A spatiotemporal graph neural network-based wind power cluster short-term power prediction method and device
By using a spatiotemporal graph neural network method that integrates physical information, a graph structure and multilayer perceptron model are constructed. Combined with wake and blockage effect models for two-stage training, the problem of insufficient accuracy and efficiency in short-term power prediction of wind power clusters is solved, and more accurate power prediction of wind power clusters is achieved.
Patent Information
- Application Number
- CN202511627017.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-07
- Publication Date
- 2026-11-10
- Estimated Expiration
- 2045-11-07
AI Technical Summary
Existing wind power prediction methods are insufficient in terms of accuracy and efficiency. In particular, the complex spatiotemporal coupling characteristics caused by wake and blockage effects between wind farms make it difficult to accurately predict the short-term power of wind power clusters.
A spatiotemporal graph neural network method incorporating physical information is adopted. By acquiring relevant information data of wind power clusters, a graph structure is constructed. Combining the wake analysis model and the blockage effect model, a graph convolutional neural network module and a gated temporal convolutional network module are constructed. A two-stage training method is adopted, using physical prior data and dynamic adjacency matrix for model training to predict the short-term power of wind power clusters.
This improved the accuracy and efficiency of short-term power prediction for wind power clusters, enabling more accurate prediction of the total power of wind farm clusters and verifying the effectiveness of introducing physical information in improving model accuracy and prediction efficiency.
Smart Images

Figure CN121710168B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of wind power cluster technology, and in particular to a method, apparatus and computer equipment for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information. Background Technology
[0002] As wind energy, as a clean energy source, gradually increases its share in the energy structure, accurate prediction of wind power output is crucial for the dispatch and operation of power systems. However, wind power prediction faces numerous challenges, primarily due to the complex spatiotemporal coupling characteristics caused by wake and congestion effects between wind farms. Existing prediction methods include physical modeling, data-driven modeling, and a fusion of both. Physical models rely on accurate simulations of physical processes, while data-driven models depend on large amounts of historical data. However, existing methods still have shortcomings in terms of prediction accuracy and efficiency. Summary of the Invention
[0003] Therefore, it is necessary to provide a method, device, and computer equipment for short-term power prediction of wind power clusters that integrates spatiotemporal graph neural networks with physical information, which can improve prediction accuracy and efficiency, in response to the above-mentioned technical problems.
[0004] In a first aspect, this application provides a method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information, including: Step 1: Obtain relevant information data of each wind farm in the wind power cluster, including wind turbine parameters, geographical location information, historical operation data and meteorological data, and preprocess the relevant information data of each wind farm in the wind power cluster. Step 2: Based on the relevant information data of each wind farm in the preprocessed wind power cluster, a physical prior dataset is formed by integrating the wake analysis model and the blocking effect model into an engineering analytical model. Step 3: Using each wind farm as a node in the graph, construct a graph structure for wind farm power prediction. Update the adjacency matrix daily based on the similarity of wind speed sequences using a dynamic time warping algorithm to form a dynamic adjacency matrix. Step 4: Construct a spatiotemporal graph neural network model architecture that includes a graph convolutional neural network module, a gated temporal convolutional network module, and a multilayer perceptron; Step 5: A two-stage training process is adopted. First, the spatiotemporal graph neural network model architecture is pre-trained using the physical prior dataset. Then, the pre-trained spatiotemporal graph neural network model architecture is formally trained based on historical power data and dynamic adjacency matrix to obtain the spatiotemporal graph neural network model. Step 6: Input the wind speed for the predicted day, and use the spatiotemporal neural network model to predict the active power of the entire wind power cluster for 24 hours on the predicted day.
[0005] Secondly, this application also provides a short-term power prediction device for wind power clusters that integrates spatiotemporal graph neural networks with physical information. This device employs the method described in any embodiment of the first aspect, including: The acquisition unit is used to execute step 1; The physical information modeling unit is used to perform step 2; Graph structure building unit, used to perform step 3; The spatiotemporal graph neural network modeling unit is used to perform step 4; Training unit, used to perform step 5; The prediction unit is used to perform step 6.
[0006] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of a short-term power prediction method for wind power clusters using a spatiotemporal graph neural network that integrates physical information.
[0007] The aforementioned method, apparatus, and computer equipment for short-term power prediction of wind farm clusters using a spatiotemporal graph neural network that integrates physical information proposes a method for short-term power prediction of wind farm clusters that considers the effects of wake and congestion between adjacent wind farms. On one hand, based on a coupled wake and congestion analytical model, a physical prior dataset containing simulated incoming wind conditions and calculated total power for the entire wind farm cluster is further obtained. The WB-STGNN model, which is adept at mining spatiotemporal coupling correlations, is used for prior knowledge learning to learn the implicit mechanisms and laws in the physical prior data. On the other hand, the dynamic graph topology of the wind farm cluster is generated daily using the DTW algorithm to characterize the complex interactions between wind farms. The parameters of the WB-STGNN model are adjusted using an appropriate amount of observation data to achieve satisfactory prediction results in actual power production scenarios. Comparative experimental results based on example data show that, compared with existing models, the proposed WB-STGNN model can fully utilize the spatiotemporal information of the wind farm cluster for more accurate power prediction, while verifying the effectiveness of introducing physical information in improving model accuracy and prediction efficiency. Attached Figure Description
[0008] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings involved in the description of the embodiments or the prior art will be briefly introduced below. It should be noted that the following drawings are only some embodiments described in the present invention, and those skilled in the art can obtain other drawings based on these drawings without creative effort.
[0009] Figure 1 This is a flowchart of a short-term power prediction method for wind power clusters using a spatiotemporal graph neural network that integrates physical information, as described in one embodiment. Figure 2 This is a flowchart illustrating the implementation of a short-term power prediction method for wind power clusters using a spatiotemporal graph neural network that integrates physical information, as described in one embodiment. Figure 3 This is a framework diagram of the Pywake wind farm simulation tool in one embodiment; Figure 4 This is a diagram showing the overall structure of the WB-STGNN model in one embodiment; Figure 5 This is a diagram showing the wind turbine layout of a near-shore wind power cluster in one embodiment; Figures 6(a) and 6(b) show the wind speed and wind direction angle in the autocorrelation calculation results of the wind power sequence in one embodiment, respectively. Figures 7(a) and 7(b) are comparison charts of MERRA-2 wind speed data and wind direction angle data at heights of 2m, 10m and 50m in one embodiment, respectively. Figures 8(a) and 8(b) are a comparison of MERRA-2 and NWP wind speed data and wind direction angle data at a height of 10m in one embodiment; Figures 9(a), 9(b), 9(c), and 9(d) are, respectively, the measured wind speed diagram, the simulated incoming wind speed diagram, the measured total power diagram, and the calculated total power diagram in a comparison diagram of measured data and physical prior data in one embodiment; Figures 10(a), 10(b), 10(c), and 10(d) show the power prediction errors of different models in four seasons in one embodiment, representing spring, summer, autumn, and winter respectively. Figures 11(a), 11(b), 11(c), and 11(d) show the power prediction results of different models in one embodiment for the four representative weeks of May 20-26, August 20-26, October 19-25, and February 22-28, respectively. Figure 12 The wind direction angle and wind speed are shown in the comparison diagram of power prediction errors of different models in one embodiment, where (a) is the wind direction angle and (b) is the wind speed. Figures 13(a), 13(b), and 13(c) are the average wind speed DTW and PCC results for wind farms 1 and 2, 2 and 3, and 1 and 3, respectively, in one embodiment. Detailed Implementation
[0010] This invention aims to provide a method, apparatus, device, and medium for short-term power prediction of wind power clusters that integrates physical information. By combining physical models and deep learning methods, it effectively improves the accuracy and efficiency of wind power cluster power prediction. To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application. This invention provides a method for short-term power prediction of wind power clusters based on spatiotemporal graph neural networks, which effectively solves the problems in the background art by integrating physical information and deep learning algorithms.
[0011] In one exemplary embodiment, a short-term power prediction method for wind power clusters based on a spatiotemporal graph neural network that integrates physical information is provided, comprising: Step 1: Obtain relevant information data of each wind farm in the wind power cluster, including wind turbine parameters, geographical location information, historical operation data and meteorological data, and preprocess the relevant information data of each wind farm in the wind power cluster. Specifically, the system acquires turbine parameters, geographical location information, historical operating data, and meteorological data from each wind farm within the wind power cluster, and performs data preprocessing. Turbine parameters include hub height, rotor diameter, power curve, and thrust coefficient. Historical operating data includes wind speed, power, and temperature from the SCADA (Supervisory Control and Data Acquisition) system. Meteorological data includes wind speed and direction from MERRA-2 reanalysis data, and NWP forecast data.
[0012] Step 2: Based on the relevant information data of each wind farm in the preprocessed wind power cluster, a physical prior dataset is formed by integrating the wake analysis model and the blocking effect model into an engineering analytical model. Specifically, an engineering analytical model that integrates the wake analysis model (Jensen model) and the blockage effect model (VC model, i.e., Viscous-Cavity model, viscous cavity model) is adopted. The flow field of the wind power cluster is simulated through the Pywake platform, and the simulated incoming wind speed and theoretical calculated power of each wind farm are calculated to form a physical prior dataset.
[0013] Step 3: Using each wind farm as a node in the graph, construct a graph structure for wind farm power prediction. Update the adjacency matrix daily based on the similarity of wind speed sequences using a dynamic time warping algorithm to form a dynamic adjacency matrix. Specifically, each wind farm is used as a node in the graph to construct graph structure data for wind farm power prediction. The adjacency matrix is updated daily based on the similarity of wind speed sequences using the Dynamic Time Warping (DTW) algorithm to quantify the coupling strength of wake and blockage effects between wind farms and form a dynamic adjacency matrix.
[0014] Step 4: Construct a spatiotemporal graph neural network model architecture that includes a graph convolutional neural network module, a gated temporal convolutional network, and a multilayer perceptron; Specifically, a spatiotemporal graph neural network (WB-STGNN) model is constructed, which includes a graph convolutional network (GCN) module, a gated temporal convolutional network (GTCN) module, and a multilayer perceptron (MLP) module.
[0015] Step 5: A two-stage training process is adopted. First, the spatiotemporal graph neural network model architecture is pre-trained using the physical prior dataset. Then, the pre-trained spatiotemporal graph neural network model architecture is formally trained based on historical power data and dynamic adjacency matrix to obtain the spatiotemporal graph neural network model. Specifically, a two-stage training strategy is adopted: first, the model architecture is pre-trained using a physical prior dataset, and then the model architecture is formally trained using a portion of the observation data to adjust the model parameters and improve prediction accuracy.
[0016] Step 6: Input the wind speed and historical power data for the predicted day, and use the spatiotemporal neural network model to predict the active power of the entire wind power cluster in the next 24 hours. Specifically, the wind speed for the predicted day is input, and the model predicts the active power output of the entire wind power cluster for 24 hours on that predicted day.
[0017] In one exemplary embodiment, such as Figure 1 and Figure 2As shown, this application relates to the field of wind power prediction technology, and in particular to a short-term power prediction method for wind power clusters based on a spatiotemporal graph neural network that integrates physical information. Based on a coupled analytical model of wake and blockage effects and a graph neural network model, a short-term power prediction framework integrating physical information is proposed for wind power clusters. The method includes: acquiring wind turbine parameters, geographical location information, historical operation data, and meteorological data of each wind farm in the wind power cluster, and performing data preprocessing; modeling the flow field between each wind farm in the wind power cluster based on a physical model of wake effect and blockage effect, obtaining the simulated incoming wind speed of each wind farm and the theoretical calculated power based on the wind power curve, forming a physical prior dataset; constructing graph structure data, taking each wind farm as a node of the graph, and updating the adjacency matrix daily according to the similarity of wind speed sequences using the Dynamic Time Warping (DTW) algorithm, forming a dynamic adjacency matrix; constructing a spatiotemporal graph neural network WB-STGNN model that includes a graph convolutional network (GCN), a gated temporal convolutional network (GTCN), and a multilayer perceptron (MLP); adopting a two-stage training strategy, first pre-training the model using the physical prior dataset, and then formally training it using historical operation data and the dynamic adjacency matrix; inputting the wind speed and historical power data of the prediction day, and predicting the active power output of the entire wind power cluster in the next 24 hours through the model.
[0018] In an exemplary embodiment, in step 1, the wind turbine parameters include hub height, rotor diameter, power curve, and thrust coefficient; historical operating data includes wind speed, power, and temperature from the SCADA system; and meteorological data includes NWP forecast data and MERRA-2 reanalysis data. Remove shutdown data and various power limitation point data from the relevant information data of each wind farm in the wind power cluster; Replace all negative values with zero. Adjust the data time resolution to the preset precision and complete any missing data; The zonal and latitude wind speed data from MERRA-2 reanalysis data and NWP forecast data are converted into the form of total wind speed V and wind direction angle θ, and then converted according to the wind turbine hub height. The formula is defined as follows: ; ; ; In the formula, u and v represent the meridional wind speed and the zonal wind speed, respectively, and arctan2(u, v) is the two-parameter arctangent function, V height and α height V represents the incoming wind speed at the height of the wind turbine hub and the wind speed conversion factor relative to the reference height ref, respectively. ref Wind speed at reference altitude.
[0019] Furthermore, the physical prior dataset was constructed by inputting MERRA-2 atmospheric reanalysis data and numerical weather prediction (NWP) data into an engineering analytical model that incorporates wake and blocking effects. The external inputs to this model include the hub height, rotor diameter, and geographical distribution information of the wind turbines, as well as the wind resource characteristics of the wind power cluster area. These wind resource characteristics include wind speed, wind direction, and turbulence intensity at a reference location, all of which can be obtained from the MERRA-2 dataset.
[0020] In one exemplary embodiment, step 2 includes: Construct an engineering analytical model that integrates the wake analysis model and the blocking effect model; The preprocessed MERRA-2 reanalysis data (including wind speed, wind direction, and turbulence intensity at the reference location of the wind farm cluster), NWP forecast data, turbine hub height, rotor diameter, and geographical location information, as well as the wind resource characteristics of the wind farm cluster, are input into the engineering analytical model. The flow field of the wind farm cluster is simulated using the Pywake platform to calculate the simulated incoming wind speed and theoretical calculated power for each wind farm, forming a physical prior dataset. The wind resource characteristics of the wind farm cluster, including wind speed, wind direction, and turbulence intensity at the reference location of the wind farm cluster area, are obtained from the MERRA-2 reanalysis data. The framework diagram of the Pywake wind farm simulation tool is shown below. Figure 3 As shown.
[0021] For example, MERRA-2 atmospheric reanalysis data, numerical weather prediction (NWP) data, wind turbine hub height, rotor diameter and their geographical distribution information, and wind resource characteristics of the wind power cluster area are input into an engineering analytical model that includes wake and blockage effects to form a physical prior dataset. The wind resource characteristics include wind speed, wind direction and turbulence intensity at the reference location, which are obtained from the MERRA-2 dataset.
[0022] In an exemplary embodiment, an engineering analytical model coupling the Jensen wake model and the VC blocking model is established, including: In existing wake calculation models, a first-order Jensen wake analytical model is derived based on the law of mass conservation. It is assumed that the velocity distribution in the wake region downstream of the wind turbine approximately satisfies a uniform "cap-shaped" distribution, and that the initial wake region diameter is the same as the wind turbine diameter. Based on the derived first-order Jensen wake analytical model, the calculation expressions for the wake region wind speed Vw(l) and diameter Dw(l) are as follows: ; ; In the formula, l is the distance downstream of the fan from the rotor. C is the free-flow velocity at the height of the wind turbine hub. TR is the thrust coefficient of the wind turbine, R is the radius of the wind turbine, and k is the thrust coefficient of the wind turbine. w The wake spread coefficient represents the linear diffusion rate of the wake influence range at a distance l from the rotor plane downstream of the wind turbine. Its magnitude mainly depends on factors such as local surface roughness and atmospheric stability. This scheme adopts a constant value. To assess the blocking effect of downstream wind turbines on upstream wind turbines, a VC blocking effect model is used. The geometry of the downstream wake region is defined as a semi-infinite cylinder (γt) with a cross-sectional radius equal to the rotor radius R and a constant tangential vorticity γt. t (Related to the thrust coefficient of the wind turbine and the free-flow velocity), and the Jensen wake analytical model is introduced to calculate the wake velocity. Considering that the wake velocity dominates downstream of the wind turbine, and that the VC model ignores wake diffusion and expansion, the cylindrical cavity is set as a pure wake region, i.e., the induced velocity within this region is constant at 0. The induced velocity in the upstream region of the wind turbine affected by the blockage effect is simplified to the form obtained by expanding the Taylor formula as follows: ; In the formula, The average axial inductance coefficient of the wind turbine plane. .
[0023] In an exemplary embodiment, step 3, constructing graph structure data to form a feature matrix, includes: A non-Euclidean graph data structure is constructed to characterize the influence relationships between wind farms. It consists of a set of vertices (also called nodes) V and a set of edges E, represented as follows: ,in , representing N wind farms in a wind power cluster, v N Represents the Nth node, v i Let represent the characteristic matrix of the i-th wind farm node, i=1, 2, …, N. (i=1, 2, …, N) consists of information such as the station's geographical location, external meteorological variables, and real-time operating parameters, where F is the input feature dimension. Reflecting the coupling relationship between wind farms, where e k The characterization represents the coupling relationship between the k-th wind farm and its neighboring wind farms, k=1, 2, …,M. The existence of edge connections between nodes is determined based on the electrical connection between wind farms via submarine cables, geographical adjacency, wake and blockage effects, and data correlation between wind farm operating parameters. Edge information includes the geographical distance between wind farms, the strength of correlation between meteorological conditions and wind farm operating parameters, etc., and a feature matrix of edges is formed based on this information.
[0024] Treating each wind farm in the wind power cluster as a node in a graph structure, the Dynamic Time Warping (DTW) algorithm is used to calculate the minimum warped distance between the wind speed time series of any two wind farms i and j on the predicted day d. This distance measures the difference in wind conditions across adjacent time segments, thus quantitatively characterizing the dynamic coupling strength between wind farms caused by wake and blocking effects. A weighted adjacency matrix of the wind power cluster is constructed using the average of the minimum warped distances over the 24 hours of that day, and the elements of the weighted adjacency matrix are dynamically updated daily to improve the graph neural network's ability to model non-stationary physical coupling relationships between wind farms. (Adjacency matrix) For a symmetric matrix, its elements are defined as follows: ; In the formula, Let represent the wind speed sequence of the i-th wind farm on day d. The edge weight between node i and node j is defined as: ; In the formula, The DTW distance between wind speed sequences. The normalization coefficient is used; the DTW distance is defined as: ; In the formula, and These represent the i-th and j-th wind farms on day d at time t. m ,t n The simulated incoming wind speed time series, with day-ahead hourly forecasts, m and n are set to 24, and π represents the time series from the initial time point (t1, t2) to the end time point (t3, t4). m ,t n ) is the set of all valid paths that satisfy the dynamic regularization constraint. For example, an valid path is (1,1)→(1,3)→(3,24)→(24,24).
[0025] In one exemplary embodiment, step 4 includes: Using the predicted wind speed (NWP) for the 24-hour forecast day as input and the active power of each wind farm node in the wind power cluster for the 24-hour forecast day as output, a spatiotemporal graph neural network (WB-STGNN) model architecture is constructed. The WB-STGNN model architecture includes three gated temporal convolutional network (GTCN) modules, one graph convolutional neural network (GCN) module approximated by a first-order Chebyshev polynomial, and a multilayer perceptron. The total power prediction value of the wind power cluster is obtained by summing the active power of each wind farm node in the wind power cluster for the 24-hour forecast day.
[0026] Specifically, the spatiotemporal graph neural network model is a short-term power prediction model for wind power clusters that integrates physical information (WB-STGNN). The input is a feature map containing spatial aggregation information of each wind farm node and its surrounding neighboring nodes, including the predicted wind speed for each wind farm on the prediction day and the adjacency matrix of the wind farm cluster. The intermediate prediction model mainly consists of three gated temporal convolutional network (GTCN) modules (denoted as GTCN-1, GTCN-2, and GTCN-3, respectively) and one graph convolutional neural network (GCN) module approximated by a first-order Chebyshev polynomial. GTCN-1 and GTCN-2 are used for temporal correlation extraction, while GCN performs information aggregation and feature mining for each wind farm in the spatial dimension. The obtained spatiotemporal feature data is then processed by GTCN-3 and several identical multilayer perceptron (MLP) models for feature aggregation and dimension transformation. The output feature map represents the active power of each wind farm node within the wind power cluster for the next day. The sum of these features is the total predicted power of the wind power cluster. The overall structure diagram of the WB-STGNN model is shown below. Figure 4 As shown.
[0027] In an exemplary embodiment, the two-stage training strategy in step 5 includes: First, in the physical prior pre-training stage, the spatiotemporal graph neural network (WB-STGNN) model architecture is trained using the physical prior dataset obtained through the wake analysis model and the blockage effect model, with theoretically calculated power as the supervision signal, to learn the spatiotemporal coupling characteristics between wind farms based on physical mechanisms and complete the initial fitting of model parameters; then, in the observation data formal training stage, measured wind speed and historical power data are used as inputs, and actual observed power is used as the label, to further fine-tune the model parameters based on the pre-training, thereby improving the prediction accuracy of the model under actual wind conditions; wherein, the pre-training stage uses a static adjacency matrix constructed based on the geographical proximity relationship of wind farms, and the formal training stage uses a dynamic adjacency matrix updated daily and constructed based on the dynamic time warping algorithm, to more accurately represent the dynamic coupling relationship between wake and blockage effect between wind farms. The specific steps are as follows: Step 5.1: Construct the adjacency matrix A for learning physical prior knowledge of the WB-STGNN model architecture of the spatiotemporal graph neural network. pre It considers the geographical information of wind farms and establishes the simulated incoming wind speed of wind farms based on the proposed wind power cluster flow field model. With calculated power The nonlinear mapping relationship between them is used to fully learn the physical mechanism of wake and blockage effect between wind farms, and the spatiotemporal graph neural network WB-STGNN model architecture is pre-trained.
[0028] Step 5.2 involves performing "model transfer" on the pre-trained spatiotemporal graph neural network (WB-STGNN) model architecture from Step 5.1, and constructing a dynamic adjacency matrix A for the formal training of the WB-STGNN model. d By using a suitable amount of measured wind speed and historical power data of wind power clusters, the network parameters of each layer of the pre-trained spatiotemporal graph neural network (WB-STGNN) model architecture are adjusted to obtain the WB-STGNN model, thereby achieving the integration of physical mechanisms and data-driven approaches. This ensures that the model output conforms to physical laws while also taking into account the actual operating conditions and environmental characteristics of the wind power clusters.
[0029] In an exemplary embodiment, the loss functions for the two-stage training of the WB-STGNN model are as follows: In step 5.1, the loss function in the pre-training phase is calculated using the Huber loss function. The Huber loss function combines the advantages of mean squared error and mean absolute error, flexibly adjusting the loss calculation method through a preset threshold. It has strong resistance to outliers, thereby improving the model's robustness and prediction accuracy. It is defined as follows: ; In the formula: and They are respectively and The elements in the array, where m is the number of samples.
[0030] Therefore, the loss function expression for the WB-STGNN model is: ; In the formula, and These are the calculated power values of N wind farms obtained from the coupled analytical wake and blockage effect model, and the predicted values obtained from the pre-trained spatiotemporal graph neural network (WB-STGNN) model architecture, respectively. Therefore, we have δ controls the threshold for transforming the Huber loss function from its quadratic form to its linear form; let's set it to 5. In step 5.2, the loss function in the formal training process is defined as the predicted power of all wind farms through the formally trained spatiotemporal graph neural network (WB-STGNN) model architecture. With actual power The average deviation between them is expressed as: .
[0031] At least some steps in the flowcharts involved in the embodiments described above may include multiple steps or multiple stages, which may be executed at different times, or may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.
[0032] Based on the same inventive concept, this application also provides a wind power cluster short-term power prediction device using a spatiotemporal graph neural network that integrates physical information, for implementing the aforementioned method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the wind power cluster short-term power prediction device using a spatiotemporal graph neural network that integrates physical information provided below can be found in the limitations of the wind power cluster short-term power prediction method using a spatiotemporal graph neural network that integrates physical information described above, and will not be repeated here.
[0033] In one exemplary embodiment, a short-term power prediction device for wind power clusters based on a spatiotemporal graph neural network that integrates physical information is provided, comprising: The acquisition unit is used to execute step 1; The physical information modeling unit is used to perform step 2; Graph structure building unit, used to perform step 3; The spatiotemporal graph neural network modeling unit is used to perform step 4; Training unit, used to perform step 5; The prediction unit is used to perform step 6.
[0034] Specifically, the data acquisition unit acquires wind resource data (wind speed, wind direction, etc.), wind turbine parameters (hub height, rotor diameter, etc.) and historical operating power data for the wind power cluster area; The physical information modeling unit constructs a physical model of wake and blockage effects based on the collected data, outputs simulated wind speed and theoretical power, and forms a physical prior dataset. The graph structure building unit treats wind farms as graph nodes and uses the DTW algorithm to calculate the similarity of wind speed sequences to construct a daily-scale dynamic weighted adjacency matrix. The spatiotemporal graph neural network modeling unit is used to construct a spatiotemporal graph neural network WB-STGNN model that integrates a fused graph convolutional neural network module (GCN), a gated temporal convolutional network (GTCN), and a multilayer perceptron (MLP) to model the variation of wind power cluster power over time and space. The training unit is used to perform two-stage training on the WB-STGNN model. The first stage is pre-training with the theoretical power output by the physical information model as the label, and the second stage is transfer learning and fine-tuning with the actual observation data as the label to optimize the model's prediction accuracy. The prediction unit is used to output the short-term active power prediction results of the wind power cluster for the next 24 hours based on the wind speed and power data and dynamic adjacency matrix before the prediction date using the trained WB-STGNN model after training.
[0035] The modules in the aforementioned spatiotemporal graph neural network-based wind power cluster short-term power prediction device, which integrates physical information, can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the computer device's memory as software, so that the processor can call and execute the corresponding operations of each module.
[0036] This application describes a computer device comprising a processor and a memory. The memory stores computer programs that the processor can execute, and when the processor runs the computer program, it performs the methods mentioned above.
[0037] This application also relates to a storage medium on which a computer program is stored. Once the program is run by a processor, the above-described method will also be executed.
[0038] In one exemplary embodiment, a short-term power prediction method for wind power clusters, including a spatiotemporal graph neural network that integrates physical information, includes the following steps: Step S1: The wind power cluster studied in this application includes three wind farms, generally arranged in a rectangular shape oriented northeast. A diagram showing the wind turbine layout of a near-shore wind power cluster is shown below. Figure 5 As shown in Table 1, Wind Farm No. 1 is equipped with 15 doubly-fed wind turbines (SL1-SL15) with a rated capacity of 3MW each, for a total installed capacity of 45MW. Wind Farm No. 2 is equipped with 21 wind turbines (SWT1-SWT21) with a rated capacity of 2.38MW each, 4 doubly-fed wind turbines (SL16-SL19) with a rated capacity of 3MW each, and 2 wind turbines (GW19 and GW20) with a rated capacity of 2.5MW each, for a total installed capacity of approximately 67MW. Wind Farm No. 3 is equipped with 18 wind turbines (GW1-GW18) with a rated capacity of 2.5MW each, for a total installed capacity of 45MW. The main parameters of each type of wind turbine are shown in Table 1.
[0039] Table 1. Main parameters of turbines in a near-shore wind power cluster
[0040] The raw data acquisition and monitoring (SCADA) dataset for the wind power cluster records data from 00:00 on November 1, 2014 to 23:50 on October 31, 2015 (UTC+08:00, i.e., Beijing time; UTC is the abbreviation for Coordinated Universal Time), a total of 365 days. It includes five external environmental and operational characteristics for each wind turbine: wind speed, temperature, nacelle yaw angle, reactive power, and active power output. The sampling time step is 10 minutes. The specific processing steps are as follows: (1) When cleaning up outliers, on the one hand, the active power output corresponding to the two types of shutdown data for each fan, namely, those below the cut-in wind speed and those above the cut-out wind speed, needs to be set to 0. On the other hand, negative active power values and zero active power values will be filled based on the theoretical wind power curve provided by the fan manufacturer, while negative reactive power values will be directly replaced with zero values. When the recorded unit wind speed shows obvious abnormalities such as excessive values, the normal wind speed value should be deduced from the corresponding output power and theoretical wind power curve.
[0041] (2) Data missing problems usually occur in two situations: one is that one or some features of a certain wind turbine data are missing within a certain period of time, and the other is that all feature data of the wind farm are missing at the same time. In this application, normal data of the unit with the closest geographical location to the wind turbine are used to fill in the missing data, and all data with completely missing features are deleted.
[0042] After data preprocessing, the original SCADA dataset was converted into hourly data. After removing several missing data sets within the same time period, 8592 samples were obtained over 358 days.
[0043] This application employs the maximum information coefficient correlation analysis method for key feature selection. In addition to wind speed, historical power data is also a major input variable of the proposed model. The optimal historical time window length for the active power output of each wind farm can be determined based on the autocorrelation coefficient. The calculation formula is defined as follows: ; In the formula: M is the number of samples in the dataset, l represents the number of lag time steps, and p i It is the active power output of the wind farm at the i-th moment. It is the average value of the active power output of the wind farm, and the denominator represents the variance of the sample.
[0044] The autocorrelation calculation results of the wind power series are shown in Figures 6(a) and 6(b), where Figure 6(a) shows the wind speed and Figure 6(b) shows the wind direction angle. The curves showing the autocorrelation coefficients of the historical power series of the three wind farms as a function of the lag time step are also presented. It can be seen that within a lag of approximately 8 hours, the autocorrelation coefficient of the power data remains above 0.6; however, as the lag time step increases, the autocorrelation shows a rapid decreasing trend. When the lag time reaches 60 hours, the autocorrelation coefficient fluctuates slightly around 0, indicating a very weak correlation between historical and real-time power. Therefore, this application selects the wind speed of each wind farm on the prediction day and the historical power of the wind farm on the previous day as the key input variables for the model during physical prior knowledge learning and formal training. At this time, the input time step T=24 and the input feature dimension F=2, i.e., the input... , Output , .
[0045] Step S2: Considering that the wind speed and direction data obtained from the SCADA system may have some deviation compared with the actual values, and that incoming wind speed data for the wind farm area is lacking, this application selects the MERRA-2 dataset to overcome the shortcomings of single-source data. Simultaneously, it can obtain large-scale meteorological change patterns, providing accurate atmospheric information for forecasting. Based on the actual geographical location of the wind farm cluster, hourly wind speed data in latitude and longitude at 2m, 10m, and 50m altitudes within the range of 32.5°~33°N and 120.625°~121.25°E are obtained from the MERRA-2 “M2T1NXSLV” dataset. The comparison diagrams of MERRA-2 wind speed and wind direction angle data at 2m, 10m, and 50m altitudes are shown in Figure 7(a) and Figure 7(b), respectively. The comparison diagrams of MERRA-2 and NWP wind speed and wind direction angle data at 10m altitude are shown in Figure 8(a) and Figure 8(b), respectively. Considering the time zone differences, the time span is selected from 16:30 on December 31, 2013 to 23:30 on October 31, 2015 (UTC+00:00, i.e., Coordinated Universal Time).
[0046] Since the wind power cluster studied in this application does not provide corresponding hourly day-ahead NWP data, the NWP data used in this application comes from the meteorological forecast data at a height of 10m at 32.5°N, 121°E in the THORPEX Interactive Grand Global Ensemble (TIGGE) system. The control forecast results of the European Centre for Medium-range Weather Forecasts (ECMWF) are selected. The spatial resolution of the forecast data is 0.5°×0.5°, and UTC+00:00 is selected as the forecast start time, with a forecast lead time of 24 hours.
[0047] For the zonal and latitude wind speed data in MERRA-2 and NWP, it is necessary to convert them into the form of total wind speed V and wind direction angle θ (which needs to be converted into angle), and then perform the calculation based on the wind turbine hub height. The relevant formulas are defined as follows: ; ; ; In the formula: u and v represent the meridional wind speed and the zonal wind speed, respectively; arctan2(u, v) is the two-parameter arctangent function; V height and α height V represents the incoming wind speed at the height of the wind turbine hub and the wind speed conversion factor relative to the reference height ref, respectively. ref Wind speed at reference altitude.
[0048] As shown in Table 1, the hub height of the wind turbines is 90m. Therefore, the wind speed data for MERRA-2 needs to be adjusted, while the wind direction angle data should remain consistent with that at a height of 50m. Based on relevant research in the literature on wind speed conversion methods at different heights in the same sea area as the wind farm in this application, the wind speed at a height of 10m is used as the reference wind speed, i.e., ref = 10m. The wind speed conversion factor α at a height of 90m is calculated. 90m =1.484.
[0049] Based on the wind speed and wind direction angle curves at altitudes of 2m, 10m, and 50m in the MERRA-2 dataset for the first two weeks of January 2014 (UTC+08:00), it can be seen that although wind speeds differ at different altitudes and increase with altitude, they exhibit a strong correlation and a nearly consistent trend. Meanwhile, except for times 276 to 282, the wind direction angle remains approximately constant for most of the time, indicating that the magnitude of the wind direction angle at different altitudes hardly changes and can be considered a constant value. Based on the wind speed and wind direction angle curves from the MERRA-2 and NWP datasets for a forecast period in 2014, the two datasets show good consistency, indicating that ECMWF meteorological forecast data has high accuracy.
[0050] Step S3 involves simulating the wind farm flow field based on the established first-order Jensen wake model and VC blockage model. The focus is on the interaction between wakes and blockages between adjacent wind farms, using the average simulated incoming wind speed of each turbine in each wind farm to represent the overall simulated incoming wind conditions of that wind farm. Then, the total power of the site is derived from the theoretical wind power curves of its turbines, providing physical prior knowledge for the spatiotemporal neural network model.
[0051] The first two weeks of November 2014 were selected from the processed SCADA measured dataset and the physical prior dataset. The average wind speed and total active power output curves for each wind farm are plotted in Figures 9(a), 9(b), 9(c), and 9(d). The blue line represents wind farm 1, the orange line represents wind farm 2, and the green line represents wind farm 3. From the measured data in Figures 9(a) and 9(c), it can be seen that although the three wind farms are geographically close and located in the same sea area, there are certain differences in wind speed levels between wind farm 2 and the other wind farms during the low wind speed and low power output periods (65-85 and 237-260) and the high wind speed and high power output periods (25-50). This indicates that the wake and blockage effects between the wind farms are significant during this period. Figures 9(b) and 9(d) show that although the first-order Jensen wake model and the VC blockage model can simulate the general trend of wind speed changes in the area where the wind farm is located, the physical prior data obtained still have certain deviations compared with the measured data due to the influence of empirical parameters and various simplification steps. For example, the overall incoming wind speed level of the wind farm is approximately the same, and the calculated total power value is too high. Therefore, the measured data are needed for further correction and improvement.
[0052] Step S4: When learning the physical prior knowledge, the physical prior data from November 2014 to August 2015 is selected to form the training set. The September data is used as the validation sample, and the October data is used for model testing. Considering that the simulated incoming wind speeds at the wind farms are almost consistent, the adjacency matrix in this training phase is set to a matrix Apre with all diagonals set to 0 and all other positions set to 1, based solely on the geographical proximity of the wind farms, as shown in the following formula: ; During formal training and prediction, the model's predictive performance was validated seasonally: December 2014 and January-February of the following year were winter; March-May 2015 were spring; June-August 2015 were summer; and November 2014 and September-October of the following year were autumn. The adjacency matrix was determined according to the dynamic association topology generation method described in Section 4.2.2. All training processes of the proposed model employed the Adam optimization algorithm, with a learning rate of 0.0001 and 300 training epochs. An early stopping strategy was also introduced: if the model's predictive performance on the validation set did not significantly improve within a given tolerance epoch (i.e., the validation loss changed between adjacent epochs), the model would be stopped. Less than the minimum change L s Training is terminated early regardless of whether the preset number of training epochs has been reached to avoid overfitting and save unnecessary training time. The optimal parameters of the model are then saved. In this application, the tolerance epochs for learning prior physical knowledge are set to 20, while the tolerance epochs for formal training are set to 10, and Ls is set to 0.0001. Step S5: To evaluate the predictive performance of the proposed model, this application selected MLP, CNN-GRU, GTCN-MLP, and GCN-MLP models for comparative experiments. The training and dataset partitioning strategies were exactly the same as those for WB-STGNN. After learning physical prior knowledge, WB-MLP, WB-CNN-GRU, WB-GTCN-MLP, and WB-GCN-MLP were obtained respectively. Among them, MLP can only extract the time-dimensional dependence of input variables and does not consider spatial relationships. Therefore, a "one-site-one-test" approach was adopted to model the power prediction for the three wind farms separately, and then the total predicted power of the wind power cluster was obtained by summing. CNN-GRU is a combined prediction model that can simultaneously mine the spatiotemporal correlation of Euclidean structure data. It consists of a two-dimensional convolutional layer, a GRU layer, and an MLP model stacked together. GTCN-MLP consists of two GTCN modules and an MLP model, while GCN-MLP contains two GCN modules and an MLP model. The construction of the adjacency matrix is consistent with the proposed model.
[0053] To quantify the accuracy of the proposed model's prediction results, this application uses five evaluation metrics—MAE, mean absolute percentage error (MAPE), root mean square error (RMSE), normalized root mean-square error (NRMSE), and coefficient of determination R²—to comprehensively compare the model's power prediction performance. To avoid situations where the actual power value is 0 or close to 0, causing the results to tend towards infinity when calculating MAPE and NRMSE, this application uses the average value of the actual power value as the denominator in the MAPE and NRMSE formulas. The relevant formula definitions are as follows: ; ; ; ; ; In the formula: and These are the actual observed value and the predicted value of wind power at time i, respectively. It is the average value of the wind power series, n p The total number of test samples. The smaller the LMAE, LMAPE, LRMSE, and LNRMSE metrics, the better the R-value. 2 The closer it is to 1, the higher the accuracy of the model's predictions.
[0054] Table 2 shows the error indices of the five models on the entire physical prior data test set. It can be seen that the predicted power output values of different models are very close to the calculated power values in the physical prior data. However, WB-STGNN shows the best fit to the physical prior data, with varying degrees of improvement across all evaluation metrics: compared to WB-MLP, the LMAPE index decreased by 8.087%, the LRMSE index improved by 3.464MW, and the R² increased by 4.115%, demonstrating the effectiveness of WB-STGNN in centralized spatiotemporal prediction of wind power clusters, while also reducing the workload of prediction modeling; compared to WB-CNN- GRU reduced the LMAPE index by 4.136%, improved the LRMSE index by 2.47MW, and increased the R2 by 2.771%, further indicating that CNN-GRU, as a classic spatiotemporal prediction model, still has limitations in extracting spatiotemporal features from non-Euclidean structure data. In contrast to WB-GCN-MLP and WB-GTCN-MLP, the proposed model captures and aggregates high-dimensional features such as the average simulated wind speed and calculated total power of adjacent wind farms through the GCN module, and introduces the GTCN module to further mine the temporal correlation of the aggregated features in the spatial dimension. Therefore, WB-STGNN achieved better prediction results.
[0055] Table 2. Comparison of prediction performance of the five models on the physical prior data test set.
[0056] After learning the prior physical knowledge, the five models were further trained according to different seasons. The dataset partitioning strategy was as follows: the first two months of each season were used as training samples, and the first two weeks and the last two weeks of the third month were used for the validation and testing phases, respectively. The prediction error statistics are shown in Figures 10(a)-10(d), where red represents WB-STGNN, orange represents WB-GTCN-MLP, blue represents WB-GCN-MLP, green represents WB-CNN-GRU, and pink represents WB-MLP. LMAE(pu) and LRMSE(pu) are values normalized to 10% (i.e., 15.7MW) of the total installed capacity of the wind power cluster. In addition, one representative week was selected from the test sets of the four seasons, and the actual active power output of the wind power cluster and the predicted wind power of different models were plotted as shown in Figures 11(a)-11(d). In the figure, black represents the actual power, green represents WB-CNN-GRU, orange represents WB-GTCN-MLP, pink represents WB-MLP, blue represents WB-GCN-MLP, and red represents WB-STGNN.
[0057] Combining Figures 10(a)-10(d) and 11(a)-11(d), it can be clearly seen that after formal training, the predicted values obtained by the five models are basically consistent with the actual values in most cases. Specifically, whether during the rapid rise and fall of wind power output on May 20th in spring, or during the low output fluctuation period from October 19th to 20th, the WB-STGNN model can accurately fit the active power output change trend of wind power clusters in actual operating scenarios. Moreover, the prediction accuracy in spring and winter is higher than other comparative methods. The LMAE index is reduced by 127kW~3562kW compared with other models, the LMAPE index is reduced by 1.141%~10.596%, and the LRMSE is also reduced by 390kW~4595kW accordingly, with an average R2 of 92.882%. The average LMAPE in summer and autumn is 20.618%, slightly lower than the 20.654% of WB-GTCN-MLP, further demonstrating the superiority of the WB-STGNN model for wind power cluster power prediction. Secondly, leveraging gating mechanisms and two-dimensional convolutional networks to extract local spatial features, WB-GTCN-MLP also demonstrates impressive prediction performance. Particularly noteworthy are the two consecutive peaks observed on October 24th and 25th, and the high-output oscillation period of the wind power cluster on February 27th during winter, where WB-GTCN-MLP more closely approximates the measured power curves. WB-CNN-GRU's prediction performance in summer is slightly inferior to other comparative models. Compared to WB-STGNN, its LMAE increased by 2.033 MW, and its LNRMSE improved by 14.682%, indirectly reflecting the advantage of WB-STGNN in handling complex spatiotemporal coupling relationships between wind farms. The WB-GCN-MLP performs well in summer and autumn, achieving the best prediction result when the power peak occurs on August 24, almost matching the actual power. However, in spring and winter, the overall error of the WB-GCN-MLP is larger, with significant over-prediction in some periods, such as the continuous power fluctuations on February 22 and May 21. This may be because although GCN can uncover the spatial relationship between three irregularly distributed wind farms, it lacks the structure to process time series. According to the power prediction curve trend of WB-MLP, the MLP model is still insufficient to capture the time series dependence of large and long-term fluctuating wind power output after learning prior knowledge. Therefore, the prediction accuracy of WB-GCN-MLP is greatly affected by the MLP model's ability to extract time series features.
[0058] Based on the prediction errors of the five models under different incoming wind directions and speeds, LMAE(pu) and LRMSE(pu) are normalized values based on 10% (i.e., 15.7MW) of the total installed capacity of the wind power cluster. Analysis shows that WB-STGNN has the lowest overall LMAE index. Compared to WB-GTCN-MLP, it reduces LMAE by approximately 321.3kW and 181.8kW at the highest wind direction frequency of 80°~100° and the highest wind speed frequency of 4~6m / s, respectively. Compared to WB-GCN-MLP, it reduces LMAE by approximately 4426.6kW and 2306.2kW, respectively, demonstrating WB-STGNN's strong adaptability and prediction stability under most wind conditions. In particular, when the incoming wind direction angle is between 240° and 260° or when encountering extreme wind speeds of 16 to 18 m / s (these wind conditions occur least frequently, with only 3 occurrences), the performance of WB-STGNN is relatively average. This may be due to the limited number of training samples. Future work will supplement the samples for low-frequency wind direction and speed ranges to improve the generalization performance of the WB-STGNN model under all possible wind conditions.
[0059] Step S6: To further verify the effectiveness of the proposed dynamic adjacency matrix and model for learning physical prior data, which takes into account the wake and blocking effects between wind farms, this application designed three comparative models: STGNN-PCC, STGNN-E, and STGNN. Wind speeds and historical power data from November 2014 to August 2015 were used as the training set, September data were used as validation samples, and October data were used for model testing. Specific details of each model are described below: (1) STGNN-E: No prior physical knowledge is learned, and formal training is carried out directly. The adjacency matrix adopts Apre, that is, only the spatial proximity relationship of the station is considered. (2) STGNN-PCC: No prior physical knowledge is learned, and formal training is carried out directly. The Pearson product-moment correlation coefficient (PCC) is used to reflect the interaction between wake and blockage effects between stations, and the adjacency matrix elements are updated daily. (3) STGNN: No prior physical knowledge is learned, and formal training is carried out directly. The dynamic adjacency matrix construction method of wind power cluster proposed in this application is adopted. (4) WB-STGNN: Physical prior knowledge is learned, and only the wind speed and historical power from June to August 2015 are used for formal training. The dynamic adjacency matrix construction method of wind power clusters proposed in this application is adopted.
[0060] The test results are shown in Table 3. The last column represents the total time for the model to learn prior physical knowledge and undergo formal training. Figure 12 Figures (a) and (b) show a comparison of power prediction errors for different models (with different wind angles and speeds). Figures 13(a)-13(c) show the curves of changes in the average measured wind speed and incoming wind direction at the stations during the first two weeks of July 2015, as well as heatmaps of the daily calculation results for DTW and PCC. Analysis leads to the following conclusions: Compared to STGNN-E, which only considers the constant graph topology, both STGNN and the WB-STGNN model proposed in this application achieve higher prediction accuracy, demonstrating the importance of considering the dynamic spatiotemporal correlations between stations. Compared to STGNN-PCC, STGNN performs better in prediction, with a reduction of 1.726MW in the LRMSE index. This indicates that when two wind farms are geographically close, DTW can measure the difference in the overall incoming wind speed under the same wind direction at adjacent time points by using the minimum regular distance between time series, thus better characterizing the strength of the wake and blockage effects between wind farms. In contrast, PCC can only measure the linear relationship between sequences, with a calculation result close to 1. Therefore, STGNN-PCC and STGNN-E have similar prediction accuracy, but the former has a longer training time because calculating the adjacency matrix once a day significantly increases the computational load of the model, and the same applies to STGNN. It is worth mentioning that although WB-STGNN uses the least amount of observation data for training, and therefore requires the shortest training time, its prediction performance is better than STGNN, which is trained using only actual observation data. This indicates that through learning from physical prior data, WB-STGNN has mastered a series of physical laws and patterns for each wind farm, from capturing incoming wind speed to generating wind power. As a result, the model can adapt to the observation data more quickly during formal training, thus improving prediction performance.
[0061] Table 3. Validation considering physical information and dynamic adjacency matrix
[0062] For those skilled in the art, various modifications and improvements can be made without departing from the concept of this application, and these all fall within the scope of protection of this application. Therefore, the scope of protection of this application shall be determined by the appended claims.
[0063] Those skilled in the art will understand that all or part of the steps involved in implementing the methods of the above embodiments can be implemented by a program instructing related hardware, and the program can be stored in a computer-readable storage medium. When the program is executed, it will include one or a combination of the steps of the method embodiments.
[0064] The storage medium mentioned above can be a read-only memory, a disk, an optical disk, etc. Although embodiments of the present invention have been shown, it should be understood that the above embodiments are merely exemplary and should not be considered as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of the present invention.
Claims
1. A method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information, characterized in that, The method includes: Step 1: Obtain relevant information data of each wind farm in the wind power cluster, including wind turbine parameters, geographical location information, historical operation data and meteorological data, and preprocess the relevant information data of each wind farm in the wind power cluster. Step 2: Based on the relevant information data of each wind farm in the preprocessed wind power cluster, a physical prior dataset is formed by integrating the wake analysis model and the blocking effect model into an engineering analytical model. The construction process of the engineering analytical model that integrates the wake analytical model and the blocking effect model is as follows: A first-order Jensen wake analytical model is derived based on the law of conservation of mass. It is assumed that the velocity distribution in the wake region downstream of the wind turbine approximately follows a uniform cap-shaped distribution, and that the initial wake region diameter is the same as the rotor diameter. The wake region wind speed is then determined based on the derived first-order Jensen wake analytical model. and diameter The calculation expression is: , , In the formula, This represents the distance from the downstream of the fan to the rotor plane. The free-flow velocity at the height of the wind turbine hub. This represents the thrust coefficient of the wind turbine. Where is the radius of the wind turbine. is the wake spread factor, which characterizes the linear diffusion rate of the wake influence range downstream of the wind turbine. The VC blocking effect model is used to evaluate the blocking effect of downstream wind turbines on upstream wind turbines. The geometry of the downstream wake region of the wind turbine is defined as having a cross-sectional radius equal to the rotor radius. And it has a constant tangential vorticity A semi-infinite cylinder is used, and the Jensen wake analytical model is introduced to calculate the wind speed in the wake region, as well as the induced wind speed in the upstream region of the wind turbine affected by the blockage effect. The simplified form obtained from the Taylor formula expansion is: , In the formula, The average axial inductance coefficient of the wind turbine plane. ; Step 3: Using each wind farm as a node in the graph, construct a graph structure for wind farm power prediction. Update the adjacency matrix daily based on wind speed sequence similarity using a dynamic time warping algorithm to form a dynamic adjacency matrix; including: A non-Euclidean graph structure is constructed to characterize the influence relationships between wind farms, consisting of a set of nodes. The set of edges Composition, represented as ,in Represents wind power clusters One wind farm, Representing the One wind farm, , No. The characteristic matrix of a wind farm is represented as follows: , For the input feature dimension, It reflects the coupling relationship between wind farms. Representing the Characterization of the coupling relationship between a wind farm and its neighboring wind farms. ; The dynamic time warping algorithm is used to calculate the time warping of any two wind farms at the twentieth moment. The minimum regularization distance between the intraday wind speed time series predicted by the weather forecast A weighted adjacency matrix for the wind power cluster is constructed using the average of the minimum normalized distances over a 24-hour period. The elements of this weighted adjacency matrix are dynamically updated daily. Let be a symmetric matrix, defined as follows: , In the formula, , They represent the first Heavenly The, the Wind speed sequence of a wind farm For the first Heavenly The and the first The edge weights between wind farms are defined as follows: , In the formula, The minimum normalized distance between wind speed sequences These are the normalization coefficients; Defined as: , In the formula, Indicates the first Heavenly Each wind farm at any time Simulated incoming wind speed time series Indicates the first Heavenly Each wind farm at any time Simulated incoming wind speed time series and The value is 24. For two time series starting from the initial time point By the end time point The set of all valid paths that satisfy the dynamic regularization constraint; Step 4: Construct a spatiotemporal graph neural network model architecture that includes a graph convolutional neural network module, a gated temporal convolutional network module, and a multilayer perceptron; Step 5: A two-stage training process is adopted. First, the spatiotemporal graph neural network model architecture is pre-trained using the physical prior dataset. Then, the pre-trained spatiotemporal graph neural network model architecture is formally trained based on historical power data and dynamic adjacency matrix to obtain the spatiotemporal graph neural network model. Step 6: Input the wind speed for the predicted day, and use the spatiotemporal neural network model to predict the active power of the entire wind power cluster for 24 hours on the predicted day.
2. The method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information, as described in claim 1, is characterized in that... In step 1, the wind turbine parameters include hub height, rotor diameter, power curve and thrust coefficient; historical operating data includes wind speed, power and temperature from the SCADA system; and meteorological data includes NWP forecast data and MERRA-2 reanalysis data. Preprocessing of relevant information and data from each wind farm in the wind power cluster includes: Remove shutdown data and various power limitation point data from the relevant information data of each wind farm in the wind power cluster; Replace all negative values with zero. Adjust the data time resolution to the preset precision and complete any missing data; The latitude and longitude wind speed data from the MERRA-2 reanalysis data and NWP forecast data were converted into full wind speed data. and wind direction angle The formula is defined as follows: (The formula is in the form of...) , , , In the formula, and These represent the meridional wind speed and the zonal wind speed, respectively. It is a two-parameter arctangent function. and These are the incoming wind speed at the height of the wind turbine hub and the speed relative to the reference height. Wind speed conversion factor, Wind speed at reference altitude.
3. The method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information, as described in claim 2, is characterized in that... Step 2 includes: Construct an engineering analytical model that integrates the wake analysis model and the blocking effect model; The preprocessed MERRA-2 reanalysis data, NWP forecast data, wind turbine hub height, rotor diameter and geographical location information are input into the engineering analytical model. The flow field of the wind power cluster is simulated through the Pywake platform to calculate the simulated incoming wind speed and theoretical calculated power of each wind farm, forming a physical prior dataset. The preprocessed MERRA-2 reanalysis data includes the wind speed, wind direction and turbulence intensity at the reference location of the wind power cluster area.
4. The method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information, as described in claim 1, is characterized in that... Step 4 includes: Using the predicted wind speed of the 24-hour NWP (Non-Power Flow) as input and the active power of each wind farm in the wind power cluster as output, a spatiotemporal graph neural network model architecture is constructed. The spatiotemporal graph neural network model architecture includes three gated temporal convolutional network modules, one graph convolutional neural network module approximated by a first-order Chebyshev polynomial, and a multilayer perceptron. The total power prediction value of the wind power cluster is obtained by summing the active power of each wind farm in the wind power cluster during the 24-hour prediction period.
5. The method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information, as described in claim 1, is characterized in that... Step 5 includes: Step 5.1: Construct an adjacency matrix for learning physical prior knowledge of the spatiotemporal graph neural network model architecture. Establish simulated incoming wind speed for wind farms With calculated power The nonlinear mapping relationship between them is used to pre-train the spatiotemporal graph neural network model architecture; Step 5.2 involves transferring the pre-trained spatiotemporal graph neural network model architecture from Step 5.1 and constructing a dynamic adjacency matrix for formal training. By using measured wind speed and historical power data from wind power clusters, the network parameters of each layer of the pre-trained spatiotemporal graph neural network model architecture are adjusted to obtain the spatiotemporal graph neural network model.
6. The method for short-term power prediction of wind power clusters using a spatiotemporal graph neural network that integrates physical information, as described in claim 5, is characterized in that... The loss function in the pre-training process in step 5.1 Using the Huber loss function, the expression is: , In the formula, and These were obtained from the engineering analytical model. The calculated power of a wind farm and the predicted value obtained from the pre-trained spatiotemporal graph neural network model architecture; The loss function in the formal training process in step 5.2 Defined as the power predicted by a formally trained spatiotemporal graph neural network model for all wind farms. With actual power The average deviation between them is expressed as: 。 7. A short-term power prediction device for wind power clusters based on a spatiotemporal graph neural network incorporating physical information, characterized in that, The apparatus employs the method as described in any one of claims 1 to 6, comprising: The acquisition unit is used to execute step 1; Physical information modeling unit, used to perform step 2; Graph structure building unit, used to perform step 3; The spatiotemporal graph neural network modeling unit is used to perform step 4; Training unit, used to perform step 5; The prediction unit is used to perform step 6.
8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.
Citation Information
Patent Citations
Wind power plant cluster power prediction method and system based on multi-scale space-time correlation
CN119674966A
Wind power ultra-short-term prediction method and device based on graph convolutional neural network
CN120433175A
