A wind direction wake bias dynamic graph reconstruction method and system for a sparse observation wind farm

Abstract

The application discloses a wind direction wake bias dynamic graph reconstruction method and system for a sparse observation wind farm. The method obtains part of observation state data and spatial coordinates of multiple wind turbines in the wind farm at the current moment, generates initial prior values of the whole field by inverse distance weighted interpolation, constructs a fixed space near neighbor graph according to the spatial coordinates, and constructs a multi-head learning dynamic graph with distance bias, wind direction alignment bias and wake overlap bias according to input features. After performing graph message passing on the fixed graph branch and the dynamic graph branch respectively, a fusion representation is generated through a gate fusion network, and the initial prior values are corrected by using a residual correction method, and the reconstruction result of the target variable of the unobserved wind turbine is output. The application can improve the state completion accuracy of the wind turbine under the same moment sparse observation scene, and enhance the physical consistency between the graph structure and the wind direction propagation and the wake influence.

Description

Technical Field

[0001] This invention relates to the field of wind farm operation monitoring and data completion technology, and in particular to a method and system for reconstructing dynamic diagrams of wind direction wake bias for sparsely observed wind farms. Background Technology

[0002] During wind farm operation, factors such as communication link fluctuations, sensor failures, abnormal data acquisition, and limited edge computing resources often result in situations where only some wind turbines have valid observations at any given time. If the operating status of the unobserved wind turbines cannot be reliably completed at the same time, it will affect subsequent operations such as wind farm operation analysis, status assessment, power statistics, and fault diagnosis.

[0003] In existing technologies, one type of scheme mainly uses spatial interpolation methods such as inverse distance weighted interpolation and Kriging interpolation to estimate the impact of wind turbines on unobserved wind turbines. This type of scheme is simple to implement and has high computational efficiency, but it mainly relies on fixed geometric proximity relationships and is difficult to reflect the dynamic changes in the upstream to downstream influence paths between wind turbines after wind direction changes.

[0004] Another approach uses graph neural networks or graph attention networks to learn the relationships between wind turbine units. While this approach can improve nonlinear modeling capabilities, if it lacks structural constraints that match the wind propagation mechanism and wake influence, it is prone to learning connections with insufficient physical meaning, thus affecting the model's stability, interpretability, and generalization ability in out-of-distribution scenarios.

[0005] Therefore, it is necessary to propose a wind farm sparse observation reconstruction scheme that can simultaneously combine stable spatial priors, fixed local neighborhood structure, dynamic graph learning structure, and wind direction wake physical bias. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for reconstructing dynamic graphs of wind direction wake bias for sparsely observed wind farms, in order to solve the technical problems in the prior art where fixed spatial interpolation cannot fully utilize dynamic propagation relationships, and pure data-driven graph learning cannot take into account both physical consistency and reconstruction stability.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: First, a full-field inverse distance weighted prior is generated based on the target variables and spatial coordinates of the observed wind turbines; then, a fixed spatial nearest neighbor graph is constructed based on the spatial coordinates; simultaneously, a multi-head learning dynamic graph is constructed based on the input features, and distance bias, wind direction alignment bias, and wake overlap bias are introduced during the attention scoring process of the dynamic graph; subsequently, graph message passing is performed on the fixed graph branch and the dynamic graph branch respectively; then, the two types of branch representations are adaptively fused through a gated fusion network; finally, the inverse distance weighted prior is corrected using a residual correction method, and the reconstruction results of the unobserved wind turbines are output.

[0008] Compared with existing technologies, this invention has the following advantages: using inverse distance-weighted priors as initial values ​​for reconstruction provides a stable estimation basis for sparse observation scenarios; through collaborative modeling of fixed graphs and multi-head learning dynamic graphs, it balances local spatial continuity and dynamic propagation relationships; through distance bias, wind direction alignment bias, and wake overlap bias, it improves the consistency between the dynamic graph structure and the propagation laws of wind power physics; and through gating fusion and residual correction mechanisms, it improves the accuracy and generalization ability of state completion for unobserved wind turbine units. Attached Figure Description

[0009] Figure 1 This is a system structure block diagram according to an embodiment of the present invention.

[0010] Figure 2 This is a flowchart of the method described in an embodiment of the present invention.

[0011] Figure 3 This is a schematic diagram illustrating the construction of the fixed diagram and the dynamic diagram of wind direction wake offset as described in an embodiment of the present invention.

[0012] Figure 4 This is a schematic diagram of dual-path encoding, gated fusion, and residual output as described in an embodiment of the present invention. Detailed Implementation

[0013] The present invention will be further described below with reference to the accompanying drawings and specific embodiments. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention. Equivalent substitutions or modifications can be made by those skilled in the art without departing from the concept of the present invention, and all such substitutions or modifications should fall within the scope of protection of the present invention.

[0014] Example 1: Data Acquisition and Input Construction. Partial observational state data of N wind turbines in a wind farm at a certain moment, along with the two-dimensional spatial coordinates of each turbine, are acquired. For each wind turbine, an input feature vector is constructed, consisting of observed values, an observation mask, inverse distance weighted prior values, and normalized coordinates. For unobserved wind turbines, their observed values ​​are represented using a preset null value code or a zero value code, and their observation availability is indicated by an observation mask. The method input stage corresponding to this example can be combined with... Figure 2 Understanding that the relevant data access relationships can be combined Figure 1 understand.

[0015] Example 2: Inverse distance weighted prior generation. For unobserved wind turbines, initial prior values ​​are calculated using inverse distance weighted interpolation based on the target variable and spatial coordinates of observed wind turbines. For observed wind turbines, their actual observation values ​​are retained as prior values. These prior values ​​are used to provide stable initial estimates for subsequent residual reconstruction.

[0016] Example 3: Construction of a Fixed Spatial Nearest Neighbor Graph. Based on the two-dimensional spatial coordinates of each wind turbine, the Euclidean distance between any two wind turbines is calculated. A preset number of k nearest neighbor wind turbines are selected for each target wind turbine to construct a fixed spatial nearest neighbor graph. This fixed graph branch is used to describe the continuity characteristics of the local spatial structure of the wind farm. (See diagram for the construction of the fixed spatial nearest neighbor graph.) Figure 3 .

[0017] Example 4: Construction of a Dynamic Graph for Wind Direction and Wake Offset. For each attention head, a query vector, key vector, and value vector are calculated based on the input features, and a basic attention score is obtained. Distance offset, wind direction alignment offset, and wake overlap offset are then superimposed on the basic attention score. The distance offset enhances local proximity connections, the wind direction alignment offset describes the matching degree between the relative direction of the source wind turbine pointing to the target wind turbine and the current wind propagation direction, and the wake overlap offset enhances candidate edges that satisfy the wake propagation law based on the projected distance along the wind direction and the crosswind offset. Subsequently, a preset number of candidate adjacent edges with the highest scores are retained for each attention head, forming a sparse multi-head learning dynamic graph. The construction process of the wind direction and wake offset dynamic graph and its relationship with the fixed graph are illustrated in the following figure. Figure 3 .

[0018] Example 5: Dual-path graph encoding and gated fusion. Input features are fed into a fixed graph branch and a dynamic graph branch, each branch comprising at least two layers of graph message passing network. The fixed graph branch outputs a fixed graph representation, and the dynamic graph branch outputs a dynamic graph representation. Subsequently, the fixed graph representation, the dynamic graph representation, and the inverse distance weighted prior value are input into a gated fusion network to generate gated fusion coefficients, which are then used to perform weighted fusion of the fixed graph representation and the dynamic graph representation. The dual-path graph encoding and gated fusion process is illustrated in the diagram. Figure 4 .

[0019] Example 6: Residual Reconstruction Output. The fused representation, dynamic graph representation, and original input features are input into the residual decoding network. The output residual correction is added to the inverse distance weighted prior value to obtain the final reconstruction result. Preferably, the reconstruction output is read only at the location corresponding to the unobserved wind turbine as the state completion result at the current moment. (See attached diagram of the residual reconstruction output process.) Figure 4 .

[0020] Example 7: Effect Verification. In one example, SCADA data from an SDWPF wind farm was used to verify the scheme of the present invention. This data includes 134 wind turbine generators, 45 days of operation records, and a 10-minute sampling interval; error indices were calculated only for unobserved wind turbine generators at observation ratios of 70%, 50%, 30%, and 10%. The verification results are shown in Tables 1 to 6.

[0021] Table 1 Comparison of Active Power Patv Reconstruction Results

[0022] method Average MAE Average RMSE MAE improvement (%) relative to IDW RMSE improvement (%) relative to IDW IDW 54.89 99.22 0 0 IDW-Residual MLP 52.49 96.37 4.39 2.88 Unstructured bias learning graph 51.38 95.15 6.44 4.07 IDW-Residual GAT 50.21 94.42 8.64 4.83 Bias learning graph of the present invention 50.03 94.17 8.95 5.05

[0023] As shown in Table 1, when active power Patv is taken as the target variable, the average MAE and average RMSE of the inverse distance-weighted baseline are 54.89 and 99.22, respectively. The average MAE and average RMSE of the proposed solution are 50.03 and 94.17, respectively, representing improvements of 8.95% and 5.05% in average MAE and average RMSE compared to the baseline. This indicates that the proposed solution can stably improve reconstruction accuracy in random sparse observation scenarios.

[0024] Table 2. Ablation results constructed from the figures.

[0025] variants Average MAE Average RMSE MAE improvement (%) relative to IDW RMSE improvement (%) relative to IDW Residual MLP 52.49 96.37 4.39 2.88 Fixed spatial diagram 51.87 95.53 5.54 3.71 Wind map 52.48 96.71 4.46 2.54 Dual-gate wind direction chart 51.72 95.25 5.80 3.97 Dual-gate wake diagram 51.73 95.32 5.80 3.91 Unstructured bias learning graph 51.38 95.15 6.44 4.07 GAT baseline 50.21 94.42 8.64 4.83 Bias learning graph of the present invention 50.03 94.17 8.95 5.05

[0026] As shown in Table 2, in the graph construction ablation verification, the scheme of the present invention is superior to the learning graph structure that only uses a fixed spatial graph, only uses a wind direction graph, only uses a wake graph, and does not have a structural bias. This indicates that the combined use of a fixed spatial structure, a multi-head learning structure, and wind direction and wake bias can achieve better reconstruction results.

[0027] Table 3 Test results outside the wind direction distribution area

[0028] method Average MAE increase (%) Average RMSE improvement (%) Wind map 3.55 1.33 Wake diagram 3.72 1.53 Unstructured bias learning graph 5.02 2.18 GAT baseline 7.14 2.55 Bias learning graph of the present invention 7.45 2.89

[0029] Table 4. Results of tests for missing time periods and regions.

[0030] Distributed external scene Leave out objects Target number of wind turbines MAE increased (%) RMSE improvement (%) Time period distribution Section 1 Missing mask corresponding to the target wind turbine 5.63 0.26 Time period distribution Section 2 Missing mask corresponding to the target wind turbine 7.30 1.18 Time period distribution Section 3 Missing mask corresponding to the target wind turbine 8.79 3.48 Time period distribution Section 4 Missing mask corresponding to the target wind turbine 8.08 3.82 Region missing 0 34 2.78 0.62 Region missing 1 33 5.82 2.98 Region missing 2 27 6.43 3.93 Region missing 3 40 7.99 6.13

[0031] As shown in Tables 3 and 4, the proposed solution still shows positive improvement in tests with wind direction distribution outside the test, time period distribution outside the test, and regional missing test, indicating that the proposed solution still has good generalization ability under the conditions of wind direction change, time period change and local area observation missing.

[0032] Table 5 Results of the three random seed stability test

[0033] Observation ratio MAE Mean MAE Standard Deviation RMSE mean RMSE standard deviation MAE increased (%) RMSE improvement (%) 70% 45.49 0.05 87.15 0.43 10.82 5.39 50% 47.23 0.02 90.56 0.47 9.60 4.46 30% 50.14 0.32 94.14 0.31 8.09 4.74 10% 57.04 0.23 103.74 0.57 7.63 6.68

[0034] As shown in Table 5, the scheme of the present invention maintains small error fluctuations under different observation ratios and different random seeds, indicating that its training and inference processes have good stability.

[0035] Table 6. Statistical results of graph structure interpretability

[0036] Wind sector Average side distance Consistency of average wind direction Downstream side proportion Overlap with fixed nearest neighbor graph overlap with wind direction map overlap with wake diagram 5 1214.28 0.05 0.52 0.97 0.50 0.51 7 1245.47 -0.01 0.48 0.96 0.50 0.49 10 1335.23 -0.10 0.43 0.95 0.41 0.43

[0037] Table 6 illustrates the interpretability of the graph structure learned in this invention. By statistically analyzing the average edge distance, average wind direction consistency, downstream edge ratio, and overlap with the fixed nearest neighbor graph, wind direction graph, and wake graph under different wind direction sectors, it is shown that the biased dynamic graph constructed in this invention is not randomly connected. Instead, it adaptively adjusts the wind direction propagation relationship and wake influence relationship while maintaining local spatial proximity relationships. This demonstrates that the graph structure constructed in this invention has good physical consistency and interpretability.

Claims

1. A method for reconstructing dynamic wind direction wake bias diagrams for sparsely observed wind farms, characterized in that, Includes the following steps: (1) Obtain partial observation status data, wind direction information and spatial coordinates of each wind turbine at the current moment for multiple wind turbines in the wind farm, and construct the input features of each wind turbine. (2) Based on the observed target variables and spatial coordinates of the wind turbine units, inverse distance weighted interpolation is used to generate initial prior values, where the weights satisfy the following conditions: The initial prior values ​​satisfy ; in, This represents the interpolation weight of the observed wind turbine j with respect to the target wind turbine i. This represents the Euclidean distance between wind turbine i and wind turbine j. This represents a positive decimal used to avoid a denominator of zero, where p represents the distance decay exponent. This represents the observed values ​​of the target variable for wind turbine j. This represents the inverse distance-weighted initial prior value of wind turbine i; (3) Construct a fixed spatial nearest neighbor graph based on the spatial coordinates of each wind turbine; (4) Construct a multi-head learning dynamic graph based on the input features, where the edge score of the h-th attention head satisfies And for each attention head, retain a preset number of candidate adjacent edges with the highest scores; in, This represents the comprehensive edge score of the h-th attention head's source wind turbine j pointing to the target wind turbine i. This represents the basic attention score calculated from the query vector and the key vector. , and represents the learnable weight coefficients of distance offset, wind direction alignment offset, and wake overlap offset in the h-th attention head, respectively. Indicates the distance offset term. This indicates the wind direction alignment offset term. Indicates the wake overlap bias term; (5) Perform graph message passing based on the fixed spatial nearest neighbor graph and the multi-head learning dynamic graph respectively to obtain the fixed graph branch representation. and dynamic graph branch representation ; (6) Generate gated fusion coefficients based on the fixed graph branch representation, the dynamic graph branch representation, and the initial prior value, wherein the gated fusion coefficients satisfy... And based on this, a fusion characterization is obtained. ; (7) Output residual correction amount based on fusion characterization The residual correction is then added to the initial prior value to obtain the target variable reconstruction result for the unobserved wind turbine units. The reconstruction result satisfies... ; in, This represents the final reconfiguration result of the target wind turbine i. This represents the residual correction amount output by the residual decoding network.

2. The method according to claim 1, characterized in that, The input features mentioned in step (1) include at least the target variable observation value, observation mask, inverse distance weighted prior value, and wind turbine normalized spatial coordinates.

3. The method according to claim 1, characterized in that, The distance between the wind turbine units in steps (2) and (3) satisfies For each target wind turbine, a preset number of k nearest neighbor wind turbines are selected to construct the fixed spatial nearest neighbor graph.

4. The method according to claim 1, characterized in that, The wind direction alignment offset in step (4) satisfies ; in This indicates the relative directional angle between the source wind turbine and the target wind turbine. It indicates the direction of wind propagation at the current moment.

5. The method according to claim 1, characterized in that, The wake overlap offset in step (4) is based on the projected distance along the wind direction. and crosswind offset jointly determined, ; in, Indicates the projected distance along the wind direction. Indicates the crosswind offset. , Let i and n represent the planar coordinate components of the target wind turbine i. , These represent the planar coordinate components of the source wind turbine j.

6. The method according to claim 1, characterized in that, The distance bias in step (4) is used to enhance local proximity connections, the wake overlap bias is used to enhance candidate adjacent edges that satisfy the wake propagation law, and the multi-head learning dynamic graph forms a sparse dynamic graph connection structure by performing Top-K edge screening on each attention head.

7. The method according to claim 1, characterized in that, The reconstruction task involves completing the target variables for only the unobserved wind turbines at any given time. The target variables include one or more of the following: active power, wind speed, rotational speed, and pitch angle.

8. A wind direction wake bias dynamic diagram reconstruction system for sparsely observed wind farms, characterized in that, include: The module includes a data access module, a priori generation module, a fixed graph construction module, a dynamic graph construction module, a bi-branch coding module, and a gated fusion and residual output module. The data access module is used to acquire partial observation status data, wind direction information, and spatial coordinates; the prior generation module is used to... and Generate inverse distance weighted initial prior values; the fixed graph construction module is used to generate a fixed spatial nearest neighbor graph; the dynamic graph construction module is used to generate a multi-head learning dynamic graph with distance offset, wind direction alignment offset, and wake overlap offset; The dual-branch encoding module is used to extract fixed graph branch representations and dynamic graph branch representations respectively; the gated fusion and residual output module is used to extract according to... Generate fused representations and output the reconstructed target variables of unobserved wind turbine units; in, This represents the gating fusion coefficient corresponding to the target wind turbine i. and Let represent the weight parameters and bias parameters of the gated fusion network, respectively. This represents the output of a fixed graph branch. This represents the output of a dynamic graph branch. Represents the integrated representation after fusion, symbol This indicates a vector concatenation operation.

9. An electronic device comprising a processor and a memory, characterized in that, The memory stores a computer program, which, when executed by the processor, implements the method of claim 1.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method of claim 1.

Images