure the embedding representation of each node.
[0132] The model takes the node static features as input and the context node as the target output. Since the static feature lengths of different node types are different, shorter features need to be padded before using the node static features. Among the node types in Step 1.2, the feature length of the secondary step-down transformer is 17, which is the longest type. To maintain consistency, the features of other types are padded with "0" to a length of 17. For example, according to Step 1.1, the node The normalized eigenvector is , fill with "0" to get the feature vector [ ] .
[0133] Through the above steps, we get the embedded representation of each node, namely The embedding representation of Subsequently, the node embedding representations are aggregated by type using Equation (4).
[0134] For regional power grids , meta path Contains node types ,in The nodes of type , The nodes of type , The nodes of type . By formula (4), we can find , , In the actual operation, the length of the vector remains unchanged, and each digit is added and averaged. For example, .
[0135] Through the above operations, the embedding representations of each type of node under each meta-path in other regional power grids are calculated in turn. Then, by calculating the meta-path difference between each region , some calculation results are shown in Table 2:
[0136] Table 2 Differences of each meta-path between regions
[0137]
[0138] Step 1.3 Regional power grid sample selection:
[0139] make For any subset of the set of grids in a given region, calculate The difference between any two regional power grids in the , and the sum of all regional differences is obtained The overall difference of all regional power grids in:
[0140] (6)
[0141] in, Represents a regional power grid collection Any two different regional power grids;
[0142] make is a subset of the set of power grids in a given region, and , if it does not exist Make , then it is called is the subset with the maximum difference, that is The difference between the regional power grids in is the largest; Formula 1-6 satisfies the submodularity theory, As the objective function, the greedy method is used to efficiently solve the optimal subset ; In the process of generating regional power grid samples, one is selected each time so that Regional grid with the largest gain , so that Maximum, process as attached Figure 4 As shown in the figure, the specific sample generation steps include:
[0143] ① Initialization , used to store the selected regional power grid;
[0144] ② Randomly select an initial regional power grid , add it to middle;
[0145] ③For each unselected regional grid , calculate and add it After the gain ;
[0146] ④Select Regional grid with the largest gain , add it to the collection middle;
[0147] ⑤ Repeat steps ③ and ④ until the selected regional grids reach 55% to 65% of the total number of regional grids to form , and finally obtain the regional power grid sample set with the largest difference ;
[0148] According to Step 1.3, select the subset with the largest difference among the 100 regions. First initialize the empty set S and divide the regions Add S, and the remaining area is to make The largest area is added to the set S, and then the remaining areas are selected. Some of the generated samples are shown in Table 3:
[0149] Table 3. Results of generating regional power grid samples
[0150]
[0151] Finally, the optimal regional power grid subset is obtained ,| |=60. Yes Renumbering of regional power grids in .
[0152] Step 2, regional power grid structural feature extraction: regional power grid structural feature extraction: take the substation and distribution line as nodes, and the spatial connectivity relationship between them as the edge between nodes, construct the regional power grid structure diagram, use R-GCN to extract the regional power grid structural features, establish the READOUT function based on hierarchical aggregation, and represent the extracted structural features as low-dimensional vectors;
[0153] Step 2.1 Construction of regional power grid structure diagram:
[0154] In order to effectively extract the structural characteristics of the regional power grid, the substations in the regional power grid are divided into and distribution line assembly Considered as a core node, is the number of stations, is the number of distribution lines, and the regional power grid structure diagram is constructed based on the physical connection relationship between nodes ,in, is a node set, is the edge set, is a collection of node types, is a set of edge types;
[0155] In order to more comprehensively describe the characteristics of the distribution area and distribution lines, the area characteristic matrix is introduced and the distribution line characteristic matrix , respectively save the initial line loss characteristic information related to the substation and distribution line, is the number of characteristics of the nodes in the area, is the number of characteristics of the distribution line nodes; in order to and Integrate it into the regional power grid structure diagram and use a fully connected neural network to and Convert to a feature matrix with the same dimensions :
[0156] (7)
[0157] Among them, X is or , is the Sigmoid activation function, is the weight matrix, is the bias vector;
[0158] Will and Enter formula (7) respectively, and the two feature dimensions are converted to The matrix is then concatenated to obtain a new feature matrix. ;
[0159] According to Step 2.1, regional power grid For example, Constructed as a trainable regional power grid structure diagram , vertex set , Contains substations and distribution line nodes, It includes the power supply and distribution relationship between the distribution line nodes and the substation nodes, the load transfer relationship between the substation nodes and the substation nodes, and the parallel operation relationship between the distribution line nodes and the distribution line nodes. Constructing the area feature matrix , distribution line characteristic matrix and the adjacency matrix According to formula (7), and Input the fully connected neural network separately, calculate the features of different types of nodes and fuse them to get the feature representation of the same dimension, and splice along the row direction to get the feature matrix of the nodes in the whole area .
[0160] Step 2.2 Regional power grid structure diagram feature fusion:
[0161] The regional power grid structure graph contains various types of nodes and edges, which is a complex graph structure with rich relationship information. The extended R-GCN that can fully represent the relationship information captures the node characteristics and topological structure in the regional power grid structure graph. layer, Through the aggregation node The neighbor information of Feature representation of the layer :
[0162] (8)
[0163] in, Representation Node On the edge The neighbor set corresponding to type, For Node In the The hidden state of the layer, and The neural network The weight matrix of the layer;
[0164] In the regional power grid structure diagram, in order to overcome the problem of hierarchical information loss when representing the power grid structure characteristics, a READOUT function based on hierarchical aggregation is designed to retain the hierarchical structure and dependency relationship of different node types in the regional power grid, thereby more accurately reflecting the structural characteristics of the entire regional power grid; the first-level READOUT function aggregates the characteristic information of the substation node to the distribution line node connected to it:
[0165] (9)
[0166] in, Indicates the area node The characteristic vector of Represents a distribution line node The neighbor set of express The updated vector;
[0167] The characteristics of the substation nodes are aggregated to the distribution line nodes to enrich the information of the distribution line nodes so that they contain the characteristic information of themselves and the connected substation nodes; the second-layer READOUT function aggregates all the distribution line nodes processed by the first-layer READOUT function to obtain the overall structural characteristics of the regional power grid. :
[0168] (10)
[0169] According to Step 2.2, the message transmission of different relationship types between nodes is calculated according to formula (8). conduct After updating, the matrix , some results are shown in Table 4:
[0170] Table 4 The feature vector of each node updated by R-GCN
[0171]
[0172] According to formula (9), the characteristic information of the substation nodes is aggregated to the distribution line nodes. Figure 5 Given The local topological structure of Aggregate its neighbor nodes , Aggregate its neighbor nodes , , , , Aggregate its neighbor nodes , ,Right now , , Table 5 shows the updated , and Then, according to formula (2-4), the characteristic information of distribution line nodes is aggregated to obtain the grid structure characteristics: ,Right now Table 6 shows some of the grid structure characteristics after dimensionality reduction. .
[0173] Table 5 and The node’s feature representation after being updated in Step 2
[0174]
[0175] Table 6. Grid structure characteristics after dimensionality reduction
[0176]
[0177] Step 3 Regional power grid time series feature extraction: Regional power grid time series feature extraction: Construct a double-column GRU model based on the attention mechanism, process the time series data of the substation and distribution line respectively, calculate the influence of different attributes and their previous moments on the line loss at the current moment, and obtain the time series feature vectors of each substation and distribution line;
[0178] Step 3.1 Attribute attention calculation:
[0179] Considering the different effects of different attributes of the substation and distribution line at any time on the line loss at the current time, an attention mechanism is designed to calculate the influence weights of different attributes:
[0180] (11)
[0181] in, is the query matrix of the substation and distribution line attributes, is the key matrix of the substation and distribution line attributes, is the value matrix of the substation and distribution line loss;
[0182] The influence weights of different attributes on line loss at different times are calculated by formula (11), and the attribute representation vectors are weighted summed and spliced to the original substation and distribution line representation vectors as the updated feature vector representation of the substation and distribution line at the current time. ;
[0183] Regional power grid Taiwan area Node feature tensor Contains 8 dynamic attributes, the attribute vector length is 5, and there are 61 timing steps in total; distribution line Node feature tensor It contains 5 dynamic attributes, the attribute vector length is 7, and there are 61 timing steps in total; the line loss value matrices corresponding to the substation area and distribution line nodes are .
[0184] According to Step 3.1, calculate the attribute attention of the substation and distribution line. For example, Split into , the line loss value vector corresponding to the first 60 attribute matrices is From the second moment on, the area attribute query matrix , the attribute key matrix of the station area , line loss value vector , the influence weights of different attributes on the line loss value at the current moment are calculated by formula (3-1). Table 7 gives the input of attention calculation at some moments:
[0185] Table 7 Area Nodes Input of attribute attention calculation at each moment
[0186]
[0187] Then, the weights of the above attributes affecting the line loss value are accumulated, and then the substation nodes are The attribute matrix of each moment is expanded into a vector, the accumulated results are concatenated to the expanded attribute vector, and the updated feature vector representation of the station area is used , , …, .
[0188] Step 3.2 Temporal attention calculation:
[0189] Based on the current time obtained in Step 3.1, the updated feature vector representation of the substation and distribution line , further considering the characteristic vectors of the substation and distribution line at the previous moment , expand the two-column GRU model to update the representation of the current station area and distribution line:
[0190] (12)
[0191] in, Indicates element-wise multiplication, update gate Determines the characteristics of the previous moment Features of the current moment the extent of the impact;
[0192] Based on formula (12), the current time of the regional power grid and distribution line is expressed as , and then calculate Always The attention weight at each moment is used to obtain the influence of historical features on current features:
[0193] (13)
[0194] in, is the query matrix of the substation and distribution line at different preceding times, is the key matrix of the substation and distribution line at different preceding moments, for The transposed matrix of is the value matrix of the substation and distribution line at the current moment, for The dimension of the vector in
[0195] The dual-column GRU model based on the attention mechanism processes the substation nodes and distribution line nodes respectively, and extracts the time series features of all nodes through Step 3.1 and Step 3.2.
[0196] According to Step 3.2, the temporal attention is calculated. The updated feature vector obtained in Step 3.1 and the attribute feature vector at each moment are used as the input of formula (12) to calculate 1~ The output of the station area at the time , , and concatenate them into a key matrix and a value matrix , .make = , calculate the temporal attention through formula (13) and get the station area The time series feature vector representation of Finally, the time series characteristics of 101 substation nodes are obtained. .
[0197] The distribution line node is on another GRU model based on the attention mechanism. After the same process, the time series features of 210 distribution lines can be obtained. .
[0198] Step 4, line loss prediction: The time series features of each node in the regional power grid structure diagram are obtained through Step 3, and then they are integrated with the overall structural features of the regional power grid obtained in Step 2 to obtain the spatiotemporal feature vector of each node. Finally, two FNNs are used to achieve accurate prediction of the line losses of the two types of nodes in the power grid structure diagram, the substation and the distribution line.
[0199] The regional power grid structure diagram The time series characteristics of each node in the grid are combined with the overall structural characteristics of the regional power grid to obtain the spatiotemporal characteristic vector of each node. , whose length is the sum of the length of the time series feature vector and the length of the overall structure feature of the regional power grid. The fused vector is used as the input of the FNN, and the output is the line loss of the node calculated by formula (14):
[0200] (14)
[0201] in, is the weight matrix from the input layer to the hidden layer, each row corresponds to the weight vector of a hidden layer neuron, is the bias vector of the hidden layer, is the Sigmoid activation function, is the weight matrix from the hidden layer to the output layer, is the bias vector of the output layer, is the line loss prediction value;
[0202] Through Step 2, the overall structural characteristic vector of the regional power grid with a length of 15 is obtained. , through Step 3, we get the time series feature vector of each area node with a length of 40 , and the time series feature vector of the distribution line node with a length of 35 . The time series feature vector of each node is compared with The spatiotemporal feature vectors are obtained by concatenation. The spatiotemporal feature vectors of the substation nodes are used as the input of the FNN with an input neuron number of 55, and the spatiotemporal feature vectors of the distribution line nodes are used as the input of the FNN with an input neuron number of 50. After setting the hyperparameters determined in Step 4, training can be started. The line loss prediction values of the substation and distribution line are obtained through forward propagation. The model loss value is calculated according to formula (15). The network parameters are continuously optimized through back propagation until the loss function value converges. After the model converges, the model parameters are fixed and saved.
[0203] When training the model, the loss weight , training iterations, learning rate, threshold It can be set according to the actual situation. Generally, the loss weight can be Set to 0.72, the number of training iterations is set to 200, the learning rate is set to 0.005, and the threshold Set to 2.
[0204] The trained model is used to predict the line loss of 40 regional power grids without samples. According to formula (14), the line loss of the model for the substation and distribution line is statistically analyzed. The average loss of the substation for these 40 samples is 0.057, and the average loss of the transmission line is 0.012, indicating that the average line loss prediction error of the model for the substation is 5.7%, and the average line loss prediction error for the distribution line is 1.2%.
Claims
1. A regional power grid line loss prediction method integrating spatiotemporal characteristics, characterized in that The steps include: Step 1, regional power grid sample generation: According to the transmission characteristics of electricity from the power generation end to the power consumption end, a metapath construction method is designed, and the Metapath2vec model is used to obtain the embedded representation of the regional power grid. A measurement function of the regional power grid difference is designed, with the goal of maximizing the difference between different regional power grids. A submodule function is designed and representative samples are generated based on a greedy strategy; Step 2, Extraction of regional power grid structure features: Taking the substations and distribution lines as nodes, and the spatial connectivity between them as the edges between nodes, construct a regional power grid structure diagram, use R-GCN to extract regional power grid structure features, establish a READOUT function based on hierarchical aggregation, and represent the extracted structural features as low-dimensional vectors; In the regional power grid structure diagram, a READOUT function based on hierarchical aggregation is designed. The first-level READOUT function aggregates the characteristic information of the substation nodes to the distribution line nodes connected to them: (9); in, Indicates the area node The characteristic vector of Represents a distribution line node The neighbor set of express The updated vector; The second-layer READOUT function aggregates all the distribution line nodes processed by the first-layer READOUT function to obtain the overall structural characteristics of the regional power grid. : (10); Step 3, regional power grid time series feature extraction: construct a double-column GRU model based on the attention mechanism, process the time series data of the substation and distribution line respectively, calculate the influence of different attributes and their previous moments on the line loss at the current moment, and obtain the time series feature vectors of each substation and distribution line; Step 4, line loss prediction: Each node in the regional power grid structure diagram obtains the time series characteristics through Step 3, and then integrates it with the overall structural characteristics of the regional power grid obtained in Step 2 to obtain the spatiotemporal feature vector of each node. Finally, two FNNs are used to accurately predict the line losses of the two types of nodes, the substation and the distribution line, in the power grid structure diagram.
2. The regional power grid line loss prediction method integrating spatiotemporal features according to claim 1 is characterized by: The step 1 of generating the regional power grid sample specifically includes the following operations: Step 1.1 Data normalization: The original power data has different dimensions and value ranges. First, the power data is normalized and converted to values between 0 and 1 using minimum-maximum scaling. and minimum value , for each value in the sequence Perform normalization: (1); in, is the original value in the feature sequence, is the normalized value; Step 1.2 Calculation of regional power grid differences: First, we construct a power transmission network diagram. During the power transmission process, the power is transmitted from the power generation end to the power consumption end through six types of equipment, namely, high-voltage transmission lines, main step-down transformers, high-voltage distribution lines, secondary step-down transformers, low-voltage distribution lines, and power consumption ends, which are respectively recorded as , , , , and , these categories of devices are regarded as nodes in the graph, then M={ } represents a set of node types, and then construct edges according to the physical connection relationship between device nodes to obtain the power transmission network diagram ,in, is a node set, is the set of edges; The above-constructed power transmission network graph is embedded and represented, and the graph representation learning method Metapath2vec based on metapath is extended. The Skip-gram model is used to maximize the co-occurrence probability of nodes and their context nodes, and a low-dimensional vector representation of nodes in the power transmission network graph is generated. By defining the power grid metapath and performing random walks, the distribution samples of power grid equipment are obtained, and the Skip-gram model is used to train the obtained Vector representation of midpoints; Then, follow the meta path right Nodes in Perform random walks to obtain the random walk sequence of power grid equipment nodes : (2); in, for The maximum number of visits to each node in ; Then, based on the generated random walk sequence, the corresponding negative samples are obtained by randomly selecting non-element path nodes, and the Skip-gram model based on the negative sample generation strategy is trained to obtain the embedded representation of each power grid device: (3); in, is the size of the context window, is the dimension of the node vector, Embedding representation matrix for all devices in the power grid; The representation vector of each node is passed Functions are aggregated to uniformly represent the node types under the regional power grid meta-path: (4); in, is the meta-path of the regional power grid, and ; Metapath Included class node, and ; for The set of nodes contained in ; Represents a meta path The grid nodes below; Finally, in the meta path The cosine mean of the node type representation vectors in any two regional power grids 𝒜 and ℬ is calculated to measure the metapath The difference between 𝒜 and ℬ in the lower region is: (5); in, is the regional power grid element path The set of node types contained. for The specific types included in express Regional Grid Metapath Down A vector representation of type , and , for Regional Grid Metapath Down A vector representation of type , and , regional power grid 𝒜 and In the meta path The greater the difference in characteristics, The larger the value of is, the smaller it is; Step 1.3 Regional power grid sample selection: make For any subset of the set of grids in a given region, calculate The difference between any two regional power grids in the , and the sum of all regional differences is obtained The overall difference of all regional power grids in: (6); in, Represents a regional power grid collection Any two different regional power grids; make is a subset of the set of power grids in a given region, and , if it does not exist Make , then it is called is the subset with the maximum difference, that is The difference between the regional power grids in is the largest; Formula 1-6 satisfies the submodularity theory, As the objective function, the greedy method is used to solve the optimal subset .
3. The method for predicting regional power grid line loss by integrating spatiotemporal features according to claim 1 is characterized in that: The regional power grid structure feature extraction in Step 2 specifically includes the following operations: Step 2.1 Construction of regional power grid structure diagram: The substation area in the regional power grid and distribution line assembly Considered as a core node, is the number of stations, is the number of distribution lines, and the regional power grid structure diagram is constructed based on the physical connection relationship between nodes ,in, is a node set, is the edge set, is a collection of node types, is a set of edge types; Introducing the area characteristic matrix and the distribution line characteristic matrix , respectively save the initial line loss characteristic information related to the substation and distribution line, is the number of characteristics of the nodes in the area, is the number of characteristics of the distribution line nodes; in order to and Integrate it into the regional power grid structure diagram and use a fully connected neural network to and Convert to a feature matrix with the same dimensions : (7); Among them, X is or , is the Sigmoid activation function, is the weight matrix, is the bias vector; Will and Enter formula (7) respectively, and the two feature dimensions are converted to The matrix is then concatenated to obtain a new feature matrix. ; Step 2.2 Regional power grid structure diagram feature fusion: The regional power grid structure graph contains various types of nodes and edges. The R-GCN that represents the relationship information is extended to capture the node characteristics and topological structure in the regional power grid structure graph. layer, Through the aggregation node The neighbor information of Feature representation of the layer : (8); in, Representation Node On the edge The neighbor set corresponding to type, For Node In the The hidden state of the layer, and The neural network The weight matrix of the layer.
4. The method for predicting regional power grid line loss by integrating spatiotemporal features according to claim 1 is characterized in that: The regional power grid time series feature extraction in Step 3 specifically includes the following operations: Step 3.1 Attribute attention calculation: Design an attention mechanism to calculate the influence weights of different attributes: (11); in, is the query matrix of the substation and distribution line attributes, is the key matrix of the substation and distribution line attributes, is the value matrix of the substation and distribution line loss; The influence weights of different attributes on line loss at different times are calculated by formula (11), and the attribute representation vectors are weighted summed and spliced to the original substation and distribution line representation vectors as the updated feature vector representation of the substation and distribution line at the current time. ; Step 3.2 Temporal attention calculation: Based on the current time obtained in Step 3.1, the updated feature vector representation of the substation and distribution line , further considering the characteristic vectors of the substation and distribution line at the previous moment , expand the two-column GRU model to update the representation of the current station area and distribution line: (12); in, Indicates element-wise multiplication, update gate Determines the characteristics of the previous moment Features of the current moment the extent of the impact; Based on formula (12), the current time of the regional power grid and distribution line is expressed as , and then calculate Always The attention weight at each moment is used to obtain the influence of historical features on current features: (13); in, is the query matrix of the substation and distribution line at different preceding times, is the key matrix of the substation and distribution line at different preceding moments, for The transposed matrix of is the value matrix of the substation and distribution line at the current moment, for The dimension of the vector in The dual-column GRU model based on the attention mechanism processes the substation nodes and distribution line nodes respectively, and extracts the time series features of all nodes through Step 3.1 and Step 3.
2.
5. The method for predicting regional power grid line loss by integrating spatiotemporal features according to claim 1, characterized in that: The line loss prediction of Step 4 specifically includes the following operations: The regional power grid structure diagram The time series characteristics of each node in the grid are combined with the overall structural characteristics of the regional power grid to obtain the spatiotemporal characteristic vector of each node. , whose length is the sum of the length of the time series feature vector and the length of the overall structure feature of the regional power grid. The fused vector is used as the input of the FNN, and the output is the line loss of the node calculated by formula (14): (14); in, is the weight matrix from the input layer to the hidden layer, each row corresponds to the weight vector of a hidden layer neuron, is the bias vector of the hidden layer, is the Sigmoid activation function, is the weight matrix from the hidden layer to the output layer, is the bias vector of the output layer, is the line loss prediction value; Two FNNs are used to construct the loss functions of the substation and distribution line respectively to predict the line losses of substation nodes and distribution line nodes. Enhance the model's robustness to outliers and introduce loss weights In order to balance the impact of the two, the following loss function is designed: (15); in, and are the predicted line loss values of the substation and distribution line, and are the true values of line losses in the substation and distribution lines in the training data, is the preset threshold, and .
6. The regional power grid line loss prediction method integrating spatiotemporal features according to claim 2 is characterized in that: The definition of the grid element path in Step 1.2 is to design the following 4 types of regional grid element paths : :High-voltage transmission line—main step-down transformer—high-voltage distribution line, describing the impact of voltage conversion within the power grid and the layout structure of the transmission line on the power grid differentiation; :Main step-down transformer - high voltage distribution line - secondary step-down transformer, describing the role of transformers of different levels in the power grid and the impact of the voltage level inside the power grid on the difference of the power grid; :High voltage distribution line—secondary step-down transformer—low voltage distribution line, describing the impact of the layout structure of different levels of distribution lines within the power grid on the grid differentiation; : Secondary step-down transformer - low voltage distribution line - power consumption end, describing the impact of the access of end users within the power grid and the layout structure of distribution lines on the grid differentiation.
7. The method for predicting regional power grid line loss by integrating spatiotemporal features according to claim 2 is characterized in that: The greedy method in Step 1.3 is used to solve the optimal subset In the process of generating regional power grid samples, one is selected each time so that Regional grid with the largest gain , so that Maximum; the specific sample generation steps include: ① Initialization , used to store the selected regional power grid; ② Randomly select an initial regional power grid , add it to middle; ③For each unselected regional grid , calculate and add it After the gain ; ④Select Regional grid with the largest gain , add it to the collection middle; ⑤ Repeat steps ③ and ④ until the selected regional grids reach 55% to 65% of the total number of regional grids to form , and finally obtain the regional power grid sample set with the largest difference .
Citation Information
Patent Citations
Line loss prediction method and device based on LSTM network, equipment and medium
CN114548602A
Transformer area line loss prediction method based on feature selection and GRU network
CN115796336A
A theoretical line loss rate prediction model based on deep transfer learning
CN109102126A
Line loss prediction method and system based on LSTM neural network
CN111639783A
Electricity stealing type detection method and system based on multi-label co-occurrence network discrimination
CN116226748A
