Wind power plant power prediction method and system based on digital twinning

Abstract

The invention discloses a wind power plant power prediction method and system based on digital twinning, and relates to the technical field of wind power plant power prediction. According to the method, ConvLSTM + GCN is used as a baseline model for improvement, a single-layer convolution layer with the step length equal to the size of a convolution kernel is cascaded behind ConvLSTM, and a feature map output by the ConvLSTM is compressed; through flattening operation and position coding, the geographic position information of the fan can still be reserved after the data is subjected to spatial dimension reduction of a convolutional layer, so that the model can sense and utilize the spatial position relationship; an algorithm of an adjacent matrix in a traditional GCN is improved, the adjacent matrix is dynamically updated through cross correlation calculation, and the real-time correlation strength between fans is quantified; meanwhile, 3D modeling is carried out on the fan, a virtual-real fusion visual monitoring system is constructed in combination with the digital twinborn technology, and the method is suitable for intelligent operation and maintenance, predictive maintenance and optimal scheduling of a wind power plant.

Description

Technical Field

[0001] The present invention relates to the technical field of wind farm power prediction, and in particular to a wind farm power prediction method and system based on digital twins. Background Art

[0002] As an important form of clean energy generation, wind farms not only effectively reduce greenhouse gas emissions such as carbon dioxide and mitigate global warming, but also provide a reliable and clean solution for energy supply. Therefore, efficient management and optimization of wind farms, especially accurate power forecasting, are key to improving wind energy utilization and reducing operating costs. With the development of digital twin (DT) technology, its application potential in wind farms is becoming increasingly prominent. By constructing a virtual model corresponding to the physical wind farm, this technology enables real-time monitoring, optimized scheduling, and accurate forecasting of wind farms, further improving the operational efficiency and reliability of wind farms.

[0003] Digital twins have been evolving continuously since they were first proposed by Professor Michael Grieves at the University of Michigan in 2003. In 2011, the U.S. Air Force Research Laboratory introduced the concept of "digital twins" to fighter aircraft maintenance, propelling the technology into engineering practice. After 2014, the scope of digital twins expanded to encompass the entire product lifecycle, from design simulation to production optimization and operations management. In 2021, ZTE released its AR point cloud digital twin platform, marking the maturity of high-precision modeling and real-time data fusion technology. Currently, digital twins are deeply integrated with AI, enabling autonomous decision-making through deep learning algorithms. Application scenarios have expanded from single devices to city-level systems, making them a common technology driving digital transformation across various industries.

[0004] Digital twin technology enables digital management of wind farms throughout their entire lifecycle by creating virtual representations of physical entities. Its core lies in the deep integration of the real and the virtual: At the real level, power generation forecasts are achieved using time series prediction methods, such as deep learning algorithms like ConvLSTM and graph convolutional networks. At the virtual level, 3D visualizations are constructed and rendered in real time using technologies like Unity, 3DsMAX, and GIS. This organic integration of the two opens up innovative avenues for intelligent wind farm operation and maintenance, fault warning, and optimized scheduling.

[0005] With the rapid development of deep learning technology, neural network models have demonstrated significant advantages in processing time series data. Recurrent Neural Networks (RNNs), proposed in 1982, and Long Short-Term Memory (LSTMs), introduced in 1997, are particularly prominent. LSTMs effectively control the flow of information by introducing a gating mechanism, extending the retention of valid information from 20 steps in RNNs to over 200 steps. To further integrate spatiotemporal features, researchers proposed the Convolutional Long Short-Term Memory (ConvLSTM) network. Its core innovation is the integration of convolution operations into the LSTM gating mechanism, resulting in an 18% improvement in prediction accuracy compared to traditional LSTM models.

[0006] In the field of time series prediction, traditional models focus on capturing temporal dependencies, but have limitations in modeling spatial dependencies. Especially in prediction tasks involving multiple spatial units (such as power equipment clusters), effectively capturing the spatial correlation between units becomes the key. To this end, researchers proposed a model that combines LSTM with spatiotemporal graphs: capturing spatial dependencies through graph convolutional networks (GCNs), using LSTM to learn temporal dependencies, and realizing collaborative processing of spatiotemporal dimensional dependencies, thereby improving prediction accuracy. However, the existing technology still has defects in prediction accuracy and efficiency. To address this problem, the present invention proposes a wind farm power prediction method and system based on digital twins. Summary of the Invention

[0007] The purpose of the present invention is to propose a wind farm power prediction method and system based on digital twins to solve the problems raised in the background technology. The present invention integrates deep learning algorithms (ConvLSTM, GCN) and three-dimensional modeling technology, and realizes accurate prediction of wind farm power through an improved spatiotemporal collaborative model. It also combines digital twin technology to construct a virtual and real integrated visual monitoring system, which is suitable for intelligent operation and maintenance, predictive maintenance and optimized scheduling of wind farms.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] A wind farm power prediction method based on digital twins includes the following steps:

[0010] S1. Design a spatiotemporal collaboration algorithm to optimize time series prediction and spatial structure, improve ConvLSTM and graph neural networks, and use the improved spatiotemporal collaboration model to predict wind turbine power, overcoming the local perception limitations of traditional ConvLSTM networks and the static constraints of traditional graph neural network models.

[0011] S2. Based on 3D modeling technology, construct a virtual model of the wind turbines in the wind farm;

[0012] S3. Combine S1 and S2, build a wind farm power prediction model based on digital twin technology, train the constructed model, and use the trained model to complete the wind farm power prediction work.

[0013] Preferably, the S1 specifically includes the following contents:

[0014] S1.1. Based on the traditional ConvLSTM network, a single convolutional layer with a stride equal to the convolution kernel size is cascaded behind it. The convolutional layer is used to compress the spatial dimension and condense the feature maps output by the ConvLSTM, aggregating each local spatiotemporal feature into a single feature point.

[0015] S1.2. By flattening and position encoding, the feature points obtained in S1.1 are added with geographic location information to facilitate the model to perceive and utilize the spatial position relationship of the feature points. After processing through the linear layer, the feature enhancement matrix is obtained.

[0016] S1.3. Construct a complete graph based on the locations of wind turbines in the wind farm, improve the traditional GCN algorithm, dynamically update the adjacency matrix through cross-correlation calculation, and quantify the real-time correlation strength between wind turbines in the wind farm;

[0017] S1.4. Based on the improved GCN algorithm, graph convolution is performed on the two-dimensional feature enhancement matrix and the adjacency matrix, and the obtained results are processed by the linear layer and then output as the predicted power of the wind turbines in the wind farm.

[0018] Preferably, all wind turbines in the complete graph described in S1.3 have connecting edges between each other, and the number of edges is calculated as follows:

[0019]

[0020] Where m represents the number of edges and n represents the number of fans.

[0021] Preferably, the method of dynamically updating the adjacency matrix by cross-correlation calculation described in S1.3 specifically includes the following:

[0022] Assume that the eigenvectors of wind turbines i and j are L i and L j , the cross-correlation calculation formula between fans is:

[0023]

[0024] in, is the dot product of wind turbines i and j, indicating their similarity; ||L i|| and ||L j || are the eigenvectors L i and L j The modulus length; through the above cross-correlation calculation, we get the n×n dimensional adjacency matrix C, each element C ij represents the similarity between wind turbines i and j, C ij The larger it is, the stronger the spatial dependence between wind turbines i and j is. The adjacency matrix C is used to represent the spatial relationship between wind turbines and represents the connection strength between their nodes in the space-time graph.

[0025] Preferably, the 3D modeling tool in S2 includes Unity.

[0026] A wind farm power prediction system based on digital twins includes the following modules:

[0027] The spatiotemporal collaboration module improves ConvLSTM and graph neural networks for wind turbine power prediction;

[0028] 3D modeling module, used to build virtual models of wind turbines in wind farms;

[0029] The digital twin module, coupled with the spatiotemporal collaboration module and the 3D modeling module, is used to build a wind farm power prediction model based on digital twins and complete the power prediction of wind turbines in the wind farm.

[0030] Preferably, the development tool of the system includes PyQt5.

[0031] The present invention further protects a computer device, which includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the above-mentioned wind farm power prediction method based on digital twins.

[0032] The present invention further protects a computer-readable storage medium, which stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by a processor to implement the above-mentioned wind farm power prediction method based on digital twins.

[0033] Compared with the existing technology, the present invention provides a wind farm power prediction method and system based on digital twins, which has the following beneficial effects:

[0034] (1) This paper uses "ConvLSTM+GCN" as the baseline model. By adding different components (convolutional layer, cross-correlation) in the ablation experiment, we can clearly see the performance gain brought by each improved part to the algorithm, and clarify the necessity of components in spatiotemporal feature extraction and dependency modeling.

[0035] (2) The comparative charts generated based on the training logs in the embodiments of the present invention clearly demonstrate the dual advantages of the improved model in terms of training loss convergence and validation loss stability, fully verifying its comprehensive improvement in feature learning, anti-overfitting and generalization performance, and providing more reliable performance support for the application of the model in practical scenarios;

[0036] (3) This paper developed a wind farm digital twin system and conducted the following functional tests on each module: the login module verifies the user's login response; the data transmission module checks the stability of the WebSocket connection and data integrity; and the visualization module tests the smoothness of 3D scene rendering and the real-time nature of data updates. Testing showed that login verification took less than 1 second, data transmission latency was within 5 seconds, and the Unity runtime frame rate remained stable at 50 fps, meeting the requirements for real-time monitoring and forecasting of wind farms.

[0037] In summary, the wind farm power prediction method and system based on digital twins proposed in the present invention integrate deep learning algorithms (ConvLSTM, GCN) and three-dimensional modeling technology, realize accurate prediction of wind farm power by improving the spatiotemporal collaborative model, and combine digital twin technology to build a virtual-reality integrated visual monitoring system, which is suitable for intelligent operation and maintenance, predictive maintenance and optimized scheduling of wind farms. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a schematic diagram of the overall network structure mentioned in Example 1 of the present invention;

[0039] Figure 2 This is the 3D model diagram of the wind turbine mentioned in Example 1 of the present invention;

[0040] Figure 3 This is a schematic diagram of the main interface of the system mentioned in Example 1 of the present invention;

[0041] Figure 4 This is a schematic diagram of the ConvLSTM memory unit mentioned in Example 1 of the present invention;

[0042] Figure 5 This is a schematic diagram of the ConvLSTM layer mentioned in Example 1 of the present invention;

[0043] Figure 6 Schematic diagram of the feature aggregation process through a single convolutional layer mentioned in Example 1 of the present invention;

[0044] Figure 7 This is a schematic diagram of the flattening operation mentioned in Example 1 of the present invention;

[0045] Figure 8 This is a schematic diagram of the improved GCN that introduces cross-correlation mentioned in Example 1 of the present invention;

[0046] Figure 9 Schematic diagram of the undirected graph G and its adjacency matrix mentioned in Example 1 of the present invention;

[0047] Figure 10 This is a schematic diagram of the adjacency matrix calculation code mentioned in Example 1 of the present invention;

[0048] Figure 11 This is a schematic diagram of the wind turbine construction mentioned in Example 1 of the present invention;

[0049] Figure 12 The real fan and the model fan mentioned in Example 1 of the present invention;

[0050] Figure 13 This is a schematic diagram of the WebSocket server configuration mentioned in Example 1 of the present invention;

[0051] Figure 14 This is a schematic diagram of the WebSocket client configuration mentioned in Example 1 of the present invention;

[0052] Figure 15 This is the system flow chart mentioned in Example 1 of the present invention;

[0053] Figure 16 This is a schematic diagram of the main interface of the system mentioned in Example 1 of the present invention;

[0054] Figure 17 Schematic diagram of the ablation experiment results mentioned in Example 2 of the present invention;

[0055] Figure 18 This is a loss comparison diagram mentioned in Example 2 of the present invention;

[0056] Figure 19 is the prediction error distribution histogram mentioned in Example 2 of the present invention;

[0057] Figure 20 This is a scatter plot of predicted values and true values mentioned in Example 2 of the present invention;

[0058] Figure 21 This is the overall system architecture mentioned in Example 3 of the present invention;

[0059] Figure 22 This is a schematic diagram of the login interface mentioned in Example 3 of the present invention;

[0060] Figure 23 This is a schematic diagram of the error message prompt mentioned in Example 3 of the present invention;

[0061] Figure 24 This is a schematic diagram of the registration interface mentioned in Example 3 of the present invention;

[0062] Figure 25 This is a schematic diagram of real-time data viewing mentioned in Example 3 of the present invention;

[0063] Figure 26 This is a schematic diagram of the range mentioned in Example 3 of the present invention;

[0064] Figure 27 This is a schematic diagram of wind turbine deployment mentioned in Example 3 of the present invention;

[0065] Figure 28 This is a schematic diagram of the prediction data mentioned in Example 3 of the present invention;

[0066] Figure 29 This is a schematic diagram of the wind turbine twins mentioned in Example 3 of the present invention;

[0067] Figure 30 This is a schematic diagram of user management mentioned in Example 3 of the present invention;

[0068] Figure 31 This is a schematic diagram of password modification mentioned in Example 3 of the present invention. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

[0070] Example 1:

[0071] This paper proposes a wind farm power prediction method based on deep learning and designs a wind farm digital twin system in combination with 3D modeling technology. The main contents are as follows:

[0072] The algorithm for power prediction of wind farms has been improved. The power prediction of wind farms often needs to take into account the temporal and spatial characteristics of wind farms. Therefore, the present invention uses "ConvLSTM+GCN" as the baseline model for improvement. The local perception ability of ConvLSTM is usually limited by the size of the convolution kernel, so a single-layer convolution layer with a step size equal to the convolution kernel size is cascaded behind ConvLSTM to compress the feature map output by ConvLSTM, which can better extract local features, reduce the amount of calculation, and fuse features at different times to obtain a more effective feature representation; through flattening operations and position encoding, the geographical location information of the wind turbines can still be retained after the data passes through the spatial dimensionality reduction of the convolution layer, allowing the model to perceive and utilize these spatial position relationships; using GCN to jointly process the spatial structure of the wind farm, the present invention improves the algorithm of the adjacency matrix in the traditional GCN, dynamically updates the adjacency matrix through cross-correlation calculation, and quantifies the real-time correlation strength between wind turbines (such as wake effect). Based on the above improvements, an improved spatiotemporal collaborative model (ConvLSTM+single-layer convolution layer+cross-correlation calculation+GCN) is obtained, such as Figure 1 shown.

[0073] The wind turbine was 3D modeled. The 3D modeling of the wind turbine was completed using the Unity platform. Using the Goldwind Technology GW150-4.0MW wind turbine as a reference, its size was scaled at a ratio of 10:1 and then 3D modeled. The 3D model incorporated the wind turbine's physical properties, rotational characteristics, and UI display interface, which can be applied to the digital twin system. Figure 2 shown.

[0074] A digital twin system for wind farms was designed. The improved spatiotemporal collaborative model was used as a predictor to predict the wind farm power for the next half hour. After obtaining the prediction results, cross-platform data transmission between Pycharm and Unity was carried out through Websocket, and the current data and predicted data of the wind turbines were displayed in real time in Unity. After testing, the data transmission delay was within 5s. A visual interface was also developed, which can view various current data of the wind farm in real time in the form of a line graph. Such as: wind speed, cabin temperature, ambient temperature and other indices; the main interface of the system is as follows: Figure 3 shown.

[0075] Based on the above content, the following is an explanation with reference to relevant drawings and specific examples. The specific content is as follows.

[0076] (1) Improved wind power prediction algorithm

[0077] In terms of time series prediction, traditional ConvLSTM networks face the problem of spatiotemporal decoupling in wind farm cluster power prediction: although local spatial features are captured through convolution operations and time series modeling is performed in combination with LSTM units, it is limited by the fixed-size convolution kernel receptive field (such as 3×3 or 5×5), and cannot effectively capture long-distance spatial dependencies between wind turbines. In addition, time dimension modeling relies only on recurrent connections, which makes it difficult to cope with dynamic features such as non-stationary wind speed mutations. In terms of spatial structure, although traditional graph neural network models achieve global spatial correlation modeling through static graph structures (such as graphs built based on geographic distance or historical correlation), they have significant defects: first, the graph structure relies on prior knowledge and cannot dynamically reflect real-time wind field changes; second, the lack of an explicit time dimension processing mechanism makes it difficult to capture periodic fluctuations and trend evolution in time series.

[0078] To address these issues, the present invention proposes a spatiotemporal collaborative algorithm that improves both ConvLSTM and graph neural networks. Module 1 utilizes a cascaded structure of ConvLSTM and a single convolutional layer. While preserving the spatial structure of the original scene graph, it adds a single convolutional layer to efficiently abstract local spatiotemporal features. Module 2, based on complete graph construction and node cross-correlation calculation, uses graph convolution to capture real-time spatial correlations and time series trends between feature clusters. This design overcomes the local perception limitations of ConvLSTM and the static constraints of traditional graph neural network models, improving prediction accuracy.

[0079] (1) Module 1: Cascading a single convolutional layer after ConvLSTM

[0080] See also Figure 4 , Figure 4 The ConvLSTM model consists of two main parts: the temporal structure follows the typical RNN architecture; the spatial structure follows the CNN feature extraction method. Simply put, the ConvLSTM model is equivalent to replacing all fully connected structures in the LSTM with convolutional structures. The motivation for the ConvLSTM model is to overcome the limitations of traditional RNNs and CNNs by combining CNNs and LSTMs. While traditional LSTMs use fully connected layers to process input data, ConvLSTMs introduce convolution operations on top of LSTMs. Convolutional layers are used to process spatial data.

[0081] See also Figure 5 , Figure 5This is the ConvLSTM layer in module 1. Its input is a three-dimensional tensor X, which contains data about the wind farm at different points in time. The input at each moment is a two-dimensional image containing various features, such as wind speed and direction, representing the spatial layout of the wind turbine cluster and the characteristics of each turbine at that moment. The ConvLSTM layer extracts the spatiotemporal features of X shallowly, outputting a three-dimensional matrix X'.

[0082] After ConvLSTM outputs the three-dimensional matrix X', the present invention performs spatial dimension compression and feature aggregation on it through a cascaded single-layer convolution layer after ConvLSTM. Figure 6 As shown in the figure, this convolutional layer uses a 5×5 kernel size and a stride of 5, compressing the original feature map by a 5:1 ratio. By covering independent regions with non-overlapping windows, each 5×5 local spatiotemporal feature is aggregated into a single feature point, reducing spatial redundancy while preserving regional spatial context. This design leverages the local receptive field characteristics of the convolution operation and ensures computational integrity through a fixed step size, enabling each output feature map to maintain its temporal dimension while also streamlining its spatial dimension. The resulting three-dimensional feature matrix retains the spatiotemporal correlation information of the original input while highlighting key features through abstract processing. This improves computational efficiency while providing an effective feature representation for subsequent analysis.

[0083] After the spatial features of the data are concentrated, each local block will be aggregated into a feature extraction point. When the data enters the module two operation, these feature points will be treated equally by the model, and the relative position spatial information of the two-dimensional image will be lost, that is, each point does not reflect the spatial position difference that should be in the image in the calculation (such as a point originally located in the upper left corner of the image, and another point located in the lower right corner). However, the relative position of each element in the image itself contains important spatial information (such as the distribution layout and spatial structural relationship of the wind turbine), which is crucial to understanding the image content and analyzing spatial logic. When the positional relationship of the feature points is equalized, these key spatial position information is lost. To make up for this defect, it is necessary to artificially supplement the geographic location information of the feature points through flattening operations and position encoding technology, so that the model can perceive and utilize these spatial position relationships, so as to complete subsequent tasks more accurately. Flattening operations such as Figure 7 As shown in the figure, after flattening, the geographic location information of each point will be added through position encoding, and the feature enhancement matrix L will be obtained after processing through the linear layer.

[0084] (2) Module 2: Improve the GCN algorithm of the adjacency matrix to strengthen the connection between wind turbines

[0085] See also Figure 8 , Figure 8This is the structure of module 2, which processes the enhanced matrix L output by module 1. First, a complete graph is constructed based on the location of the wind turbines, that is, all wind turbines have connecting edges between them. The edge number calculation process is as follows:

[0086]

[0087] Where m is the number of edges and n is the number of wind turbines.

[0088] After constructing the complete graph, the model needs to calculate the cross-correlation between the feature enhancement matrix L output by module 1 and its transpose to obtain the dynamic spatial adjacency matrix C between wind turbines. Traditional spatiotemporal graph models usually rely on prior knowledge (such as physical proximity) to preset node connection edges, resulting in limited modeling capabilities for non-local associations. Figure 9 It is the undirected graph G and its original adjacency matrix A, which represents the nodes connected to each node. In Model 2, the present invention improves the calculation method of the graph convolution adjacency matrix and generates the adjacency matrix by dynamically calculating the cross-correlation of node features. Cross-correlation is a common method to measure the similarity between two vectors. Suppose the eigenvectors of wind turbines i and j are L i and L j , the cross-correlation calculation process between fans is as follows:

[0089]

[0090] in, represents the dot product of wind turbines i and j, indicating their similarity; ||L i || and ||L j || are the eigenvectors L i and L j The mold length.

[0091] Through the above cross-correlation calculation, we get the n×n dimensional adjacency matrix C, where each element C ij represents the similarity between wind turbines i and j, C ij The larger the value, the stronger the spatial dependence between wind turbine i and wind turbine j. The adjacency matrix C is used to represent the spatial relationship between wind turbines, that is, the connection strength between their nodes in the space-time graph.

[0092] The calculation code of the adjacency matrix is as follows Figure 10 shown.

[0093] After completing the construction of the complete graph and obtaining the adjacency matrix C, the model performs graph convolution on the two-dimensional feature enhancement matrices L and C based on GCN, and outputs the predicted power of the wind turbine after processing the obtained results through the linear layer.

[0094] (2) 3D modeling of the wind turbine and achieving virtual-real integration

[0095] (1) Construction of virtual model of wind turbine

[0096] As the "virtual" part of the digital twin, 3D modeling of wind turbines is essential. As the industry's leading real-time 3D development engine, Unity, with its efficient real-time rendering capabilities, integrated physics engine (such as simulation of physical effects such as wind and rotation), convenient interactive development tools, and cross-platform deployment advantages, can accurately build virtual models of wind turbines, support dynamic data access and scene interaction, and become an ideal technical carrier for realizing digital twins of wind turbines. Therefore, this paper chooses Unity to complete the 3D modeling of wind turbines.

[0097] Table 1 Component dimensions and scaling

[0098]

[0099] Using Goldwind's GW150-4.0MW wind turbine as a reference, we converted the actual dimensions (hub height 100m, blade length 73.5m) into model dimensions using engineering modeling specifications. This scale was set at a 1:10 ratio, converting the actual dimensions (hub height 100m, blade length 73.5m) into model dimensions (10m, 7.35m), as shown in Table 1. This ratio balances scene realism and rendering performance, meeting the typical modeling requirements of a digital twin system. The hub height and blade length are directly referenced from Goldwind's publicly available specifications, while the tower and nacelle dimensions are derived from the product's technical documentation.

[0100] First, use the extruded geometry and surface editing tools in Unity to create the blade. First, build the cross-sectional profile of the blade. Generate the three-dimensional structure of the blade through extrusion operations. Adjust the length to 7.35m (1:10 scale). Then copy two blades, a total of three blades. The blades are as follows Figure 11 (a); then use the 3D geometry creation tool to generate a cylinder as the tower body, and accurately adjust the parameters of the tower body through the components: the height is 10m, the radius is 0.45m; finally, create a rectangular block as the base, adjust the length, width and height to 3m, 3m and 0.8m respectively, and align the tower and the base through the position parameters (X:0, Y:0, Z:0) to form a stable support structure. After the blades, tower and base are completed, use the rotation tool to set the blade installation angle. For the cabin part, the cylinder and sphere are combined for modeling, and the size is adjusted to 1.95m in length and 1.2m in diameter. Finally, the three blades are assembled with the cabin to ensure that they are evenly distributed around the center of the tower to form a complete wind turbine appearance (as shown in Figure 2). Figure 11 (b) shows the structure after the blades and the nacelle are combined).

[0101] After adding various components (rigidity, rotation, etc.), exclusive materials are assigned to parts such as towers, blades, and nacelles. The tower uses metal material, sets the metalness to 0.4 and the roughness to 0.65 to simulate the texture of the metal surface; the blades use composite material, adjust the metalness to 0.15 and the roughness to 0.5, and add a normal map (strength 1.0) to enhance the surface details; the nacelle is given a high metalness material (metalness 0.6) to highlight the characteristics of the mechanical components. Preview the effect in real time through the material ball panel, and gradually fine-tune it until it matches the appearance of the real wind turbine (such as Figure 12 As shown in the figure, the material performance of the real fan and the optimized fan model).

[0102] (2) Data transmission combining virtual and real

[0103] WebSocket technology is used for real-time data transmission. The reason for choosing WebSocket technology is its full-duplex communication feature. It allows the server and client to transmit data in both directions at any time after the connection is established. It has the advantages of low latency and high efficiency, which is very suitable for the real-time data transmission requirements of this system.

[0104] On the server side, the present invention uses Python language combined with asyncio and Websockets libraries to build a WebSocket server. The main function of the server is to generate prediction data for wind farms as a predictor through an improved algorithm, and send these data to connected clients in real time. During the development and testing phase, in order to simplify network configuration and ensure the stability and security of the system, the present invention uses the local loopback address 127.0.0.1. In the initialization function of the WindFarmServer class, the host parameter is set to 127.0.0.1 by default, and the port is set to 8765. This means that the server will listen to the address and port on the local computer, waiting for the client to connect. When the server starts, it will continue to obtain prediction data for the wind farm and send this data to the connected client. If the client connection is interrupted, the server will perform corresponding cleanup operations and remove the client. The python code is as follows Figure 13 shown.

[0105] On the Unity side, you also need to implement a WebSocket client to connect to the Python server and receive real-time data. Since development and testing are done on the same computer, the Unity client can also use the local loopback address 127.0.0.1 to connect to the Python server. In the Unity C# script, create a WebSocket object and set its connection address to ws: / / 127.0.0.1:8765. The C# code is as follows Figure 14 shown.

[0106] (3) Wind farm power prediction system based on digital twin

[0107] The present invention chooses to use PyQt5 to develop the digital twin system. When designing a digital twin system, PyQt5 has multiple advantages as a graphical user interface (GUI) framework. First, PyQt5 provides a wealth of controls and components that can meet the needs of various complex interfaces, such as real-time data visualization, status monitoring, control panels, alarm systems, etc.; secondly, PyQt5 also supports the design of modern interfaces, and can create responsive and visually appealing user interfaces to ensure a smooth user experience. PyQt5's good compatibility with the Python ecosystem is also a highlight. It can seamlessly integrate other Python libraries, such as NumPy, Pandas, Matplotlib, etc., to facilitate data processing and analysis, especially in digital twin systems. When it comes to real-time data acquisition and complex data analysis, Python's powerful computing power is crucial; furthermore, PyQt5's signal and slot mechanism makes the interaction between the interface and the system background more concise and efficient, supports event-driven programming, and can respond to user operations and system status changes in real time. A flexible layout manager also enables the interface to adapt to different screen sizes and devices, ensuring system adaptability. Finally, PyQt5 has extensive documentation and community support, making it easy for developers to find relevant information and solutions, reducing development challenges. In summary, PyQt5 provides a powerful, flexible, and efficient development platform for digital twin systems, enabling developers to quickly build complex and highly interactive user interfaces, improving system usability and scalability.

[0108] The present invention not only meets the needs of real-time data query of wind farms, but also meets the needs of active power prediction of wind turbines 30 minutes later. The data acquisition part directly uses the data in the SDWPF data set, and inputs the data of 134 wind turbines in the first 48 hours into the system. The active power 30 minutes later is predicted by predicting it, and the predictor algorithm calls the improved wind farm power prediction algorithm. After running the system and logging in successfully, the user will jump to the function interface, where he can click the phase button to view the corresponding data. All data are displayed in the form of a line graph. The system flow chart is as follows: Figure 15 shown.

[0109] When the system is running, users must log in with their registered account. After successfully logging in and entering the functional interface, five functional modules will be displayed: "Wind Turbine Real-Time Data," "Wind Turbine Deployment," "Forecast Data," "Wind Turbine Twin," and "User Management." Different real-time data is displayed depending on the module selected.

[0110] The main interface of the system is as follows Figure 16 shown.

[0111] Example 2:

[0112] Based on Example 1, but different in that an experiment is designed to characterize the performance of the wind farm power prediction method and system based on deep learning proposed in the present invention, the specific contents are as follows.

[0113] (1) Power prediction part

[0114] Taking "ConvLSTM+GCN" as the baseline model, by adding different components (convolutional layer, cross-correlation) in the ablation experiment, we can clearly see the performance gain brought by each improvement part to the algorithm, and clarify the necessity of components in spatiotemporal feature extraction and dependency modeling. Figure 17 The results of the ablation experiment. Figure 17 The four models in are replaced by a, b, c, and d respectively.

[0115] Table 2 Ablation experiment

[0116]

[0117] From the experimental results in Table 2, we can see that ConvLSTM+GCN, as the baseline model, has an RMSE of 0.2783 and a MAE of 0.1635. The MAPE is higher at 103.76, and R 2 The value is 0.9405, which reflects the basic spatiotemporal modeling capability of the fusion of ConvLSTM and GCN. After the convolution layer is introduced separately on the baseline model, the RMSE increases to 0.2812 (increase of 1.04%), the MAE increases to 0.1698 (increase of 3.85%), and the R 2 The performance dropped to 0.9392 (a decrease of 0.14%). Adding convolutional layers alone failed to optimize the performance and instead degraded it. This may be because the local feature extraction of the convolutional layer overlaps with the function of ConvLSTM, or it introduces redundant calculations, interfering with the core spatiotemporal feature learning.

[0118] After introducing cross-correlation alone on the baseline model, RMSE increased slightly to 0.2795 (increase of 0.43%), MAE increased to 0.1660 (increase of 1.53%), but R 2 It rose slightly to 0.9410 (up 0.05%). This shows that cross-correlation has a certain positive effect on spatial feature modeling and can slightly improve the ability to capture spatial dependencies, but its effect is limited when used alone and fails to significantly optimize the overall performance.

[0119] After introducing convolutional layers and cross-correlation on the baseline model, the final improved spatiotemporal collaborative model was obtained. This full-component model achieved the best performance, with RMSE reduced to 0.2683 (a decrease of 3.6%), MAE reduced to 0.1582 (a decrease of 3.24%), and MAPE reduced to 101.25 (a decrease of 2.42%). 2 This indicates that the combination of convolutional layers and cross-correlation enhances the extraction of spatiotemporal features: the convolutional layers enhance temporal feature learning, while cross-correlation deepens spatial correlation modeling. The two complement each other and synergistically improve the ability to capture the spatiotemporal dependencies of wind power data, verifying the effectiveness of multi-component fusion.

[0120] Depend on Figure 18 It can be seen that the RMSE, MAE, and MAPE of the improved model are all lower than those of the baseline model. 2 Improved compared to the baseline model.

[0121] The ablation experiment clarified the role of components: adding a convolutional layer alone may have a negative impact, and the improvement of cross-correlation alone is limited; however, when the convolutional layer and cross-correlation are added together, the model performance is significantly optimized through complementary advantages, proving the necessity of multi-component fusion to enhance spatiotemporal feature modeling, verifying the superiority of the improved model in wind power cluster power prediction, and illustrating the effectiveness of the design of the present invention.

[0122] Based on the log files of the training process, extract the training loss and validation loss data of each round of the baseline model and the improved model, and generate a loss comparison chart (such as Figure 18 As shown in the figure). Dropout was used in the experiment to hide some neurons in the network according to a certain probability value, so that they do not participate in the training process to prevent overfitting during network training. The following is an analysis of the chart based on the log content:

[0123] Figure 18 (a) shows that in the initial stage, both models rapidly reduced their loss using a decreasing learning rate of 0.01. However, the log records show that the baseline model's loss decline slowed significantly after 30 epochs, ultimately stabilizing at approximately 0.068. While the improved model experienced brief fluctuations around 30 epochs, its loss continued to decline, ultimately stabilizing at approximately 0.060. This difference suggests that the improved model learns data features more efficiently. Despite a slightly higher initial loss, its structural advantages (a better feature extraction mechanism or network architecture) enable it to more fully exploit the underlying patterns in the data. Combined with the log records of the improved model's continued decline in training loss, this further demonstrates that the decreasing learning rate strategy and the dropout rate of 0.3 work together more effectively in the improved model. The high learning rate in the early stages accelerates convergence, while the small learning rate in the later stages enables refined optimization. Furthermore, dropout effectively suppresses overfitting, ensuring a continued decline in loss in the later stages of training.

[0124] Figure 18 (b) shows that the validation loss of the baseline model (yellow curve) fluctuates wildly, reflecting its insufficient generalization ability and the risk of overfitting. The improved model (red curve) shows significantly less fluctuation and exhibits a more stable overall downward trend. The more stable validation loss of the improved model recorded in the log directly confirms the effectiveness of its structural optimizations (such as more rational network layer design and feature fusion methods) combined with a dropout rate of 0.3. By randomly deactivating neurons, this effectively improves model robustness and reduces overfitting to the training data, making the model more adaptable to unseen data and continuously optimizing validation loss.

[0125] In summary, the comparison charts generated based on training logs clearly demonstrate the dual advantages of the improved model in terms of training loss convergence and validation loss stability, fully verifying its comprehensive improvements in feature learning, anti-overfitting, and generalization performance, and providing more reliable performance support for the application of the model in practical scenarios.

[0126] Figure 19 The prediction error distribution of the improved model is shown, with the horizontal axis representing the prediction error range (-3 to 3kW). The mean of the error is close to 0, the standard deviation is 0.33, the peak near 0 is sharp, and the absolute value of the error is less than 1.5kW. Figure 17 This shows that the improved model has better control over errors and stronger prediction stability.

[0127] exist Figure 20 The scatter plot shows the improved model comparing true values to predicted values over a range of -1.0 to 3.0 kW. The points are closely distributed around the diagonal line, with fewer points deviating from the diagonal. Most of these points are located in the extreme value regions, where the true values are above 2.5 kW or below 0.5 kW. Figure 18 The prediction accuracy of the improved model is demonstrated.

[0128] (2) System implementation

[0129] A wind farm digital twin system was developed using PyQT5, and the following functional tests were conducted on each module: the login module verified the user's login response; the data transmission module checked the stability of the WebSocket connection and data integrity; and the visualization module tested the smoothness of 3D scene rendering and the real-time nature of data updates. Testing showed that login verification took less than 1 second, data transmission latency was kept within 5 seconds, and the Unity runtime frame rate remained stable at 50 fps, meeting the requirements for real-time wind farm monitoring and forecasting.

[0130] Example 3:

[0131] Based on Example 1-2, but with the difference that, the wind farm power prediction method and system based on deep learning proposed in the present invention are described below with a specific example, the specific content is as follows.

[0132] (1) Data preparation

[0133] This paper uses the SDWPF (Spatial Dynamic Wind Power Forecasting) public dataset, which contains real-world data from Longyuan Group Co., Ltd. (the largest wind power producer in China and Asia). This dataset was acquired using the wind farm's Supervisory Control and Data Acquisition (SCADA) system. The wind farm used for the dataset consists of 134 turbines. The SCADA system samples data from each turbine every 10 minutes for 245 days, resulting in a total of 4,727,520 data points (6 × 24 × 245 × 134).

[0134] (2) Model training

[0135] The split ratio of the training set and the test set is 1:1, and the wind power output is predicted for the next 30 minutes. The final algorithm output Y is in the shape of R 1×134 A one-dimensional vector, where each element corresponds to the predicted power value of a wind turbine 30 minutes later, is used to achieve batch prediction of the active power of 134 wind turbines, namely:

[0136] Y=[Patv1,Patv2,...,Patv 134 ].

[0137] (3) System integration testing

[0138] The overall architecture of this system includes four parts: front-end GUI interactive interface, data query, system software environment and hardware environment. Figure 21 The GUI includes a login and registration interface and a main operation interface. The main operation interface includes real-time data of wind turbines, wind turbine deployment status, forecast data, wind turbine twins, and user management. This interface allows users to query real-time and forecast data for 134 wind turbines.

[0139] 1. Login and Registration

[0140] After starting the system, first enter the system login interface, such as Figure 22 .

[0141] The user needs to enter the registered username and password correctly, otherwise an error message will pop up, such as Figure 23 .

[0142] New users need to click the "Register" button and can log in normally after completing the registration. Figure 24 The registration interface shown.

[0143] 2. Functional interface

[0144] After the user logs in successfully and enters the main operation interface, five function buttons will be displayed, namely "Wind Turbine Real-time Data", "Wind Turbine Deployment", "Forecast Data", "Wind Turbine Twin" and "User Management".

[0145] Click "Real-time Wind Turbine Data" and you can view the current six types of real-time data of 134 wind turbines through the drop-down box in the upper left corner, including "wind speed", "angle of wind direction and nacelle", "ambient temperature", "internal temperature of turbine generator nacelle", "reactive power" and "active power". These data are updated every 10 minutes. When you click to view a certain data type (for example, wind speed), the data of 134 wind turbines will be displayed by default, such as Figure 25 As shown in the figure, the wind speed at 11:10 on March 13, 2025 is displayed. If you want to view the data of a wind turbine separately, you can enter the wind turbine ID range you want to view in the range selection box on the right side of the drop-down box. Click "Query" to display the current data of the corresponding wind turbine. Figure 26 , view the current wind speed for ID numbers 30-45.

[0146] Click "Wind Turbine Deployment" to view the current location of each wind turbine. Each wind turbine corresponds to a point and coordinates. Just like real-time data, you can query the location of a wind turbine individually by entering a range. Figure 27 shown.

[0147] Click "Forecast Data" to view the predicted active power and total power 30 minutes later (you can also enter a range to view). Figure 28 .

[0148] Click on "Wind Turbine Twin" to jump to the Unity platform, which displays the 3D model of the wind turbine rotating with wind speed and the current and predicted data of the corresponding wind turbine, such as Figure 29 Click "User Management" to display the registered user name, password, corresponding authority and registration time. Figure 30 The currently logged in user can change the password by clicking "Change Password". Figure 31 . Click "Exit System" to return to the login interface.

[0149] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A wind farm power prediction method based on digital twin, characterized in that: The steps include: S1. Design a spatiotemporal collaboration algorithm based on time series prediction and wind farm spatial structure, improve ConvLSTM and graph neural network, and use the improved spatiotemporal collaboration model to predict wind turbine power; S2. Based on 3D modeling technology, construct a virtual model of the wind turbines in the wind farm; S3. Combine S1 and S2, build a wind farm power prediction model based on digital twin technology, train the constructed model, and use the trained model to complete the wind farm power prediction work.

2. A wind farm power prediction method based on digital twin according to claim 1, characterized in that: The S1 specifically includes the following contents: S1.

1. Based on the traditional ConvLSTM network, a single convolutional layer with a stride equal to the convolution kernel size is cascaded behind it. The convolutional layer is used to compress the spatial dimension and condense the feature maps output by the ConvLSTM, aggregating each local spatiotemporal feature into a single feature point. S1.

2. By flattening and position encoding, the feature points obtained in S1.1 are added with geographic location information to facilitate the model to perceive and utilize the spatial position relationship of the feature points. After processing through the linear layer, the feature enhancement matrix is obtained. S1.

3. Construct a complete graph based on the locations of wind turbines in the wind farm, improve the traditional GCN algorithm, dynamically update the adjacency matrix through cross-correlation calculation, and quantify the real-time correlation strength between wind turbines in the wind farm; S1.

4. Based on the improved GCN algorithm, graph convolution is performed on the two-dimensional feature enhancement matrix and the adjacency matrix, and the obtained results are processed by the linear layer and then output as the predicted power of the wind turbines in the wind farm.

3. A wind farm power prediction method based on digital twin according to claim 2, characterized in that: In the complete graph described in S1.3, there are connecting edges between all two wind turbines. The formula for calculating the number of edges is: Where m represents the number of edges and n represents the number of fans.

4. A wind farm power prediction method based on digital twin according to claim 3, characterized in that: The dynamic update of the adjacency matrix by cross-correlation calculation described in S1.3 specifically includes the following: Assume that the eigenvectors of wind turbines i and j are L i and L j , the cross-correlation calculation formula between fans is: in, is the dot product of wind turbines i and j, indicating their similarity; ||L i || and ||L j || are the eigenvectors L i and L j The modulus length; through the above cross-correlation calculation, we get the n×n dimensional adjacency matrix C, each element C ij represents the similarity between wind turbines i and j, C ij The larger it is, the stronger the spatial dependence between wind turbines i and j is. The adjacency matrix C is used to represent the spatial relationship between wind turbines and represents the connection strength between their nodes in the space-time graph.

5. The wind farm power prediction method based on digital twin according to claim 1, characterized in that: The 3D modeling tool described in S2 is Unity.

6. A wind farm power prediction system based on digital twins for implementing the method according to any one of claims 1 to 5, characterized in that: Includes the following modules: The spatiotemporal collaboration module improves ConvLSTM and graph neural networks for wind turbine power prediction; 3D modeling module, used to build virtual models of wind turbines in wind farms; The digital twin module, coupled with the spatiotemporal collaboration module and the 3D modeling module, is used to build a wind farm power prediction model based on digital twins and complete the power prediction of wind turbines in the wind farm.

7. A wind farm power prediction system based on digital twin according to claim 6, characterized in that: The development tools of the system include PyQt5.

8. A computer device, characterized in that: The computer device includes a processor and a memory, wherein the memory stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the wind farm power prediction method based on digital twins as described in any one of claims 1 to 5.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores at least one instruction, at least one program, code set or instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to implement the wind farm power prediction method based on digital twins as described in any one of claims 1 to 5.

Images