Multi-modal offshore wind power ultra-short-term prediction method

Abstract

The invention discloses a multi-modal offshore wind power ultra-short-term prediction method in the field of offshore wind power plant cluster power prediction, and aims to solve the technical problems of spatial-temporal feature splitting and insufficient dynamic dependency relationship modeling. The method comprises the steps of performing anomaly detection and restoration on fan data, and generating a corrected wind power cluster data set; extracting a mean value, a standard deviation and a latest value of core operation data of each fan through a dynamic time window, and constructing a multi-dimensional node feature; a static geographic similarity matrix is generated based on geographic coordinates, a basic wake effect matrix is generated in combination with real-time wind direction data, correction is carried out through the maximum mutual information quantization time-delay effect, and then a dynamic adjacency matrix is obtained through self-adaptive fusion; and integrating the multi-dimensional node features and the dynamic adjacency matrix into a space-time diagram sequence data architecture, inputting the space-time diagram sequence data architecture into a multi-scale wake flow perception diagram space-time prediction model, and outputting a multi-fan power prediction value. According to the invention, high-precision multi-fan power prediction can be realized.

Description

Technical Field

[0001] This invention relates to a multimodal method for ultra-short-term prediction of offshore wind power, belonging to the field of offshore wind farm cluster power prediction technology. Background Technology

[0002] Offshore wind power, as a core pillar of the clean energy transition, plays an irreplaceable role in promoting the strategic goal of energy transformation. However, offshore wind farms generally face the dual challenges of complex and variable meteorological environments and strong spatial coupling among multiple turbines, resulting in highly nonlinear and uncertain power output. In recent years, Graph Neural Networks (GNNs) have been introduced to model wind turbine network structures, but existing hybrid GNN-based schemes have significant drawbacks: traditional GNNs rely on predefined fixed adjacency matrices, which cannot dynamically adapt to real-time spatial dependencies caused by wind direction changes; conventional graph convolution operations treat all neighbor nodes equally, failing to distinguish the differential influence of key turbine nodes (such as upstream turbines in the dominant wind direction); and the oversmoothing effect of message passing in deep GNNs weakens the effective association of distant turbines. To overcome these technical bottlenecks, this invention proposes an innovative two-stage solution. Summary of the Invention

[0003] The purpose of this invention is to overcome the three major defects in existing offshore wind power prediction technologies: rigid spatial modeling, fragmented spatiotemporal features, and sluggish response to transient fluctuations. It provides a multimodal offshore wind power ultra-short-term prediction method based on a dual-modal dynamic graph fusion mechanism and a multi-scale wake perception graph spatiotemporal prediction model prediction framework.

[0004] To solve the above-mentioned technical problems, the present invention is implemented using the following technical solution.

[0005] On the one hand, this invention provides a multimodal offshore wind power ultra-short-term prediction method, including: Obtain the original wind turbine data of the offshore wind farm cluster, and perform anomaly detection and repair processing on the original wind turbine data of the offshore wind farm cluster to generate a corrected wind power cluster dataset. Based on the corrected wind power cluster dataset, the operating data of each wind turbine within the window is extracted using a dynamic time window mechanism, and the operating data of all wind turbines are aggregated into multi-dimensional node features. A static geographic similarity matrix is ​​generated based on the wind turbine's geographic coordinates, and a basic wake effect matrix is ​​generated based on real-time wind direction data. The basic wake effect matrix is ​​modified to obtain the modified wake effect matrix, and the static geographic similarity matrix and the modified wake effect matrix are adaptively weighted and fused to generate a dynamic adjacency matrix. The dynamic adjacency matrix and multi-dimensional node features are integrated to form a spatiotemporal graph sequence data architecture arranged in chronological order. The spatiotemporal graph sequence data architecture is input into the pre-constructed multi-scale wake perception graph spatiotemporal prediction model, and the predicted power values ​​of multiple wind turbines are output.

[0006] In conjunction with the first aspect, the generation of the corrected wind power cluster dataset further includes: The isolated forest algorithm is used to detect abnormal power values ​​of the wind turbine data of the original offshore wind farm cluster in combination with the physical rules of the wind power field, and the abnormal power values ​​are marked. Construct a lightweight gradient booster model and train it with data that is not labeled as anomalous power values; The non-power features corresponding to the labeled outlier power values ​​are input into the trained lightweight gradient booster model to obtain the outlier power prediction values. The power values ​​of the outliers are replaced with the predicted power values ​​of the outliers, and a physical upper limit constraint is applied to the predicted power values ​​of the outliers to obtain the corrected wind power cluster dataset.

[0007] In conjunction with the first aspect, further, the aggregation of all wind turbine operating data into multi-dimensional node features includes: Starting from the current time t, slide forward a time window of fixed length W; Calculate the Pearson correlation coefficient between each operating variable of each wind turbine and the wind turbine power p within the time window W; The absolute values ​​of all Pearson correlation coefficients are compared with a preset threshold. Compare and filter out those whose absolute value is greater than a preset threshold. Variables; Calculate the mean, standard deviation, and latest value of each variable for each wind turbine within the time window W; The mean, standard deviation, and latest value of all variables for each wind turbine are concatenated in the order of the variables to form the feature vector of each wind turbine; Stack the feature vectors of all wind turbines at the current moment to form multi-dimensional node features.

[0008] In conjunction with the first aspect, the expression for the Pearson correlation coefficient between each operating variable of each wind turbine and the wind turbine power p is as follows: ; in, X represents the Pearson correlation coefficient between the input feature variable and the wind turbine power P; X represents the input feature variable; P represents the wind turbine output power. The standard deviation of a variable; This represents the standard deviation of the fan power. Indicates wind speed; Indicates the pitch angle; Indicates rotational speed; This represents the covariance between the input feature variable and the wind turbine power P.

[0009] In conjunction with the first aspect, further, the generation of a static geographic similarity matrix based on the wind turbine's geographic coordinates and the generation of a basic wake effect matrix based on real-time wind direction data include: Based on the latitude and longitude coordinates of each wind turbine, a static geographic similarity matrix between wind turbines is calculated using a Gaussian kernel function; Based on real-time wind direction data, the upstream and downstream relationship between wind turbines is determined by the wind direction angle; for wind turbine pairs with upstream and downstream relationships, the wake influence intensity is calculated based on the cosine attenuation model, and a dynamic wake effect matrix is ​​generated. Based on the static geographic similarity matrix and the dynamic wake effect matrix, a basic wake effect matrix is ​​formed.

[0010] In conjunction with the first aspect, the expression for the basic wake effect matrix is ​​further as follows: ; in, This represents the basic wake effect matrix between wind turbine i and wind turbine j; This represents the azimuth angle of fan j relative to fan i; Represents the latitude and longitude coordinates of wind turbine i; Represents the latitude and longitude coordinates of wind turbine j; denoted by , where represents the wake attenuation coefficient; U indicates that fan i has a wake effect on fan j.

[0011] In conjunction with the first aspect, the generation of the dynamic adjacency matrix further includes: Set sliding time window Within the sliding time window, calculate the maximum mutual information coefficient of power delay between the two wind turbines. Using the aforementioned time-delay maximum mutual information coefficient, the basic wake effect matrix of the current time step is corrected according to the nonlinear correlation of power data between wind turbines under time delay, to obtain the corrected dynamic wake effect matrix. The weights of the static geographic similarity matrix are calculated using the wind direction change rate as the weighting function. and the weights of the dynamic wake effect matrix ; The static geographic similarity matrix and the modified dynamic wake effect matrix are weighted and fused according to their corresponding weights to generate a dynamic adjacency matrix.

[0012] In conjunction with the first aspect, further, the weights of the static geographic similarity matrix... and the weights of the dynamic wake effect matrix The calculation expression is: ; in, The weights represent the static geographic similarity matrix. Represents the weights of the dynamic wake effect matrix; Indicates the rate of change of wind direction; This represents the sensitivity function.

[0013] In conjunction with the first aspect, the formation of the spatiotemporal graph sequence data architecture arranged in chronological order includes: The multidimensional node feature matrix of each time step is combined with the dynamic adjacency matrix of the corresponding time step after time delay correction to construct a graph sequence on discrete time steps; The graph sequence at multiple consecutive time steps is constructed as a sample for training the graph neural network using a sliding window mechanism, and the rate of change of the dynamic adjacency matrix within each sliding window is calculated. Based on the change rate of the dynamic adjacency matrix, topological transition events caused by sudden changes in wind conditions are marked; based on the samples of the sliding window and their corresponding topological transition marks, a standardized spatiotemporal graph sequence dataset supporting multi-scale time window sampling is constructed, thereby forming a graph sequence data architecture arranged in chronological order and capable of representing spatiotemporal evolution patterns.

[0014] In conjunction with the first aspect, the multi-scale wake perception map spatiotemporal prediction model further includes a multi-scale convolutional module, a wake perception map attention network, and a bidirectional long short-term memory network. The multi-dimensional node features of each time step in the spatiotemporal graph sequence data architecture are input into the multi-scale convolution module to extract the multi-scale transient features of the power sequence of each wind turbine. In the wake perception graph attention network, the multi-scale transient features and dynamic adjacency matrix of the power sequence are transformed to obtain the attention mask of physical constraints. The attention mask of the physical constraints is input into a bidirectional long short-term memory network to obtain fused spatiotemporal features; A fully connected graph decoder is used to decode the fused spatiotemporal features and synchronously output the power prediction values ​​of all wind turbines at future times.

[0015] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: This invention performs anomaly detection and repair processing on the original offshore wind farm cluster's turbine data to generate a corrected wind farm cluster dataset. Then, based on a dynamic time window mechanism, it aggregates multi-scale statistical features of each turbine's operating data within the window, forming multi-dimensional node features describing the turbine's "trend-fluctuation-instantaneous state." Simultaneously, it constructs a static geographical similarity matrix based on the turbine's geographical coordinates and generates a basic wake effect matrix using a cosine decay model based on real-time wind direction. Furthermore, it introduces a time delay compensation mechanism to correct the wake matrix and adaptively fuses it with the geographical matrix using wind direction change rate as weight, generating a dynamic adjacency matrix. Finally, it integrates the dynamic adjacency matrix and node features into a time-series graph sequence, which is input into a multi-scale wake perception spatiotemporal prediction model (Inception-GAT-BiLSTM) that integrates multi-scale convolution (Inception), wake perception graph attention network (GAT), and bidirectional long short-term memory network (BiLSTM). This simultaneously outputs high-precision, physically interpretable multi-turbine power predictions, effectively improving prediction timeliness and robustness. Attached Figure Description

[0016] Figure 1 The image shown is a feature aggregation diagram of statistical window nodes provided in an embodiment of the present invention; Figure 2 The diagram shown is a schematic diagram of static geographic similarity matrix calculation provided in an embodiment of the present invention. Figure 3 The diagram shown is a flowchart of the dynamic wake effect adjacency matrix calculation provided in an embodiment of the present invention. Figure 4 The image shown is a schematic diagram of a spatiotemporal heterogeneous graph dataset provided in an embodiment of the present invention; Figure 5 The diagram shown is an Inception architecture based on depthwise separable convolution provided by an embodiment of the present invention. Figure 6 The diagram shown is an LSTM cell structure provided in an embodiment of the present invention; Figure 7 The diagram shown is a BiLSTM network structure provided in an embodiment of the present invention. Figure 8 The diagram shown is a framework diagram of offshore wind power prediction based on Incepton-GAT-BiLSTM provided in an embodiment of the present invention. Detailed Implementation

[0017] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0018] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship. Example 1

[0019] This embodiment introduces a multimodal method for ultra-short-term prediction of offshore wind power. First, the wind turbine data of the original offshore wind farm cluster is subjected to anomaly detection and repair processing. By combining the isolated forest algorithm with the physical rules of the wind power field, the abnormal power values ​​are accurately identified. The abnormal power values ​​are replaced by a lightweight gradient booster model. At the same time, physical upper limit constraints based on historical data are applied, thereby generating a corrected wind power cluster dataset with strong physical consistency and high data quality. Subsequently, based on the dynamic time window mechanism, the operating data of each wind turbine that is strongly correlated with the power prediction are extracted, and the statistics (mean, standard deviation and latest value) within the window are calculated to aggregate and construct a multi-dimensional node feature that can simultaneously characterize the long-term trend, fluctuation characteristics and instantaneous state of the wind turbine.

[0020] Next, a basic wake effect matrix is ​​generated by fusing geospatial information and real-time wind direction data; To further improve the accuracy of spatial dependency characterization, a time delay compensation mechanism based on the maximum mutual information coefficient is introduced to modify the basic wake effect matrix in order to capture the non-instantaneous characteristics of wake propagation. Furthermore, based on the wind direction change rate, static geographical association and dynamic wake effect are adaptively weighted and fused to generate a dynamic adjacency matrix that can accurately reflect the time-varying spatial relationship between wind turbines.

[0021] Furthermore, the dynamic adjacency matrix and multi-dimensional node features are integrated according to timestamps to construct a spatiotemporal graph sequence data architecture that can be sampled by a sliding window. This architecture effectively organizes the spatiotemporal evolution information of the wind turbine cluster.

[0022] Finally, the image data architecture is input into the spatiotemporal prediction model of the multi-scale wake perception map, and the predicted values ​​of multiple wind turbine power are output, realizing high-precision and interpretable ultra-short-term wind power prediction with deep integration of physical mechanism and data-driven approach. Example 2

[0023] Step S1, Dataset Outlier Detection and Reconstruction: Obtain the original wind turbine data of the offshore wind farm cluster, and perform anomaly detection and repair processing on the original wind turbine data of the offshore wind farm cluster to generate a corrected wind power cluster dataset. The process of generating the corrected wind power cluster dataset includes the following steps: Step S11: Use the isolated forest algorithm combined with the physical rules of the wind power field to detect abnormal power values ​​of the original offshore wind farm cluster's wind turbine data, and mark the abnormal power values. Specifically, the isolated forest algorithm and the physical rules in the wind power field work together to detect abnormal data in the original offshore wind farm cluster's turbine data that deviate from the main body and do not conform to physical laws. In other words, a large amount of noise in the dataset is labeled to provide a data foundation for subsequent reconstruction of abnormal power values. Its basic principle is as follows: (1) Anomaly power value detection based on the isolated forest algorithm The Isolation Forest algorithm is based on the fact that outliers are sparsely distributed in the feature space and far from the population center, so they can be "isolated" with fewer segmentation operations. That is, each sample is treated as a point in a multi-dimensional space, and multiple binary trees (Isolation Trees) are recursively constructed by randomly selecting features and segmentation values.

[0024] Then, the results of all binary trees are integrated, and the average path length of sample point x in all trees is calculated. Since the average path length of a sample point in a completely randomly generated binary tree is related to the sample size n, a standardization factor is introduced to eliminate the influence of sample size on path length and to standardize outlier scores. It is defined as the expected estimate of the path length in a binary search tree (BST) constructed from n samples, and the formula is as follows: (1) (1) in, Represents the harmonic number and , represents Euler's constant, typically taken as 0.5772156649; n represents the total sample size. This represents the standardization factor.

[0025] Furthermore, the average path length is converted into an anomaly score between 0 and 1 using formula (1), as follows: (2) in, Indicates the average path length; Represents a fraction.

[0026] It should be noted that, If it approaches 0 (it is easily isolated). It then approaches 1 (abnormal); If it approaches (the path length of a normal point) , It then approaches 0.5; If it exceeds (the point that is difficult to isolate) , Then it is less than 0.5 (normal).

[0027] In this embodiment of the invention, an anomaly score threshold is set. At that time, The sample was identified as a suspected outlier.

[0028] (2) Domain rules assist in judgment It is necessary to define physical or logical constraint rules based on knowledge of the wind power field, verify the anomalies identified in the isolated forest analysis, correct misjudgments or enhance missed judgments, and ultimately generate accurate anomaly markers. The main rules include power limit exceeding rules and power deviation anomaly rules: (3) in, This represents the measured power. Indicates the rated power of the fan. This represents the theoretical power obtained by looking up the table based on wind speed; This indicates the set threshold.

[0029] According to formula (3), if the measured power is greater than the rated power of the fan, it may be due to a power sensor malfunction or fan overload, requiring further investigation. If the measured power curve deviates from the theoretical value, i.e. exceeds the threshold, it is marked as an abnormal point, which may be due to problems such as decreased blade efficiency or sensor drift.

[0030] Step S12: Construct a lightweight gradient boosting machine model and train it with data that is not labeled as outliers. Lightweight Gradient Boosting Machine (LightGBM) is an efficient improvement framework for Gradient Boosting Decision Tree (GBDT). Its core is the iterative construction of an additive model, where each iteration combines weak learners by minimizing the loss function L. (Usually a decision tree), the output of the model in the Mth iteration is as follows: (4) in, Indicates the learning rate; This represents the cumulative predictions of the first M-1 models; This represents the cumulative prediction of the current model; This represents a decision tree composed of x sample points.

[0031] The learning objective of this model is to make the goal of each newly constructed decision tree fit the negative gradient (i.e., pseudo-residual) of the current model on training sample i. Its expression is as follows: (5) in, This represents the negative gradient backlash of the current model on training sample i; This represents the loss function, used to measure the true value. Compared with the predicted value The error between the two values; represents the model's prediction for the i-th sample; Indicates the model's predicted value; This represents the partial derivative of the loss function with respect to itself; This represents the cumulative model predictions after the first M−1 iterations.

[0032] In the regression task of wind power forecasting, the commonly used loss functions are mean squared error (MSE) or minimum absolute error (MAE), which are defined as follows: (6) Where MSE represents mean square error; MAE represents minimum absolute error; and y represents target power.

[0033] LightGBM, building upon traditional GBDT, introduces gradient-based one-sided sampling (GOSS) and mutually exclusive feature binding (EFB) to reduce data volume and feature dimensionality, thereby significantly improving efficiency and reducing memory consumption. The negative gradient of the loss function described above is calculated as follows: (7) in, This represents the pseudo-residual under MSE loss; This represents the predicted value of the model for sample i in the M-1th round; This represents the pseudo-residual under MAE loss.

[0034] Based on the above model construction, the LightGBM model is trained using only labeled normal sample data. The model's input feature variables... Select variables that are highly correlated with power (such as wind speed, wind direction, paddle pitch angle, rotational speed, temperature, etc.), and the target power y is the power value P at the corresponding time.

[0035] Step S13: Identify the non-power characteristics corresponding to the marked abnormal power values. The values ​​are input into the trained lightweight gradient booster model to obtain the predicted power values ​​of outliers; For sample data points marked as outliers, their corresponding non-power features are input into the trained lightweight gradient boosting machine model, and the model outputs the predicted power value of the outlier. This is the estimated value for the repair of the anomaly.

[0036] Step S14: Replace the power value of the outlier with the predicted power value of the outlier and impose a physical upper limit constraint on the predicted power value of the outlier to obtain the corrected wind power cluster dataset.

[0037] To prevent the model's repair estimates In extreme cases, exceeding the physical limits of wind turbine operation, such as rated power. This introduces an empirical upper bound constraint based on historical data. The 99th percentile of all normal power data for the wind farm cluster within a representative historical period is calculated. This quantile represents the power level that has not been exceeded in 99% of historical data points, typically close to or equal to the rated power. The expression for applying an upper limit constraint to each repair value is: (8) in, This represents the estimated outlier repair value after applying upper bound constraints. Represents the 99th percentile of all normal power data; represents the repaired estimate without upper limit constraints.

[0038] This upper limit constraint operation ensures that the repair power value ultimately used for model training and prediction both conforms to the trend of data-driven model prediction and is strictly limited to the physically possible range.

[0039] Finally, after the above-mentioned anomaly detection, labeling, repair, and constraint processes, a corrected wind power and operating status dataset is generated. This dataset effectively removes unreliable outliers and replaces them with reasonable predicted values ​​that meet physical constraints, significantly improving data quality and physical consistency, and laying a reliable data foundation for subsequent accurate spatiotemporal modeling and power prediction.

[0040] Step S2, Spatiotemporal Graph Node Feature Construction: Based on the corrected wind power cluster dataset, the operating data of each wind turbine within the window is extracted using a dynamic time window mechanism, and the operating data of all wind turbines are aggregated into multi-dimensional node features.

[0041] See Figure 1 Wind power clusters typically consist of multiple wind turbines, each with temporal characteristics and coupling relationships. Therefore, the core objective of this step is to generate a multi-dimensional feature vector for each wind turbine (a node in the spatiotemporal graph). This vector must simultaneously include: the historical operating trend of the wind turbine in the long-term time window (long-term evolution), the instantaneous fluctuation characteristics of the wind turbine in the short-term time window (short-term changes), and the real-time status information of the wind turbine at the current time step (latest dynamics). Furthermore, when constructing the node features of its spatiotemporal graph, it is necessary not only to consider the long-term evolution, short-term changes, and latest dynamics of the wind turbine, but also to perform feature variable filtering to obtain a spatiotemporal graph dataset that adequately reflects the dynamic characteristics of the wind turbine cluster without redundancy. The aggregation process is as follows: Step S21: Using the current time t as the endpoint, slide forward a time window of fixed length W; To construct the node features of the spatiotemporal graph, a dynamic time window needs to be built. That is, the input features are the original time-series data of each wind turbine at time step t (such as power, wind speed, blade pitch angle, rotational speed, temperature, etc.), and a historical time window of a fixed length is slid forward from the current time as the endpoint. Window length Adjustable. Step S22: Calculate the Pearson correlation coefficient between each operating variable of each wind turbine and the wind turbine power p within the time window W; For variables related to power prediction within the window, calculate the value of each variable (e.g., wind speed). Pitch angle Rotation speed The Pearson correlation coefficient between the wind turbine power P and the wind turbine power P is expressed as follows: (9) in, X represents the Pearson correlation coefficient between the input feature variable and the wind turbine power P; X represents the input feature variable; P represents the wind turbine output power. The standard deviation of variables that are highly correlated with power; This represents the standard deviation of the fan power. Indicates wind speed; Indicates the pitch angle; Indicates rotational speed; This represents the covariance between the input feature variable and the wind turbine power P.

[0042] Step S23: Compare the absolute values ​​of all Pearson correlation coefficients with a preset threshold. Compare and filter out those whose absolute value is greater than a preset threshold. (Right now The variable X is the core variable. Step S24: Calculate the mean, standard deviation, and latest value of each variable for each wind turbine within the time window W; The calculation expressions for the above three sets of statistics are as follows: (10) in, The mean within the window is used to characterize the long-term trend of the variable within the window; k represents the length of the window; t represents the time step; and i represents the sample data point. This represents the value of the variable at the i-th time step within the window; It represents the standard deviation within the window and is used to quantify the volatility of a variable; This represents the latest value within the window, i.e., the value at the current time step, and is used to capture the instantaneous state of the variable; The variable representing time t.

[0043] Step S25: Concatenate the mean, standard deviation, and latest value of all variables for each wind turbine in the order of the variables to form the feature vector of each wind turbine, the expression of which is as follows: (11) in, This represents the eigenvector of wind turbine i at time step t; This represents the average wind speed of fan i within a window of time step t; This represents the standard deviation of the wind speed of fan i within the window of time step t; This represents the current value of the wind speed of fan i at time step t; This represents the average pitch angle of wind turbine i within a window at time step t; This represents the standard deviation of the pitch angle of wind turbine i within a window at time step t; This represents the current value of the pitch angle of wind turbine i at time step t.

[0044] Step S26: Assess the feature vectors of all wind turbines at the current moment. Stack them to form multi-dimensional node features.

[0045] At this point, the vector dimension equals the number of core variables multiplied by 3 (mean / standard deviation / latest value). Then, the feature vectors of all wind turbines at time step t are stacked to form a 3D feature vector, the expression of which is as follows: (12) in, This represents a multidimensional feature vector, i.e., a 3D feature vector; denoted as the node feature matrix; D represents the feature dimension of each node; N represents the total number of wind turbines.

[0046] The resulting 3D feature vector retains variables closely related to power physics while comprehensively describing the "inertia-fluctuation-abrupt change" characteristics of wind turbine operation. This step generates a multi-dimensional feature vector. The generated dynamic adjacency matrix (describing the spatial relationship between wind turbines) is combined and input into the Inception-GAT-BiLSTM framework (multi-scale wake perception spatiotemporal prediction model).

[0047] Step S3, Basic Wake Effect Matrix: Generate a static geographic similarity matrix based on the wind turbine's geographic coordinates and a basic wake effect matrix based on real-time wind direction data. The generation process of this basic wake effect matrix is ​​as follows: Step S31: Calculate the static geographical similarity matrix between wind turbines using the Gaussian kernel function based on the latitude and longitude coordinates of each wind turbine. See Figure 2Geographically proximate wind turbines are affected by similar meteorological environments (such as wind speed and turbulence). Therefore, this invention uses a Gaussian kernel function to quantify the spatial attenuation effect: the closer the distance, the higher the similarity. The closer the similarity is to 1, the closer it is to 0; the farther the distance, the closer the similarity is to 0. The formula for calculating the elements of the static geographic similarity matrix is ​​as follows: (13) in, This represents the static geographical similarity matrix between wind turbine i and wind turbine j; Latitude and longitude coordinates of wind turbine i; Represents the latitude and longitude coordinates of wind turbine j; This represents the spatial attenuation coefficient.

[0048] Because the operating environment of offshore wind farm clusters has significant time dynamic characteristics: the continuous changes in instantaneous wind direction and speed will cause the wind turbine yaw angle to be adjusted in real time, which will lead to the dynamic evolution of the upstream and downstream relationship and wake effect between wind turbines. Traditional modeling methods based on static geographical similarity matrix can no longer accurately characterize this time-varying feature. Therefore, in addition to obtaining the static response between wind turbines, a dynamic response model also needs to be established, as shown in step S32 below.

[0049] Step S32: Based on real-time wind direction data, determine the upstream and downstream relationship of the wind turbines by the wind direction angle; for wind turbine pairs with upstream and downstream relationships, calculate the wake influence intensity based on the cosine attenuation model and generate a dynamic wake effect matrix. This invention uses the wind direction at each moment as the driving force to determine the upstream and downstream relationships between wind turbines, and then establishes a basic dynamic wake effect matrix based on a cosine attenuation model. Let the wind direction angle at each moment be... The azimuth angle from fan i to fan j is The rule for determining the upstream and downstream relationship of wind turbines driven by wind direction is as follows: If and When wind turbine i is in the upwind fan-shaped region relative to wind turbine j, it is considered that the wind turbine has a wake effect on the wind turbine. This rule is denoted as U. In other cases, the upstream and downstream relationship does not hold. Then, the element expression of the wake effect matrix calculated based on the cosine attenuation model is as follows: (14) in, This represents the basic wake effect matrix between wind turbine i and wind turbine j; This represents the azimuth angle of fan j relative to fan i; Represents the latitude and longitude coordinates of wind turbine i; Represents the latitude and longitude coordinates of wind turbine j; denoted by , where represents the wake attenuation coefficient; U indicates that fan i has a wake effect on fan j.

[0050] Step S33: Based on the static geographic similarity matrix and the tail dynamic tail effect matrix, form the basic tail effect matrix.

[0051] Step S4, Dynamic Adjacency Matrix Lag Correction: Correct the basic wake effect matrix to obtain the corrected wake effect matrix, and then adaptively weight and fuse the static geographic similarity matrix and the corrected wake effect matrix to generate a dynamic adjacency matrix. See Figure 3 This step aims to establish dynamic spatial dependencies between wind turbine nodes, overcoming the limitations of traditional fixed adjacency matrices. The dynamic adjacency matrix is ​​constructed through a time-delay compensation mechanism based on the maximum mutual information coefficient, integrating static geographical association and dynamic wake effects, as follows: (1) Time Delay Effect Compensation Mechanism While the wake effect matrix constructed based on the cosine decay model can preliminarily characterize the influence relationship between wind turbines, its modeling results still have two key limitations: First, there is a significant time delay effect in the physical system; the wake effect of the upstream wind turbine on the downstream wind turbine needs to undergo a propagation process, exhibiting typical non-instantaneous, short-term dynamic characteristics. Second, the calculation results based on the physical model are easily affected by environmental disturbances (such as turbulence intensity, wind shear, etc.), which may introduce systematic errors and reduce modeling accuracy. Therefore, this invention innovatively adopts the maximum mutual information coefficient (MIC) algorithm to establish a dynamic time delay compensation mechanism by quantifying the nonlinear correlation and time delay characteristics between the wind turbine output sequences within the current MIC calculation time window. This mechanism complements and integrates with the basic wake effect matrix, constructing a hybrid modeling framework with time-series awareness capabilities, ultimately realizing dynamic wake matrix modeling that integrates data-driven and physical mechanisms. The specific process includes: Step S41: Set the sliding time window Within the sliding time window, calculate the maximum mutual information coefficient of power delay between the two wind turbines. The wake dependency quantification based on the MIC historical window differs from the statistical time window; instead, a new MIC calculation window is established to calculate the nonlinear time-delay dependency between wind turbines. That is, within the sliding time window... Inside, calculate the power of fan i. For the power of fan j The maximum mutual information coefficient with time delay is expressed as follows: (15) in, Represents the maximum mutual information coefficient with time delay; Indicates the time delay offset; Indicates the maximum time delay window; This indicates that fan i has moved forward by the current time step t. The power sequence afterwards; Let X represent the power sequence of wind turbine i at the current time step t; X and Y represent two random variables or sequences; G represents the grid divided on the value range of X and Y. Indicates the number of rows in grid G; Indicates the number of columns in grid G; This represents the mutual information value after grid binning.

[0052] Step S42: Using the maximum mutual information coefficient with time delay, measure the nonlinear correlation of power data between wind turbines under time delay, and then correct the basic wake effect matrix of the current time step to obtain the dynamic wake effect matrix that incorporates the time delay effect, the expression of which is as follows: (16) in, Indicates the time-delay gain coefficient; This represents the dynamic wake matrix incorporating time-delay effects, indicating that... Historical changes and The current changes have a strong causal relationship; The dynamic wake effect matrix representing the fusion time delay effect; This represents the corrected dynamic wake effect matrix; This indicates the power of fan i; This indicates the power of fan j.

[0053] (2) Bimodal dynamic adjacency matrix After obtaining the geographic similarity matrix and the corrected dynamic wake matrix, a dynamic adjacency matrix that integrates geographic location and time-delay wake effects is constructed using the wind direction change rate as a weighting function, as follows: Step S43: Calculate the weights of the static geographic similarity matrix using the wind direction change rate as the weighting function. and the weights of the dynamic wake effect matrix ; The weighting function for the rate of change of wind direction is: (17) in, Represents the sensitivity function; Indicates the instantaneous rate of change of wind direction; The weights represent the static geographic similarity matrix. This represents the weights of the dynamic wake effect matrix.

[0054] Step S44: The static geographic similarity matrix and the dynamic wake effect matrix are weighted and fused according to their corresponding weights to generate a dynamic adjacency matrix.

[0055] When the wind direction is stable ( ), This is achieved through a balanced integration of the wake effect and the overall dynamic adjacency matrix. When wind direction changes abruptly, the wake effect dominates (reflecting the restructuring of upstream and downstream relationships). When wind direction changes gradually, geographical similarity dominates. Ultimately, a dual-modal dynamic adjacency matrix integrating data-driven and physical mechanisms is realized, expressed as follows: (18) in, Represents a dynamic adjacency matrix; This represents the static geographic similarity matrix.

[0056] Step S5: Spatiotemporal heterogeneous graph dataset generation: The dynamic adjacency matrix and multi-dimensional node features are integrated to form a spatiotemporal graph sequence data architecture arranged in chronological order. This spatiotemporal graph sequence data architecture enables graph neural networks to extract spatiotemporal evolution patterns through a sliding window mechanism and detect topological transitions caused by sudden changes in wind conditions, so as to quantify the wake delay propagation process and support multi-scale feature learning.

[0057] join Figure 4 This step integrates the scattered, time-separated graph structures into a chronologically ordered graph data architecture. Thus, the graph data at each time step includes the node features, adjacency matrix, and global context information (such as wind direction and speed) for that moment, forming a standard spatiotemporal graph dataset. The generated dataset can be directly used for sliding window sampling to construct samples for training neural networks. The core implementation is as follows: (1) Define the spatiotemporal graph sequence Step S51: Combine the multidimensional node feature matrix of each time step with the time-delay corrected dynamic adjacency matrix of the corresponding time step to construct a graph sequence on discrete time steps, the expression of which is as follows: (19) in, Represents a graph sequence; Represents the grid at time step t; Indicates the first time step; Indicates the second time step; This represents the time step at time T.

[0058] Furthermore, the graph structure for each time step is as follows: (20) in, This represents a fixed set of nodes (a set of wind turbines). Represents a time-varying edge set (defined by a dynamic adjacency matrix); It represents the node feature matrix, which includes statistical features such as wind speed and pitch angle; This represents the wake adjacency matrix corrected by MIC.

[0059] (2) Handling topology transition events Step S52: Using a sliding window mechanism, the graph sequence of multiple consecutive time steps is constructed into samples for training the graph neural network, and the rate of change of the dynamic adjacency matrix within each sliding window is calculated. The sampling function is defined using a sliding window, which is used to construct samples for training the graph neural network: (twenty one) in, This represents the set of data samples captured by the k-th window, containing graph-structured data for L consecutive time steps; This represents the start timestamp of the k-th window; Represents timestamp The corresponding graph structure data; Represents timestamp The corresponding graph structure data; Represents timestamp The corresponding graph structure data.

[0060] Furthermore, to account for wind fluctuations, it is necessary to quantify the topological evolution within the window, i.e., calculate the rate of change of the adjacency matrix within the window: (twenty two) in, This indicates the window for marking topological transition events; L represents the window length, i.e., the number of time steps. This represents the (i+1)th dynamic wake matrix within the window that incorporates time delay effects; This represents the i-th dynamic wake matrix within the window that incorporates time delay effects.

[0061] Furthermore, based on the rate of change of the dynamic adjacency matrix, topological transition events caused by sudden changes in wind conditions are marked.

[0062] (3) Output dataset structure Step S53: Based on the samples of the sliding window and their corresponding topological transition markers, construct a standardized spatiotemporal graph sequence dataset that supports multi-scale time window sampling, thereby forming a graph sequence data architecture that is arranged in chronological order and can characterize spatiotemporal evolution patterns.

[0063] Specifically, during model training, a multi-level time-scale learning architecture can be constructed to achieve more transparent wind power prediction. For example, for samples... Three timescale windows can be constructed, namely short-term window, medium-term window and long-term window. Therefore, the samples used by graph neural networks can be expanded to: (twenty three) Where D represents a multi-scale graph structure data sample; Indicates a short-term window; Indicates the intermediate window; Indicates a long-term window.

[0064] It should be noted that each sample includes a node feature sequence. Adjacency matrix sequence and topological transition markers The label represents the wind power cluster power value for the next time step. Therefore, the graph sequence data architecture generated in step S5 serves as a bridge connecting front-end data processing and back-end graph neural network training. It integrates temporal, spatial, and attribute information in a structured manner, providing standardized input for subsequent deep learning models and forming a key data foundation for learning the spatiotemporal patterns of wind farms.

[0065] Step S6, Model Construction and Prediction: Input the image data architecture into the multi-scale wake perception map spatiotemporal prediction model and output the predicted power values ​​of multiple wind turbines; wherein, the multi-scale wake perception map spatiotemporal prediction model includes an Inception module, a wake perception map attention (GAT) network and a bidirectional long short-term memory network (forward LSTM and backward LSTM, i.e., BiLSTM).

[0066] The model first uses a depthwise separable convolutional Inception module to extract multi-scale transient features of node power sequences. Second, based on a dynamic adjacency matrix constructed from real-time wind direction, a mask-constrained graph attention network is used to calculate spatial dependency coefficients to aggregate neighbor features. Third, a bidirectional LSTM is used to model the forward meteorological-power causal chain and the reverse wake lag effect, respectively. Finally, the bidirectional state is fused and input into a fully connected graph decoder to synchronously output multi-turbine power predictions. The main working mechanism of this model is as follows: Step S61: Input the multi-dimensional node features of each time step in the graph sequence data architecture into the Inception module to extract the multi-scale transient features of the power sequence of each wind turbine; See Figure 5 The Inception module is a variant of a convolutional neural network (CNN) that captures multi-scale information by using multiple convolutional kernels of different sizes in parallel within a single module. The core idea of ​​the Inception module is to increase the diversity of the network's receptive fields while reducing computational cost. First, its multi-branch convolutions are computed: (twenty four) in, Represents the power timing characteristics of a node; This represents the learnable weight matrix and bias terms; Indicates the bias term. Indicates the activation function; This indicates a multi-scale branch of the input feature map.

[0067] Then calculate its depthwise separable convolution and pointwise convolution; (25) in, This represents depthwise separable convolution; This represents pointwise convolution; This indicates a splicing operation along the channel dimension; This indicates a convolutional branch with a kernel size of 3×3; This indicates a convolutional branch with a kernel size of 5×5; This indicates a convolutional branch with a kernel size of 7×7; Z represents the input feature map. This represents the output feature map of a depthwise convolution; This represents multi-scale transient features, i.e., the final output feature map.

[0068] Since the actual power is affected by multiple scale features, this invention learns weight parameters through pointwise convolution and fuses these features into multi-scale transient features of the output node power sequence. This provides more comprehensive input information for subsequent time series modeling.

[0069] Step S62: In the wake perception graph attention network, the multi-scale transient features of the power sequence and the dynamic adjacency matrix are transformed to obtain the attention mask of physical constraints, which is the output that integrates the multi-scale transient features output by the Inception module. Obtaining the multi-scale transient characteristics of nodes The main steps of using graph attention mechanisms to enhance key features include: (1) Query-key-value vector calculation: (26) in, Represents the query weight matrix; Represents the key weight matrix; Q represents the value weight matrix; K represents the query value; V represents the key value; and V represents the value vector. Representing the multi-scale temporal features of graph nodes (2) Physically constrained attention: (27) in, Indicates attention score; Indicates the reduction factor; The classification function is represented by the mask matrix, which is defined by the dynamic adjacency matrix, and its expression is as follows: (28) in, This represents the mask matrix between wind turbine i and wind turbine j; A matrix indicating the positional relationship of the wind turbines.

[0070] (3) Information aggregation: (29) in, This represents mapping the original state to a transferable energy feature; This indicates the intensity of the wake effect of fan i on fan j; The set of neighbor matrices representing a node; This represents the output of the scale transient characteristics.

[0071] Step S63: Input the attention mask of physical constraints into the bidirectional long short-term memory network BiLSTM to obtain the fused spatiotemporal features; join Figure 6 and Figure 7 BiLSTM is an LSTM unit, widely used for processing temporal information, based on scale transient features. The information update of LSTM cells mainly includes the following formulas: Input Gate: (30) in, Indicates the input gate; The weight matrix of the input gate is represented; the hidden state of the previous time step is represented. This represents the input gate bias term.

[0072] Forgotten Gate: (31) in, Indicates the Gate of Oblivion; The weight matrix representing the forget gate; This represents the forget gate bias term.

[0073] Output gate: (32) in, Indicates the output; This represents the weight matrix of the output gate; This indicates the output gate bias term.

[0074] Cell state: (33) in, Indicates cell state; A weight matrix representing cell states; Indicates cell state bias; This indicates the cell state at the previous moment; Hidden state: (34) in, Indicates a hidden state.

[0075] That is, the hidden state from the previous time step is updated to the hidden state at the current time step using the above LSTM unit formula: (35) in, This indicates the current hidden state.

[0076] BiLSTM typically consists of two independent LSTM layers: one processes the sequence from front to back (forward LSTM) and the other processes the sequence from back to front (backward LSTM). The hidden state at each time step t... It is the hidden state of the forward LSTM Hidden state of reverse LSTM The splicing or combination of elements. BiLSTM's two-stream timing modeling can be represented as: (36) in, This represents the element-wise sum of the forward and backward output components; This represents the hidden state of a forward LSTM; Indicates the hidden state of the reverse LSTM; This represents the positive input to the graph features; Represents the previous forward LSTM hidden state; represents the reverse input of the graph features; This represents the hidden state of the reverse LSTM at a future time.

[0077] See Figure 8 Finally, a fully connected graph decoder is used to decode the fused spatiotemporal features and synchronously output the power prediction values ​​of all wind turbines at future moments.

[0078] In summary, the advantages of the Inception-GAT-BiLSTM hybrid model proposed in this invention are as follows: (1) Strong spatiotemporal decoupling and physical interpretability The extraction of temporal features, modeling of spatial relationships, and modeling of temporal dependencies are processed in stages to avoid information confusion caused by mixed modeling. Inception captures the multi-scale time patterns (short-term fluctuations, medium-term trends, and long-term cycles) unique to wind farms. GAT dynamically learns the physical interactions between wind turbines (such as wind direction blocking and wake effects). Attention weights have real-time interpretability. BiLSTM integrates temporal information from historical and future states to solve the problem of time-delay coupling in the physical system.

[0079] (2) Excellent multi-scale spatiotemporal feature fusion capability The Inception module uses multi-scale convolutional kernels to effectively extract local features of power sequences at different time scales from different nodes;

[0080] The GAT module adaptively aggregates the features of key neighbor nodes using an attention mechanism and models dynamic spatial dependencies based on the constructed dynamic tail effect adjacency matrix, enabling the organic integration of data-driven features and physical constraints at the feature level.

[0081] (3) Computational efficiency and parameter optimization Inception employs depthwise separable convolutions to significantly reduce parameters, GAT greatly reduces parameters (compared to fully connected systems) through a shared attention mechanism, and BiLSTM optimizes its structure using a parameter reuse mechanism. The overall architecture offers significant advantages in terms of model parameter count and inference speed, meeting the demands of real-time prediction.

[0082] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A multi-modal offshore wind power ultra-short-term prediction method, characterized in that, The method comprises the following steps: acquiring wind turbine data of an original offshore wind farm cluster, and performing abnormality detection and repair processing on the wind turbine data of the original offshore wind farm cluster to generate a corrected wind power cluster data set; based on the corrected wind power cluster data set, using a dynamic time window mechanism to extract the operation data of each wind turbine within the window, and aggregating the operation data of all wind turbines into multi-dimensional node features; generating a static geographical similarity matrix based on wind turbine geographical coordinates and a basic wake effect matrix based on real-time wind direction data; correcting the basic wake effect matrix to obtain a corrected wake effect matrix, and adaptively weighting and fusing the static geographical similarity matrix and the corrected wake effect matrix to generate a dynamic adjacency matrix; integrating the dynamic adjacency matrix and the multi-dimensional node features to form a time-space graph sequence data architecture arranged in time sequence; inputting the time-space graph sequence data architecture into a pre-constructed multi-scale wake-aware graph space-time prediction model to output multi-wind turbine power prediction values.

2. The multi-modal offshore wind power ultra-short-term prediction method according to claim 1, characterized in that, The method of generating a corrected wind power cluster data set comprises: detecting abnormal power values of the wind turbine data of the original offshore wind farm cluster using an isolation forest algorithm combined with physical rules in the wind power field, and marking the abnormal power values; constructing a light gradient boosting machine model, and training the light gradient boosting machine model using data that is not marked as abnormal power values; inputting non-power features corresponding to the marked abnormal power values into the trained light gradient boosting machine model to obtain abnormal point power prediction values; replacing the power values of the abnormal points with the abnormal point power prediction values, and performing physical upper limit constraint on the abnormal point power prediction values to obtain a corrected wind power cluster data set.

3. The multi-modal offshore wind power ultra-short term prediction method according to claim 1, characterized in that, The method of aggregating the operation data of all wind turbines into multi-dimensional node features comprises: sliding a fixed length time window W forward from the current time t as the end point; calculating the Pearson correlation coefficient of each operating variable of each wind turbine and the wind turbine power p within the time window W; comparing the absolute values of all Pearson correlation coefficients with a preset threshold value and screening out variables whose absolute values are greater than the preset threshold value ​ calculating the mean, standard deviation and latest value of each variable within the time window W for each wind turbine respectively; concatenating the mean, standard deviation and latest value of all variables of each wind turbine in sequence to form a feature vector of each wind turbine; stacking the feature vectors of all wind turbines at the current time to form multi-dimensional node features.

4. The multi-modal offshore wind power ultra-short-term prediction method according to claim 3, characterized in that, The expression of the Pearson correlation coefficient of each operating variable of each wind turbine and the wind turbine power p is: ； wherein, represents the Pearson correlation coefficient of the input characteristic variable and the fan power P; X represents the input characteristic variable; P represents the fan output power; represents the standard deviation of the variable; represents the standard deviation of the fan power; represents the wind speed; represents the pitch angle; represents the rotational speed; represents the covariance of the input characteristic variable and the fan power P.

5. The multi-modal offshore wind power ultra-short term prediction method according to claim 1, characterized in that, The method of generating a static geographical similarity matrix based on wind turbine geographical coordinates and a basic wake effect matrix based on real-time wind direction data comprises: calculating the static geographical similarity matrix between wind turbines using a Gaussian kernel function according to the latitude and longitude coordinates of each wind turbine; determining the upstream and downstream relationship between wind turbines according to real-time wind direction data; wherein, for wind turbine pairs with upstream and downstream relationship, the wake influence intensity is calculated based on a cosine attenuation model, and a dynamic wake effect matrix is generated; forming a basic wake effect matrix according to the static geographical similarity matrix and the dynamic wake effect matrix.

6. The multi-modal offshore wind power ultra-short-term prediction method according to claim 5, characterized in that, The expression of the basic wake effect matrix is: ； wherein, represents the base wake effect matrix between wind turbine i and wind turbine j; represents the azimuth angle of wind turbine j relative to wind turbine i; represents the latitude and longitude coordinates of wind turbine i; represents the latitude and longitude coordinates of wind turbine j; represents the wake decay coefficient; U represents that wind turbine i has a wake effect on wind turbine j.

7. The multi-modal offshore wind power ultra-short term prediction method according to claim 1, characterized in that, The method of generating a dynamic adjacency matrix comprises: Setting a sliding time window Within the sliding time window, a time-lagged maximum mutual information coefficient of power between the two fans is calculated. The time-delay maximum mutual information coefficient is used to correct the basic wake effect matrix of the current time step according to the nonlinear correlation of the power data between wind turbines under time delay, and a corrected dynamic wake effect matrix is obtained. The weight of the static geographical similarity matrix is calculated by taking the wind direction change rate as a weight function and the weight of the dynamic wake effect matrix ; The static geographical similarity matrix and the corrected dynamic wake effect matrix are weighted and fused according to their corresponding weights to generate a dynamic adjacency matrix.

8. The multi-modal offshore wind power ultra-short-term prediction method according to claim 7, characterized in that, a weight of the static geographical similarity matrix a weight of the dynamic wake effect matrix the computational expression of which is ； wherein, denotes a weight of the static geographical similarity matrix; denotes a weight of the dynamic wake effect matrix; denotes a wind direction change rate; denotes a sensitivity function.

9. The multi-modal offshore wind power ultra-short term prediction method according to claim 1, characterized in that, The time-sequentially arranged space-time graph sequence data architecture includes: The multi-dimensional node feature matrix of each time step is combined with the dynamic adjacency matrix of the corresponding time step after time delay correction to construct a graph sequence on a discrete time step; Through a sliding window mechanism, the graph sequence of a plurality of continuous time steps is constructed into a sample for graph neural network training, and the dynamic adjacency matrix change rate in each sliding window is calculated; According to the dynamic adjacency matrix change rate, a topological transition event caused by a sudden change in wind conditions is marked; According to the sample of the sliding window and its corresponding topological transition mark, a standardized space-time graph sequence dataset supporting multi-scale time window sampling is constructed, thereby forming a graph sequence data architecture arranged in time sequence and capable of representing space-time evolution patterns.

10. The multi-modal offshore wind power ultra-short term prediction method according to claim 1, characterized in that, The multi-scale wake-aware graph space-time prediction model includes a multi-scale convolution module, a wake-aware graph attention network, and a bidirectional long short-term memory network; The multi-dimensional node features of each time step in the space-time graph sequence data architecture are input into the multi-scale convolution module to extract multi-scale transient features of the power sequence of each wind turbine; Through the wake-aware graph attention network, the multi-scale transient features of the power sequence and the dynamic adjacency matrix are transformed to obtain a physically constrained attention mask; The physically constrained attention mask is input into the bidirectional long short-term memory network to obtain fused space-time features; The fused space-time features are decoded by using a fully connected graph decoder to synchronously output the power prediction values of all wind turbines at future time points.

Images