A section transmission power probability prediction method, device, equipment and storage medium
By converting the power grid topology into a power system graph, and utilizing graph neural networks and hybrid density networks, the problem of traditional methods being unable to efficiently process unstructured data in power systems is solved, achieving more accurate and efficient power prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA SOUTHERN POWER GRID COMPANY
- Filing Date
- 2025-02-27
- Publication Date
- 2026-06-26
AI Technical Summary
Traditional methods cannot efficiently process unstructured data in power systems and cannot adapt to dynamically changing power grid topologies, resulting in low accuracy in power prediction.
The power grid topology is converted into a power system graph. Graph neural networks are used to extract node embedding vectors. Combined with a spatiotemporal feature extraction model and a hybrid density network, a cross-sectional transmission power probability distribution is constructed.
It improves the accuracy and efficiency of power probability prediction, better reflects the load and energy transmission status of the power grid, and supports the security analysis and dispatch of the power system.
Smart Images
Figure CN120165366B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of power prediction technology, and in particular to a method, apparatus, device and storage medium for predicting cross-sectional transmission power probability. Background Technology
[0002] In power systems, to ensure the transmission and distribution of electrical energy, it is necessary to monitor and control the transmission power at different cross-sections. Cross-sectional transmission power refers to the power of electrical energy passing through a specific set of equipment (such as transmission lines, transformers, etc.) in a power system. The cross-section is usually a node or line in the power grid, and the power passing through that cross-section can reflect the load and energy transmission status of that part of the power grid.
[0003] Accurate prediction of cross-sectional transmission power is crucial for the safe and stable operation of the power system. By monitoring and analyzing the transmission power of different cross-sections in real time, grid operators can promptly identify and resolve potential problems, avoid overloads and faults, and ensure the reliable operation of the grid.
[0004] However, traditional methods model power systems based on matrices or vectors, which makes it difficult to efficiently process unstructured data and adapt to dynamically changing power grid topologies, thus reducing the accuracy of power prediction results. Summary of the Invention
[0005] The purpose of this invention is to provide a method, apparatus, device, and storage medium for predicting cross-sectional transmission power probability, which can improve the accuracy and efficiency of power probability prediction, thereby assisting in the safety analysis and power dispatching of power systems.
[0006] To achieve the above objectives, embodiments of the present invention provide a method for probabilistic prediction of cross-sectional transmission power, comprising:
[0007] Input the power grid topology, and convert the power grid topology into a power system diagram with the busbars in the power grid topology as nodes and the lines as edges;
[0008] The power system diagram is input into a graph neural network to obtain node embedding vectors;
[0009] Initial values are generated based on the power system diagram, and power flow calculations are performed on the initial values to construct a training dataset; wherein, the initial values include load demand and unit output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power;
[0010] The node embedding vector is input into the spatiotemporal feature extraction model to obtain spatiotemporal feature information;
[0011] The spatiotemporal feature information and the training dataset are input into a hybrid density network to obtain the cross-sectional transmission power probability distribution.
[0012] As an improvement to the above scheme, the spatiotemporal feature extraction model includes a convolutional neural network and a long short-term memory network, and the spatiotemporal feature information is obtained in the following way:
[0013] The node embedding vector is input into the convolutional neural network to obtain intermediate features;
[0014] The intermediate features and the training dataset are input into the long short-term memory network to obtain spatiotemporal feature information.
[0015] As an improvement to the above scheme, the cross-sectional transmission power probability distribution is as follows:
[0016]
[0017] Among them, P ′ (y|x) represents the cross-sectional transmission power probability distribution; y represents the target vector; x represents the spatiotemporal feature information; N(y|μ) i (x),σ i (x)) represents the conditional density of the target vector of the i-th kernel; g represents the total number of kernels in the hybrid density network; μ i (x) represents the center of the i-th nucleus; σ i (x) represents the common variance.
[0018] As an improvement to the above scheme, N(y|μ i (x),σ i The expression for (x) is shown below:
[0019]
[0020] Where y represents the target vector; x represents spatiotemporal feature information; π i (x) represents the mixing coefficient; N(y|μ) i (x),σ i (x)) represents the conditional density of the target vector of the i-th kernel; μ i (x) represents the center of the i-th nucleus; σ i (x) represents the common variance; c represents the dimension of the target vector.
[0021] As an improvement to the above scheme, the training dataset is constructed in the following way:
[0022] Initialize the load demand power of each side of the power system diagram and the unit output data of each bus node; wherein, the unit output data includes the active power and reactive power injected by the generator into the bus node;
[0023] Based on the unit output data and the load demand power, a power balance equation is constructed.
[0024] The power balance equation is solved iteratively to obtain the cross-sectional transmission power under steady state.
[0025] The unit output data, the load demand power, and the corresponding cross-sectional transmission power are added as a set of data to the training dataset.
[0026] As an improvement to the above scheme, in each iteration, the voltage amplitude correction and voltage phase angle correction of each bus node are calculated. Based on the voltage amplitude correction and voltage phase angle correction, the voltage amplitude and voltage phase angle of each bus node are corrected, and then a new iteration is entered.
[0027] When both the voltage amplitude correction and the voltage phase angle correction are less than the preset error upper limit, the cross-sectional transmission power between the bus nodes is calculated based on the current voltage amplitude of each bus node and the current voltage phase angle difference between the bus nodes.
[0028] As an improvement to the above scheme, the step of inputting the power system graph into a graph neural network to obtain node embedding vectors includes:
[0029] Input the initial feature matrix and adjacency matrix of the power system diagram;
[0030] Based on the adjacency matrix, the nodes in the power system graph are aggregated to obtain hidden layer features;
[0031] Linear regression prediction is performed on the hidden layer features to obtain the node embedding vector.
[0032] To achieve the above objectives, embodiments of the present invention also provide a cross-sectional transmission power probability prediction device, comprising:
[0033] The graph construction module is used to input the power grid topology and convert the power grid topology into a power system graph, with the busbars in the power grid topology as nodes and the lines as edges.
[0034] The graph feature extraction module is used to input the power system graph into a graph neural network to obtain node embedding vectors.
[0035] The training data generation module is used to generate initial values based on the power system diagram, and to perform power system flow calculations on the initial values to construct a training dataset; wherein, the initial values include load demand and unit output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power;
[0036] The spatiotemporal feature information extraction module is used to input the node embedding vector into the spatiotemporal feature extraction model to obtain spatiotemporal feature information;
[0037] The power probability prediction module is used to input the spatiotemporal feature information and the training dataset into the hybrid density network to obtain the cross-sectional transmission power probability distribution.
[0038] To achieve the above objectives, embodiments of the present invention also provide a cross-sectional transmission power probability prediction device, including a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, it implements the cross-sectional transmission power probability prediction method as described in any of the above embodiments.
[0039] To achieve the above objectives, embodiments of the present invention also provide a computer-readable storage medium, the computer-readable storage medium including a stored computer program, wherein, when the computer program is executed, it controls the device where the computer-readable storage medium is located to perform the cross-sectional transmission power probability prediction method as described in any of the above embodiments.
[0040] Compared with existing technologies, the present invention provides a method, apparatus, device, and storage medium for predicting cross-sectional transmission power probabilistically. The method involves inputting a power grid topology, using buses as nodes and lines as edges, to convert the topology into a power system graph. The power system graph is then input into a graph neural network to obtain node embedding vectors. Initial values are generated based on the power system graph, and power flow calculations are performed on these initial values to construct a training dataset. The initial values include load demand and generator output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power. The node embedding vectors are input into a spatiotemporal feature extraction model to obtain spatiotemporal feature information. Finally, the spatiotemporal feature information and the training dataset are input into a hybrid density network to obtain the cross-sectional transmission power probability distribution. Compared with existing technologies, the present invention can improve the accuracy and efficiency of power probability prediction. Attached Figure Description
[0041] Figure 1 This is a flowchart of a cross-sectional transmission power probability prediction method provided in an embodiment of the present invention;
[0042] Figure 2 This is a schematic diagram of a graph convolutional neural network and a multi-task regression module provided in an embodiment of the present invention;
[0043] Figure 3 This is a schematic diagram of a self-balancing loss training module provided in an embodiment of the present invention;
[0044] Figure 4This is a schematic diagram of the structure of a cross-sectional transmission power probability prediction device provided in an embodiment of the present invention;
[0045] Figure 5 This is a schematic diagram of the structure of a cross-sectional transmission power probability prediction device provided in an embodiment of the present invention. Detailed Implementation
[0046] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and 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.
[0047] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, unless otherwise stated, "a plurality of" means two or more.
[0048] In the description of this application, it should be noted that, unless otherwise expressly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.
[0049] See Figure 1 This is a flowchart of a cross-sectional transmission power probability prediction method provided in an embodiment of the present invention, including steps S1 to S5:
[0050] S1. Input the power grid topology, and convert the power grid topology into a power system diagram with the busbars in the power grid topology as nodes and the lines as edges.
[0051] S2. Input the power system diagram into a graph neural network to obtain node embedding vectors;
[0052] S3. Generate initial values based on the power system diagram, and perform power system flow calculations on the initial values to construct a training dataset; wherein, the initial values include load demand and unit output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power;
[0053] S4. Input the node embedding vector into the spatiotemporal feature extraction model to obtain spatiotemporal feature information;
[0054] S5. Input the spatiotemporal feature information and the training dataset into the hybrid density network to obtain the cross-sectional transmission power probability distribution.
[0055] In step S1, it can be understood that the line can be understood as a transmission line. In the power grid topology, the line connects the busbars. Therefore, in this embodiment of the invention, the busbars are used as nodes in the power system diagram, and the lines are used as edges in the power system diagram. It is worth noting that, in the following text, the "node" will also be referred to as the "busbar node".
[0056] Furthermore, in step S2, by inputting the power system diagram into the graph neural network, not only can the efficiency of feature extraction be improved, but also the features of the power grid topology can be extracted globally. Compared with the prior art which only extracts local features, the present invention can extract the global features of the power grid topology. Moreover, the features extracted by the present invention include not only the power grid connection relationship, but also the attribute information of each element in the power grid topology, thereby improving the accuracy and reliability of subsequent power probability prediction.
[0057] Furthermore, in step S3, the training dataset includes multiple sets of data. Each set of data includes generator output data, load demand, and cross-sectional transmission power between bus nodes. The generator output data and load demand need to be randomly initialized, while the cross-sectional transmission power between bus nodes is calculated based on the generator output data and load demand. Furthermore, the generator output data includes the active and reactive power injected by the generators into the bus nodes, and the load demand is the power required by the load.
[0058] Furthermore, in step S4, the spatiotemporal feature extraction model may include a Convolutional Neural Network (CNN) and a Long Short-Term Memory (LSTM) network. In this embodiment of the invention, the temporal information of graph features is extracted by CNN and LSTM, and spatiotemporal information fusion is performed.
[0059] Furthermore, in step S5, the future cross-sectional power level is predicted based on Mixture Density Networks (MDN). The spatiotemporal feature information extracted by LSTM is used as input, and the output is the cross-sectional transmission power probability distribution, thus more comprehensively reflecting the uncertainty and risk of power on the transmission line and providing more accurate information for power system operation.
[0060] Furthermore, after obtaining the cross-sectional transmission power probability distribution model, in some implementations, the cross-sectional transmission power probability distribution model can be used to perform power grid security analysis or power dispatching, which is not limited here.
[0061] Compared with existing technologies, the embodiments of the present invention can more accurately capture the features of the power grid topology by introducing graph neural networks. Combined with spatiotemporal feature extraction models and hybrid density networks, the accuracy and efficiency of power probability prediction can be improved.
[0062] As one optional implementation, in step S2, the power system graph is input into a graph neural network to obtain node embedding vectors, including:
[0063] Input the initial feature matrix and adjacency matrix of the power system diagram;
[0064] Based on the adjacency matrix, the nodes in the power system graph are aggregated to obtain hidden layer features;
[0065] Linear regression prediction is performed on the hidden layer features to obtain the node embedding vector.
[0066] Understandably, after obtaining the power system graph, graph feature extraction is required. In some implementations, graph neural networks mainly include three modules: a graph convolutional network (GCN) module, a multi-task regression module, and a self-balancing loss training module.
[0067] See Figure 2 This is a graph convolutional neural network and multi-task regression module provided in an embodiment of the present invention (i.e., Figure 2 A schematic diagram of the "multi-branch fully connected layer" in [the diagram]. Figure 2 As can be seen, the graph convolutional neural network (Graph Convolutional Neural Network) module includes an input module and intermediate layers. Its function is to learn the implicit information between nodes using a pre-defined adjacency matrix A, and generate hidden layer features using the aggregation rules of spectral domain convolution. These hidden layer features contain structural information between nodes. Linear regression prediction is then performed through a multi-branch fully connected layer to obtain the node embedding vectors of different nodes. It is understandable that the Graph Convolutional Neural Network module has good scalability and can adapt to more complex power system scenarios. Furthermore, through the Graph Convolutional Neural Network module and the multi-task regression module, information between different nodes is fully transferred and fused, improving the model's performance. See also... Figure 3 This is a schematic diagram of a self-balancing loss training module provided in an embodiment of the present invention. Figure 3In the diagram, solid arrows indicate backpropagation, while other solid arrows indicate forward propagation. The self-balancing loss training module is introduced to further improve model robustness and parameter identification accuracy. It modifies the original loss function that averages each node, preventing the intermediate layer feature distribution of the graph convolutional neural network from becoming excessively sharp due to differences between nodes, thus achieving outlier reduction. Compared to existing technologies, this embodiment of the invention improves model robustness by introducing the self-balancing loss training module, avoids overfitting and noise interference, and enhances the model's generalization ability.
[0068] Furthermore, by Figure 2 As can be seen from the left box, unlike image data, the power grid branch parameter data source is represented as a two-dimensional matrix. Each row represents a time-series data record, while the number of columns indicates the number of operating parameters measured and recorded by the SCADA (Supervisory Control And Data Acquisition) system. Furthermore, based on the requirements of graph neural networks for the characteristics of input data, the input layer data needs to be structured into graph data.
[0069] The multi-task regression module uses an MLP (Multilayer Perceptron) as the regressor. Specifically, based on the dimensions of the node embedding vectors, the sigmoid function is selected as the activation function to prevent the appearance of a large number of outliers. This model can predict multiple node embedding vectors simultaneously, thereby improving the model's efficiency and accuracy.
[0070] The input is divided into two parts: an input feature matrix X with dimension (N, F0), where N is the number of nodes in the network and F0 is the number of input features for each node; the graph structure is represented by the adjacency matrix A ∈ R. N×N Therefore, the feature H of the hidden layer after the input passes through the graph neural network can be expressed as:
[0071] H (l+1) =f(H (l) A) (1)
[0072] Among them, H (0) =X, hidden layer feature matrix H (l) The dimension of the graph neural network is N×F(1), and f(·) is the information aggregation rule of the graph neural network. In each layer, the graph neural network aggregates this information using the propagation rule f(·) to form the features of the next layer. Within this framework, the various variants of the graph neural network differ only in the propagation rule f(·).
[0073] In machine learning, computational complexity implies stronger fitting ability and robustness. Our model needs to consider the branch topology and cannot ignore the influence of the original branch characteristics on parameter identification. GCN uses convolution kernel operations to convolve the data with the connection relationships of the power grid branches. Define the Laplacian matrix of the graph as L = DA, normalize the Laplacian matrix to Δ, and decompose the normalized matrix Δ.
[0074]
[0075] Where △ represents the normalized matrix; I represents the identity matrix; D represents the degree matrix; and A represents the adjacency matrix.
[0076] Furthermore, Chebyshev polynomials are introduced to replace the convolution kernel to reduce the time complexity of the model based on equation (2). According to the definition of Chebyshev polynomials, the convolution of a graph neural network can be approximately defined as shown in equation (5):
[0077]
[0078] Where X represents the input feature; This represents the convolution kernel or filter in graph convolution operations, used to extract local features of the graph; The parameters representing the convolution kernel; express The degree matrix; Let A be the substitution matrix for the adjacency matrix A, and I∈R n×n It is an identity matrix.
[0079] It is worth noting that using The adjacency matrix A is replaced by an identity matrix I because when the input features increase exponentially, calculating only the weighted sum of the features of all the neighbors of a node would ignore the features of the node itself. Therefore, an identity matrix I is added to A to replace the adjacency matrix A.
[0080] Thus, the convolution aggregation rule on the graph is obtained, and the final hidden layer features can be expressed as shown in equation (4):
[0081]
[0082] in, Represents the convolution aggregation rule of a graph neural network; W (l) Let represent the training weight matrix of the l-th layer; σ represents the nonlinear activation function.
[0083] As one alternative implementation, the training dataset is constructed in the following manner:
[0084] Initialize the load demand power of each side of the power system diagram and the unit output data of each bus node; wherein, the unit output data includes the active power and reactive power injected by the generator into the bus node;
[0085] Based on the unit output data and the load demand power, a power balance equation is constructed.
[0086] The power balance equation is solved iteratively to obtain the cross-sectional transmission power under steady state.
[0087] The unit output data, the load demand power, and the corresponding cross-sectional transmission power are added as a set of data to the training dataset.
[0088] As one possible implementation, in each iteration, the voltage amplitude correction and voltage phase angle correction of each bus node are calculated, and the voltage amplitude and voltage phase angle of each bus node are corrected based on the voltage amplitude correction and voltage phase angle correction, and then a new iteration is entered.
[0089] When both the voltage amplitude correction and the voltage phase angle correction are less than the preset error upper limit, the cross-sectional transmission power between the bus nodes is calculated based on the current voltage amplitude of each bus node and the current voltage phase angle difference between the bus nodes.
[0090] For example, the `np.random.uniform` function from the NumPy library is used to randomly generate generator output data and load demand power, and then power flow calculations are performed. The power flow calculation aims to calculate the active power, reactive power, voltage amplitude, and voltage phase angle of each node in the power grid under steady state, using given partially known grid parameters, i.e., initial power flow values.
[0091] In practical power flow calculations, buses are often classified into three types: constant power buses, voltage stabilizing buses, and balancing buses. First, a bus carrying generators is selected as the balancing bus to balance the system power. The voltage amplitude V at this bus is... i With phase angle θ i Given a value, the active power injection P i With reactive power injection Q i This is given by calculation and is therefore also called the Vθ node. The remaining nodes with generators act as voltage stabilizing buses. These buses need adjustable reactive power capacity to maintain their set voltage amplitude. The active power injected at this bus is P. i With voltage amplitude V i Given a value, the reactive power Q i With voltage phase angle θ i The required parameters are to be determined, hence the name PV node. Finally, the remaining nodes are set as constant power buses, with their active power P...i With reactive power Q i Given a value, the voltage amplitude V i With phase angle θ i It is called a PQ node because it is pending.
[0092] The power flow calculation first uses Kirchhoff's Law to derive a set of nonlinear power balance equations for the target power grid, and then solves them to obtain the steady-state solution of the system's power flow problem. In polar coordinates, the power balance equations are shown in equation (5):
[0093]
[0094] Where N is the total number of busbars in the power grid; Inject active power into bus i. and These represent the active power of the generator and the load at bus i, respectively. Inject reactive power at bus i. and These represent the reactive power of the generator and the load at that node, respectively; V i θ represents the voltage amplitude at bus i; ij =θ i -θ j G represents the difference in voltage phase angle between bus i and bus j; ij With B ij These are the conductance and susceptance values of the conductors between bus i and bus j, respectively; for a power system including source-load-grid-storage, the injected power at bus i is... in It is the active power stored at bus i, where This indicates that the energy storage is in a charging state. Early power flow calculation methods relied on manual calculations and could only calculate power flow in small power grids.
[0095] Furthermore, the power balance power flow equation (6) is solved iteratively using the NR method. In the k-th iteration, let:
[0096]
[0097] Wherein, the superscript (k) represents the value of the variable in the k-th iteration; and These represent the corrected imbalance amounts of active and reactive power in the k-th iteration, respectively.
[0098] Furthermore, by performing a Taylor expansion on equation (6), we obtain the corrected system of equations as shown in equation (7):
[0099]
[0100] in, and These represent the voltage amplitude and phase angle correction at bus i in the k-th iteration, respectively.
[0101] Furthermore, (7) can be simplified into a matrix form as shown in equation (8):
[0102]
[0103] Where J is the Jacobian matrix.
[0104] Furthermore, after calculating the values of each element of the Jacobian matrix, the voltage magnitude correction at each node is solved. and phase angle correction amount And calculate the new voltage values at each bus as shown in equation (9):
[0105]
[0106] Furthermore, after obtaining the new values of the voltage at each node, substitute them back into equation (6) for the (k+1)th iteration to obtain the corrected imbalance between the new active power and reactive power. and And check whether it meets the convergence condition, that is, check whether it is less than the set upper limit of error as shown in equation (10):
[0107]
[0108] Where ε is the set upper limit of error.
[0109] If the convergence condition is met, the calculation result is output; if the convergence condition is still not met after the number of iterations reaches the set upper limit, the system power flow is judged to be non-convergent.
[0110] Furthermore, based on the iterative calculation results, the active power (cross-sectional transmission power) P of bus i and bus j in state t is calculated. ij (t), as shown in equation (11):
[0111]
[0112] Among them, Z ij This represents the impedance value of the conductor between busbar i and busbar j.
[0113] Furthermore, a training dataset is constructed based on load demand, unit output data, and corresponding cross-sectional transmission power.
[0114] As one optional implementation, the spatiotemporal feature extraction model includes a convolutional neural network and a long short-term memory network. The spatiotemporal feature information is then obtained in the following manner:
[0115] The node embedding vector is input into the convolutional neural network to obtain intermediate features;
[0116] The intermediate features and the training dataset are input into the long short-term memory network to obtain spatiotemporal feature information.
[0117] Specifically, in CNN, the node embedding vector is taken as input, and a one-dimensional convolution window of size 1×7 is used to perform convolution operation on the node embedding vector in a downward stepping manner to ensure that the feature matrix of the hidden layer contains the input features of other branches. Finally, a Linear layer is used for prediction.
[0118] In LSTM, considering the historical time series relationships of line nodes in the training dataset, the first layer of LSTM is used to obtain the latent features of line nodes for prediction, and the second layer of LSTM is used to predict branch parameters.
[0119] This invention uses 80% of the training dataset as the training set, 10% as the validation set, and the last 10% as the test set. To reduce storage overhead, the batch size of data fed to the model is set to 4. The model will be trained for 200 generations in a PyTorch environment, using the most popular Adam algorithm as the optimization function, with an initial learning rate of 0.002 and a learning rate decay of 0.9 every 10 generations.
[0120] As one optional implementation, the cross-sectional transmission power probability distribution is shown in the following formula:
[0121]
[0122] Among them, P ′ (y|x) represents the cross-sectional transmission power probability distribution; y represents the target vector; x represents spatiotemporal characteristic information; π i (x) represents the mixing coefficient; N(y|μ) i (x),σ i (x)) represents the conditional density of the target vector of the i-th kernel; g represents the total number of kernels in the hybrid density network; μ i (x) represents the center of the i-th nucleus; σ i (x) represents the common variance.
[0123] As one alternative implementation, N(y|μ i (x),σ i The expression for (x) is shown below:
[0124]
[0125] Where y represents the target vector; x represents spatiotemporal feature information; π i(x) represents the mixing coefficient; N(y|μ) i (x),σ i (x)) represents the conditional density of the target vector of the i-th kernel; μ i (x) represents the center of the i-th nucleus; σ i (x) represents the common variance; c represents the dimension of the target vector.
[0126] It is worth noting that hybrid density networks are a type of neural network used to model complex probability distributions, primarily handling multi-peak (multiple peaks) or non-Gaussian distributions. Unlike traditional neural networks with a single output node, the output of a hybrid density network is a mixed distribution composed of multiple distributions, making it superior in handling tasks with uncertainty. Cross-sectional power transmission is uncertain data, meaning that multiple different power flow scenarios may exist in the future. Hybrid density networks have the ability to model multimodalities, outputting a probability density function rather than a single point prediction, thus better capturing multiple possibilities and adapting to situations with different probability distributions. Therefore, this embodiment of the invention employs a hybrid density network for probabilistic power flow prediction.
[0127] Hybrid density networks represent the probability density of the target vector y as a linear combination of kernel functions. By parameterizing the output layer, a mixture normal distribution is used to describe the probability distribution. Consider a mixture Gaussian distribution with g components, the probability density function of each component can be expressed as shown in equation (12):
[0128]
[0129] Among them, P ′ (y|x) represents the cross-sectional transmission power probability distribution; y represents the target vector (the target vector refers to the vector composed of active power from different buses); x represents spatiotemporal characteristic information; π i (x) represents the mixing coefficient, and represents the weight of the i-th component (the i-th component is also called the i-th kernel function). It can be regarded as the prior probability of y given x, and π i (x) is a function of x that satisfies N(y|μ i (x),σ i (x)) represents the conditional density of the target vector of the i-th kernel; g represents the total number of kernels in the hybrid density network; μ i (x) represents the center of the i-th nucleus; σ i (x) represents the common variance.
[0130] Furthermore, N(y|μ i (x),σ i(x) is the conditional density of the target vector y of the i-th kernel. There are many choices for the kernel function. In this embodiment of the invention, a normal distribution is chosen, as shown in equation (13):
[0131]
[0132] Where y represents the target vector; x represents spatiotemporal feature information; π i (x) represents the mixing coefficient; N(y|μ) i (x),σ i (x) represents the conditional density of the target vector of the i-th kernel; g represents; μ i (x) represents the center of the i-th nucleus; σ i (x) represents the common variance; c represents the dimension of the target vector.
[0133] Compared with existing technologies, the cross-sectional transmission power probability prediction method provided in this invention involves inputting a power grid topology, using buses as nodes and lines as edges, and converting the topology into a power system graph. The power system graph is then input into a graph neural network to obtain node embedding vectors. Initial values are generated based on the power system graph, and power flow calculations are performed on these initial values to construct a training dataset. The initial values include load demand and generator output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power. The node embedding vectors are input into a spatiotemporal feature extraction model to obtain spatiotemporal feature information. Finally, the spatiotemporal feature information and the training dataset are input into a hybrid density network to obtain the cross-sectional transmission power probability distribution. Compared with existing technologies, this invention can improve the accuracy and efficiency of power probability prediction.
[0134] See Figure 4 This invention also provides a cross-sectional transmission power probability prediction device 10, comprising:
[0135] Graph construction module 11 is used to input the power grid topology and convert the power grid topology into a power system graph with the busbars in the power grid topology as nodes and the lines as edges.
[0136] The graph feature extraction module 12 is used to input the power system graph into a graph neural network to obtain node embedding vectors.
[0137] The training data generation module 13 is used to generate initial values based on the power system diagram, and to perform power system flow calculations on the initial values to construct a training dataset; wherein, the initial values include load demand and unit output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power;
[0138] The spatiotemporal feature information extraction module 14 is used to input the node embedding vector into the spatiotemporal feature extraction model to obtain spatiotemporal feature information;
[0139] The power probability prediction module 15 is used to input the spatiotemporal feature information and the training dataset into the hybrid density network to obtain the cross-sectional transmission power probability distribution.
[0140] The cross-sectional transmission power probability prediction device provided in this embodiment of the invention can realize all the process steps of the cross-sectional transmission power probability prediction method described in the above embodiments. The functions and technical effects of each module and unit in the device are the same as the functions and technical effects of the cross-sectional transmission power probability prediction method described in the above embodiments. The specific implementation method will not be described in detail here.
[0141] See Figure 5 This invention also provides a cross-sectional transmission power probability prediction device 20, including a processor 21, a memory 22, and a computer program stored in the memory 22 and configured to be executed by the processor 21. When the processor 21 executes the computer program, it implements the steps as described in the cross-sectional transmission power probability prediction method embodiments above, for example... Figure 1 The steps S1 to S6 described above; or, when the processor 21 executes the computer program, it implements the functions of each module in the above-described device embodiments.
[0142] The cross-sectional transmission power probability prediction device can be a computing device such as a desktop computer, laptop, handheld computer, or cloud server. The cross-sectional transmission power probability prediction device may include, but is not limited to, a processor and memory. Those skilled in the art will understand that the schematic diagram is merely an example of a cross-sectional transmission power probability prediction device and does not constitute a limitation on the device. It may include more or fewer components than illustrated, or combine certain components, or use different components. For example, the cross-sectional transmission power probability prediction device may also include input / output devices, network access devices, buses, etc.
[0143] The processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor. The processor is the control center of the cross-sectional transmission power probability prediction device, connecting all parts of the device via various interfaces and lines.
[0144] The memory can be used to store the computer program and / or modules. The processor implements various functions of the cross-sectional transmission power probability prediction device by running or executing the computer program and / or modules stored in the memory, and by calling the data stored in the memory. The memory may mainly include a program storage area and a data storage area. The program storage area may store the operating system, at least one application program required for a function, etc.; the data storage area may store data created according to the use of the controller, etc. In addition, the memory may include high-speed random access memory, and may also include non-volatile memory, such as hard disk, memory, plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, at least one disk storage device, flash memory device, or other volatile solid-state storage device.
[0145] If the module integrated into the cross-sectional transmission power probability prediction device is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the above embodiments of the present invention can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, a recording medium, a USB flash drive, a portable hard drive, a magnetic disk, an optical disk, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium, etc.
[0146] Compared with existing technologies, the cross-sectional transmission power probability prediction device, equipment, and storage medium provided in this invention input a power grid topology, using buses in the topology as nodes and lines as edges, to convert the topology into a power system graph; input the power system graph into a graph neural network to obtain node embedding vectors; generate initial values based on the power system graph, and perform power flow calculations on the initial values to construct a training dataset; wherein the initial values include load demand and generator output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power; input the node embedding vectors into a spatiotemporal feature extraction model to obtain spatiotemporal feature information; input the spatiotemporal feature information and the training dataset into a hybrid density network to obtain the cross-sectional transmission power probability distribution. Compared with existing technologies, this invention can improve the accuracy and efficiency of power probability prediction.
[0147] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.
Claims
1. A method for probabilistic prediction of cross-sectional transmission power, characterized in that, include: Input the power grid topology, and convert the power grid topology into a power system diagram with the busbars in the power grid topology as nodes and the lines as edges; The process of inputting the power system graph into a graph neural network to obtain node embedding vectors includes: inputting an initial feature matrix and an adjacency matrix of the power system graph; aggregating the nodes in the power system graph based on the adjacency matrix to obtain hidden layer features; and performing linear regression prediction on the hidden layer features to obtain node embedding vectors. Initial values are generated based on the power system diagram, and power flow calculations are performed on the initial values to construct a training dataset; wherein, the initial values include load demand and unit output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power; The node embedding vector is input into the spatiotemporal feature extraction model to obtain spatiotemporal feature information; The spatiotemporal feature information and the training dataset are input into a hybrid density network to obtain the cross-sectional transmission power probability distribution.
2. The cross-sectional transmission power probabilistic prediction method as described in claim 1, characterized in that, The spatiotemporal feature extraction model includes a convolutional neural network and a long short-term memory network. The spatiotemporal feature information is obtained through the following methods: The node embedding vector is input into the convolutional neural network to obtain intermediate features; The intermediate features and the training dataset are input into the long short-term memory network to obtain spatiotemporal feature information.
3. The cross-sectional transmission power probabilistic prediction method as described in claim 1, characterized in that, The probability distribution of the cross-sectional transmission power is shown in the following equation: in, This represents the probability distribution of cross-sectional transmission power. Represents the target vector; Represents spatiotemporal characteristic information; Indicates the mixing coefficient; Indicates the first Conditional density of the target vector of each kernel; This represents the total number of cores in a hybrid density network. Indicates the first The center of each core; This represents the common variance.
4. The cross-sectional transmission power probabilistic prediction method as described in claim 3, characterized in that, The expression is shown in the following formula: in, Represents the target vector; Represents spatiotemporal characteristic information; Indicates the mixing coefficient; Indicates the first Conditional density of the target vector of each kernel; Indicates the first The center of each core; Indicates the common variance; This represents the dimension of the target vector.
5. The cross-sectional transmission power probabilistic prediction method as described in claim 1, characterized in that, The training dataset was constructed in the following manner: Initialize the load demand power of each side of the power system diagram and the unit output data of each bus node; wherein, the unit output data includes the active power and reactive power injected by the generator into the bus node; Based on the unit output data and the load demand power, a power balance equation is constructed. The power balance equation is solved iteratively to obtain the cross-sectional transmission power under steady state. The unit output data, the load demand power, and the corresponding cross-sectional transmission power are added as a set of data to the training dataset.
6. The cross-sectional transmission power probabilistic prediction method as described in claim 5, characterized in that, In each iteration, the voltage amplitude correction and voltage phase angle correction of each bus node are calculated. Based on the voltage amplitude correction and voltage phase angle correction, the voltage amplitude and voltage phase angle of each bus node are corrected, and then a new iteration begins. When both the voltage amplitude correction and the voltage phase angle correction are less than the preset error upper limit, the cross-sectional transmission power between the bus nodes is calculated based on the current voltage amplitude of each bus node and the current voltage phase angle difference between the bus nodes.
7. A cross-sectional transmission power probability prediction device, characterized in that, include: The graph construction module is used to input the power grid topology and convert the power grid topology into a power system graph, with the busbars in the power grid topology as nodes and the lines as edges. The graph feature extraction module is used to input the power system graph into a graph neural network to obtain node embedding vectors. Specifically, it includes: inputting the initial feature matrix and adjacency matrix of the power system graph; aggregating the nodes in the power system graph based on the adjacency matrix to obtain hidden layer features; and performing linear regression prediction on the hidden layer features to obtain node embedding vectors. The training data generation module is used to generate initial values based on the power system diagram, and to perform power system flow calculations on the initial values to construct a training dataset; wherein, the initial values include load demand and unit output data, and the training dataset includes the initial values and the corresponding cross-sectional transmission power; The spatiotemporal feature information extraction module is used to input the node embedding vector into the spatiotemporal feature extraction model to obtain spatiotemporal feature information; The power probability prediction module is used to input the spatiotemporal feature information and the training dataset into the hybrid density network to obtain the cross-sectional transmission power probability distribution.
8. A cross-sectional transmission power probability prediction device, characterized in that, It includes a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor, wherein the processor executes the computer program to implement the cross-sectional transmission power probabilistic prediction method as described in any one of claims 1 to 6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the cross-sectional transmission power probability prediction method as described in any one of claims 1 to 6.