Power quality whole-network observability research method based on space-time diagram neural network

Abstract

The invention discloses a power quality whole-network observability research method based on a space-time diagram neural network, and belongs to the field of power quality analysis methods. According to the method, space and time characteristics among power quality data of monitoring nodes are extracted by utilizing an attention mechanism-based space-time diagram neural network (ASTGCN), a coupling relationship among the nodes is analyzed, time sequence data of a part of nodes are utilized to estimate time sequence data of a current window of the other one or more nodes in a power grid, and the time sequence data of the current window of the other one or more nodes in the power grid are analyzed. Whole-network observability of electric energy quality is realized, and an implicit adjacency matrix existing between data is extracted through an adaptive network to adapt to changes existing in a network structure. The method is helpful for continuously complementing the time sequence data of the nodes of the whole network under the condition that the number of the monitoring devices of the power system is insufficient, and is of great significance to the whole network observability of power quality, a smart power grid and a digital twin power grid.

Description

Technical Field

[0001] The present invention relates to a method for studying the observability of the entire power quality network based on a spatiotemporal graph neural network, and belongs to the field of power quality analysis methods. Background Art

[0002] The power system is one of the most important infrastructures in modern society. With the improvement of the power industry, new energy sources are gradually changing the grid form, operation characteristics and power quality characteristics of the power system. The source-grid-load power electronics has enhanced the coupling relationship between the disturbance source and the highly sensitive load, increased the correlation between the disturbances, and made the relationship between the grid nodes more complex. In a complex power grid, a disturbance event at a single node may affect the power quality of the remaining multiple nodes. The power quality analysis method based on a single node can no longer effectively analyze and predict the complex power grid. The full network observability of power quality is the key to analyzing the coupling relationship of multiple nodes. To achieve the full network observability of power quality, small real-time monitoring devices can be installed at the nodes of the entire network to obtain data from the entire network. However, this requires the equipment to be small in size, low in power consumption, and able to self-collect energy and self-organize the network, which does not meet the actual requirements. During the installation of the monitoring equipment, some nodes will not be able to install monitoring equipment for a long time, and at this time, the data of the entire network cannot be fully obtained.

[0003] Traditional power system state estimation, such as distribution network state estimation and harmonic state estimation, is mostly based on mechanism models. State estimation based on mechanism models requires sufficient input conditions to obtain output through the model, and the design of the model tends to be idealized. Traditional mechanism-based power grid simulation calculation and estimation have great difficulty in processing the massive, high-dimensional, nonlinear, and complex spatiotemporal data generated by new power systems. The use of data-driven deep learning models can deeply mine the coupling relationship between data. The topology of the power grid is a typical graph structure, so graph neural networks can be used to construct the spatial relationship between nodes in the power grid topology. Graph neural networks are a method that can process large-scale, high-dimensional data and have been applied in the field of power systems.

[0004] This paper proposes a method for studying the observability of the entire power quality network based on spatiotemporal graph neural network with adaptive matrix and attention mechanism. First, the model is trained on the data of the monitoring nodes with monitoring devices installed to obtain the coupling model between some nodes. Then the monitoring device is installed on other nodes and the model is trained using the same steps to obtain the coupling model of each node, thereby obtaining the overall estimation model. The basic model needs to divide and expand the data set according to the time window of each node data of the same time series. The basic model can contain multiple sub-models. Different sub-modules extract features from data of different time windows. Finally, the features are fused and the time series features of the target node of the current time window are estimated. Each sub-model contains multiple spatiotemporal attention layers and graph convolutional network layers. The model makes up for the shortcomings of traditional mechanism-based state estimation, and also solves the limitation that general graph neural networks cannot extract time and space features at the same time, and adds attention mechanism and adaptive matrix to improve the stability and generalization ability of the model. Summary of the invention

[0005] The purpose of the present invention is to provide a data-driven method for studying the observability of the entire power quality network based on a spatiotemporal graph neural network with an adaptive matrix and attention mechanism, in view of the fact that traditional state estimation methods are difficult to extract the complex coupling relationships of nodes in complex power grids and in the case of insufficient monitoring devices.

[0006] The present invention adopts the attention-based spatiotemporal graph neural network (ASTGCN) with powerful extraction capabilities for spatiotemporal features of graph structured data as the main architecture, and combines the adaptive adjacency matrix to extract the coupling relationship between power network node data with complex features. This extraction of spatiotemporal dimension features relying on the spatiotemporal attention mechanism focuses on the important features in the data and better extracts coupling relationships from large amounts of data. Relying on the adaptive matrix can better adapt to changes in the topology of the power grid during the data acquisition process, which is also common in the power grid environment. The present invention also uses multiple sub-models to extract the coupling relationship between different time periods and the target, so that the model has good generalization capabilities.

[0007] A method for studying the observability of the whole network of power quality based on spatiotemporal graph neural network with adaptive matrix and attention mechanism. The method can be divided into three major processes. The first is to train the model until the coupling relationship between the data is obtained, so that accurate power quality state estimation can be performed within the time period to which the data set belongs. Secondly, the coupling relationship of different node data is trained in each subnet with the same training steps. Finally, each model is combined with the installation location of the monitoring point to estimate the state of the whole network node, so as to achieve the observability of the whole network of power quality. The specific steps are as follows:

[0008] Step 1: Perform data collection and processing. Suppose there are N nodes and M monitoring devices in total, and M < N. Divide the M nodes into two subsets M1 and M2. M1 are key nodes, which are equipped with monitoring devices most of the time, and M2 are mobile monitoring nodes. In the training of a single model, use the monitoring devices to collect the monitoring data of M monitoring points at τ time points within T time. After normalizing the data, divide all the data according to time windows to obtain a time series data sample set in units of time windows. The data of the M nodes is converted into a matrix X, where

[0009] X = (X1, X2, …, X τ ) T ∈R M×3×τ , and the corresponding one for M2 is the Y matrix, where Y = (Y1, Y2, …, Y τ ) T ∈R M×3×τ . And set the standardized data corresponding to M2 in X to 0, indicating that these nodes are nodes to be estimated;

[0010] Step 2: Generate an initial spatial adjacency matrix A according to the power grid topology diagram, and store it in the form of a triple storage. Generate a temporal adjacency matrix S according to the relationship between time windows;

[0011] Step 3: Extract spatio-temporal features from the X matrix in Step 1 through multiple spatio-temporal convolutional layers, and perform spatial graph convolution operations on the spatial adjacency matrix A to obtain spatial graph convolution features Z k1 , and perform temporal graph convolution operations on the spatial graph convolution features Z k1 to obtain spatio-temporal graph convolution features Z k2 . Introduce an attention mechanism, and use attention weights to perform a weighted operation on the spatio-temporal graph convolution features Z k2 to obtain Z k3 . Apply an activation function to the spatio-temporal graph convolution features Z k3 and perform normalization processing as the input of the subsequent model;

[0012] Step 4: Re-divide the sample set in Step 1 according to historical time into X t1 , X t2 …X tn , and design the same structure as in Steps 2 and 3;

[0013] Step 5: Perform feature fusion on the outputs of each sub-module generated in Step 4, and use the target matrix Y in Step 1 to guide the learning of the overall model parameters;

[0014] Step 6: Repeat steps 3, 4, and 5 to train the model until Nash equilibrium is reached. The model learns the coupling relationship between nodes M1 and M2, and finally inputs matrix X to the model to output data that is extremely close to matrix Y. Then a set of model training is completed.

[0015] Step 7: Remove the monitoring devices of some mobile monitoring nodes in step 1, and install them on the corresponding number of the remaining R points, where R = NM, and repeat steps 2 to 6 to train the model corresponding to the set. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is the overall flow chart of the observability research model of power quality in the whole network.

[0017] Figure 2 This is a diagram of a spatiotemporal graph neural network model based on the attention mechanism.

[0018] Figure 3 It is a single-layer structure diagram of the model extracting spatiotemporal features. DETAILED DESCRIPTION

[0019] The following is combined with Figure 1 , Attachment Figure 2 , Attachment Figure 3 The embodiments and examples provide a detailed introduction to the method for studying the observability of the entire power quality network based on the spatiotemporal graph neural network, but the implementation of the present invention is not limited to this.

[0020] Example:

[0021] like Figure 1 As shown, the overall process of the power quality whole network observability research model in this embodiment is divided into multiple steps. Figure 1 Each step includes a set of nodes in the power grid. In the first step, all nodes are divided into nodes to be processed, key nodes, and movable monitoring nodes. The movable monitoring nodes include one or more nodes. When extracting the coupling relationship, the nodes can be selected as pre-estimated nodes. After the model training is completed, these nodes become extracted nodes, and the next step of model training is carried out. Finally, the coupling relationship of all non-key nodes is extracted, and the power quality of the whole network is observable.

[0022] like Figure 2 As shown, this embodiment firstly performs a spatial-temporal convolution on the feature matrix X and the historical feature matrix X t1 , X t2 …X tn Features are extracted separately, and their respective features are output through the output layer in each sub-model. Finally, they are fused and the loss of the node to be estimated is calculated before the next round of iterative training. The model optimizer uses the Adam optimizer, and the loss function uses gluon.loss.L2Loss.

[0023] like Figure 3 As shown, this embodiment first calculates the adaptive matrix through deep learning, generates a spatial adjacency matrix A in combination with the topological structure, and inputs it into the spatial graph convolution layer together with the spatial weight matrix. After obtaining the spatial features, they are input into the temporal graph convolution layer together with the temporal adjacency matrix and the temporal weight matrix to obtain the spatiotemporal features. The attention mechanism is introduced to obtain the final features and activated by the activation function. The output is obtained after normalization, which is the input of the next layer.

[0024] The present invention provides a method for studying the observability of the entire power quality network based on a spatiotemporal graph neural network, which belongs to the field of power quality analysis methods. The method uses an attention-based spatiotemporal graph neural network (Attention Based Spatial-Temporal Graph Convolutional Networks, ASTGCN) to extract the spatial and temporal characteristics between the power quality data of the monitoring nodes, analyze the coupling relationship between the nodes, and use the time series data of some nodes to estimate the time series data of the current window of one or more other nodes in the power grid, so as to achieve the observability of the entire power quality network, and extract the implicit adjacency matrix between the data through an adaptive network to adapt to the changes in the network structure. This method helps to continuously complete the time series data of the entire network nodes when the number of monitoring devices in the power system is insufficient, which is of great significance to the observability of the entire power quality network, smart grids and digital twin grids.

Claims

1. A method for studying the observability of the entire power quality network based on spatiotemporal graph neural network, characterized in that: The spatiotemporal graph neural network in deep learning is used to extract the time and space features of the time series data of the monitoring nodes in a certain time window in combination with the adjacency matrix of the monitoring nodes, and the characteristic relationship of the nodes in the power grid is obtained, so that the time series data of the remaining one or more nodes can be predicted through the data of some of the nodes in the future time window. The specific steps of this method are as follows: Step 1: Conduct data collection and processing. Assume there are N nodes and M monitoring devices in total, and M < N. Divide the M nodes into two subsets, M1 and M2. M1 are key nodes, which are equipped with monitoring devices most of the time, and M2 are mobile monitoring nodes. In the training of a single model, use the monitoring devices to collect monitoring data of M monitoring points at τ time points. After normalizing the data, divide all the data according to time windows to obtain a time series data sample set with time windows as units. The data of the M nodes is converted into a matrix X, where X = (X1, X2, …, X τ ) T ∈R M×3×τ , and the corresponding one for M2 is the Y matrix, where Y = (Y1, Y2, …, Y τ ) T ∈R M×3×τ , and set the standardized data corresponding to M2 in X to 0, indicating that these nodes are nodes to be estimated; Step 2: Generate an initial spatial adjacency matrix A according to the power grid topology diagram, which is stored in the form of triples, and generate a temporal adjacency matrix S according to the relationship between time windows; Step 3: Pass the X matrix in step 1 through multiple spatiotemporal convolutional layers to extract spatiotemporal features and calculate a new adaptive adjacency matrix; Step 4: Re-divide the sample set in step 1 into X according to historical time t1 , X t2 …X tn , and design the same structure as in steps 2 and 3; Step 5: Perform feature fusion on the outputs of each submodule generated in step 4, and use the target matrix Y in step 1 to guide the learning of the overall model parameters; Step 6: Repeat steps 3, 4, and 5 to train the model until Nash equilibrium is reached. The model learns the coupling relationship between nodes M1 and M2, and finally inputs matrix X to the model to output data that is extremely close to matrix Y. Then a set of model training is completed. Step 7: Remove the monitoring devices of some mobile monitoring nodes in step 1, and install them on the corresponding number of the remaining R points, where R = NM, and repeat steps 2 to 6 to train the model corresponding to the set.

2. According to claim 1, a method for studying the observability of the entire power quality network based on a spatiotemporal graph neural network is characterized by: The judgment conditions of M1 and M2 in step 1 can be adjusted manually, which has no effect on the training method of the model.

3. According to claim 1, a method for studying the observability of the entire power quality network based on spatiotemporal graph neural network is characterized in that: The specific steps of spatiotemporal feature extraction in step 3 are: Step 1: The spatial adjacency matrix A is subjected to a spatial graph convolution operation to obtain the spatial graph convolution feature. The spatial graph convolution can be expressed as Z k1 =σ(AXW s ), where σ represents the nonlinear activation function, W s is the weight matrix of the spatial graph convolution; Step 2: Convolutional feature Z k1 Through the time graph convolution operation, the spatiotemporal graph convolution feature Z is obtained k2 , the spatial graph convolution can be expressed as Z k2 =σ(TZ k1 W t ), where T is the time adjacency matrix, W t is the weight matrix of the temporal graph convolution; Step 3: Introduce the attention mechanism and use the attention weight to convolve the spatiotemporal graph feature Z k2 Perform weighted operation, which can be expressed as Z k3 =α·Z k2 +β·Z k2 ; Step 4: Convolutional feature Z on spatiotemporal graph k3 Apply activation function and normalize it as input for subsequent models; Step 5: Based on the data of each node, deep learning is used to calculate the adaptive spatial adjacency matrix from a macro and micro perspective.

4. According to claim 1, a method for studying the observability of the entire power quality network based on a spatiotemporal graph neural network is characterized by: The number of submodules in step 4 can be freely controlled. The historical time data of each submodule is determined according to the actual situation, usually not exceeding 6 hours. When using historical event data, the data set length τ in step 1 k Less than τ.

5. According to claim 1, a method for studying the observability of the entire power quality network based on a spatiotemporal graph neural network is characterized by: In step 5, using the output of each sub-model for feature fusion can better extract the temporal relationship between the target node and the input node. In step 6, the fused features and matrix Y are used to calculate the loss to achieve Nash equilibrium.

6. According to claim 1, a method for studying the observability of the entire power quality network based on a spatiotemporal graph neural network is characterized by: Step 7 uses a mobile monitoring device to train the coupling model between different nodes, so as to achieve the power quality of the entire power grid.

Images