CN-122020001-A - Power missing data space-time correlation completion method based on deep neural network

CN122020001ACN 122020001 ACN122020001 ACN 122020001ACN-122020001-A

Abstract

The invention discloses a space-time correlation completion method of electric power missing data based on a deep neural network, which comprises the steps of constructing graph structured data according to the topology of an electric power system, carrying out preliminary filling on the missing data by adopting a space-time local weighted average method, inputting the processed data into the deep neural network formed by stacking a plurality of space-time processing modules, embedding graph attention networks in each module, connecting gate-control circulating units in series, outputting a predicted data matrix containing completion values, and carrying out end-to-end training on the neural network by constructing a composite loss function consisting of data fidelity loss and physical residual loss. According to the invention, through the optimization paradigm of data and physical dual-drive, the problem that the completion result may violate the physical law in the prior art is effectively solved, meanwhile, the modeling accuracy of the dynamic characteristics of the power grid is improved by using a graph attention mechanism, and the accuracy and physical consistency of the power deficiency data completion are improved.

Inventors

WU QUAN
TANG XIAOLAN
LI LETIAN
Xiang Jiaoying

Assignees

西安卓俊建设工程有限公司

Dates

Publication Date: 20260512
Application Date: 20260205

Claims (10)

1. The power missing data space-time correlation completion method based on the deep neural network is characterized by comprising the following steps of: constructing graph structured data according to a topological structure of the power system, wherein the graph structured data comprises node characteristics extracted from time sequence data of the power system; inputting the graph structured data into a deep neural network, wherein the deep neural network is configured to fuse spatial information in the graph structure and time information in the time sequence data and output a predicted data matrix containing missing data complement values; and constructing and jointly optimizing a data fidelity loss and a physical residual loss, and training parameters of the deep neural network, wherein the physical residual loss is used for quantifying the deviation degree of the predicted data matrix to a preset power grid physical law.
2. The deep neural network-based power loss data spatiotemporal correlation completion method of claim 1, wherein the building map structured data comprises: and performing preliminary filling on missing positions in the original time sequence data to form a complete input matrix, wherein the preliminary filling adopts a space-time local weighted average method.
3. The deep neural network-based power loss data spatio-temporal correlation completion method of claim 2, wherein the spatio-temporal local weighted averaging method includes: weighted average calculations are performed based on values of adjacent known data points of the missing points in the time dimension and the space dimension.
4. The deep neural network-based power deficiency data space-time correlation completion method of claim 1, wherein the deep neural network is formed by stacking a plurality of space-time processing modules, each space-time processing module comprises a graph neural network layer for spatial dimension information fusion and a cyclic neural network layer for temporal dimension information fusion, and residual connection is adopted inside the modules and among the modules.
5. The deep neural network-based power loss data spatiotemporal correlation completion method of claim 1, wherein fusing spatial information in the graph structure with temporal information in the time series data comprises: Dynamically calculating the spatial correlation weight among the nodes through a graph attention mechanism, and updating node representation by aggregating neighbor node information; and inputting the updated node representation into a gating circulation unit, and capturing the evolution rule of the node state on a time sequence.
6. The deep neural network-based power loss data spatiotemporal correlation completion method of claim 5, wherein the graph attention mechanism is a multi-head graph attention mechanism which learns dynamic spatial correlation among nodes from different representation subspaces by setting a plurality of independent attention heads in parallel and integrating outputs of the heads.
7. The deep neural network-based power loss data spatiotemporal correlation completion method of claim 1, wherein constructing the physical residual loss comprises: Extracting node voltage and phase angle information from a prediction data matrix output by the deep neural network; combining a preset power grid node admittance matrix, and calculating theoretical power injection values of all nodes according to a power flow equation of the power system; and comparing the theoretical power injection value with the node power injection value in the predicted data matrix to obtain the physical residual loss.
8. The deep neural network-based power loss data spatiotemporal correlation completion method of claim 7, further comprising performing differential constraint according to a preset type of bus in a power system when constructing the physical residual loss: Simultaneously restraining physical residual errors of active power and reactive power of a PQ bus; the physical residual error of the active power of the PV bus is only restrained to the PV bus; no power injection constraints are imposed on the balanced bus.
9. The deep neural network-based power loss data spatiotemporal correlation completion method of claim 1, wherein the weight of the physical residual loss is a variable dynamically adjusted with the progress of deep neural network training in the joint optimization process, the variable gradually increasing with the progress of deep neural network training.
10. The deep neural network-based power loss data spatiotemporal correlation completion method of claim 1, wherein the data fidelity loss is obtained by calculating an average absolute error between the predicted data matrix and known observations in the power system time series data.

Description

Power missing data space-time correlation completion method based on deep neural network Technical Field The invention relates to the technical field of power data processing, in particular to a power missing data space-time correlation completion method based on a deep neural network. Background With the rapid development of smart grids, monitoring devices represented by a Wide Area Measurement System (WAMS) and an advanced measurement system (AMI) are widely deployed in a power system, and massive and high-dimensional power time series data are generated. These data are the basis for implementing accurate state estimation, dynamic safety assessment and optimal control decisions of the power grid, and the integrity of the data is of vital importance. In order to cope with the common data missing problem caused by communication faults, sensor failures and the like, a data complement technology is developed. Early completion methods relied primarily on statistical models, such as mean filling, regression interpolation, etc., but these methods have difficulty capturing the complex nonlinear relationships inherent in the power data. In recent years, artificial intelligence technology represented by deep neural networks has been introduced in this field, in which cyclic neural networks (RNNs) are capable of efficiently processing time dependence of data, while Graph Neural Networks (GNNs) are good at exploiting the topology of the power grid to exploit spatial correlation between nodes. And by combining the space-time diagram neural network models of the two, the comprehensive modeling of the space-time coupling characteristic of the electric power data is realized, and remarkable progress is made in the completion precision. However, there are still significant limitations to using the neural network model architecture described above. Firstly, the existing space-time diagram neural network model mainly regards complementation as a pure statistical fitting task, and the essence of the space-time diagram neural network model is a 'black box' model, and the physical rule of the power grid behind the data is unknown. This may lead to that the complement result may perform well in the statistical index, but may not be established in the physical level, for example, the complement voltage and power value may severely violate the basic physical constraints such as the tide equation, and generate pseudo data which is not self consistent in physics, thus bringing great hidden trouble to the subsequent power grid safety analysis and decision. Secondly, a static adjacency matrix based on physical topology is generally adopted to define the spatial relationship among nodes in the method based on the graph neural network model, and the objective fact that the electric coupling strength of a power grid can dynamically change along with the operation working condition is ignored. Because the traditional fixed graph structure cannot adaptively capture the correlation of the dynamic change, the spatial modeling accuracy of the model under the complex working condition is limited. Therefore, how to integrate the prior knowledge of the power grid physics into the deep learning model and realize the accurate capture of the dynamic space-time correlation so as to ensure the physical consistency of the completion result is a difficult problem to be solved in the prior art. Disclosure of Invention This section is intended to outline some aspects of embodiments of the application and to briefly introduce some preferred embodiments. Some simplifications or omissions may be made in this section as well as in the description of the application and in the title of the application, which may not be used to limit the scope of the application. The present invention has been made in view of the above-described problems occurring in the prior art. Therefore, the invention provides a power missing data space-time correlation completion method based on a deep neural network, which is used for solving the problems in the background technology. In order to solve the technical problems, the invention provides the following technical scheme that the power missing data space-time correlation completion method based on the deep neural network comprises the following steps: constructing graph structured data according to a topological structure of the power system, wherein the graph structured data comprises node characteristics extracted from time sequence data of the power system; inputting the graph structured data into a deep neural network, wherein the deep neural network is configured to fuse spatial information in the graph structure and time information in the time sequence data and output a predicted data matrix containing missing data complement values; and constructing and jointly optimizing a data fidelity loss and a physical residual loss, and training parameters of the deep neural network, wherein the physical residual loss is used for