Search

CN-121980115-A - Wind power plant data synchronization correction method based on view topology and mask graph neural network

CN121980115ACN 121980115 ACN121980115 ACN 121980115ACN-121980115-A

Abstract

The invention discloses a wind power plant data synchronous correction method based on view topology and mask graph neural network, which belongs to the field of wind power plant data processing and artificial intelligence space-time sequence prediction, and comprises the steps of firstly obtaining historical space-time observation data of unit nodes, constructing space-time feature tensors, generating mask matrixes, extracting space-time waveform association features of a plurality of unit nodes, constructing a global multi-view adjacency matrix, inputting the space-time feature tensors and the global multi-view adjacency matrix into the space-time graph neural network, performing blocking and shielding operation in a message transmission stage of the network, extracting full-field features, outputting an initial deduction sequence, obtaining boundary physical residual errors, reversely compensating the boundary physical residual errors into the initial deduction sequence, and generating and outputting final correction data. The invention realizes high-fidelity error correction of test data under high concurrency loss rate and strong noise interference, and remarkably reduces the space-time diagram calculation memory overhead of the bottom layer deep learning framework by designing a static topology caching mechanism.

Inventors

  • YUAN JIANHAO
  • DU JIE

Assignees

  • 南京信息工程大学

Dates

Publication Date
20260505
Application Date
20260403

Claims (10)

  1. 1. A wind power plant data synchronization correction method based on view topology and mask graph neural network is characterized by comprising the following steps: Step 1, acquiring historical space-time observation data of a plurality of unit nodes in a wind power plant, and constructing a space-time characteristic tensor according to the historical space-time observation data; Step 2, scanning and evaluating the effectiveness of the historical space-time observation data, and dynamically generating a mask matrix matched with the space-time characteristic tensor dimension; step 3, extracting the space and time sequence waveform association characteristics of a plurality of unit nodes, and carrying out weighted fusion and normalization to construct a global multi-view adjacency matrix; Step 4, inputting the space-time feature tensor and the global multi-view adjacency matrix into a space-time graph neural network, and in the message transmission stage of the network, performing blocking and shielding operation on hidden layer features of nodes by using a mask matrix, synchronously and alternately extracting full-field features and outputting an initial deduction sequence; And 5, acquiring a boundary physical residual error between the initial deduction sequence and the restored real observation data, reversely compensating the boundary physical residual error into the initial deduction sequence based on a preset smooth distribution mechanism, and generating and outputting final wind power plant set correction data.
  2. 2. The method for synchronously correcting wind power plant data based on view topology and mask pattern neural network according to claim 1, wherein step 1 comprises the steps of obtaining wind speed test data of all wind power units in a wind power plant in a preset history time window, and splicing the wind speed test data into an initial space-time characteristic tensor in space and time dimensions Performing Z-score normalization preprocessing on the initial feature tensor to generate standard space-time feature tensor for eliminating magnitude difference of sensor dimensions of different units and accelerating graph network convergence The formula for the Z-score normalization is as follows: ; Wherein, the And Respectively represent the first And the average value and standard deviation of the observed wind speed of the platform unit in the historical time window.
  3. 3. The method for synchronously correcting wind farm data based on view topology and mask graph neural network according to claim 1, wherein step 2 comprises dynamically generating binary mask matrix completely matched with space-time characteristic tensor dimension for nodes with data missing or strong noise abnormal extremum by real-time scanning of test data quality of whole-farm machine set Wherein, the method comprises the steps of, The representation represents the total number of wind turbines contained within the wind farm, And representing a historical time window, wherein the corresponding position of the effective observation data is marked as 1, and the corresponding position of the missing or abnormal data is marked as 0.
  4. 4. The method for synchronously correcting the wind farm data based on the view topology and mask pattern neural network according to claim 1, wherein the step 3 specifically comprises the following steps: Step 3.1, extracting the spatial longitude and latitude coordinates of the full-field unit, and calculating the spherical great circle physical distance of the two units on the surface of the earth by using a semi-normal vector formula ; Distance of sphere from sphere Mapping to a geographic adjacency matrix via a Gaussian kernel function The calculation formula is as follows: ; wherein p is a bandwidth parameter controlling the spatial correlation decay rate; step 3.2, extracting a historical stable wind speed time sequence, calculating morphological similarity by using a Dynamic Time Warping (DTW) algorithm to generate a waveform similarity matrix, and calculating pearson correlation coefficients among the historical wind speed sequences to generate a linear correlation matrix The calculation formula is as follows: ; Wherein, the And (3) with Respectively represent units And Historical time window The average value of the wind speed in the wind power generation system, Indicating machine set The observed value of the wind speed at the time t, Indicating machine set A wind speed observation at time t; Step 3.3, weighting and fusing according to preset weights, meeting the constraint condition that the sum of the weights is 1, normalizing the longitude matrix to generate a global multi-view adjacency matrix, extracting a historical stable wind speed time sequence, and setting a set Is the historical wind speed sequence of (1) Unit set Is of the historical sequence of Searching the best matching path between two sequences by using dynamic time warping DTW algorithm ; Step 3.4, constructing an accumulated distance matrix Wherein S represents the total number of time steps of the wind speed time sequence, taking For the minimum accumulated alignment distance of two wind speed waveform sequences, generating waveform similarity matrix through exponential decay mapping The calculation formula is as follows: ; Wherein, the Is a smoothing coefficient.
  5. 5. The method for synchronously correcting the wind farm data based on the view topology and mask pattern neural network according to claim 1, wherein the step 4 specifically comprises the following steps: Step 4.1, inputting space-time characteristic tensors into a space-time diagram neural network comprising residual connection, breaking through the limitation of space-time dimension alternating convolution, taking a global multi-view adjacency matrix as a physical and time sequence coupling path between full-field nodes, and carrying out global tensor operation and synchronous cross extraction on the space wake correlation between units and the dynamic evolution of wind condition time sequence in a network hidden layer; step 4.2, in the message transmission stage of the space-time diagram neural network, the binary mask matrix is formed Performing element-by-element multiplication shielding operation with the node hidden layer feature tensor, blocking missing nodes or abnormal nodes from transmitting error features to the full-field associated nodes so as to avoid error avalanche, wherein the calculation formula is as follows: ; Wherein, the Is the first The nodes of the layer hide layer feature tensors; represent the first An output feature tensor for the layer; representing the normalized global multi-view adjacency matrix; Representing a binary mask matrix; element-wise multiplicative Hadamard Product masking operations representing tensors; Is the first A layer's learnable weight parameter matrix; Representing a nonlinear activation function; and meanwhile, carrying out dynamic characteristic compensation on the missing node with the mask mark of 0 by utilizing the characteristic information converged by the healthy neighbor nodes based on the global topological network, and synchronously outputting an initial corrected wind speed sequence of the full-field unit.
  6. 6. The method for synchronously correcting the wind farm data based on the view topology and mask pattern neural network according to claim 1, wherein the step 5 specifically comprises the following steps: Step 5.1, in the model training stage, a mask loss function is constructed, a loss gradient of the missing node is filtered by utilizing a binary mask matrix M, only the effective node with the mask mark of 1 is extracted, and the mean square reconstruction loss of the deduction value and the real observation value is calculated; Step 5.2, calculating the boundary physical residual error between the initial deduction wind speed value and the real observation wind speed value output by the abutting point model when the deduction missing time window is ended and the right boundary abutting point of the real observation data is restored Constructing a dynamic weight distribution vector which monotonically increases along with the time step length, and carrying out boundary physical residual error And after the dynamic weight distribution vector is multiplied, the dynamic weight distribution vector is reversely overlapped on the initial deduction wind speed sequence in the missing time window, the step mutation cliff generated at the butt joint boundary due to independent deduction is smoothed, and final high-fidelity reconstruction data meeting the continuity of the first derivative is output.
  7. 7. A wind farm data synchronization correction system based on a view topology and a mask graph neural network for implementing the method of claim 1, comprising: the space-time feature and mask generation module is used for acquiring historical space-time observation data of a plurality of unit nodes in the wind power plant, constructing a space-time feature tensor according to the historical space-time observation data, scanning and evaluating the validity of the historical space-time observation data, and dynamically generating a mask matrix matched with the space-time feature tensor dimension; the multi-view topology construction module is used for extracting the space and time sequence waveform association characteristics of a plurality of unit nodes, and carrying out weighted fusion and normalization to construct a global multi-view adjacency matrix; The graph neural network mask deduction module is used for inputting the space-time characteristic tensor and the global multi-view adjacency matrix into the space-time graph neural network, and in the message transmission stage of the network, the mask matrix is used for carrying out blocking and shielding operation on hidden layer characteristics of the nodes, and the full-field characteristics are synchronously and alternately extracted and an initial deduction sequence is output; And the boundary residual error compensation module is used for acquiring a boundary physical residual error between the initial deduction sequence and the restored real observation data, reversely compensating the boundary physical residual error into the initial deduction sequence based on a preset smooth distribution mechanism, and generating and outputting final wind power plant unit correction data.
  8. 8. A computer device comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to carry out the steps of the method of claim 1.
  9. 9. A computer readable storage medium having stored thereon a computer program/instruction which when executed by a processor performs the steps of the method of claim 1.
  10. 10. A computer program product comprising computer programs/instructions which, when executed by a processor, implement the steps of the method of claim 1.

Description

Wind power plant data synchronization correction method based on view topology and mask graph neural network Technical Field The invention belongs to the field of wind power plant data processing and artificial intelligence space-time sequence prediction, and particularly relates to a wind power plant data synchronization correction method based on view topology and mask map neural networks. Background Wind farms are typically deployed in complex and harsh natural environments such as mountains, gobi, or offshore. In the actual running and new unit testing process, sensor data faults and abnormal extreme noise often occur in a wind power plant SCADA (data acquisition and monitoring control) system under the influence of extreme freezing disasters, typhoons or faults of a large-scale communication base station and other irresistible forces. Unlike sporadic single-point short-time data loss, such faults often appear as "multiple-machine-position, long-time-sequence" concurrency loss, i.e., a plurality of wind turbine generators highly coupled in space simultaneously lose data for up to several hours or even days in the same time period. Accurate and complete high-fidelity wind speed and power test data are the basic preconditions of subsequent wind farm power prediction, microscopic site selection verification and unit grid-connected scheduling. However, the existing data interpolation and error correction techniques have the following serious technical drawbacks when dealing with the above-mentioned extreme concurrent missing scenario: 1. The local space reference failure and single machine limitation is that most of the existing wind speed correction methods (such as traditional BP, LSTM neural network or KNN neighbor algorithm) highly depend on the time sequence data of a single fan or search for a plurality of local neighbor nodes with the nearest physical distance. When multi-machine-position concurrent communication interruption occurs, local reference nodes of a target unit are simultaneously disabled, and the method ignores the space wake effect and global relevance among a plurality of fans in the whole wind power plant, so that the model is seriously degraded when the loss occurs. 2. The space-time evolution splitting and joint deduction capability is lacking, and when the traditional space-time diagram neural network (such as STGCN) is applied to the wind power field, a stripping type calculation strategy of alternately carrying out 'space diagram convolution' and 'time sequence convolution' is generally adopted. When the architecture processes the high-frequency dynamic change of the wind power plant and deals with a large amount of NaN (non-numerical value) incomplete data, characteristic information delay or interruption is easy to generate, and the integral hydrodynamic cooperative fluctuation characteristic of the wind power plant cannot be maintained in long-time-sequence multi-node concurrent deduction. 3. The memory overload and computation bottleneck caused by complex graph computation is that in the forward propagation of a large-scale global space-time graph network, the dynamic computation of graph node characteristics and Gao Weibian characteristics can cause extremely high memory occupation. Particularly, under the dynamic graph or static graph compiling mode of the current localization deep learning framework (such as MindSpore), the existing algorithm lacks a bottom memory optimizing mechanism for decoupling static physical topological relation and dynamic wind condition sequence, so that a great amount of repeated calculation overhead exists in forward propagation, and efficient industrial-level deployment is difficult to realize on a localization NPU computing platform. In summary, in the prior art, under a high-miss-rate test environment where multiple nodes fail concurrently, the underlying computing framework is efficiently utilized, and high-fidelity spatial-temporal data joint reconstruction and error correction are completed based on the global topology network. Disclosure of Invention The invention aims to provide a wind farm data synchronization correction method based on view topology and mask pattern neural network, which aims to break the limitation of single machine correction and local neighbor correction, realize test data high-fidelity error correction under high concurrency missing rate and strong noise interference through a space-time feature extraction mechanism and dynamic mask compensation of global synchronization, and simultaneously remarkably reduce the space-time diagram calculation memory overhead of a bottom deep learning frame through designing a static topology buffer mechanism. The wind farm data synchronization correction method based on the view topology and mask map neural network comprises the following steps: Step 1, acquiring historical space-time observation data of a plurality of unit nodes in a wind power plant, and constructing a space-time