Wind power plant wind speed prediction method and device based on deep learning, equipment and medium

Abstract

The invention discloses a wind power plant wind speed prediction method and device based on deep learning, equipment and a medium, and relates to the technical field of intelligent operation and maintenance of wind power plants, and the method comprises the steps: obtaining meteorological data and fan operation data in a preset time window, and obtaining fan static data; preprocessing the data, then constructing a three-dimensional input tensor, inputting the three-dimensional input tensor into a wind speed prediction model for wind speed prediction, and obtaining a wind speed prediction result of a prediction time period; a coding module of the wind speed prediction model obtains an initial node state matrix based on three-dimensional input tensor coding; the dynamic physical construction module constructs and obtains a dynamic physical information graph based on the three-dimensional input tensor and fan static data; and the iterative graph information transmission module performs iterative calculation to obtain a node state matrix at each moment in a prediction time period, and the decoding module performs decoding to obtain a wind speed prediction result. The high-precision fan-level ultra-short-term wind speed prediction provided by the embodiment of the invention can provide key technical support for high-order intelligent cooperative control of a wind power plant.

Description

Technical Field

[0001] This application relates to the field of intelligent operation and maintenance technology for wind farms, and in particular to a wind speed prediction method, device, equipment and medium for wind farms based on deep learning. Background Technology

[0002] Wind power, as a key renewable energy technology, occupies an increasingly important position in the global energy system. However, the inherent intermittency and volatility of wind energy pose challenges to the stable operation of power systems. High-precision ultra-short-term wind speed forecasting technology is crucial for achieving smooth control and optimized dispatch of wind power.

[0003] Wind speed prediction methods in related technologies can be broadly categorized into three types: physical models, statistical models, and artificial intelligence models. Physical models, such as those based on Numerical Weather Prediction (NWP), predict large-scale meteorological changes by solving atmospheric physical equations. These methods have long-term predictive capabilities, but require complex data fusion and correction models to handle data errors from different NWP sources. Relying solely on raw NWP data is insufficient to meet the demands of high-precision, minute-level ultra-short-term predictions within wind farms (hundred-meter scale). Statistical models, such as the Autoregressive Integral Moving Average (ARIMA) model, are simple and easy to implement, but typically require assumptions of linearity or stationarity in time-series data, making it difficult to effectively capture the strong nonlinear and non-stationary characteristics of wind speed sequences. Furthermore, artificial intelligence models can use graph neural networks (GNNs) to capture spatial dependencies between wind turbines; however, existing artificial intelligence models still suffer from the following problems: Question 1: The static nature of graph structures and the superficial integration of physical mechanisms. The construction of graph networks is based on the fixed geographical location information (such as latitude and longitude differences) between wind turbines to form a static graph topology. Although this static graph introduces the concept of spatial proximity, it fails to dynamically couple with the core physical mechanism of the interaction between wind turbines—the wake effect. The propagation path and influence range of the wake change drastically with the real-time wind direction. Static graphs cannot capture this directional and time-varying physical causal relationship driven by wind direction.

[0004] Problem 2: Lack of explicit modeling of physical laws. Artificial intelligence models in related technologies rely on the model itself to inductively learn the correlations between wind turbines from massive amounts of data. However, the model structure itself does not explicitly embed prior aerodynamic knowledge about how wakes form, deflect, and attenuate. This approach not only requires a huge amount of data, but the model may only learn statistical correlations rather than true physical causality. When wind farms encounter new operating conditions not covered in the training data, the model's predictive reliability and generalization ability will face challenges due to the lack of strong constraints from physical laws.

[0005] In summary, the relevant technologies have limitations in the accuracy of ultra-short-term collaborative prediction of multiple wind turbine units, either due to insufficient spatiotemporal resolution, overly simplified model assumptions, or failure to dynamically and explicitly integrate spatial physical mechanisms into the model architecture. Summary of the Invention

[0006] This application provides a method, apparatus, equipment, and medium for predicting wind speed in wind farms based on deep learning, in order to overcome the shortcomings of the aforementioned related technologies. The technical solution is as follows: Firstly, this application provides a deep learning-based method for predicting wind speed in wind farms, including: Obtain meteorological data and wind turbine operation data within a preset time window prior to the current moment, and obtain static data for each wind turbine; The wind turbine operation data and the meteorological data are preprocessed, and a three-dimensional input tensor is constructed based on the preprocessed data; The three-dimensional input tensor and the static data of the wind turbine are input into the trained wind speed prediction model to predict the wind speed and obtain the wind speed prediction results for the prediction period. The wind speed prediction model includes an encoding module, a dynamic physics construction module, an iterative graph information transmission module, and a decoding module. The encoding module is used to obtain the initial node state matrix at the current time based on the three-dimensional input tensor encoding; The dynamic physics construction module is used to construct dynamic physics information graphs at various times based on the three-dimensional input tensor and the wind turbine static data. The iterative graph information transmission module is used to perform iterative calculations based on the initial node state matrix and the dynamic physical information graph to obtain the node state matrix at each time point within the prediction period. The decoding module is used to decode based on the node state matrix at each time point to obtain the wind speed prediction result for the prediction period.

[0007] In one alternative embodiment of the first aspect, the preprocessing of the wind turbine operating data and the meteorological data includes: The time axes of the wind turbine operation data and the meteorological data are aligned to obtain an aligned data sequence, and the aligned data sequence is resampled to a preset time resolution. Outlier detection and removal are performed on the resampled data sequence, and missing values ​​are filled by interpolation calculation to obtain the cleaned data sequence. Each variable in the cleaned data sequence is converted to the preset input format of the wind speed prediction model to obtain preprocessed time series data; The construction of the three-dimensional input tensor based on the preprocessed data includes: Based on the preprocessed data, node features related to a single wind turbine and global features related to the entire wind farm are extracted. The node features of each wind turbine are sequentially concatenated with the global features to construct the three-dimensional input tensor.

[0008] In one alternative embodiment of the first aspect, the training process of the wind speed prediction model includes: Historical data for a historical period is acquired, including historical meteorological data and historical wind turbine operation data. Based on the historical data, the preprocessing steps are performed to obtain preprocessed time series data. The data sequence of the first time period is extracted from the preprocessed time series data as the sample input, and the data sequence of the second time period is extracted from the preprocessed time series data to obtain the wind speed sequence of the second time period. The wind speed sequence is used as the sample label. The step of constructing a three-dimensional input tensor is performed on the sample input, and the corresponding three-dimensional input tensor is input into the wind speed prediction model to obtain the wind speed prediction result for the second time period output by the wind speed prediction model. Based on the wind speed prediction results and the corresponding sample labels, a loss function is constructed. The value of the loss function is then propagated to adjust the parameters of the wind speed prediction model until the model converges, thus obtaining a trained wind speed prediction model.

[0009] In one alternative embodiment of the first aspect, obtaining the initial node state matrix at the current time based on the three-dimensional input tensor encoding includes: The three-dimensional input tensor is decomposed into node feature sequences and global feature sequences; The global feature sequence is processed by time pooling to obtain the global physical state vector; The global physical state vector is input into the gating network to generate a gating vector; Based on the gate vector, the node feature sequence is subjected to gate modulation operation. The node feature sequence after the gate modulation operation is processed by the Transformer encoder to capture complex temporal dependencies. The output sequence of the last time step is extracted to construct the initial node state matrix.

[0010] In one alternative embodiment of the first aspect, the dynamic physical information graph includes an adjacency matrix and an edge feature matrix, and the process of constructing the adjacency matrix includes: Based on the analysis of the three-dimensional input tensor, the real-time parameters at the corresponding time are obtained, including global wind direction, global turbulence intensity, inflow wind speed and yaw angle of each wind turbine; Based on the static data of the wind turbines, the position coordinates of each wind turbine and the relative positional relationship between each pair of wind turbines are determined. Based on the relative positional relationship and the global wind direction at the corresponding time, the upstream wind turbine corresponding to each downstream wind turbine is determined. Calculate the inflow velocity loss caused by the influence of each upstream fan on each downstream fan, and construct a loss matrix based on the inflow velocity loss of each fan pair consisting of each downstream fan and a corresponding upstream fan. The values ​​in the loss matrix are normalized to obtain the adjacency weights of the degree of influence of each upstream wind turbine on the downstream wind turbine, thus obtaining the adjacency weight matrix. The process of constructing the edge feature matrix includes: The directed edges corresponding to each wind turbine pair are determined, the multidimensional physical features of each directed edge are extracted, and the edge feature vector is constructed. The edge feature matrix is ​​constructed based on the edge feature vector of each directed edge.

[0011] In one alternative embodiment of the first aspect, the construction yields a dynamic physical information graph at each time step, comprising: Based on the meteorological data, the forecast wind condition parameters for each moment within the forecast period are derived. Based on the forecast wind condition parameters for each moment, the adjacency matrix and the edge feature matrix for the corresponding moment are updated respectively, thus obtaining the dynamic physical information map for each moment.

[0012] In one alternative embodiment of the first aspect, the iterative calculation based on the initial node state matrix and the dynamic physical information graph to obtain the node state matrix at each time point within the prediction period includes: The node embedding vector of each wind turbine is extracted based on the initial node state matrix, and the scaled dot product attention of each wind turbine pair is calculated based on the node embedding vector of each wind turbine. Extract the adjacency weights and edge feature vectors of each wind turbine pair in the dynamic physical information graph at the corresponding time. The attention score for each wind turbine pair is calculated based on scaled dot product attention, adjacency weights, and edge feature vectors. Based on all the wind turbine pairs corresponding to the downstream wind turbines, the adjacency weight, edge feature vector and node embedding vector of the upstream wind turbine are weighted and fused according to the attention score. The node embedding vector of the downstream wind turbine at the next time step is updated based on the weighted fusion result. The node state matrix at the corresponding time step is obtained based on the node embedding vector of each wind turbine at the next time step. Based on the updated node state matrix, proceed to the step of extracting the node embedding vector for each wind turbine, until the node state matrix for each time period within the prediction period is updated.

[0013] Secondly, this application also provides a wind speed prediction device for wind farms based on deep learning, comprising: The data acquisition module is used to acquire meteorological data and wind turbine operation data within a preset time window before the current moment, and also to acquire static data of each wind turbine. The data processing module is used to preprocess the wind turbine operation data and the meteorological data, and construct a three-dimensional input tensor based on the preprocessed data; The prediction module is used to input the three-dimensional input tensor and the static data of the wind turbine into the trained wind speed prediction model to perform wind speed prediction and obtain the wind speed prediction result for the prediction period. The wind speed prediction model includes an encoding module, a dynamic physics construction module, an iterative graph information transmission module, and a decoding module. The encoding module is used to obtain the initial node state matrix at the current time based on the three-dimensional input tensor encoding; The dynamic physics construction module is used to construct dynamic physics information graphs at various times based on the three-dimensional input tensor and the wind turbine static data. The iterative graph information transmission module is used to perform iterative calculations based on the initial node state matrix and the dynamic physical information graph to obtain the node state matrix at each time point within the prediction period. The decoding module is used to decode based on the node state matrix at each time point to obtain the wind speed prediction result for the prediction period.

[0014] Thirdly, this application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the method provided by the first aspect of this application or any implementation thereof.

[0015] Fourthly, this application also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method provided by the first aspect of this application or any implementation thereof.

[0016] The beneficial effects of the technical solution provided in this application include at least the following: (1) This application overcomes the fundamental defect of static graph models based on geographical distance that ignore or simplify spatial physical coupling by dynamically constructing dynamic physical information graphs that reflect aerodynamic causal relationships. Based on the explicit integration of physical wake models, this application provides strong physical prior knowledge constraints for deep wind speed prediction models, enabling wind speed prediction models to more accurately capture the spatiotemporal correlation between wind turbines driven by complex phenomena such as wake propagation, deflection and merging.

[0017] (2) This application enables the model to capture the dynamic changes of the wake propagation path in real time by recalculating and constructing the dynamic physical information map based on the real-time wind conditions at each time step, thereby accurately modeling the physical interaction under different operating conditions.

[0018] (3) The high-precision, high-fidelity wind turbine-level ultra-short-term wind speed prediction provided in the embodiments of this application can provide key feedforward decision inputs for the intelligent control body of each wind turbine. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is one of the flowcharts illustrating a deep learning-based wind speed prediction method for wind farms provided in this application embodiment; Figure 2 This is the second flowchart of a wind farm wind speed prediction method based on deep learning provided in the embodiments of this application; Figure 3 This is one of the schematic diagrams showing the relative positional relationship of the fans provided in the embodiments of this application; Figure 4 This is a second schematic diagram illustrating the relative positional relationship of the fans provided in the embodiments of this application; Figure 5 This is a schematic diagram of the simulation experiment effect of the model provided in the embodiment of this application; Figure 6 This is a schematic diagram of the structure of a wind farm wind speed prediction device based on deep learning provided in an embodiment of this application; Figure 7 This is a schematic diagram of the structure of the electronic device provided in the embodiments of this application. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0022] The terms "comprising" and "having," and any variations thereof, in the specification, claims, and accompanying drawings of this application are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or modules is not limited to the steps or modules listed, but may optionally include steps or modules not listed, or may optionally include other steps or modules inherent to such process, method, product, or apparatus.

[0023] It should be noted that the terms "first" and "second" used in this application are merely to distinguish similar objects and do not represent a specific ordering of the objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in an order other than those described or illustrated herein.

[0024] The present application will now be described in detail with reference to specific embodiments.

[0025] Next, combine Figure 1 This paper introduces a deep learning-based wind speed prediction method for wind farms, as provided in an embodiment of this application. For details, please refer to... Figure 1 , Figure 1 This diagram illustrates a flowchart of a deep learning-based wind speed prediction method for wind farms, as provided in an embodiment of this application. Figure 1 As shown, the method includes the following steps: S101: Obtain meteorological data and wind turbine operation data within a preset time window prior to the current moment, and obtain static data of each wind turbine; S102, preprocess the wind turbine operation data and the meteorological data, and construct a three-dimensional input tensor based on the preprocessed data; S103, input the three-dimensional input tensor and the static data of the wind turbine into the trained wind speed prediction model to perform wind speed prediction and obtain the wind speed prediction result for the prediction period.

[0026] Specifically, the wind turbine operation data in S101 is Supervisory Control And Data Acquisition (SCADA) data, which includes data such as active power, rotor speed, pitch angle, yaw angle, nacelle wind speed and nacelle wind direction for each wind turbine in the wind farm within a preset time window before the current moment.

[0027] Specifically, the global meteorological data in S101 can be recorded by meteorological measurement equipment (such as wind measurement towers or lidar) deployed in the wind farm, including high-frequency data such as ambient wind speed, wind direction, and atmospheric turbulence intensity.

[0028] It should be noted that when predicting the inflow wind speed of each wind turbine in the wind farm, it is necessary to collect meteorological data and wind turbine operation data for a preset time window before the current moment. The length of the preset time window can be set according to actual needs, and the current moment is the last time step of the preset time window.

[0029] Specifically, the static data of wind turbines in S101 includes the precise geographical coordinates of all wind turbines in the wind farm (e.g., coordinates in the RGF93 coordinate system), hub height (80 meters), and rotor diameter (82 meters). This information is readily available from publicly available materials such as technical manuals.

[0030] In some embodiments, S102 may preprocess the wind turbine operating data and the meteorological data, including: Time alignment and resampling: The time axes of the wind turbine operation data and the meteorological data are aligned to obtain an aligned data sequence, and the aligned data sequence is resampled to a preset time resolution; Data cleaning: Outlier detection and removal are performed on the resampled data sequence, and missing values ​​are filled by interpolation calculation to obtain the cleaned data sequence; Feature variable derivation: Each variable in the cleaned data sequence is transformed to the preset input format of the wind speed prediction model. That is, the original variables are transformed and derived to adapt to the model input, resulting in preprocessed time series data. For example, wind direction (angle value) is decomposed into sine and cosine components to solve its periodic discontinuity problem near 0 / 360 degrees.

[0031] Furthermore, in S102, a three-dimensional input tensor can be constructed based on the preprocessed data, including the following steps: Based on the preprocessed data, node features related to a single wind turbine (such as wind turbine power and wind turbine inflow velocity) and global features related to the entire wind farm (such as global wind direction and turbulence intensity) are extracted. Specifically, in order to construct the input tensor suitable for subsequent graph neural network models, the global features are copied and expanded along the spatial dimension representing the wind turbine, and then sequentially concatenated with the node-level features of each wind turbine.

[0032] To construct the input tensor suitable for the subsequent wind speed prediction model, the global features are replicated and expanded along the spatial dimension representing the wind turbine, and then sequentially concatenated with the node features of each wind turbine. In this way, a feature sequence with consistent dimensions is constructed for each wind turbine, which integrates its own independent state and includes a unified environmental background, ultimately forming a three-dimensional input tensor.

[0033] In some embodiments, the wind speed prediction model in S103 includes an encoding module, a dynamic physics construction module, an iterative graph information transmission module, and a decoding module; The encoding module is used to obtain the initial node state matrix at the current time based on the three-dimensional input tensor encoding. The dynamic physics construction module is used to construct dynamic physics information graphs at various times based on the three-dimensional input tensor and the wind turbine static data. The iterative graph information transmission module is used to perform iterative calculations based on the initial node state matrix and the dynamic physical information graph to obtain the node state matrix at each time point within the prediction period. The decoding module is used to decode based on the node state matrix at each time point to obtain the wind speed prediction result for the prediction period.

[0034] In some embodiments, such as Figure 2 As shown, the process of wind speed prediction based on the wind speed prediction model includes the following steps: S201, Based on the three-dimensional input tensor encoding, the initial node state matrix at the current moment is obtained; S202, Based on the three-dimensional input tensor and the static data of the wind turbine, a dynamic physical information diagram at each moment is constructed; S203, based on the initial node state matrix and the dynamic physical information graph, iterative calculation is performed to obtain the node state matrix at each time point within the prediction period; S204, based on the node state matrix at each time point, decode to obtain the wind speed prediction result for the prediction period.

[0035] Specifically, in S201, the initial node state matrix at the current time is obtained based on the three-dimensional input tensor encoding, which specifically includes: The three-dimensional input tensor is decomposed into node feature sequences and global feature sequences; The global feature sequence is processed by time pooling to obtain the global physical state vector; The global physical state vector is input into a gate network to generate a gate vector g. Gating modulation of the node feature sequence based on the gating vector g is expressed as follows: ; in, Represents the node feature sequence. Let g represent the node feature sequence after the gated modulation operation, and g be the gate vector. For Hadamah accumulation; The node feature sequence after the gated modulation operation is processed by a Transformer encoder to capture complex temporal dependencies, and the output sequence of the last time step is extracted to construct the initial node state matrix. .

[0036] Specifically, the initial node state matrix The shape is Where N is the number of wind turbines, For the hidden layer dimension.

[0037] Specifically, in S202, the dynamic physical information graph includes an adjacency matrix and an edge feature matrix, and the process of constructing the adjacency matrix includes: S2021, Based on the analysis of the three-dimensional input tensor, the real-time parameters at the corresponding moment are obtained, including the global wind direction. Global turbulence intensity TI, inflow velocity of each fan and yaw angle .

[0038] Where i represents the serial number of the wind turbine.

[0039] S2022, Based on the static data of the wind turbines, determine the position coordinates of each wind turbine and the relative positional relationship between every two wind turbines, and based on the relative positional relationship and the global wind direction at the corresponding time, determine the upstream wind turbine corresponding to each downstream wind turbine.

[0040] For example, such as Figure 3 As shown, Figure 3 This example illustrates the distribution and relative positions of multiple wind turbines in a wind farm, with the global wind direction being... Figure 3 In the case shown, the upstream and downstream relationships between wind turbines can be determined along the wind direction. For example, wind turbine 1 is the upstream wind turbine of wind turbines 2 and 3, and wind turbines 2 and 3 are both downstream wind turbines of wind turbine 1. Each wind turbine can affect multiple downstream wind turbines along the wind direction, and each downstream wind turbine will also be affected by multiple upstream wind turbines in the direction of the incoming wind. Each upstream wind turbine and a corresponding downstream wind turbine can be recorded as a wind turbine pair.

[0041] S023 calculates the inflow velocity loss caused by the influence of each upstream wind turbine on each downstream wind turbine, specifically including: Combination Figure 4 As shown, Figure 4 Given a wind turbine, consider the relative positions of the upstream and downstream turbines. Let the upstream turbine be denoted as i, and the downstream turbine as j. Known quantities include the yaw angle of the upstream turbine i, denoted as yaw angle. The yaw angle of the downstream wind turbine j is denoted as... The straight-line distance between upstream wind turbine i and downstream wind turbine j is denoted as . .

[0042] The following parameters can be calculated based on the known parameters, including: the wake width generated by upstream fan i on downstream fan j. The Euclidean distance between upstream wind turbine i and downstream wind turbine j The lateral relative distance between upstream wind turbine i and downstream wind turbine j in the wind direction coordinate system. The wake deflection angle caused by the incoming airflow passing through the upstream fan i , This represents the radial distance between the upstream fan i and the downstream fan i.

[0043] Based on the real-time parameters obtained from S101, in terms of wake expansion calculation, the model will use the wake expansion coefficient of upstream wind turbine i. The model is a linear function related to the global turbulence intensity TI, i.e.: ; in, and For calibration constants, The global turbulence intensity is TI.

[0044] Specifically, the wake width σ increases linearly with the distance x between the wind turbines, that is: ; in, The initial wake width can be represented by the inflow velocity. Thrust coefficient below Calculated.

[0045] Specifically, the upstream wind turbines can be approximated based on the Jiménez model. The initial deflection angle of the generated wake centerline downstream Apply the formula: ; Therefore, the wake width generated by upstream wind turbine i on downstream wind turbine j can be calculated. .

[0046] Furthermore, the maximum loss amplitude on the center line is calculated. Apply the formula: ; Where D is the diameter of the wind turbine rotor, which can be obtained from the static data of the wind turbine.

[0047] The inflow velocity loss caused by upstream fan i to any downstream point j It is described by a Gaussian function centered on the centerline of the deflected wake, using the formula: ; in, This indicates the cumulative lateral offset of the wake centerline at downstream fan j. It can be obtained by integrating a nonlinear trajectory equation that precisely describes the decay of the deflection angle as the wake expands.

[0048] S024, based on the inflow velocity loss of each downstream wind turbine and its corresponding upstream wind turbine, constitutes a loss matrix, specifically representing the velocity loss of all wind turbine pairs. Organize the data into an N×N loss matrix and normalize the values ​​in the loss matrix to obtain the adjacency weights of each downstream wind turbine j in a wind turbine pair, representing the degree of influence of each upstream wind turbine i on each downstream wind turbine. This ensures that the sum of the adjacency weights of all upstream wind turbines received by each downstream wind turbine is 1. The resulting adjacency weight matrix is ​​represented as follows: ; Furthermore, the process of constructing the edge feature matrix includes: S2025, determine the directed edges corresponding to each wind turbine pair ij. Extract each directed edge The multidimensional physical features are extracted and filled into the corresponding directed edges to construct the edge feature vector. The edge feature matrix is ​​constructed based on the edge feature vector of each directed edge. .

[0049] Among them, multidimensional physical features include five-dimensional features: (1) Inflow wind speed loss ; (2) Global wind direction With wind turbine The cosine of the angle between the static geographical lines represents the alignment relationship between the wind direction and the geometric position of the wind turbine. (3) The normalized Euclidean distance between upstream wind turbine i and downstream wind turbine j This characterizes spatial proximity; (4) The straight-line distance between upstream wind turbine i and downstream wind turbine j is a key independent variable in wake evolution; (5) The lateral relative distance between upstream fan i and downstream fan j in the wind direction coordinate system is the basis for calculating wake deflection and radial distance.

[0050] It should be noted that the wind direction and speed of the incoming wind may change over time. Under westerly wind conditions, the two wind turbines that are upstream and downstream of each other may have a completely reversed relationship under easterly wind conditions. Therefore, the dynamic physical information diagram above needs to be updated according to the wind direction.

[0051] In some embodiments, the forecast wind condition parameters (wind speed and wind direction) for each moment within the forecast period can be derived from the meteorological data. Based on the forecast wind condition parameters for each moment, the adjacency matrix and the edge feature matrix for the corresponding moment are updated respectively to obtain the dynamic physical information map for each moment.

[0052] This allows for dynamically updated physical information graphs, resulting in a dynamic graph sequence that provides a solid physical foundation for subsequent computational processes in the model.

[0053] Furthermore, by executing S203 through the iterative graph information transmission module, during the graph information transmission phase, the initial node state matrix of all nodes at the current moment is... Along with the dynamic physical diagram at that moment The input multi-head graph attention network layer (EdgeAttnMessage) is used to aggregate and update spatial information.

[0054] Specifically, it includes: S2031, based on the initial node state matrix The node embedding vector of each wind turbine is extracted, and the scaled dot product attention of each wind turbine pair is calculated based on the node embedding vector of each wind turbine.

[0055] For example, for a pair of wind turbines ij, the node embedding vector of the upstream wind turbine i can be extracted. The node embedding vector of downstream wind turbine j The calculated scaled dot product attention is represented as .

[0056] S2032, Extract the adjacency weight of each wind turbine pair in the dynamic physical information graph at the corresponding time. and edge feature vectors ; S2033 calculates the attention score for each wind turbine pair based on scaled dot product attention, adjacency weights, and edge feature vectors, using the following formula: ; in, It is a learnable scalar parameter. These are learnable vector parameters used to adaptively adjust the importance of physical topology and physical edge features, respectively. This represents the function used to calculate the weighted contribution of physical edge features.

[0057] S2034, based on all wind turbine pairs corresponding to the downstream wind turbines, the adjacency weights, edge feature vectors, and node embedding vectors of the upstream wind turbines for each wind turbine pair are weighted and fused according to the attention scores. The node embedding vector of the downstream wind turbine at the next time step is updated based on the weighted fusion result, using the formula: ; in, This indicates that the node embedding vector of upstream wind turbine i has undergone a linear transformation. The subsequent feature vectors, This indicates normalization.

[0058] S2035, based on the node embedding vector of each wind turbine at the next time step, obtain the node state matrix at the corresponding time step. ; S2036, based on the updated node state matrix, proceed to the step of extracting the node embedding vector of each wind turbine, until the node state matrix of each time point in the prediction period is updated.

[0059] For example, the preset time period has a total of T prediction times, including Let the current time be t=0, and the node state matrix at the current time is the initial node state matrix. Based on the initial node state matrix The node state matrix at the next time step t=0+1 can be calculated by performing steps S2031-S2035 on the dynamic physical information graph at the current time t=0. Based on the node state matrix Execute step S2036, perform iterative loops, based on the node state matrix. By performing steps S2031-S2035 on the dynamic physical information graph at time t=1, the node state matrix for the next time step can be obtained, until the node state matrix for T prediction times within the prediction period is calculated.

[0060] It should be noted that the embodiments of this application are more suitable for ultra-short-term wind speed forecasts. For example, the forecast period can be set to 10 minutes, 20 minutes, or 1 hour, which can be set according to the actual situation. The forecast step size (the interval between each forecast moment) between each forecast moment can be set according to the calculation accuracy requirements.

[0061] Furthermore, step S204 can be executed through the decoding module to decode based on the node state matrix at each time point, thereby obtaining the wind speed prediction result for the prediction period.

[0062] Specifically, it includes: First, the hidden states are aggregated through a pooling layer (such as average pooling) to generate a graph-level representation that captures global information at the current time step. This representation is then fed into a decoder of a gated recurrent unit (GRU), which is responsible for maintaining the temporal continuity and dependencies of the predicted sequence and outputting an updated hidden state.

[0063] Finally, the hidden state is fed into a node-specific Kolmogorov-Arnold Network (NP-KAN) output head for final nonlinear mapping and wind speed prediction. This output head employs an innovative dual-path architecture: one path is a shared KAN backbone network responsible for learning the universal physical evolution laws followed by all wind turbines; the other path consists of N independent, parameter-unshared, personalized linear residual layers, each learning its unique response bias (e.g., biases caused by equipment aging or micro-topography) specifically for a single wind turbine. The outputs of the two paths are weighted and fused through a learnable scalar gating system to generate a predicted inflow wind speed for each wind turbine at time t. N represents the total number of wind turbines, and n represents the ordinal number of each wind turbine.

[0064] In some embodiments, the core objective during the training of the wind speed prediction model is to minimize the difference between the wind speed sequence predicted by the model and the actual wind speed sequence.

[0065] During the training process, historical data for historical time periods are first acquired. The historical data includes historical meteorological data and historical wind turbine operation data. Based on the historical data, the preprocessing steps are performed to obtain preprocessed time series data. The preprocessing process is the same as that in S102, and will not be described again here.

[0066] Furthermore, a data sequence of the first time period is extracted from the preprocessed time series data as a sample input, and a data sequence of the second time period is extracted from the preprocessed time series data to obtain the wind speed sequence of the second time period, which is used as the sample label; wherein, the end time of the first time period and the start time of the second time period are adjacent, and the second time period is after the first time period.

[0067] The sample input is processed by the step of constructing a three-dimensional input tensor as described in S102. The corresponding three-dimensional input tensor is then input into the wind speed prediction model. The wind speed prediction model is then processed by the steps of S201-S204 to obtain the wind speed prediction result for the second time period output by the wind speed prediction model. Based on the wind speed prediction results and corresponding sample labels Construct the loss function: ; Where B is the batch size, H is the prediction step size, N is the number of wind turbines, ||·||² represents the L2 norm (sum of squares of all elements), and Σ represents the summation of errors for all batches, all step sizes, and all wind turbines.

[0068] The calculated loss value is then propagated throughout the computation graph via the backpropagation algorithm to calculate the gradients of all learnable parameters in the model (including all weights and biases in the GateTransformer, Graph Attention Network, and NP-KAN decoder).

[0069] Finally, an advanced optimizer, such as adaptive moment estimation, is employed to update all model parameters based on the calculated gradients. To prevent gradient explosion during training, a gradient pruning strategy can also be introduced. The entire training process is repeated multiple times on all training data, and an early stopping strategy is implemented by monitoring the model's loss performance on the validation set. That is, training is stopped when the loss on the validation set no longer decreases significantly, to prevent overfitting and preserve the optimal model parameters, resulting in a well-trained wind speed prediction model.

[0070] In some specific embodiments, the deep learning-based wind speed prediction method for wind farms provided in this application has the following advantages compared to related technologies: (1) Significantly improves prediction accuracy and achieves deep integration of physical mechanisms and data-driven approaches.

[0071] By dynamically constructing a dynamic physical information graph that reflects aerodynamic causal relationships, the fundamental shortcomings of existing technologies (such as pure time-series models or static graph models based on geographical distance) are overcome, which neglect or simplify spatial physical coupling. The explicit integration of the dynamic physical information graph provides powerful physical prior knowledge constraints for deep learning models, enabling them to more accurately capture the spatiotemporal correlations between wind turbines driven by complex phenomena such as wake propagation, deflection, and merging.

[0072] Comparative experiments based on the publicly available French SMV commercial wind farm dataset demonstrate that the method described in this application achieves comprehensive performance advantages over various advanced benchmark models, such as... Figure 5 As shown, compared with a standard Transformer baseline model that does not consider spatial relationships, the model provided in this application (i.e. Figure 5 The mean absolute error (MAE) of the Gateformer-PhysGNN-NPKAN model was reduced from 0.3810 to 0.3138, a reduction of 17.6%. Even compared with the CEEMDAN-Transformer-BiLSTM baseline model, which employs complex signal decomposition techniques, the model provided in this application still achieved a significant reduction of 8.1% in the mean absolute error (MAE) (0.3138 vs 0.3416). This demonstrates the significant superiority of the technical solution in deeply integrating physical mechanisms and data-driven approaches.

[0073] (2) The model's adaptability and robustness to dynamic operating conditions of wind farms have been enhanced. Static diagrams or independent prediction models of related technologies cannot adapt to the drastic changes in the influence relationships between wind turbines caused by real-time variations in wind direction and speed. This application recalculates and reconstructs a dynamic physical information diagram based on real-time wind parameters at each time step, enabling the model to capture the dynamic changes in the wake propagation path in real time, thereby accurately modeling the physical interactions under different operating conditions.

[0074] Specifically, this was verified through ablation experiments, such as... Figure 5 As shown, if the dynamic physics graph construction module of this invention is removed (i.e., changed from Gateformer-PhysGNN-NPKAN to Gateformer-NPKAN), the model's mean absolute error deteriorates by 3.63% (from 0.3138 to 0.3252), while the root mean square error (RMSE) deteriorates by 4.78% (from 0.4724 to 0.4950). It is noteworthy that this performance degradation is mainly concentrated on the downstream wind turbines affected by the wake, while the performance of wind turbine 1, which is unaffected by the upstream, remains almost unchanged. This demonstrates that the dynamic physics graph mechanism provided in the embodiments of this application is key to maintaining high accuracy and robustness of the model in complex flow fields.

[0075] (3) Improved the physical consistency and interpretability of the prediction results.

[0076] Unlike traditional "black box" deep learning models, the dynamic physical information graph provided in this application is generated by a wake analysis model with clear physical meaning, and its prediction process is closely coupled with physical laws. For example, the weights in the adjacency matrix directly correspond to the physical contribution of the inflow wind speed loss, while the edge features quantify key physical quantities such as relative distance and angle.

[0077] This makes the model's internal decision-making process somewhat "transparent" and physically intuitive. In experiments, the spatial distribution of the prediction error exhibits a reasonable physical gradient: the error peaks at the mid-level turbines, which are most severely affected by wake accumulation, and is minimized at the leading turbines in free flow. This perfectly matches the actual physics of wakes. Furthermore, the graph attention network parameters learned by the model indicate that it has learned to rely heavily on physical priors to guide information transmission. This enhanced interpretability not only improves the model's credibility in engineering applications but also provides more insightful decision-making support for subsequent wind farm operation and maintenance and control strategy optimization.

[0078] (4) It provides key technical support for high-level intelligent collaborative control of wind farms.

[0079] The accuracy and reliability bottlenecks in related technologies for inflow wind speed prediction are major obstacles restricting the realization of advanced applications such as multi-agent collaborative control and farm-level active optimization in wind farms. The high-precision, high-fidelity wind turbine-level ultra-short-term wind speed prediction provided in this application can provide key feedforward decision inputs for the intelligent control body of each wind turbine.

[0080] Accurate prediction based on the model provided in this application's embodiments enables the wind farm's collaborative control system to shift from "passive response" to "active prediction." For example, the system can pre-calculate the optimal collaborative yaw or pitch strategy to cope with wake changes in the coming minutes, thereby maximizing farm-level power generation, reducing turbine fatigue load, and smoothing power fluctuations. Therefore, this application not only improves prediction accuracy but also provides core technical support for unlocking higher levels of intelligent and refined operation and control of wind farms, providing a reliable technical foundation for improving the overall economic benefits and grid friendliness of wind farms.

[0081] The following are apparatus embodiments of this application, which can be used to execute the method embodiments of this application. For details not disclosed in the apparatus embodiments of this application, please refer to the method embodiments of this application.

[0082] Please see below. Figure 6The image below is a schematic diagram of a deep learning-based wind speed prediction device for a wind farm, provided as an exemplary embodiment of this application. The device includes: The data acquisition module is used to acquire meteorological data and wind turbine operation data within a preset time window before the current moment, and also to acquire static data of each wind turbine. The data processing module is used to preprocess the wind turbine operation data and the meteorological data, and construct a three-dimensional input tensor based on the preprocessed data; The prediction module is used to input the three-dimensional input tensor and the static data of the wind turbine into the trained wind speed prediction model to perform wind speed prediction and obtain the wind speed prediction result for the prediction period. The wind speed prediction model includes an encoding module, a dynamic physics construction module, an iterative graph information transmission module, and a decoding module. The encoding module is used to obtain the initial node state matrix at the current time based on the three-dimensional input tensor encoding; The dynamic physics construction module is used to construct dynamic physics information graphs at various times based on the three-dimensional input tensor and the wind turbine static data. The iterative graph information transmission module is used to perform iterative calculations based on the initial node state matrix and the dynamic physical information graph to obtain the node state matrix at each time point within the prediction period. The decoding module is used to decode based on the node state matrix at each time point to obtain the wind speed prediction result for the prediction period.

[0083] It should be noted that the apparatus provided in the above embodiments, when executing a deep learning-based wind farm wind speed prediction method, is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the equipment can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the apparatus and method embodiments provided in the above embodiments belong to the same concept, and their implementation process is detailed in the method embodiments, which will not be repeated here.

[0084] This application also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0085] Please see Figure 7 This is a structural block diagram of an electronic device provided in an embodiment of this application.

[0086] like Figure 7 As shown, the electronic device includes a processor and a memory.

[0087] In this embodiment, the processor is the control center of the computer system, and can be a processor of a physical machine or a processor of a virtual machine. The processor may include one or more processing cores, such as a 4-core processor, an 8-core processor, etc. The processor can be implemented using at least one hardware form of DSP (Digital Signal Processing), FPGA (Field-Programmable Gate Array), or PLA (Programmable Logic Array).

[0088] A processor can also include a main processor and a coprocessor. The main processor is used to process data in the wake-up state and is also called the CPU (Central Processing Unit). The coprocessor is a low-power processor used to process data in the standby state.

[0089] The memory may include one or more computer-readable storage media, which may be non-transitory. The memory may also include high-speed random access memory and non-volatile memory, such as one or more disk storage devices or flash memory devices. In some embodiments of this application, the non-transitory computer-readable storage media in the memory are used to store at least one instruction, which is executed by a processor to implement the methods in the embodiments of this application.

[0090] In some embodiments, the electronic device further includes a peripheral device interface and at least one peripheral device. The processor, memory, and peripheral device interface are connected via a bus or signal line. Each peripheral device is connected to the peripheral device interface via a bus, signal line, or circuit board. Specifically, the peripheral device includes: a display screen, a camera, and audio circuitry. The peripheral device interface can be used to connect at least one I / O (Input / Output) related peripheral device to the processor and memory.

[0091] In some embodiments of this application, the processor, memory, and peripheral device interfaces are integrated on the same chip or circuit board; in other embodiments of this application, any one or two of the processor, memory, and peripheral device interfaces can be implemented on separate chips or circuit boards. This application does not specifically limit the implementation in this regard.

[0092] The electronic device structural block diagrams shown in the embodiments of this application do not constitute a limitation on the electronic device. The electronic device may include more or fewer components than shown, or combine certain components, or use different component arrangements.

[0093] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the methods in any of the foregoing embodiments. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0094] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the parts that contribute to the related technology, can be embodied in the form of software products. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0095] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A wind speed prediction method for wind farms based on deep learning, characterized in that, include: Obtain meteorological data and wind turbine operation data within a preset time window prior to the current moment, and obtain static data for each wind turbine; The wind turbine operation data and the meteorological data are preprocessed, and a three-dimensional input tensor is constructed based on the preprocessed data; The three-dimensional input tensor and the static data of the wind turbine are input into the trained wind speed prediction model to predict the wind speed and obtain the wind speed prediction results for the prediction period. The wind speed prediction model includes an encoding module, a dynamic physics construction module, an iterative graph information transmission module, and a decoding module. The encoding module is used to obtain the initial node state matrix at the current time based on the three-dimensional input tensor encoding; The dynamic physics construction module is used to construct dynamic physics information graphs at various times based on the three-dimensional input tensor and the wind turbine static data. The iterative graph information transmission module is used to perform iterative calculations based on the initial node state matrix and the dynamic physical information graph to obtain the node state matrix at each time point within the prediction period. The decoding module is used to decode based on the node state matrix at each time point to obtain the wind speed prediction result for the prediction period.

2. The wind speed prediction method for wind farms based on deep learning according to claim 1, characterized in that, The preprocessing of the wind turbine operating data and the meteorological data includes: The time axes of the wind turbine operation data and the meteorological data are aligned to obtain an aligned data sequence, and the aligned data sequence is resampled to a preset time resolution. Outlier detection and removal are performed on the resampled data sequence, and missing values ​​are filled by interpolation calculation to obtain the cleaned data sequence. Each variable in the cleaned data sequence is converted to the preset input format of the wind speed prediction model to obtain preprocessed time series data; The construction of the three-dimensional input tensor based on the preprocessed data includes: Based on the preprocessed data, node features related to a single wind turbine and global features related to the entire wind farm are extracted. The node features of each wind turbine are sequentially concatenated with the global features to construct the three-dimensional input tensor.

3. The wind speed prediction method for wind farms based on deep learning according to claim 2, characterized in that, The training process of the wind speed prediction model includes: Historical data for a historical period is acquired, including historical meteorological data and historical wind turbine operation data. Based on the historical data, the preprocessing steps are performed to obtain preprocessed time series data. The data sequence of the first time period is extracted from the preprocessed time series data as the sample input, and the data sequence of the second time period is extracted from the preprocessed time series data to obtain the wind speed sequence of the second time period. The wind speed sequence is used as the sample label. The step of constructing a three-dimensional input tensor is performed on the sample input, and the corresponding three-dimensional input tensor is input into the wind speed prediction model to obtain the wind speed prediction result for the second time period output by the wind speed prediction model. Based on the wind speed prediction results and the corresponding sample labels, a loss function is constructed. The value of the loss function is then propagated to adjust the parameters of the wind speed prediction model until the model converges, thus obtaining a trained wind speed prediction model.

4. A wind speed prediction method for wind farms based on deep learning according to any one of claims 1-3, characterized in that, The process of obtaining the initial node state matrix at the current time based on the three-dimensional input tensor encoding includes: The three-dimensional input tensor is decomposed into node feature sequences and global feature sequences; The global feature sequence is processed by time pooling to obtain the global physical state vector; The global physical state vector is input into the gating network to generate a gating vector; Based on the gate vector, the node feature sequence is subjected to gate modulation operation. The node feature sequence after the gate modulation operation is processed by the Transformer encoder to capture complex temporal dependencies. The output sequence of the last time step is extracted to construct the initial node state matrix.

5. The wind speed prediction method for wind farms based on deep learning according to claim 1, characterized in that, The dynamic physical information graph includes an adjacency matrix and an edge feature matrix. The process of constructing the adjacency matrix includes: Based on the analysis of the three-dimensional input tensor, the real-time parameters at the corresponding time are obtained, including global wind direction, global turbulence intensity, inflow wind speed and yaw angle of each wind turbine; Based on the static data of the wind turbines, the position coordinates of each wind turbine and the relative positional relationship between each pair of wind turbines are determined. Based on the relative positional relationship and the global wind direction at the corresponding time, the upstream wind turbine corresponding to each downstream wind turbine is determined. Calculate the inflow velocity loss caused by the influence of each upstream fan on each downstream fan, and construct a loss matrix based on the inflow velocity loss of each fan pair consisting of each downstream fan and a corresponding upstream fan. The values ​​in the loss matrix are normalized to obtain the adjacency weights of the degree of influence of each upstream wind turbine on the downstream wind turbine, thus obtaining the adjacency weight matrix. The process of constructing the edge feature matrix includes: The directed edges corresponding to each wind turbine pair are determined, the multidimensional physical features of each directed edge are extracted, and the edge feature vector is constructed. The edge feature matrix is ​​constructed based on the edge feature vector of each directed edge.

6. The wind speed prediction method for wind farms based on deep learning according to claim 5, characterized in that, The construction yields dynamic physical information graphs at various times, including: Based on the meteorological data, the forecast wind condition parameters for each moment within the forecast period are derived. Based on the forecast wind condition parameters for each moment, the adjacency matrix and the edge feature matrix for the corresponding moment are updated respectively, thus obtaining the dynamic physical information map for each moment.

7. A wind speed prediction method for wind farms based on deep learning according to claim 5 or 6, characterized in that, The iterative calculation based on the initial node state matrix and the dynamic physical information graph yields the node state matrix at each time point within the prediction period, including: The node embedding vector of each wind turbine is extracted based on the initial node state matrix, and the scaled dot product attention of each wind turbine pair is calculated based on the node embedding vector of each wind turbine. Extract the adjacency weights and edge feature vectors of each wind turbine pair in the dynamic physical information graph at the corresponding time. The attention score for each wind turbine pair is calculated based on scaled dot product attention, adjacency weights, and edge feature vectors. Based on all the wind turbine pairs corresponding to the downstream wind turbines, the adjacency weight, edge feature vector and node embedding vector of the upstream wind turbine are weighted and fused according to the attention score. The node embedding vector of the downstream wind turbine at the next time step is updated based on the weighted fusion result. The node state matrix at the corresponding time step is obtained based on the node embedding vector of each wind turbine at the next time step. Based on the updated node state matrix, proceed to the step of extracting the node embedding vector for each wind turbine, until the node state matrix for each time period within the prediction period is updated.

8. A wind speed prediction device for wind farms based on deep learning, characterized in that, include: The data acquisition module is used to acquire meteorological data and wind turbine operation data within a preset time window before the current moment, and also to acquire static data of each wind turbine. The data processing module is used to preprocess the wind turbine operation data and the meteorological data, and construct a three-dimensional input tensor based on the preprocessed data; The prediction module is used to input the three-dimensional input tensor and the static data of the wind turbine into the trained wind speed prediction model to perform wind speed prediction and obtain the wind speed prediction result for the prediction period. The wind speed prediction model includes an encoding module, a dynamic physics construction module, an iterative graph information transmission module, and a decoding module. The encoding module is used to obtain the initial node state matrix at the current time based on the three-dimensional input tensor encoding; The dynamic physics construction module is used to construct dynamic physics information graphs at various times based on the three-dimensional input tensor and the wind turbine static data. The iterative graph information transmission module is used to perform iterative calculations based on the initial node state matrix and the dynamic physical information graph to obtain the node state matrix at each time point within the prediction period. The decoding module is used to decode based on the node state matrix at each time point to obtain the wind speed prediction result for the prediction period.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 7.

10. A non-transitory 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 steps of the method as described in any one of claims 1 to 7.

Images