An intelligent filling method for power flow data based on spatio-temporal coupling

Abstract

The present invention belongs to the technical field of power system applications, and particularly relates to an intelligent filling method for power flow data based on spatio-temporal coupling. The present invention proposes an intelligent filling method for power flow data based on spatio-temporal coupling. First, the original data is preprocessed by Hampel filtering, phase alignment, and normalization. A spatio-temporal encoder is constructed through the joint encoding of Adaptive Graph Convolution (Adaptive GCN) and Attention LSTM to capture spatio-temporal coupling features. A multi-physical constraint generative adversarial network (Phy-GAN) is designed through the physical constraints of the generator and the multi-scale discriminator to achieve data filling, ensuring that the filling results have both numerical rationality and physical feasibility.

Description

Technical Field

[0001] The present invention belongs to the technical field of power system applications, and particularly relates to an intelligent filling method for power flow data based on spatio-temporal coupling. Background Art

[0002] With the accelerating construction of the new power system, the power grid is undergoing a profound transformation from a traditional centralized architecture to a collaborative interaction mode of "source-grid-load-storage". The high proportion of renewable energy access, large-scale deployment of distributed power sources, and wide penetration of power electronic devices have made the operating conditions of the power grid show characteristics such as strong randomness, multi-time scale coupling, and complex spatio-temporal correlation. Under this background, power flow data, as the core carrier reflecting the real-time operating state of the power grid, its integrity directly affects the reliability of key functions such as state estimation, security assessment, and optimal dispatching. However, in the actual system, the problem of data loss caused by factors such as measurement device failures, communication link interruptions, and cyber attacks is becoming increasingly prominent.

[0003] Traditional filling methods based on statistical principles (such as mean filling and time series interpolation) are ineffective in dealing with complex non-linear spatio-temporal correlations. Although pure data-driven deep learning models can capture complex patterns, they often violate basic constraints such as circuit principles due to ignoring the physical laws of the power grid, which may lead to chain errors in subsequent analysis. Summary of the Invention

[0004] In view of the technical problems existing in the filling of power flow data, the present invention proposes an intelligent filling method for power flow data based on spatio-temporal coupling, which is reasonable in design, simple in method, strong in theory, and can realize the intelligence and rationality of power flow data filling.

[0005] In order to achieve the above object, the technical solution adopted by the present invention is: an intelligent filling method for power flow data based on spatio-temporal coupling, including the following steps:

[0006] S1. Preprocess the original data through Hampel filtering, phase alignment, and normalization;

[0007] S2. Construct a spatio-temporal encoder through the joint encoding of Adaptive Graph Convolution (Adaptive GCN) and Attention LSTM to capture spatio-temporal coupling features Z;

[0008] S3. Design a multi-physical constraint generative adversarial network (Phy-GAN) through generator physical constraints and multi-scale discriminators;

[0009] S4. Expand the feature dimensions of the missing mask M and the covariates C to have the same spatial dimension as the spatio-temporal coupling feature Z, and then concatenate the expanded M, C, and Z as the input to the generator of the generative adversarial network Phy-GAN. Data filling is achieved through the discriminator and adversarial training.

[0010] Preferably, in step S1, the original data is phase-aligned, and its formula is: , where represents the unwrapped continuous phase, represents the current phase, represents the phase angle at the previous moment, represents the rounding function.

[0011] Preferably, in step S2, the attention LSTM introduces a time attention mechanism in the LSTM. The spatio-temporal encoder performs gated feature fusion on the spatial feature matrix obtained by the adaptive graph convolution (Adaptive GCN) and the time feature matrix obtained by the attention LSTM, sets the gated weights, and obtains the spatio-temporal coupling feature:

[0012] , where is the spatio-temporal coupling feature, is the gated weight, is the trainable fusion weight matrix, is the spatial feature matrix, is the time feature matrix, represents element-wise multiplication of matrices, represents the Sigmoid activation function;

[0013] Preferably, in step S3, the multi-physics constrained generative adversarial network adds a physics constraint layer after the output layer of the generator, including dynamic constraints and algebraic constraints.

[0014] Preferably, in step S3, the multi-scale discriminator of the multi-physics constrained generative adversarial network is: , where is the data authenticity score, X represents the input data, s represents the downsampling scale, , represents the discriminator sub-network at the corresponding scale, represents average pooling.

[0015] Preferably, in step S4, the missing mask M is a binary matrix indicating the data missing positions, and the covariates C are vectors obtained by encoding the external variable breaker status and the weather.

[0016] Preferably, the dynamic constraint and the algebraic constraint are respectively: , , where represents the node voltage vector, represents the branch current vector, represents the nodal admittance matrix, represents the conjugate complex number of the branch current vector, represents taking the real part of a complex number, represents the actually measured active power value.

[0017] Compared with the prior art, the advantages and positive effects of the present invention are as follows: The present invention proposes an intelligent filling method for power flow data based on spatio-temporal coupling. First, the original data is preprocessed through Hampel filtering, phase alignment, and normalization. A spatio-temporal encoder is constructed through the joint encoding of Adaptive Graph Convolution (Adaptive GCN) and Attention LSTM to capture spatio-temporal coupling features. A multi-physical constraint generative adversarial network (Phy-GAN) is designed through generator physical constraints and multi-scale discriminators to achieve data filling, ensuring that the filling results have both numerical rationality and physical feasibility. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0019] Figure 1 is a schematic flow chart of an intelligent filling method for power flow data based on spatio-temporal coupling provided by an embodiment of the present invention;

[0020] Figure 2 is a schematic flow chart of data preprocessing provided by an embodiment of the present invention;

[0021] Figure 3 is a schematic structural diagram of a multi-physical constraint generative adversarial network provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0022] In order to be able to more clearly understand the above objects, features, and advantages of the present invention, the following further describes the present invention with reference to the drawings and embodiments. It should be noted that, without conflict, the embodiments of the present application and the features in the embodiments can be combined with each other.

[0023] In the following description, many specific details are set forth in order to fully understand the present invention. However, the present invention can also be implemented in other ways different from those described herein. Therefore, the present invention is not limited by the specific embodiments disclosed in the following specification.

[0024] Examples, such as Figure 1 and 2 As shown in 3, considering that traditional filling methods based on statistical principles (such as mean filling, time series interpolation) are ineffective in dealing with complex non-linear spatio-temporal correlations, while pure data-driven deep learning models can capture complex patterns but often violate basic constraints such as circuit principles due to ignoring the physical laws of the power grid, which may lead to chain errors in subsequent analysis. Therefore, an intelligent filling method for power flow data based on spatio-temporal coupling is proposed in the present invention. Considering that there are time stamp offsets, sensor noises and dimension differences in the original power data, directly inputting them into the model will cause unstable training. So in this step, through filtering, alignment and normalization operations, data heterogeneity is eliminated and the model convergence efficiency is improved.

[0025] Considering the statistical characteristics of the Hampel filter based on the median and absolute deviation, it can effectively identify and replace impulse noises (such as outliers caused by communication interference), and is more resistant to outlier interference than the mean filter. The sliding window mechanism can adapt to the noise levels in different time periods, avoid misjudgment of the fixed threshold during load fluctuations, and only replace the points that significantly deviate from the statistical characteristics of the window, while retaining real power events (such as sudden changes caused by faults). Traditional low-pass filters will smooth real fluctuations, while the Hampel filter retains key dynamic information while denoising.

[0026] Phase alignment eliminates phase jumps, that is, the jumps of the voltage / current phasor angles caused by the modulo 2π periodicity, ensures the continuity and differentiability of the time series, and avoids misjudgment by the model as a mutation. The continuous phase change is more in line with the modeling assumptions of time series networks such as LSTM and reduces gradient anomalies. The aligned phase can be directly used for complex number operations (such as power calculation) without additional processing.

[0027] Normalization eliminates dimension differences, unifies parameters with different dimensions (such as kV, MW) into per-unit values, avoids the model being biased towards large-scale features due to numerical range differences (such as power suppressing voltage), enhances the generalization ability, and the same model can be adapted to power grids with different voltage levels (such as 10 kV distribution network and 500 kV main network) without retraining. Numerical stability is especially significant in the high-dimensional parameter space.

[0028] Therefore, first, data preprocessing is performed on the original data through Hampel filtering, phase alignment and normalization. Specifically, the formula for phase alignment of the original data is: , where represents the expanded continuous phase, represents the current phase, represents the phase angle at the previous moment, represents the rounding function.

[0029] Considering that the power of grid nodes is affected by both topological connections (space) and load fluctuations (time). Traditional methods independently process the spatio-temporal dimensions, resulting in information fragmentation. This module captures spatio-temporal coupling features through the joint encoding of adaptive graph convolution and attention LSTM. Therefore, through the joint encoding of Adaptive GCN and attention LSTM, a spatio-temporal encoder is constructed to capture the spatio-temporal coupling feature Z. Specifically, the attention LSTM introduces a temporal attention mechanism into the LSTM. The spatio-temporal encoder performs gated feature fusion on the spatial feature matrix obtained from Adaptive GCN and the temporal feature matrix obtained from the attention LSTM, sets the gated weights, and obtains the spatio-temporal coupling feature: , where, is the spatio-temporal coupling feature, is the gated weight, is the trainable fusion weight matrix, is the spatial feature matrix, is the temporal feature matrix, denotes element-wise multiplication of matrices, denotes the Sigmoid activation function. Since this fusion is not an ordinary addition, but a selective fusion using a matrix with a gated mechanism, more useful information can be retained, noise information can be removed, and information loss can be reduced. The joint encoding of adaptive graph convolution and attention LSTM generates an asymmetric adjacency relationship through a learnable embedding matrix, captures the actual electrical coupling (such as the strong correlation of high-load lines), and transcends the limitations of fixed topologies. It balances computational efficiency and receptive field, covering multi-hop neighbors (such as capturing the indirect influence of adjacent substations). The attention LSTM introduces a temporal attention mechanism into the LSTM, automatically focuses on key time periods (such as peak load), and alleviates the problem of insufficient modeling of long-distance dependencies in traditional LSTM.

[0030] Considering that generative adversarial networks can synthesize data distributions, but power data needs to strictly satisfy physical laws. This module ensures that the filling results have both numerical rationality and physical feasibility through differential-algebraic equation constraints and multi-scale discriminator design. Therefore, through the generator physical constraint and multi-scale discriminator design of the multi-physical constraint generative adversarial network Phy-GAN, a physical constraint layer is added after the output layer of the generator in the multi-physical constraint generative adversarial network, including dynamic constraints and algebraic constraints, which are respectively: 、 , where, represents the node voltage vector, represents the branch current vector, represents the node admittance matrix, represents the conjugate complex number of the branch current vector, represents taking the real part of a complex number, Represents the actually measured active power value. The multi-scale discriminator of the multi-physical constraint generative adversarial network is as follows: , where is the data authenticity score, X represents the input data, s represents the downsampling scale, , represents the discriminant sub-network corresponding to the scale, represents average pooling. The multi-scale joint training forces the generator to cover the full-dimensional feature distribution and avoid generating single-mode data.

[0031] Finally, the missing mask M and the covariates C are expanded in feature dimension to have the same spatial dimension as the spatio-temporal coupling feature Z, and then the expanded M and C and Z are concatenated as the input of the generator of the generative adversarial network Phy-GAN. Data filling is achieved through the discriminator and adversarial training. The missing mask M is a binary matrix indicating the missing positions of the data, and the covariates C are vectors after encoding the external variable breaker status and weather. The mask M (missing positions) and the covariates C (such as breaker status) are extended along the spatial dimension (number of nodes N) to the same dimension N×T as the spatio-temporal feature Z to ensure independent encoding of the missing status and external factors for each node. This application forms a closed loop from data cleaning to physical constraints, taking into account both data quality and physical laws.

[0032] The above is only a preferred embodiment of the present invention, and it is not a limitation of the present invention in other forms. Any person skilled in the art may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as it does not depart from the technical solution content of the present invention, any simple modification, equivalent change, and modification made to the above embodiments according to the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.

Claims

1. A method for intelligent filling of power flow data based on time-space coupling, characterized in that: The steps include: S1. Preprocess the raw data by Hampel filtering, phase alignment and normalization; S2, through the joint encoding of adaptive graph convolution Adaptive GCN and attention LSTM, a spatiotemporal encoder is constructed to capture the spatiotemporal coupling feature Z; S3. Design a multi-physics constraint generative adversarial network Phy-GAN through the generator physical constraint and the multi-scale discriminator. The multi-physics constraint generative adversarial network adds a physical constraint layer after the generator output layer, including dynamic constraints and algebraic constraints. The dynamic constraints and algebraic constraints are: , where V represents the node voltage vector, represents the branch current vector, represents the node admittance matrix, represents the conjugate complex number of the branch current vector, represents taking the real part of a complex number, Indicates the actual measured active power value; S4. Expand the feature dimensions of the missing mask M and the covariate C so that they have the same spatial dimension as the spatiotemporal coupling feature Z. Then concatenate the expanded M, C and Z as the generator input of the generative adversarial network Phy-GAN, and achieve data filling through the discriminator and adversarial training.

2. According to claim 1, a method for intelligently filling in power flow data based on time-space coupling is characterized in that: The S1 step performs phase alignment on the original data, and the formula is: ,in, represents the continuous phase after expansion, Indicates the current phase. represents the phase angle at the previous moment, Represents the rounding function.

3. According to claim 1, a method for intelligently filling in power flow data based on time-space coupling is characterized in that: In the step S2, the attention LSTM introduces the temporal attention mechanism into the LSTM. The spatiotemporal encoder performs gated feature fusion on the spatial feature matrix obtained by the adaptive graph convolution Adaptive GCN and the temporal feature matrix obtained by the attention LSTM, sets the gating weights, and obtains the spatiotemporal coupling features: , where Z is the spatiotemporal coupling feature, g is the gating weight, is the trainable fusion weight matrix, is the spatial feature matrix, is the time feature matrix, Indicates the multiplication of corresponding positions of matrices, Represents the Sigmoid activation function.

4. The method for intelligently filling in power flow data based on time-space coupling according to claim 1 is characterized in that: In the step S3, the multi-scale discriminator of the multi-physics constraint generative adversarial network is: ,in, Score the authenticity of the data. X represents the input data, s represents the downsampling scale, , represents the discriminant subnetwork of the corresponding scale, represents average pooling.

5. The method for intelligently filling in power flow data based on time-space coupling according to claim 1 is characterized in that: The missing mask M in step S4 is a binary matrix that identifies the location of missing data, and the covariate C is a vector of external variables, the circuit breaker status, and the weather encoding.

Images