Distributed new energy output and load separation method and system based on dynamic graph space-time attention network

Abstract

The invention discloses a distributed new energy output and load separation method based on a dynamic graph space-time attention network, and relates to the technical field of power system data processing. The method comprises the following steps: S1, constructing a dynamic graph structure, fusing multi-source data, and accurately representing the real-time operation topology of a power grid through a dynamic adjacency matrix; s2, constructing a dynamic graph space-time attention network model, capturing dynamic space dependence by using a graph attention layer, capturing time dependence by using a time convolution layer, and embedding a loss function by using a power balance law as a physical constraint; and S3, training the model and performing real-time separation. The method overcomes the defects that an existing method neglects topological dynamic change and lacks physical interpretability, self-adaption is achieved, source-load separation meeting power grid constraints is achieved, and key technical support is provided for accurate sensing and optimal dispatching of power grid data.

Description

Technical Field

[0001] This invention relates to the field of power system data processing and artificial intelligence, and in particular to a distributed renewable energy output and load separation method based on dynamic graph spatiotemporal attention network. Background Technology

[0002] With the increasing penetration of distributed renewable energy sources such as photovoltaics and wind power in distribution networks, the measured power at the distribution network bus nodes is often an aliased signal of distributed renewable energy output and user load. This "source-load" aliasing phenomenon poses significant challenges to accurate power prediction, real-time state estimation, safe and stable dispatching, and renewable energy consumption in the power grid. Therefore, accurately separating the actual output and net load of distributed renewable energy is crucial for understanding the true operating status of the power grid and improving its intelligence level.

[0003] Existing solutions mainly fall into two categories: one is based on direct measurement, which involves installing monitoring devices at all renewable energy outlets, but this method is costly and difficult to deploy on a large scale; the other is based on data-driven separation methods, such as using traditional mathematical statistics or shallow machine learning models for data mining. However, existing data-driven methods have significant shortcomings: First, most methods treat the power grid topology as static, failing to fully consider the dynamic changes in power grid operation caused by switching operations, maintenance plans, etc., and lacking accurate modeling of spatiotemporal correlations; second, the models are mostly pure "data black boxes," failing to effectively embed the physical laws of power grid operation (such as the law of power conservation), resulting in separation results that sometimes violate physical common sense, have poor interpretability, and weak generalization ability; finally, traditional models have limited ability to capture complex spatiotemporal dependencies when processing large-scale, nonlinear, high-dimensional gridded spatiotemporal data (such as meteorological data and historical power data).

[0004] Therefore, there is an urgent need in this field for an innovative technology that can deeply integrate the physical characteristics of the power grid with advanced artificial intelligence algorithms, adapt to dynamic changes in the power grid topology, and provide reliable separation results with strong physical interpretability. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for separating the output and load of distributed new energy sources with high precision, strong physical interpretability, and the ability to adapt to dynamic changes in power grid topology.

[0006] The present invention solves the above-mentioned technical problems through the following technical means: The distributed renewable energy output and load separation method based on dynamic graph spatiotemporal attention network includes the following steps: Step S1: Dynamic graph structure construction and multi-source data fusion: The distribution network is abstracted into a graph structure, and a dynamic adjacency matrix is ​​generated according to the real-time switch status and maintenance plan; at the same time, power grid measurement data and gridded meteorological data are fused to construct a node feature matrix; Step S2: Construction of dynamic graph spatiotemporal attention network model: The model includes a graph attention layer for capturing dynamic spatial dependencies, a time series learning layer for capturing time dependencies, and a physical information constraint layer for embedding the physical laws of the power grid; the model is trained using historical data, and the loss function during training includes prediction error and physical constraint violation terms; Step S3: Output separation: Real-time data is input into the trained model, and the separation results of the theoretical output and net load of distributed renewable energy are output.

[0007] Furthermore, in step S1, the dynamic adjacency matrix From the fundamental topological matrix With dynamic mask matrix Generate by element-wise product ,in Generated based on a permanent power grid wiring diagram. It is determined by real-time switch status and maintenance plan, where the symbol ⊙ represents the Hadamard product, which is the multiplication of corresponding elements of two matrices.

[0008] Furthermore, the attention coefficient of the dynamic graph attention layer in step S2... The calculation formula is:

[0009] in, a This is a learnable attention vector; W The weight matrix is ​​a learnable matrix; h i and h j For node features; || denotes vector concatenation; A t,ij Elements of a dynamic adjacency matrix; N ( i ) is a node i The neighborhood set, LeakyReLU is an activation function that allows the model to learn more complex rating relationships.

[0010] Furthermore, the time-series learning layer in step S2 is a temporal convolutional network.

[0011] Furthermore, the physical information constraint layer in step S2 is implemented using a loss function, which is:

[0012] in, L pred The mean square error between the predicted and actual values. L physics Loss due to physical constraints;

[0013] in, Indicates the true label, Indicates the predicted value. Contribute to the prediction of new energy sources at the node. Predict net load for nodes. Measure the total power of the node. λ For hyperparameters, t For a certain moment.

[0014] Furthermore, early stopping and learning rate decay strategies are employed during model training.

[0015] This invention also provides a distributed renewable energy output and load separation system based on a dynamic graph spatiotemporal attention network, comprising: a dynamic graph structure construction and multi-source data fusion module: abstracting the distribution network into a graph structure and generating a dynamic adjacency matrix based on real-time switch status and maintenance plans; simultaneously, fusing power grid measurement data and gridded meteorological data to construct a node feature matrix; a dynamic graph spatiotemporal attention network model construction module: the model includes a graph attention layer for capturing dynamic spatial dependencies, a time series learning layer for capturing time dependencies, and a physical information constraint layer for embedding the physical laws of the power grid; training the model using historical data, with the loss function during training including prediction error and physical constraint violation terms; and an output separation module: inputting real-time data into the trained model and outputting the separation results of the theoretical output and net load of distributed renewable energy.

[0016] Furthermore, dynamic adjacency matrix From the fundamental topological matrix With dynamic mask matrix Generate by element-wise product ,in Generated based on a permanent power grid wiring diagram. It is determined by real-time switch status and maintenance plan, where the symbol ⊙ represents the Hadamard product, which is the multiplication of corresponding elements of two matrices.

[0017] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method described above.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0019] The advantages of this invention are: By accurately modeling power grid topology changes using dynamic graph structures and combining spatiotemporal attention mechanisms to deeply mine the complex dynamic spatial correlations and long-term temporal dependencies between nodes, the separation accuracy and adaptability to different operating modes are improved.

[0020] By embedding the power balance law as a physical constraint into the model's loss function, the model is guided to learn results that conform to physical laws, thereby enhancing the credibility and interpretability of the results.

[0021] The model learns the spatiotemporal evolution patterns and basic physical laws behind the data, rather than simply fitting the training data. It has good generalization ability for new operating scenarios, topologies, or new energy access situations that it has not experienced before. Attached Figure Description

[0022] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a schematic diagram of the structure of the Dynamic Graph Spatiotemporal Attention Network (DGSTAN) model in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. 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.

[0024] This embodiment describes a distributed renewable energy output and load separation method based on a dynamic graph spatiotemporal attention network, including the following steps: S1. Data Preparation and Preprocessing. The goal of this step is to construct a high-quality, spatiotemporally aligned model input dataset. Specifically: S101. Multi-source data acquisition and understanding. First, data needs to be collected from four sources: Supervisory Control and Data Acquisition (SCADA) system: Acquires time-series data for each node (including busbars, renewable energy access points, and load points), with key indicators including total active power P. meas For voltage (V) and current (I), a sampling interval of 5 minutes or 15 minutes is recommended.

[0025] Power grid dispatch and management system: acquires network wiring diagrams that define power grid connection relationships, as well as switch position signals and maintenance plans that reflect real-time structural changes.

[0026] Meteorological Service Center: Acquire gridded meteorological data that is precisely matched to the geographical location of each node. The core parameters include total irradiance (GHI), ambient temperature (T), and wind speed (WS). Its spatiotemporal resolution should be consistent with that of the power grid measurement data.

[0027] Specialized monitoring and estimation data: serving as the "standard answer" (i.e., label) for model training, mainly includes two categories: one is the actual renewable energy output P directly obtained from some renewable energy stations with installed monitoring devices. dg Secondly, the "quasi-label data" is estimated by selecting special operating days such as "days with extremely low load" (such as statutory holidays like the Spring Festival, when the load is extremely low and can be approximated as all being generated by new energy sources).

[0028] S102. Data Cleaning and Quality Control. Raw data inevitably contains noise and missing information, and must undergo rigorous cleaning: Outlier handling: A sliding window method based on the 3σ criterion is used for identification. Specifically, within a time window, if a data point deviates from the mean by more than three times its standard deviation, it is identified as an outlier. For these outliers, linear interpolation is used for smoothing, that is, the values ​​of the two preceding and following normal data points are used for calculation and replacement.

[0029] Missing value handling: Different strategies are adopted according to the duration of the missing value. For short-term missing values ​​(e.g., less than 1 hour), linear interpolation is used to fill the missing value; for long-term missing values, the data for that period is directly removed to avoid introducing excessive uncertainty.

[0030] Data standardization: To eliminate the influence of different feature units and accelerate model training, Z-score standardization is performed on all input features. That is, for each feature data, its mean μ is subtracted and then divided by its standard deviation σ, so that the processed data has a mean of 0 and a standard deviation of 1.

[0031] S103, Spatiotemporal Alignment and Sample Construction Spatiotemporal alignment: Using a power grid geographic information system (GIS), the grid coordinates of meteorological data are precisely matched with the electrical nodes of the power grid, ensuring that each node has corresponding meteorological information. At the same time, data from all sources are unified to the same timestamp.

[0032] Sequence sample construction: The processed time-series data is constructed into samples that the model can learn from. A historical time window T is defined. in (e.g., 24 hours, used for models to observe historical patterns) and a prediction window T out(Usually 1, meaning predicting the next time step). Each training sample consists of two parts: the input is the past T... in The dynamic graph structure and node characteristics at each moment; the output (label) is the true value of the renewable energy output and net load at a future moment.

[0033] Dataset partitioning: The overall dataset is strictly divided into training set (70%), validation set (20%) and test set (10%) in chronological order, and random shuffling is strictly prohibited, in order to conform to the time dependence of power data and ensure the reliability of model evaluation.

[0034] S2. Construction of the dynamic graph spatiotemporal attention network model. This step is the core, aiming to build an intelligent model that can simultaneously understand the spatial structure, temporal patterns, and physical constraints of the power grid.

[0035] S201. Definition and Generation of Dynamic Graph Structures Graph Structure Abstraction: The power distribution network is modeled as a graph structure G = (V, E, A). Here, V is the set of nodes, containing all the electrical nodes of interest; E is the set of edges, representing the electrical connections between nodes; and A is the adjacency matrix, a mathematical table used to quantify the connection relationships and strength between nodes.

[0036] Dynamic adjacency matrix: We generate a matrix that reflects real-time changes in the power grid topology using a formula: .here, This is the basic topology matrix, generated based on the permanent power grid wiring diagram (connected is 1, otherwise 0). M t It is a dynamic mask matrix that is dynamically updated based on real-time switch status and maintenance plans (connected is 1, disconnected is 0). The symbol ⊙ represents the Hadamard product, which is the element-wise multiplication of two matrices at corresponding positions. This operation ensures that the graph structure can accurately capture the topology changes caused by switch operations and maintenance.

[0037] S202, Node Feature Engineering. Construct a feature vector for each node in the graph, incorporating information from multiple sources. For example, the feature vector of a typical node might include: the node's real-time measurement data (P...). meas The node feature matrix X consists of: (V, I), real-time meteorological data (GHI, T, WS) of the location, historical power sequences over a past period, and a type code to distinguish node types (e.g., bus, photovoltaic, load). All node features together constitute the node feature matrix X. t .

[0038] S203, the core mechanism of the dynamic graph spatiotemporal attention network. The model is composed of multiple identical "spatiotemporal blocks" stacked together, each block containing learning modules in both spatial and temporal dimensions.

[0039] Spatial Dimension: Dynamic Graph Attention Layer. This layer allows each node to selectively obtain information from its neighboring nodes.

[0040] The first step is to calculate the correlation: For a pair of adjacent nodes (i, j), the unnormalized correlation score between them is calculated using a small neural network. e ij This calculation incorporates the characteristics of nodes i and j themselves. .

[0041] The second step, normalization and masking: Use the softmax function to normalize the association scores between node i and all its neighbors into attention coefficients that sum to 1. The key point is that this process multiplies by the dynamic adjacency matrix A. t,ij The element. If this element is 0 (i.e., the line is broken), the attention weight is forced to 0, achieving dynamic spatial modeling that only focuses on the currently electrically connected neighbors. Attention coefficient The calculation formula is:

[0042] in, a This is a learnable attention vector; W The weight matrix is ​​a learnable matrix; h i and h j For node features; || denotes vector concatenation; A t,ij Elements of a dynamic adjacency matrix; N ( i ) is a node i The set of neighbors, where LeakyReLU is the activation function.

[0043] The third step is feature aggregation: the updated features of node i are the weighted sum of the features of all its neighboring nodes, and the weights are the attention weights calculated in the previous step. This allows the model to flexibly capture complex spatial correlations.

[0044] Temporal Dimension: Time Series Learning Layer. This layer focuses on learning the patterns of how the features of each node change over time. This invention recommends using a Temporal Convolutional Network (TCN), which, through a special "dilated causal convolution" technique, can efficiently capture long-term temporal dependencies (e.g., the time lag and diurnal cycle patterns between photovoltaic power output and solar radiation intensity) and is computationally stable.

[0045] S204. Embedding of Physical Information Constraints. This aims to integrate physics knowledge into data-driven models, enhancing their rationality and interpretability.

[0046] Joint loss function: The goal of model training is no longer simply to fit the data, but to minimize a joint loss function: L total = L pred + λ·L physics .

[0047] Predicted loss L pred : This is the mean squared error between the model output value and the true label value, ensuring the accuracy of the model prediction.

[0048]

[0049] Where, N lable The number of nodes in this set, Ω lable It is a set of labeled nodes.

[0050] Physical constraint loss L physics This is the core of the invention; it calculates the difference between the sum of the new energy output and net load output by the model and the total power actually measured at that node. This item mandates that the model's output must approximately satisfy the fundamental physical law of "power conservation."

[0051]

[0052] Hyperparameter λ: Used to balance the importance of both data fitting and physical constraints; in this embodiment, it is recommended to set it to 0.5.

[0053] S3, Model Training, Evaluation, and Practical Applications S301, Model Training and Tuning Training configuration: An adaptive learning rate optimizer, such as the AdamW optimizer (which effectively prevents overfitting), is used. The initial learning rate is set to 1e-4, and the batch size is set to 64. A cosine annealing strategy is used to dynamically adjust the learning rate, which helps the model converge to a better solution.

[0054] To prevent overfitting: Continuously monitor performance on the validation set during training. If the validation set loss stops decreasing for 10 consecutive epochs, an early stopping mechanism is triggered, and the model parameters that performed best during this period are automatically restored, thus preventing the model from over-memorizing training data and losing its generalization ability.

[0055] S302. Performance Evaluation and Validation. After the model is trained, a comprehensive evaluation is performed on an independent test set: Separation accuracy: The root mean square error (RMSE) and mean absolute error (MAE) are used to quantitatively assess the deviation between the new energy output separation results and the true values.

[0056] Physical consistency: Calculate the power balance deviation of all nodes at all times, i.e. The mean and maximum value of . In an ideal model, this indicator should be very close to zero.

[0057] S303, Real-time Output Separation Application. Deploying the trained and evaluated model to the actual power grid system: Input: The system receives real-time access to current switch status, maintenance plans, measurement data, and meteorological data, and automatically generates a dynamic graph structure (A). t ) and node features (X t ).

[0058] Inference: Input the above real-time data into the trained model.

[0059] Output: The model automatically outputs the theoretical output of distributed new energy sources for all bus nodes in the entire network. and net load The separation results.

[0060] Service: The separation results can be directly provided to the power grid dispatching system to improve the accuracy of power prediction, state estimation and optimized dispatching.

[0061] This embodiment addresses the signal separation problem in a "state-aware" context, aiming to restore underlying physical quantities that cannot be directly measured. The method provided in this embodiment can accurately separate distributed renewable energy output and net load superimposed on bus measurement data. Specifically, this embodiment is based on the power grid physical topology and real-time switching / maintenance status. Its dynamism is reflected in the hard change (0 or 1) of the adjacency matrix as physical switches are turned on and off, accurately reflecting the true electrical connection relationship of the power grid. The dynamism of this embodiment stems from changes in the physical operation mode of the power grid, and is causal and deterministic, providing a more accurate and essential characterization of the power grid's physical structure. In terms of spatial feature extraction, this embodiment uses a dynamic graph attention network (GAT) to assign different attention weights to adjacent nodes in the graph, which can capture complex spatial dependencies more flexibly and non-uniformly.

[0062] Regarding physical information fusion: This embodiment directly embeds the physical law (power balance) as the core constraint into the deep learning model. Method 1 (soft constraint): MSE(P) meas -(P dg + P load The variable P is added as a regularization term to the loss function. Method 2 (hard constraint): The model only outputs P. dg Through P load =P meas -P dg Automatically achieves power balance.

[0063] This embodiment introduces physical laws as guidance in the core stage of model training / inference, which ensures the physical rationality of the results from the "root". Regarding output and function: This embodiment outputs the distributed renewable energy theoretical output P of each node. dg and net load P load This embodiment achieves source decomposition of aliased signals; it provides high-precision, physically interpretable state estimation results, providing a cleaner and more reliable data foundation for subsequent prediction, scheduling and other applications. Based on the above innovations, this embodiment solves a completely new and more fundamental technical problem and overcomes the corresponding technical difficulties: 1. Technical problems to be solved: In complex distribution network environments characterized by dynamic changes in grid topology and deep overlap between renewable energy and load, how can we accurately and physically separate the distributed renewable energy generation output and actual load components from the total power measured at the bus without the need for large-scale installation of dedicated monitoring equipment? 2. Technical challenges overcome: Challenge 1: Conflict between "black box" models and physical laws: Purely data-driven deep learning models may produce separation results that violate the law of power conservation, resulting in low credibility and inapplicability in engineering. This embodiment overcomes the challenge of "physical unreliability" of AI models in power systems by embedding physical constraints into the model kernel. Challenge 2: Accurate Modeling of Complex Spatiotemporal Correlations and Dynamic Topologies: Traditional static graph models or dynamic graphs based solely on data correlations cannot accurately describe the abrupt changes in electrical connections in the power grid caused by switching operations, affecting the accuracy of spatial correlation learning. This embodiment overcomes the challenge of accurately capturing spatial dependencies under dynamic changes in real physical topology by introducing a dynamic graph attention network based on real-time switching states. Challenge 3: Ambiguity and Uncertainty in Separating Aliased Signals: Separating multiple source signals from a single aliased signal is an inherently ill-posed problem with a non-unique solution. This embodiment effectively reduces the uncertainty of the problem and overcomes the challenge of non-unique separation results by employing a dual strategy of "mining deep patterns through spatiotemporal attention mechanisms" + "constraining the solution space by physical laws". By implementing the above-mentioned innovative technical solution, this embodiment has achieved significant technical progress and effects not covered in the prior art: 1. Dual Improvement in Separation Accuracy and Physical Reliability: This embodiment not only fully exploits the nonlinear spatiotemporal characteristics of the data through a dynamic graph spatiotemporal attention network, but more importantly, it ensures that the separation results strictly follow the power balance law through physical information constraints. This makes the output results both data-driven and physically guided, with accuracy and reliability far exceeding those of purely data-driven models, thus solving the "reliability" bottleneck of AI models in critical power applications. 2. Adaptability to changes in power grid operation mode: Since the dynamic graph structure is directly driven by real-time switch states, the model in this embodiment can automatically adapt to changes in the topology of the power grid (such as line maintenance and network reconfiguration). It can maintain high-precision separation capability without retraining the model, and has extremely high generalization and engineering practical value. 3. Provides a high-quality data foundation for upper-level applications: The accurately separated renewable energy output and net load can greatly improve the accuracy of advanced applications such as short-term / ultra-short-term power forecasting, state estimation, and reactive power optimization. For example, using the separated, clean load data for load forecasting can avoid noise interference caused by renewable energy fluctuations, which is unmatched by load forecasting results from comparative files. 4. End-to-end automation and efficiency: This method provides an end-to-end solution, avoiding the error accumulation and complexity caused by the splicing of multiple independent modules (such as estimating power flow first and then separating) in traditional methods, and realizing efficient and automated output and load separation. The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed renewable energy output and load separation method based on dynamic graph spatiotemporal attention networks, characterized in that, Includes the following steps: Step S1: The distribution network is abstracted into a graph structure, and a dynamic adjacency matrix is ​​generated based on real-time switch status and maintenance plan; at the same time, power grid measurement data and gridded meteorological data are integrated to construct a node feature matrix; Step S2: Construct a dynamic graph spatiotemporal attention network model, which includes a graph attention layer for capturing dynamic spatial dependencies, a time series learning layer for capturing temporal dependencies, and a physical information constraint layer for embedding the physical laws of the power grid; train the dynamic graph spatiotemporal attention network model using historical data, and the loss function during training includes prediction error and physical constraint violation terms; Step S3: Input real-time data into the trained model and output the separation results of distributed renewable energy theoretical output and net load.

2. The method according to claim 1, characterized in that, In step S1, the dynamic adjacency matrix From the fundamental topological matrix With dynamic mask matrix Generate by element-wise product ,in Generated based on a permanent power grid wiring diagram. It is determined by real-time switch status and maintenance plan, where the symbol ⊙ represents the Hadamard product, which is the multiplication of corresponding elements of two matrices.

3. The method according to claim 1, characterized in that, The attention coefficient of the dynamic graph attention layer in step S2 The calculation formula is: in, a This is a learnable attention vector; W The weight matrix is ​​a learnable matrix; h i and h j For node features; || denotes vector concatenation; A t,ij Elements of a dynamic adjacency matrix; N ( i ) is a node i The set of neighbors, where LeakyReLU is the activation function.

4. The method according to claim 1, characterized in that, The time-series learning layer in step S2 is a temporal convolutional network.

5. The method according to any one of claims 1 to 4, characterized in that, The physical information constraint layer in step S2 is implemented through a loss function, which is: in, L pred The mean square error between the predicted and actual values. L physics Loss due to physical constraints; in, Indicates the true label, Indicates the predicted value. Contribute to the prediction of new energy sources at the node. Predict net load for nodes. Measure the total power of the node. λ For hyperparameters, t For a certain moment.

6. The method according to claim 1, characterized in that, In step S2, the model training employs early stopping and a learning rate decay strategy.

7. A distributed renewable energy output and load separation system based on a dynamic graph spatiotemporal attention network, characterized in that, include: Dynamic graph structure construction and multi-source data fusion module: Abstracts the distribution network into a graph structure and generates a dynamic adjacency matrix based on real-time switch status and maintenance plan; at the same time, it integrates power grid measurement data and gridded meteorological data to construct a node feature matrix; A dynamic graph spatiotemporal attention network model module is constructed: the model includes a graph attention layer for capturing dynamic spatial dependencies, a time series learning layer for capturing temporal dependencies, and a physical information constraint layer for embedding the physical laws of the power grid; the model is trained using historical data, and the loss function during training includes prediction error and physical constraint violation terms; the output separation module: real-time data is input into the trained model, and the output is the separation result of the theoretical output and net load of distributed new energy.

8. The system according to claim 7, characterized in that, The aforementioned dynamic adjacency matrix From the fundamental topological matrix With dynamic mask matrix Generate by element-wise product ,in Generated based on a permanent power grid wiring diagram. It is determined by real-time switch status and maintenance plan, where the symbol ⊙ represents the Hadamard product, which is the multiplication of corresponding elements of two matrices.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1 to 6.

Images