A wind farm power prediction method and system based on digital twinning

Abstract

The application discloses a wind farm power prediction method and system based on digital twinning, and relates to the technical field of wind farm power prediction.The application improves a baseline model of "ConvLSTM+GCN", cascades a single-layer convolution layer with a step equal to the size of a convolution kernel after the ConvLSTM, and compresses the feature map output by the ConvLSTM; through flattening operation and position coding, the geographical position information of wind turbines can be retained after the spatial dimension reduction of data through the convolution layer, so that the model can perceive and utilize the spatial position relationship; the algorithm of an adjacency matrix in a traditional GCN is improved, the adjacency matrix is dynamically updated through cross correlation calculation, and the real-time correlation strength between wind turbines is quantified; the application also simultaneously performs 3D modeling on the wind turbines, and combines digital twinning technology to construct a virtual-real fusion visual monitoring system, which is suitable for intelligent operation and maintenance, predictive maintenance and optimal scheduling of a wind farm.

Description

TECHNICAL FIELD

[0001] The present application 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 twinning. BACKGROUND

[0002] As an important clean energy generation form, wind farms not only effectively reduce greenhouse gas emissions such as carbon dioxide and alleviate global climate warming, but also provide a reliable and clean solution for energy supply. Therefore, achieving efficient management and optimization of wind farms, especially accurate power prediction, has become the key to improving wind energy utilization and reducing operating costs. With the development of digital twinning (DT) technology, its application potential in wind farms is increasingly prominent. This technology can achieve real-time monitoring, optimal scheduling and accurate prediction of wind farms by building a virtual model corresponding to the physical wind farm, further improving the operational efficiency and reliability of wind farms.

[0003] Digital twinning has been continuously developing since it was first proposed by Professor Michael Grieves at the University of Michigan in the United States in 2003. In 2011, the U.S. Air Force Research Laboratory introduced the concept of "digital twinning" into fighter aircraft maintenance, pushing the technology into the engineering practice stage. After 2014, digital twinning expanded its coverage to the entire product life cycle, from design simulation to production optimization and operation management. In 2021, ZTE released an AR point cloud digital twinning platform, marking the maturity of high-precision modeling and real-time data fusion technology. Currently, digital twinning is deeply integrated with AI, enabling autonomous decision-making through deep learning algorithms, and its application scenarios have expanded from single-device level to city-level systems, becoming a common technology for driving digital transformation in various industries.

[0004] Digital twinning technology enables digital management of the entire life cycle of wind farms by creating a virtual mapping of physical entities. Its core lies in the deep integration of "reality" and "virtuality": the "reality" layer achieves power generation prediction through time series prediction methods such as ConvLSTM, graph convolution network, and other deep learning algorithms; the "virtual" layer relies on technologies such as Unity, 3DsMAX, and GIS to build three-dimensional visual scenes and perform real-time rendering. The organic combination of the two opens up an innovative path for intelligent operation and maintenance, fault warning, and optimal scheduling of wind farms.

[0005] With the rapid development of deep learning technology, neural network models have significant advantages in time series data processing, among which the recurrent neural network (RNN) proposed in 1982 and the long short-term memory network (LSTM) born in 1997 are particularly prominent. LSTM effectively controls information flow by introducing a gating mechanism, extending the effective information retention time from 20 steps in RNN to more than 200 steps. To further integrate spatial and temporal features, researchers proposed the convolutional long short-term memory network (ConvLSTM), which embeds convolutional operations into the LSTM gating mechanism, improving the prediction accuracy of traditional LSTM models by 18%.

[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 spatial correlations between units becomes crucial. To this end, researchers proposed a model that combines LSTM with a spatio-temporal graph: through a graph convolutional network (GCN) to capture spatial dependencies, and using LSTM to learn temporal dependencies, achieving collaborative processing of spatio-temporal dimensional dependencies and improving prediction accuracy. However, existing techniques still have defects in prediction accuracy and efficiency. To address this issue, the present invention proposes a wind farm power prediction method and system based on digital twinning. SUMMARY

[0007] The present invention aims to propose a wind farm power prediction method and system based on digital twinning to solve the problems raised in the background art. The present invention combines deep learning algorithms (ConvLSTM, GCN) with three-dimensional modeling techniques, achieves accurate prediction of wind farm power through an improved spatio-temporal collaborative model, and constructs a virtual-real integrated visual monitoring system using digital twinning technology, suitable for intelligent operation and maintenance, predictive maintenance, and optimal scheduling of wind farms.

[0008] To achieve the above-mentioned purposes, the present invention adopts the following technical solutions:

[0009] A wind farm power prediction method based on digital twinning, comprising the following steps:

[0010] S1, for time series prediction and spatial structure, design a spatio-temporal collaborative algorithm, improve ConvLSTM and graph neural network, and use the improved spatio-temporal collaborative model for wind turbine power prediction to break through the local perception limitations of traditional ConvLSTM network and the static constraints of traditional graph neural network model;

[0011] S2, based on 3D modeling technology, a virtual model is constructed for the wind turbine in the wind farm;

[0012] S3, combining S1-S2, based on digital twinning technology, a wind farm power prediction model based on digital twinning is constructed, the constructed model is trained, and the trained model is used to complete the wind farm power prediction work.

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

[0014] S1.1, based on the traditional ConvLSTM network, a single-layer convolutional layer with a step equal to the convolution kernel size is cascaded at the back, and the convolutional layer is used to compress the spatial dimension and aggregate the features of the ConvLSTM output, and each local spatio-temporal feature is aggregated into a single feature point;

[0015] S1.2, by flattening and position encoding, the feature points obtained in S1.1 are added with geographical position information, so as to facilitate the model to perceive and utilize the spatial position relationship of the feature points, and a feature enhancement matrix is obtained after linear layer processing;

[0016] S1.3, according to the position of the wind turbine in the wind farm, a complete graph is constructed, the traditional GCN algorithm is improved, the adjacency matrix is dynamically updated by cross correlation calculation, and the real-time correlation strength between the wind turbines in the wind farm is quantified;

[0017] S1.4, based on the improved GCN algorithm, the two-dimensional feature enhancement matrix and the adjacency matrix are graph-convolved, and the obtained result is output after linear layer processing to output the predicted power of the wind turbine in the wind farm.

[0018] Preferably, in S1.3, all wind turbines in the complete graph have connecting edges between each other, and the number of edges is calculated according to the following formula:

[0019]

[0020] Wherein, m represents the number of edges; n represents the number of wind turbines.

[0021] Preferably, in S1.3, the adjacency matrix is dynamically updated by cross correlation calculation, which specifically includes the following contents:

[0022] Suppose the feature vectors of wind turbines i and j are L i and L j , the cross correlation calculation formula between the wind turbines is:

[0023]

[0024] Wherein, is the dot product of wind turbines i and j, which represents their similarity; ||L i|| and ||L j || are respectively feature vectors L i and L j The length of the module; through the above cross correlation calculation, the adjacency matrix C of n*n dimension is obtained, each element C ij represents the similarity between the fan i and j, C ij The greater, the stronger the spatial dependence between fan i and fan j; the adjacency matrix C is used to represent the spatial relationship between the fans, representing the connection strength between nodes in the space-time graph.

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

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

[0027] The space-time coordination module improves the ConvLSTM and the graph neural network, and is used for wind turbine power prediction.

[0028] The 3D modeling module is used for virtual model construction of the wind turbines in the wind farm.

[0029] The digital twinning module couples the space-time coordination module and the 3D modeling module, and is used for constructing a wind farm power prediction model based on digital twinning to complete the wind turbine power prediction in the wind farm.

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

[0031] The application further protects a computer device including a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the instruction, program, code set or instruction set is loaded and executed by the processor to realize the above-mentioned wind farm power prediction method based on digital twinning.

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

[0033] Compared with the prior art, the application provides a wind farm power prediction method and system based on digital twinning, which has the following beneficial effects:

[0034] (1) The application takes "ConvLSTM+GCN" as a baseline model, and through the addition of different components (convolution layer, cross correlation) in the ablation experiment, the performance gain brought by each improved part to the algorithm can be clearly seen, and the necessity of the components in the spatio-temporal feature extraction and dependency modeling is clear.

[0035] (2) The application generates a comparison chart based on the training log in the embodiment, which clearly shows the double advantages of the improved model in training loss convergence and validation loss stability, fully verifies the comprehensive improvement in feature learning, anti-overfitting and generalization performance, and provides more reliable performance support for the application of the model in actual scenarios;

[0036] (3) The application develops a wind farm digital twin system and performs the following function tests on each module of the system: the login module verifies the login response of the user; the data transmission module checks the stability of the WebSocket connection and the integrity of the data; the visualization module tests the 3D scene rendering fluency and data update real-time. After testing, the login verification time is less than 1 second, the data transmission delay is controlled within 5 seconds, and the Unity running frame rate is stable at 50fps, meeting the real-time monitoring and prediction needs of the wind farm.

[0037] In summary, the wind farm power prediction method and system based on digital twinning proposed in the application combines deep learning algorithms (ConvLSTM, GCN) and three-dimensional modeling techniques, realizes accurate prediction of wind farm power by improving the spatio-temporal collaborative model, and constructs a virtual-real integrated visualization monitoring system combined with digital twinning technology, which is suitable for intelligent operation and maintenance, predictive maintenance and optimal scheduling of wind farms. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 The network overall structure mentioned in embodiment 1 of the application is shown in the figure;

[0039] Figure 2 The fan 3D model mentioned in embodiment 1 of the application is shown in the figure;

[0040] Figure 3 The system main interface mentioned in embodiment 1 of the application is shown in the figure;

[0041] Figure 4 The ConvLSTM memory unit mentioned in embodiment 1 of the application is shown in the figure;

[0042] Figure 5 The ConvLSTM layer mentioned in embodiment 1 of the application is shown in the figure;

[0043] Figure 6 The feature aggregation process through a single convolution layer mentioned in embodiment 1 of the application is shown in the figure;

[0044] Figure 7 Schematic diagram of the flattening operation mentioned in Embodiment 1 of the present application;

[0045] Figure 8 Schematic diagram of the GCN improvement mentioned in Embodiment 1 of the present application introducing cross correlation;

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

[0047] Figure 10 Schematic diagram of the adjacency matrix calculation code mentioned in Embodiment 1 of the present application;

[0048] Figure 11 Schematic diagram of the wind machine construction mentioned in Embodiment 1 of the present application;

[0049] Figure 12 Real wind machine and model wind machine mentioned in Embodiment 1 of the present application;

[0050] Figure 13 Schematic diagram of the WebSocket server-side configuration mentioned in Embodiment 1 of the present application;

[0051] Figure 14 Schematic diagram of the WebSocket client-side configuration mentioned in Embodiment 1 of the present application;

[0052] Figure 15 System flowchart mentioned in Embodiment 1 of the present application;

[0053] Figure 16 Schematic diagram of the system main interface mentioned in Embodiment 1 of the present application;

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

[0055] Figure 18 Loss comparison chart mentioned in Embodiment 2 of the present application;

[0056] Figure 19 Prediction error distribution histogram mentioned in Embodiment 2 of the present application;

[0057] Figure 20 Scatter plot of predicted value and true value mentioned in Embodiment 2 of the present application;

[0058] Figure 21 System overall architecture mentioned in Embodiment 3 of the present application;

[0059] Figure 22 Schematic diagram of the login interface mentioned in Embodiment 3 of the present application;

[0060] Figure 23 Error information prompt schematic diagram mentioned in embodiment 3 of the present application;

[0061] Figure 24 Registration interface schematic diagram mentioned in embodiment 3 of the present application;

[0062] Figure 25 Real-time data viewing schematic diagram mentioned in embodiment 3 of the present application;

[0063] Figure 26 Range viewing schematic diagram mentioned in embodiment 3 of the present application;

[0064] Figure 27 Fan deployment schematic diagram mentioned in embodiment 3 of the present application;

[0065] Figure 28 Predicted data schematic diagram mentioned in embodiment 3 of the present application;

[0066] Figure 29 Fan twin schematic diagram mentioned in embodiment 3 of the present application;

[0067] Figure 30 User management schematic diagram mentioned in embodiment 3 of the present application;

[0068] Figure 31 Password modification schematic diagram mentioned in embodiment 3 of the present application. DETAILED DESCRIPTION

[0069] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments.

[0070] Embodiment 1:

[0071] The present application proposes a wind farm power prediction method based on deep learning, and designs a wind farm digital twin system combined with 3D modeling technology, the main contents are as follows:

[0072] The algorithm for power prediction of a wind farm is improved. The power of the wind farm is often predicted by taking into account the time characteristics and spatial characteristics of the wind farm, so the present application improves the baseline model of "ConvLSTM+GCN". 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 equal to the size of the convolution kernel is cascaded after ConvLSTM, which can better extract local features, reduce the amount of calculation, and fuse the features of different times to obtain more effective feature representation; through the flattening operation and the position encoding, the geographical position information of the wind turbine can be retained after the spatial dimension reduction of the data through the convolution layer, so that the model can perceive and utilize the spatial position relationship; the GCN is used to process the spatial structure of the wind farm, and the algorithm of the adjacency matrix in the traditional GCN is improved, the adjacency matrix is dynamically updated by cross-correlation calculation, and the real-time correlation strength (such as the wake effect) between wind turbines is quantified. Based on the above improvements, the improved spatio-temporal collaborative model (ConvLSTM+single-layer convolution layer+cross-correlation calculation+GCN) is obtained, as shown in Figure 1 .

[0073] The wind turbine is 3D modeled. The 3D modeling of the wind turbine is completed through the Unity platform. Taking the GW150-4.0MW wind turbine of Goldwind Science and Technology as a reference, the size is scaled by 10:1, and then 3D modeling is performed, the physical properties, rotating characteristics and UI display interface of the wind turbine are added to the 3D model, which can be applied to the digital twin system; as shown in Figure 2 .

[0074] The digital twin system of the wind farm is designed. The improved spatio-temporal collaborative model is used as a predictor to predict the wind farm power in the next half hour, and after obtaining the prediction result, the cross-platform data transmission between Pycharm and Unity is realized through Websocket, and the current data and prediction data of the wind turbine are displayed in real time in Unity, and the data transmission delay is within 5s after testing. A visual interface is developed, which can real-time view various current data of the wind farm in the form of a line chart, such as wind speed, cabin temperature, ambient temperature and other indexes; the main interface of the system is as shown in Figure 3 .

[0075] Based on the above content, the following will be described in combination with related drawings and specific examples, and the specific content is as follows.

[0076] (I) Improved wind power prediction algorithm

[0077] In the aspect of time series prediction, the traditional ConvLSTM network faces the problem of spatio-temporal decoupling in wind farm cluster power prediction: although the local spatial features are captured through convolution operation and the time series modeling is conducted combining LSTM units, it is limited to the fixed size of the convolution kernel receptive field (such as 3x3 or 5x5), and cannot effectively capture the long-distance spatial dependence between wind turbines, and the time dimension modeling only relies on the recurrent connection, which is difficult to deal with dynamic characteristics such as non-stationary wind speed mutation; in the aspect of spatial structure, although the traditional graph neural network model realizes the modeling of global spatial correlation through static graph structure (such as the graph constructed based on geographical distance or historical correlation), it has the following defects: first, the graph structure depends on prior knowledge and cannot dynamically reflect real-time wind field changes; second, it lacks an explicit time dimension processing mechanism, making it difficult to capture periodic fluctuations and trend evolution in time series.

[0078] To solve the above problems, the present application proposes a spatio-temporal collaborative algorithm, which improves ConvLSTM and graph neural network respectively: module one adopts the cascaded structure of ConvLSTM and single-layer convolution layer, realizes efficient abstraction of local spatio-temporal features on the basis of preserving the original scene graph spatial structure, and adds a single-layer convolution layer; module two is based on complete graph construction and node cross-correlation calculation, and uses graph convolution to capture real-time spatial correlation and time series trend between feature condensation points. This design breaks through the local perception limitation of ConvLSTM and the static constraint of traditional graph neural network model, and improves the prediction accuracy.

[0079] (1) Module one: cascading a single-layer convolution layer after ConvLSTM

[0080] Please refer to Figure 4 , Figure 4 is the recurrent memory unit of ConvLSTM. The model structure of ConvLSTM mainly includes two parts: following the typical RNN network structure in time series structure; following the feature extraction method of CNN in spatial structure. Simply speaking, the ConvLSTM model is equivalent to replacing all the fully connected structures in LSTM with convolution structures. The motivation of ConvLSTM model is to overcome the limitations of traditional RNN and CNN by combining CNN and LSTM. Traditional LSTM uses a fully connected layer to process input data, while ConvLSTM introduces convolution operation on the basis of LSTM. The convolution layer is used to process spatial data.

[0081] Please refer to Figure 5 , Figure 5is the ConvLSTM layer of module one, and the input of this layer is a three-dimensional tensor X containing the data of the wind farm at different time points. The input of each time point is a two-dimensional image containing various features such as wind speed and wind direction, representing the spatial layout of the wind turbine cluster and the characteristics of each wind turbine at a certain time. X realizes the shallow extraction of spatio-temporal features in the ConvLSTM layer, and outputs a three-dimensional matrix X'.

[0082] After the ConvLSTM outputs a three-dimensional matrix X', the present application uses a cascaded single-layer convolution layer after the ConvLSTM to compress the spatial dimension and condense the features. As shown in Figure 6 , the convolution kernel size of this convolution layer is 5x5 and the step size is set to 5, which compresses the original feature map by a ratio of 5:1. By covering independent areas without overlapping windows, each 5x5 local spatio-temporal feature is aggregated into a single feature point, reducing spatial redundancy while preserving regional spatial context information. This design takes advantage of the local receptive field characteristics of convolution operation, ensuring that each output feature map is calculated without repetition while maintaining the time dimension and reducing the spatial dimension. The final three-dimensional feature matrix not only retains the spatio-temporal correlation information of the original input, but also highlights key features through abstraction, improving computational efficiency and providing effective feature representation for subsequent analysis.

[0083] After the spatial features of the data are condensed, each local block is aggregated into a feature extraction point. When the data enters module two, these feature points are treated equally by the model, and the relative position spatial information of the two-dimensional image is lost, i.e. each point does not reflect the spatial position difference in the image in the calculation (such as the position relationship that one point is located in the upper left corner of the image and another point is located in the lower right corner). However, the relative position of elements in the image itself contains important spatial information (such as the distribution layout and spatial structure relationship of wind turbines), which is crucial for understanding image content and analyzing spatial logic. When the position relationship of the feature points is treated equally, these key spatial position information is lost. To compensate for this defect, subsequent operations are needed to supplement the geographical position information of the feature points through flattening and position encoding techniques, so that the model can perceive and utilize these spatial position relationships to more accurately complete subsequent tasks. Flattening is shown in Figure 7 , after flattening, the position encoding adds geographical position information to each point, and after processing by a linear layer, a feature enhancement matrix L is obtained.

[0084] (2) Module two: improved GCN algorithm of adjacency matrix to strengthen the relationship between wind turbines

[0085] Please refer to Figure 8 , Figure 8is the structure of module two, which is to process the enhanced matrix L output by module one. First, a complete graph is constructed according to the positions of the fans, that is, there are connecting edges between all pairs of fans. The number of edges is calculated as follows:

[0086]

[0087] where m is the number of edges and n is the number of fans.

[0088] After the construction of the complete graph, the model needs to calculate the cross-correlation between the enhanced matrix L output by module one and its transpose to obtain the dynamic spatial adjacency matrix C between the fans. Traditional spatio-temporal graph models usually rely on prior knowledge (such as physical proximity) to preset node connecting edges, which limits the modeling ability of non-local associations. For example, Figure 9 is an undirected graph G and its original adjacency matrix A, which represents the nodes connected to each node. The present application improves the calculation method of graph convolution adjacency matrix in model two, 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. Let the feature vectors of fans i and j be L i and L j , the cross-correlation calculation process between fans is as follows:

[0089]

[0090] where, represents the dot product of fans i and j, which represents their similarity; ||L i || and ||L j || are the lengths of feature vectors L i and L j .

[0091] Through the above cross-correlation calculation, an n x n dimensional adjacency matrix C is obtained, where each element C ij represents the similarity between fans i and j, and C ij is larger, indicating that the spatial dependence between fan i and fan j is stronger. Adjacency matrix C is used to represent the spatial relationship between fans, that is, the connection strength between nodes in the spatio-temporal graph.

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

[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 enhanced matrix L and C based on GCN, and outputs the predicted power of the fan after processing the obtained result through a linear layer.

[0094] (II) 3D modeling of fans and virtual-real combination

[0095] (1) Virtual model construction of the wind turbine

[0096] As the "virtual" part of digital twinning, 3D modeling of the wind turbine is essential. Unity, as a leading real-time 3D development engine in the industry, has the advantages of efficient real-time rendering capability, integrated physics engine (such as simulation of wind power, rotation, etc.), convenient interactive development tools, and cross-platform deployment advantage. It can accurately construct the virtual model of the wind turbine, support dynamic data access and scene interaction, and become an ideal technical carrier for realizing the digital twinning of the wind turbine. Therefore, the present application selects unity to complete the 3D modeling of the wind turbine.

[0097] Table 1 Component size and scale

[0098]

[0099] Referring to the GW150-4.0MW wind turbine of Goldwind Science and Technology, the physical size is converted to model size by engineering modeling specification, i.e. the scale is set to 1:10, the physical size (hub height 100m, blade length 73.5m) is converted to model size (10m, 7.35m), as shown in Table 1. This scale balances the scene realism and rendering performance, and meets the typical modeling requirements of the digital twinning system. The hub height and blade length directly refer to the public parameters of Goldwind Science and Technology, and the tower size and cabin size are derived from the product technical document.

[0100] First, create the blade in Unity using the stretch geometry and surface editing tools. Build the blade cross-sectional profile first, generate the three-dimensional structure of the blade through the stretch operation, adjust the length to 7.35m (1:10 scale), and then copy two blades, a total of three blades, as shown in 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 component: height 10m, radius 0.45m; finally create a cuboid 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 blade, tower and base are completed, use the rotation tool to set the blade installation angle. For the cabin part, model by combining cylinder and sphere, adjust the size to length 1.95m, diameter 1.2m, and finally assemble the three blades and the cabin to ensure uniform distribution around the tower center, forming the complete appearance of the wind turbine (as shown in Figure 11 (b), the structure after combining the blade and the cabin).

[0101] After adding various components (rigid, rotation, etc.), assign a dedicated material to the tower, blades, nacelle, and other components. The tower uses a metal material, with a metal degree of 0.4 and a roughness of 0.65 to simulate the texture of a metal surface. The blades use a composite material, with a metal degree of 0.15 and a roughness of 0.5, and add a normal map (intensity 1.0) to enhance the details of the curved surface. The nacelle is assigned a high-metal-degree material (metal degree 0.6) to highlight the mechanical component characteristics. Use the material sphere panel to preview the effects in real-time, and gradually fine-tune until the real fan appearance is achieved (as shown in Figure 12 , and the optimized fan model material performance).

[0102] (2) Real and virtual combined data transmission

[0103] WebSocket technology is used for real-time data transmission. The reason for choosing WebSocket technology is its full-duplex communication feature, which allows the server and client to transmit data bidirectionally at any time after establishing a connection, with low latency and high efficiency, which is very suitable for the real-time data transmission requirements of the system.

[0104] On the server side, the invention uses Python language combined with asyncio and Websockets library to build a WebSocket server. The main function of this server is to generate prediction data for wind farms by improving the algorithm as a predictor, and to send these data to the connected clients in real time. In the development and testing stage, in order to simplify the network configuration and ensure the stability and security of the system, the 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 this address and port on the local computer, waiting for the connection of the client. When the server starts, it will continuously obtain the prediction data of the wind farm and send these data to the connected clients. If the client connection is interrupted, the server will perform the corresponding cleaning operation to remove the client. The python code is as shown in Figure 13 .

[0105] On the Unity side, a WebSocket client also needs to be implemented to connect to the Python server and receive real-time data. Since it is developed and tested 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 C# script of Unity, create a WebSocket object and set its connection address to ws: / / 127.0.0.1:8765. The C# code is as shown in Figure 14 .

[0106] (III) Wind farm power prediction system based on digital twinning

[0107] The present application selects to use PyQt5 for the development of the digital twinning system. When designing the digital twinning system, PyQt5 has multiple advantages as a graphical user interface (GUI) framework. First, PyQt5 provides rich controls and components that can meet the needs of various complex interfaces, such as real-time data visualization, state monitoring, control panels, alarm systems, etc. Second, PyQt5 also supports the design of modern interfaces, enabling the creation of responsive and visually appealing user interfaces that ensure a smooth user experience. The good compatibility of PyQt5 with the Python ecosystem is also a highlight, as it can seamlessly integrate other Python libraries such as NumPy, Pandas, and Matplotlib, facilitating data processing and analysis, especially in digital twinning systems that involve real-time data acquisition and complex data analysis. The powerful computing capabilities of Python are crucial in such cases. Third, the signal and slot mechanism of PyQt5 makes the interaction between the interface and the system background more concise and efficient, supporting event-driven programming and enabling real-time responses to user operations and system state changes. The flexible layout manager also allows the interface to adapt to different screen sizes and devices, ensuring the system's applicability. Finally, PyQt5 has rich documentation and community support, making it easy for developers to find relevant information and solutions, reducing development difficulty. In summary, PyQt5 provides a powerful, flexible, and efficient development platform for digital twinning systems, enabling developers to quickly build complex and interactive user interfaces and improve the system's usability and scalability.

[0108] The present application not only meets the real-time query of wind farm data, but also meets the active power prediction of wind turbines 30 minutes later. The data acquisition part directly uses the data in the SDWPF dataset, inputs the data of 134 wind turbines in the previous 48 hours into the system, predicts the active power 30 minutes later by predicting it, and the algorithm of the predictor calls the improved wind farm power prediction algorithm. After the user runs the system and logs in successfully, he will be redirected to the function interface, where he can click on the corresponding data by clicking on the phase button. All data is displayed in the form of a line chart. The system flow chart is shown in Figure 15 .

[0109] When the system is running, the user needs to log in to the registered account. After the user logs in successfully and enters the function interface, five function modules will be displayed, namely "wind turbine real-time data", "wind turbine deployment", "prediction data", "wind turbine twin", and "user management". Different real-time data will be displayed according to the user's selected module.

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

[0111] Example 2

[0112] Based on Example 1 but with the difference that the experiment is designed to characterize the performance of the deep learning-based wind farm power prediction method and system proposed in the present application, the specific content is as follows.

[0113] (I) Power prediction part

[0114] Taking "ConvLSTM+GCN" as the baseline model, through the addition of different components in the ablation experiment (convolutional layer, cross correlation), the performance gain brought by each improved part to the algorithm can be clearly seen, and the necessity of the components in the spatio-temporal feature extraction and dependency modeling is clear. Table 2 and Figure 17 The results of the ablation experiment. Figure 17 The four models in Table 2 are replaced by a, b, c, and d.

[0115] Table 2 Ablation experiment

[0116]

[0117] As can be seen from the experimental results in Table 2, ConvLSTM+GCN as the baseline model, its RMSE is 0.2783, MAE is 0.1635, MAPE is higher, R 2 is 0.9405, which reflects the basic spatio-temporal modeling ability of the fusion of ConvLSTM and GCN. After introducing the convolutional layer alone on the baseline model, the RMSE rises to 0.2812 (an increase of 1.04%), the MAE rises to 0.1698 (an increase of 3.85%), and the R 2 drops to 0.9392 (a decrease of 0.14%). The addition of the convolutional layer alone fails to optimize but rather reduces performance, which may be because the convolutional layer's local feature extraction overlaps with the function of ConvLSTM, or introduces redundant calculations, interfering with the core spatio-temporal feature learning;

[0118] After introducing the cross correlation alone on the baseline model, the RMSE rises slightly to 0.2795 (an increase of 0.43%), the MAE rises to 0.1660 (an increase of 1.53%), but the R 2 slightly rises to 0.9410 (an increase of 0.05%). This shows that the cross correlation has a certain positive effect on spatial feature modeling and can slightly improve the spatial dependency capture ability, but the effect is limited when used alone, and the overall performance cannot be significantly optimized.

[0119] By simultaneously introducing convolutional layers and cross-correlation on the baseline model, a final improved spatiotemporal collaborative model was obtained. This fully-component model achieved optimal performance, reducing RMSE to 0.2683 (a decrease of 3.6%), MAE to 0.1582 (a decrease of 3.24%), and MAPE to 101.25 (a decrease of 2.42%). 2 The value rose to 0.9520 (an increase of 1.2%). This indicates that the combination of convolutional layers and cross-correlation enhances the extraction of spatiotemporal features: convolutional layers enhance the learning of temporal features, while cross-correlation deepens the modeling of spatial relationships. The two complement each other and work together to improve the ability to capture the spatiotemporal dependence of wind power data, verifying the effectiveness of multi-component fusion.

[0120] Depend on Figure 18 It can be seen that the improved model's RMSE, MAE, and MAPE are all lower than the baseline model, and R... 2 It is an improvement over the baseline model.

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

[0122] Based on the log files of the training process, the training loss and validation loss data of the baseline model and the improved model for each round are extracted, and a loss comparison chart is generated (e.g., Figure 18 (As shown). Dropout was used in the experiment to hide some neurons in the network with certain probability values, preventing them from participating in the training process and thus avoiding overfitting during network training. The following analysis of the charts is based on log content:

[0123] Figure 18 As shown in (a), both models initially used a decreasing learning rate of 0.01 to quickly reduce loss. However, log records show that the baseline model's loss decline slowed significantly after 30 epochs, eventually stabilizing at approximately 0.068; while the improved model experienced brief fluctuations around 30 epochs, its loss continued to decrease, eventually stabilizing at approximately 0.060. This difference indicates that the improved model is more efficient at learning data features. Although the initial loss is slightly higher, its structural advantages (a superior feature extraction mechanism or network architecture) allow it to more fully uncover the potential patterns in the data. Combined with the log records showing a continuous decrease in the improved model's training loss, this further illustrates that the decreasing learning rate strategy and the Dropout rate of 0.3 work synergistically better in the improved model. The high learning rate in the early stages accelerates convergence, while the low learning rate in the later stages achieves fine-tuning. Simultaneously, Dropout effectively suppresses overfitting, ensuring a continuous decrease in loss in the later stages of training.

[0124] Figure 18 (b) shows that the reference model (yellow curve) has a volatile validation loss, reflecting its insufficient generalization ability and the risk of overfitting; the improved model (red curve) has a significantly reduced fluctuation range and a more stable downward trend overall. The improved model's more stable validation loss is recorded in the log, which directly confirms the effect of its structural optimization (such as more reasonable network layer design and feature fusion method) combined with a Dropout rate of 0.3 - by randomly inactivating neurons, the model's robustness is effectively improved, reducing the degree of overfitting to training data, thereby making the model more adaptable to unknown data, and the validation loss is continuously optimized.

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

[0126] Figure 19 The improved model's prediction error distribution is shown, with the horizontal axis representing the prediction error range (-3 to 3 kW). The mean 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 1.5 kW, Figure 17 indicating that the improved model has good error control and strong prediction stability.

[0127] In Figure 20 the scatter plot, the improved model's comparison range based on true value and predicted value is -1.0 to 3.0 kW. The scatter points in the figure are closely distributed around the diagonal line, with fewer scatter points deviating from the diagonal line, and most of them are in the extreme value area when the true value is higher than 2.5 kW or lower than 0.5 kW. Figure 18 This indicates the prediction accuracy of the improved model.

[0128] (II) System Implementation

[0129] The wind farm digital twin system was developed using PyQT5, and the following functional tests were conducted on each module of the system: the login module verifies the user's login response; the data transmission module checks the stability of WebSocket connection and data integrity; the visualization module tests the 3D scene rendering smoothness and data update real-time performance. After testing, the login verification time is less than 1s, the data transmission delay is controlled within 5s, and the Unity running frame rate is stable at 50fps, meeting the real-time monitoring and prediction needs of wind farms.

[0130] Example 3:

[0131] Based on the embodiment 1-2 but with different, the following in a specific example of the present application proposed based on deep learning wind farm power prediction method and system are described, the specific content as follows.

[0132] (1) Data preparation

[0133] The present application uses SDWPF (Spatial Dynamic Wind Power Forecasting) public data set, is the actual data of Longyuan Group Co., Ltd. (China and Asia's largest wind power producer), which is obtained by the monitoring and data acquisition (SCADA) system of wind farm. The wind farm sampled by the data set consists of 134 turbine generators, and the SCADA system samples every ten minutes from each turbine generator in the wind farm, a total of 245 days, so the SDWPF data set has 4727520 data (6x24x245x134).

[0134] (2) Model training

[0135] The split ratio of training set and test set is 1:1, and the wind power output after 30 minutes is predicted. The final algorithm output Y is a one-dimensional vector with the shape of R 1×134 , and each element corresponds to the predicted power value of a wind turbine after 30 minutes, realizing batch prediction of the active power of 134 wind turbines, that is:

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

[0137] (3) System integration test

[0138] The overall architecture of the system includes four parts of the front-end GUI interaction interface, data query, system software environment and hardware environment, as shown in Figure 21 . The GUI interface includes login and registration interface, operation main interface; the operation main interface includes real-time data of wind turbine, deployment of wind turbine, prediction data, wind turbine twin, and user management, and the real-time data and prediction data of 134 wind turbines can be queried in the interface.

[0139] 1. Login and registration

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

[0141] The user needs to correctly input the registered username and password, otherwise an error message will be prompted, as shown in Figure 23 .

[0142] New users need to click the "register" button to log in normally after registration. As shown in the registration interface. Figure 24

[0143] 2. Functional interface

[0144] After the user successfully logs in and enters the main operation interface, five function buttons will be displayed, namely "real-time data of wind turbine", "wind turbine deployment", "prediction data", "wind turbine twin", and "user management".

[0145] Click "real-time data of wind turbine" to 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 between wind direction and nacelle", "ambient temperature", "turbine generator nacelle internal temperature", "reactive power", and "active power". These data are updated every 10 minutes. When clicking to view a certain type of data (taking wind speed as an example), the data of 134 wind turbines will be displayed by default, as shown in the figure, which shows the wind speed at 11:10 on March 13, 2025. If you want to view the data of a single wind turbine, you can enter the wind turbine ID range you want to view in the range selection box to the right of the drop-down box. Click "query" to display the current data of the corresponding wind turbine, as shown in the figure, which shows the current wind speed of ID numbers 30-45. Figure 25 Figure 26

[0146] Click "wind turbine deployment" to view the location of each wind turbine. Each wind turbine corresponds to a point and coordinates. Like real-time data, you can enter a range to query the location of a single wind turbine. As shown in the figure. Figure 27

[0147] Click "prediction data" to view the predicted active power and total power after 30 minutes (you can also enter a range to view). As shown in the figure. Figure 28

[0148] Click "wind turbine twin" to jump to the unity platform, which displays a wind turbine 3D model rotating with wind speed and the current data and prediction data of the corresponding wind turbine, as shown in the figure. Figure 29 Click "user management" to display the registered usernames, passwords, corresponding permissions, and registration times, as shown in the figure. The currently logged-in user can modify the password by clicking "modify password", as shown in the figure. Click "exit system" to return to the login interface. Figure 30 Figure 31

[0149] ​​​​​​​The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, which should be covered within the protection scope of the present application.

Claims

1. A wind farm power prediction method based on digital twinning, characterized in that, Comprise the following steps: S1, for time series prediction and wind farm spatial structure, design space-time collaborative algorithm, improve ConvLSTM and graph neural network, use the improved space-time collaborative model to carry out wind turbine power prediction, it specifically includes the following contents: S1.1, based on traditional ConvLSTM network, a single convolution layer with equal step and convolution kernel size is cascaded in its back, the spatial dimension compression and feature condensation of the feature map output by ConvLSTM are carried out by using the convolution layer, and each local space-time feature is aggregated into a single feature point; S1.2, by flattening operation and position coding, the feature points obtained in S1.1 are added with geographical position information, so that the model can perceive and utilize the spatial position relationship of the feature points, and a feature enhancement matrix is obtained after linear layer processing; S1.3, according to the position of the wind turbine in the wind farm, a complete graph is constructed, the traditional GCN algorithm is improved, the adjacency matrix is dynamically updated by cross correlation, and the real-time correlation strength between the wind turbines in the wind farm is quantified; S1.4, based on the improved GCN algorithm, the two-dimensional feature enhancement matrix and the adjacency matrix are graph convolved, and the obtained results are output after linear layer processing to output the predicted power of the wind turbine in the wind farm; S2, based on 3D modeling technology, a virtual model of the wind turbine in the wind farm is constructed; S3, combined with S1-S2, based on digital twinning technology, a wind farm power prediction model based on digital twinning is constructed, the constructed model is trained, and the trained model is used to complete the wind farm power prediction.

2. The wind farm power prediction method based on digital twinning according to claim 1, characterized in that, All wind turbines in the complete graph described in S1.3 have connecting edges between each other, and the number of edges is calculated by the formula: wherein, m represents the number of edges; n represents the number of fans.

3. The wind farm power prediction method based on digital twinning of claim 2, wherein, In S1.3, the adjacency matrix is dynamically updated by cross correlation, which specifically includes the following contents: Assume the fan i and j The eigenvectors of L i and L j The cross-correlation calculation formula between the fans is: wherein, is the fan i and j dot product, representing their similarity; L i ||and|| L j ||are the feature vectors L i and L j norms; through the above cross-correlation calculation, the n × n adjacency matrix of dimension C is obtained, each element C ij represents the similarity between the fan i and j , C ij the greater, the stronger the spatial dependence between the fan i and the fan j ; the adjacency matrix C is used to represent the spatial relationship between the fans, representing the connection strength between the nodes in the space-time graph.

4. The wind farm power prediction method based on digital twinning of claim 1, wherein, The tool for 3D modeling in S2 is Unity.

5. A digital-twin-based wind farm power prediction system for implementing the method according to any one of claims 1-4, characterized in that, Comprise the following modules: The space-time collaborative module improves ConvLSTM and graph neural network and is used for wind turbine power prediction; The 3D modeling module is used for virtual model construction of the wind turbine in the wind farm; The digital twinning module couples the space-time collaborative module and the 3D modeling module, and is used for constructing a wind farm power prediction model based on digital twinning to complete the power prediction of the wind turbine in the wind farm.

6. The wind farm power prediction system based on digital twinning of claim 5, wherein, The development tool of the system comprises PyQt5.

7. A computer device, comprising: The computer device comprises a processor and a memory, and the memory stores at least one instruction, at least one program, a code set or an instruction set, which is loaded and executed by the processor to realize the wind farm power prediction method based on digital twinning as claimed in any one of claims 1-4.

8. A computer-readable storage medium, characterized in that, The computer readable storage medium stores at least one instruction, at least one program, a code set or an instruction set, which is loaded and executed by the processor to realize the wind farm power prediction method based on digital twinning as claimed in any one of claims 1-4.

Images