Wind power prediction system applied to wind power generation

Abstract

The invention provides a wind power prediction system applied to wind power generation, and relates to the field of wind power prediction, and the system comprises a data collection module which collects real-time wind power related data; the wind power historical data acquisition module is used for acquiring and marking historical data of the wind power related data; the data processing module is used for preprocessing the received data, extracting multi-dimensional data features and fusing the multi-dimensional data features with spatio-temporal data features, and adaptively optimizing feature extraction and fusion strategies; the data storage module is used for storing the data in a data warehouse in a classified manner after receiving the data; and the large model training module is used for establishing a 3D (three-dimensional) transformer wind speed prediction model based on a visual transformer, and further training basic parameters of the large model by applying a Fine-Tuning technology. Based on the dual attention mechanism and the graph convolutional neural network, the prediction model with high semantic understanding ability is constructed, and accurate wind power prediction ability is realized.

Description

Technical Field

[0001] The present application relates to the technical field of wind power prediction, and in particular to a wind power prediction system applied to wind power generation. Background Art

[0002] In recent years, with the rapid growth of global energy demand, the development and utilization of new energy has become an important issue worldwide. In this context, wind energy has attracted much attention due to its decentralized, pollution-free and renewable characteristics, and is regarded as an important alternative to traditional energy. However, wind power generation systems face challenges such as randomness, intermittency and volatility, which make the stability and predictability of wind energy output an important consideration in the operation and management of power systems.

[0003] In the power system, accurate prediction of wind speed is crucial to ensure the stable operation of wind power generation systems. Accurate wind power forecasting can help the power system effectively manage and balance wind power output, and improve the reliability and economy of the system. According to the different forecast time spans, wind speed forecasts can be divided into ultra-short-term (minute level), short-term (hours to days), medium-term (weeks to months) and long-term (yearly). Especially in short-term forecasts, accurate prediction of wind speed is crucial for daily power dispatching, which can significantly affect the economy and operational reliability of the power system.

[0004] The existing wind prediction model is based on historical wind data, weather forecast data, and measured meteorological data, and uses a neural network to establish a prediction model to predict wind power. However, the existing wind prediction model does not consider the impact of geographical factors on wind power, and the wind prediction model has too many iterations of short-term wind prediction in the medium term, which may cause large iteration errors and cannot accurately predict wind data in complex environments.

[0005] Application Contents

[0006] This application is made in view of the above-mentioned issues, and its purpose is to provide a wind prediction system for wind power generation. This application intends to use deep neural network technology based on time and space, combined with real-time data and historical observation data of surface microscale short-term wind prediction, to improve the prediction accuracy and reliability of the wind power generation system. Based on the dual attention mechanism and graph convolutional neural network, this application extracts key features such as surface microscale wind observation data, meteorological elements, geographic information, etc. from multiple dimensions to construct a prediction model with a high degree of semantic understanding ability. The model will deeply explore the complex relationship between wind output and factors such as geographical environment, seasonal changes, and meteorological variables to form a comprehensive and accurate wind prediction capability. This application will bring new technical solutions to energy management and power system operations through accurate wind prediction capabilities, and promote the sustainable development and application prospects of new energy.

[0007] Specifically, the first aspect of the present application provides a wind prediction system for wind power generation, including:

[0008] The data acquisition module collects real-time wind-related data, including geographical environment information, seasonal changes, and meteorological element change information, and sends the data to the data processing module;

[0009] A wind power historical data collection module is used to collect historical data of wind power related data, mark the historical data, and send the data to the data processing module;

[0010] Historical data needs to be aligned with the input features of the model, which includes labeling key meteorological variables (such as 2m air temperature, dew point temperature, surface pressure, etc.), as well as relevant terrain data, wind speed components (such as 10m u wind component and v wind component) and other meteorological elements. These variables will be labeled according to time and space dimensions for model learning and prediction.

[0011] By analyzing the correspondence between historical wind data (such as wind speed and wind direction) and other relevant meteorological variables, the historical data is labeled as the input features and target outputs required for the prediction task, including setting the mapping relationship between the target variable (such as wind speed in the next 15 minutes, 30 minutes, and 10 days) and the current observation value. After the data is initially labeled using an automated algorithm, it is combined with manual verification steps to ensure the accuracy and consistency of the labeling, especially when complex meteorological events or abnormal weather conditions are involved.

[0012] The data processing module pre-processes the received data and extracts multi-dimensional data features through the dual attention mechanism. It then fuses spatiotemporal data features through the graph convolutional neural network, adaptively optimizes feature extraction and fusion strategies through joint training, and sends the processed data to the data storage module.

[0013] A data storage module, which receives data and stores the data in a data warehouse according to classification;

[0014] The large model training module builds a 3D transformer wind speed prediction model based on the visual transformer, sets model training hyperparameters, uses self-supervised training methods to train the model, inputs annotated wind history data sets to train the basic parameters of the large model, and then uses Fine-Tuning technology to further train the basic parameters of the large model until the average error and mean square error of the large model for real-time wind prediction meet expectations;

[0015] Display module: Based on the development of an interactive Web platform, the system development and visualization technology are integrated to visualize the prediction results of the large wind prediction model.

[0016] Furthermore, the preprocessing includes designing a unified spatiotemporal representation framework for meteorological, geographical, remote sensing and other data from different sources and in different formats, achieving data standardization and alignment, and eliminating scale differences and inconsistencies between different data sources through spatiotemporal interpolation and resampling.

[0017] Spatiotemporal interpolation methods estimate and interpolate missing or inconsistent data by considering the correlation of data in time and space, thereby improving the consistency and accuracy of data; resampling techniques eliminate inconsistencies in data sets by changing the time interval or spatial resolution of data. Resampling can convert time series data from one frequency to another, increase granularity through upsampling or reduce granularity through downsampling, thereby achieving data consistency.

[0018] Furthermore, the extraction of multidimensional data features through the dual attention mechanism specifically includes:

[0019] Step 1: Introduce the dual attention mechanism to adaptively learn the importance weights of different data sources and different spatiotemporal locations;

[0020] The dual attention mechanism can effectively capture the differences and correlations between feature maps by calculating the attention weights in both spatial and channel dimensions.

[0021] Step 2: Automatically adjust the weight distribution between different data sources through the channel attention module;

[0022] Step 3: Dynamically focus on key information in different spatiotemporal regions through the spatiotemporal attention module;

[0023] The dual attention mechanism can also dynamically adjust the importance weights of different data sources and spatiotemporal locations. It learns the internal correlation between channels and the spatiotemporal dependency of features through the channel attention module and the spatiotemporal attention module, respectively, thereby achieving adaptive weight allocation for different data sources and spatiotemporal locations.

[0024] Step 4: The above two attention modules cooperate with each other to achieve adaptive optimization of feature extraction.

[0025] Furthermore, the spatiotemporal data feature fusion by graph convolutional neural network specifically includes:

[0026] Each spatiotemporal location is regarded as a node of the graph, and the associations between different data sources are regarded as edges of the graph. GCN is used to propagate and aggregate information on the graph structure to achieve feature interaction and fusion between different data sources and different spatiotemporal locations, and obtain a comprehensive feature representation that integrates time and space.

[0027] Furthermore, during the joint training, reconstruction loss and adversarial loss are introduced simultaneously to ensure that the preprocessed features have good distinguishability and generalization while retaining key information.

[0028] Reconstruction loss is mainly used to measure the difference between the generated image and the original image. Common reconstruction loss functions include L1 loss and L2 loss. L1 loss is an absolute error loss, which is usually clearer and converges faster than L2 loss, i.e., mean square error loss, and is suitable for image reconstruction tasks. In some cases, L2 loss is also used for reconstruction loss. For example, in image super-resolution reconstruction, L2 reconstruction loss is used to ensure that the pixel difference between the reconstructed image and the original image is minimized.

[0029] Adversarial loss is used to train the generator and discriminator. The goal of the generator is to generate realistic samples, while the goal of the discriminator is to distinguish between real samples and generated samples. The adversarial loss function trains the generator by maximizing the error rate of the discriminator, so that it generates more realistic samples.

[0030] Furthermore, the design points of the 3D transformer wind speed prediction model specifically include: reconstructing the encoder and decoder to meet the model's ability to process highly characteristic historical wind-related data, using the Swin-Transformer sliding window attention mechanism to reduce the model's calculation amount, and using a hierarchical time domain aggregation strategy to reduce the number of model prediction iterations.

[0031] The sliding window attention mechanism in Swin-Transformer is used to limit the attention calculation to a local window, which helps reduce the computational complexity. Each local window only focuses on a limited adjacent space instead of all global positions, which significantly reduces the computational overhead. In the 3D variant transformer, this mechanism is extended to a three-dimensional window, allowing the model to efficiently learn local features at different time segments, spatial positions, and height levels.

[0032] Ordinary Transformers are mainly used for two-dimensional data (such as natural language processing, image recognition, etc.). The encoder and decoder modules of Transformers have limited processing of temporal or spatial information. In this application, the encoder and decoder are reconstructed to adapt to the data input of three-dimensional space, including:

[0033] First, the input layer is improved: the input layer is expanded to enable it to receive and process three-dimensional tensors. Specifically, the convolution layer of the input data is adjusted from 2D convolution to 3D convolution to capture the three-dimensional spatial characteristics of meteorological data.

[0034] Then, the three-dimensional extension of the self-attention mechanism: the self-attention mechanism is extended from 2D to 3D, so that the model can simultaneously pay attention to the correlation in time, space and height. Both the multi-head self-attention mechanism in the encoder and the cross-attention mechanism in the decoder are adjusted to adapt to the modeling of three-dimensional spatiotemporal relationships.

[0035] The main task of the encoder is to map the input data to a low-dimensional space. The feature extraction capability of the encoder can be improved by adding more hidden layers and adjusting the number of neurons in the hidden layers. For example, a convolutional layer can be used to extract features of image data. The improved encoder of the present application introduces a three-dimensional position encoding method, which enables the model to distinguish data points at different spatial positions and altitude levels. This encoding method combines spatial information such as altitude, latitude, and longitude to ensure that the model does not lose spatial structure information during the extraction of high-dimensional features. It can process three-dimensional features more efficiently and extract complex feature relationships of input data in space, time, and meteorological elements by using 3D convolution and 3D pooling layers.

[0036] The task of the decoder is to reconstruct the data in the low-dimensional space into the original data. The reconstruction effect can be improved by adding more decoding layers and adjusting the number of neurons in the decoding layer. This application enhances the feature back propagation capability of the decoder so that it can restore accurate wind speed prediction values ​​from high-dimensional features. The design of the decoder focuses on the output dimension of the prediction to improve the prediction accuracy of the wind speed in the target area.

[0037] The present application also uses a layered, fully connected layer with variable width (or a hybrid multi-layer perceptron MLP) to adapt to diverse three-dimensional feature inputs. This design can flexibly adjust the number of layers and neurons to improve the adaptability and processing capabilities of complex wind speed data.

[0038] Swin-Transformer introduces a sliding window operation to limit the attention calculation to a local window, thereby significantly reducing the computational complexity. Specifically, Swin-Transformer uses two sliding window operations: non-overlapping local windows and overlapping cross windows.

[0039] The core idea of ​​the hierarchical time-domain aggregation strategy is to take time as an important dimension and organize and display relevant information according to the chronological order. Through the hierarchical time-domain aggregation strategy, the model extracts and aggregates features at multiple time scales, thereby reducing the number of iterations. This hierarchical aggregation can integrate historical wind speed data hierarchically through a recursive window aggregation method, aggregate step by step from the short-term to the medium-term, and finally form a stable forecast output.

[0040] Furthermore, the model training hyperparameters include learning rate, batch size, optimizer and number of iterations.

[0041] Hyperparameters are parameters that need to be manually set before training a model. These parameters are not automatically learned during the training process, but need to be manually adjusted to optimize model performance. The adjustment of hyperparameters has an important impact on the performance of the model and the training process.

[0042] The learning rate is one of the most important parameters in the optimizer, which determines the step size of each parameter update. A reasonable learning rate can make the optimizer converge quickly, while a too large learning rate may cause instability in the training process, and a too small learning rate may lead to slow convergence. In the early stages of training, a larger learning rate is usually used. As the training progresses, the learning rate is gradually reduced to improve the stability of the model. Common learning rate adjustment strategies include learning rate decay and warm-up strategies.

[0043] Batch size refers to the number of samples used each time the model is trained. Larger batch sizes can improve training stability and memory utilization, but may require more computing resources and may reduce the generalization ability of the model. Small batch sizes can make the model more stable, but the training time is longer. The choice of batch size depends on many factors, including the complexity of the model, the available memory (GPU or CPU), the size of the training data, and the requirements of the specific task.

[0044] An optimizer is an algorithm used to adjust model parameters to minimize the loss function. Common optimizers include SGD, Adam, RMSProp, etc. Different optimizers have different characteristics and applicable scenarios. For example, the Adam optimizer improves training efficiency and stability through adaptive learning rate.

[0045] The number of iterations refers to the number of times the model completely traverses the training data. A reasonable number of iterations can improve the convergence speed and generalization ability of the model and avoid overfitting and underfitting. Usually you can start with 10 and decide whether to increase or decrease it based on the performance of the validation set. Too many iterations will increase the computational complexity and convergence speed.

[0046] Furthermore, the basic parameters of the large model are the learnable parts within the model, including weights and biases.

[0047] The weight class is responsible for connecting each neuron and adjusting the importance of signal transmission. The bias is responsible for adding a fixed value to the input signal of the neuron. These parameters are automatically updated during the model training process to optimize the performance of the model.

[0048] Furthermore, the use of Fine-Tuning technology to further train the basic parameters of the large model specifically includes:

[0049] Step 1: Collect wind-related data in a specific area and form a specific fine-tuning dataset after preprocessing steps such as data cleaning;

[0050] Specific fine-tuning datasets refer to datasets used for specific tasks or fields during the fine-tuning of large models. These datasets usually contain input data and target data. The input data is the data that the model needs to process, and the target data is the result that the model needs to predict.

[0051] Step 2: Use Fine-Tuning technology to fine-tune and optimize the wind prediction model based on the constructed specific fine-tuning dataset;

[0052] Fine-Tuning technology involves updating all parameters of the model to adapt to the data of specific tasks. This method is effective and has strong fine-tuning capabilities for the model. Wind forecasting tasks require processing a large amount of time series data and complex meteorological information. Through Fine-Tuning technology, the performance of the model in wind forecasting tasks can be significantly improved.

[0053] Step 3: Establish relevant fine-tuning technical evaluation indicators to comprehensively evaluate the accuracy and generalization performance of wind forecasting in different actual regions and areas.

[0054] Furthermore, the related fine-tuning technology evaluation indicators include the accuracy, precision and recall rate of the model.

[0055] Accuracy refers to the ratio of the number of samples correctly predicted by the model to the total number of samples. Precision refers to the ratio of all samples predicted to be positive that are actually positive. Recall refers to the ratio of all samples actually positive that are correctly predicted to be positive. Accuracy, precision and recall are important indicators for evaluating the performance of classification models, and they reflect the quality of the model from different perspectives.

[0056] In the second aspect, the present application also provides a computing device, which has the function of implementing the method described in the first aspect above. The beneficial effects can be found in the description of the first aspect, which will not be repeated here. The function can be implemented by hardware, or by executing the corresponding software implementation through hardware. The hardware or software includes one or more modules corresponding to the above functions. In one possible design, the structure of the device includes an acquisition module, a training module, and optionally, a construction module. These modules can implement the function of the training node in the method example of the first aspect above. Please refer to the detailed description in the method example for details, which will not be repeated here.

[0057] In the third aspect, the present application also provides a computing device, which is used to implement the functions of the method described in the first aspect above. The beneficial effects can be found in the description of the first aspect and will not be repeated here. The structure of the computing device includes a processor and a memory, and the memory is used to store instructions and / or data. The memory is coupled to the processor, and when the processor executes the program instructions stored in the memory, the function of the training node in the example of the first aspect above can be implemented. The structure of the computing device also includes a communication interface for communicating with other devices.

[0058] In a fourth aspect, the present application also provides a computer-readable storage medium, in which instructions are stored, and when the computer-readable storage medium is run on a computer, the computer executes the method in the above-mentioned first aspect and various possible designs of the first aspect.

[0059] In a fifth aspect, the present application also provides a computer program product comprising instructions, which, when executed on a computer, enables the computer to execute the method in the above-mentioned first aspect and various possible designs of the first aspect.

[0060] In a sixth aspect, the present application also provides a computing chip, which is connected to a memory, and the chip is used to read and execute a software program stored in the memory, and to execute the methods in the above-mentioned first aspect and various possible implementation methods of the first aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] In order to more clearly illustrate the embodiments of the present drawings or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present drawings. For ordinary technicians in this field, other drawings can be obtained based on the structures shown in these drawings without paying any creative work.

[0062] Figure 1 This is a structural diagram of this application;

[0063] Figure 2 This is a comparison chart of the wind force prediction results and actual wind force of this application.

[0064] The purpose, features and advantages of this figure will be further described in conjunction with the embodiments with reference to the accompanying drawings. DETAILED DESCRIPTION

[0065] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present application.

[0066] Obviously, the drawings described below are only some examples or embodiments of the present application. For ordinary technicians in this field, the present application can also be applied to other similar scenarios based on these drawings without creative work. In addition, it can also be understood that although the efforts made in this development process may be complicated and lengthy, for ordinary technicians in this field related to the content disclosed in this application, some changes in design, manufacturing or production based on the technical content disclosed in this application are just conventional technical means, and should not be understood as the content disclosed in this application is insufficient.

[0067] Unless otherwise specified, all embodiments and optional embodiments of the present application can be combined with each other to form a new technical solution.

[0068] Unless otherwise specified, all technical features and optional technical features of this application can be combined with each other to form a new technical solution.

[0069] If there is no special explanation, all steps of the present application can be performed sequentially or randomly, preferably sequentially. For example, the method includes steps (a) and (b), which means that the method may include steps (a) and (b) performed sequentially, or may include steps (b) and (a) performed sequentially. For example, the method may further include step (c), which means that step (c) may be added to the method in any order. For example, the method may include steps (a), (b) and (c), or may include steps (a), (c) and (b), or may include steps (c), (a) and (b), etc.

[0070] If there is no special explanation, the "include" and "comprising" mentioned in this application are open-ended or closed-ended. For example, the "include" and "comprising" may mean that other components not listed may also be included or only the listed components may be included or only the listed components may be included.

[0071] If not specifically stated, in this application, the term "or" is inclusive. For example, the phrase "A or B" means "A, B, or both A and B". More specifically, any of the following conditions satisfies the condition "A or B": A is true (or exists) and B is false (or does not exist); A is false (or does not exist) and B is true (or exists); or both A and B are true (or exist).

[0072] In order to better understand the solutions of the embodiments of the present application, some relevant terms and concepts that may be involved in the embodiments of the present application are first introduced below.

[0073] (1) Artificial intelligence (AI), also known as intelligent machinery or machine intelligence, refers to machines that are made by humans and can display intelligence. Generally speaking, AI refers to the technology that displays human intelligence through ordinary computer programs.

[0074] (2) Machine learning (ML). Machine learning is the core of artificial intelligence. The theory of machine learning mainly designs and analyzes some algorithms that allow computers to learn automatically. Machine learning algorithms are a type of algorithm that automatically analyzes data to obtain patterns and uses the patterns to predict unknown data. Therefore, the core of machine learning is data, algorithms (models), and computing power (computer computing power). The application areas of machine learning are very wide, including data mining, data classification, computer vision, natural language processing (NLP), biometric recognition, search engines, medical diagnosis, detection of credit card fraud, securities market analysis, DNA sequencing, speech and handwriting recognition, strategic games and robot applications. Machine learning is to design an algorithm model to process data and output the results that users want. Users can continuously tune the algorithm model to form more accurate data processing capabilities.

[0075] (3) Deep learning (DL). Deep learning is a method of machine learning. Its concept originates from the study of artificial neural networks. Multilayer perceptrons with multiple hidden layers are a type of deep learning structure, so deep learning is often referred to as deep neural networks. Compared with general machine learning, deep learning can automatically extract features, that is, automatically combine simple features into more complex features, and use these combinations to perform multi-layer weight learning to solve problems. The motivation for studying deep learning is to establish a neural network that simulates the human brain for analysis and learning. It imitates the mechanism of the human brain to interpret data, such as images, sounds, and text. Deep learning first emerged in image recognition, but in just a few years, deep learning has been extended to various fields of machine learning and has performed well. It has been applied in various fields such as image recognition, speech recognition, audio processing, natural language recognition, robot bioinformatics processing, search engines, human-computer games, online advertising targeting, medical automatic diagnosis, and finance.

[0076] (4) Neural networks (NN) are mathematical models that imitate the behavioral characteristics of animal neural networks and perform distributed parallel information processing. The purpose of information processing can be achieved by adjusting the interconnected relationships between a large number of nodes within the neural network. Neural networks have the ability to self-learn and adapt. Neural networks are usually used in artificial intelligence model training and data derivation processing.

[0077] Specifically, a neural network may usually include multiple layers connected end to end, such as convolutional layers, fully connected layers (FC), activation layers or pooling layers.

[0078] (5) Deep neural network (DNN), also known as multi-layer neural network, can be understood as a neural network with multiple hidden layers. According to the position of different layers, DNN can be divided into three categories: input layer, hidden layer, and output layer. Generally speaking, the first layer is the input layer, the last layer is the output layer, and the layers in between are all hidden layers. The layers are fully connected, that is, any neuron in the i-th layer must be connected to any neuron in the i+1-th layer. Although DNN looks complicated, the work of each layer is not complicated. Simply put, it is the following linear relationship expression: where is the input vector, is the output vector, is the offset vector, W is the weight matrix (also called coefficient), and α() is the activation function. Each layer simply performs such a simple operation on the input vector to obtain the output vector

[0079] (6) Convolutional neural network (CNN) is a deep neural network with a convolutional structure. A convolutional neural network contains a feature extractor consisting of a convolution layer and a subsampling layer, which can be regarded as a filter. A convolution layer refers to a neuron layer in a convolutional neural network that performs convolution processing on the input signal. In a convolutional layer of a convolutional neural network, a neuron can only be connected to some neurons in the adjacent layers. A convolutional layer usually contains several feature planes, each of which can be composed of a number of rectangularly arranged neural units. The neural units in the same feature plane share weights, and the shared weights here are convolution kernels. Shared weights can be understood as a way of extracting image information that is independent of position. The convolution kernel can be initialized in the form of a matrix of random size, and the convolution kernel can obtain reasonable weights through learning during the training process of the convolutional neural network. In addition, the direct benefit of shared weights is to reduce the connections between the layers of the convolutional neural network, while reducing the risk of overfitting.

[0080] (7) Graph Convolutional Network (GCN) is a deep learning model specifically designed for processing graph-structured data. It applies convolution operations on nodes to extract graph features and perform tasks such as classification or regression. It generalizes convolution operations from traditional Euclidean data space to non-Euclidean data structures such as graphs. Traditional convolution operations have good translation invariance in Euclidean space, but cannot maintain this property on graph-structured data. To solve this problem, GCN uses spectral domain and spatial domain methods to define graph convolution operations. The spectral domain method defines graph convolution operations by introducing filters, which are similar to filters in graph signal processing. By applying filters to graph signals, useful features can be extracted and tasks such as classification or regression can be performed. The key to this method is to select appropriate filters to better extract graph features. The spatial domain method represents graph convolution as aggregating feature information from neighbors.

[0081] (8) Transformer is a deep neural network based on the self-attention mechanism, which was originally applied in the field of natural language processing. Inspired by the powerful representation ability of Transformer, researchers proposed to extend Transformer to computer vision tasks. Compared with other network types (such as convolutional networks and recurrent networks), Transformer-based models show competitive or even better performance on various visual benchmarks.

[0082] (9) Fine-tuning is a technique used in natural language processing (NLP) to adapt a pre-trained language model to a specific task or domain. It fine-tunes the pre-trained model to improve its performance on a specific task. Fine-tuning has become an important model optimization method in the field of deep learning and is widely used in computer vision, natural language processing, speech recognition and other fields.

[0083] like Figure 1 As shown, in this embodiment, the wind prediction system applied to wind power generation includes:

[0084] The data acquisition module collects real-time wind-related data, including geographical environment information, seasonal changes, and meteorological element change information, and sends the data to the data processing module;

[0085] A wind power historical data collection module is used to collect historical data of wind power related data, mark the historical data, and send the data to the data processing module;

[0086] Historical data needs to be aligned with the input features of the model, which includes labeling key meteorological variables (such as 2m air temperature, dew point temperature, surface pressure, etc.), as well as relevant terrain data, wind speed components (such as 10m u wind component and v wind component) and other meteorological elements. These variables will be labeled according to time and space dimensions for model learning and prediction.

[0087] By analyzing the correspondence between historical wind data (such as wind speed and wind direction) and other relevant meteorological variables, the historical data is labeled as the input features and target outputs required for the prediction task, including setting the mapping relationship between the target variable (such as wind speed in the next 15 minutes, 30 minutes, and 10 days) and the current observation value. After the data is initially labeled using an automated algorithm, it is combined with manual verification steps to ensure the accuracy and consistency of the labeling, especially when complex meteorological events or abnormal weather conditions are involved.

[0088] The data processing module pre-processes the received data and extracts multi-dimensional data features through the dual attention mechanism. It then fuses spatiotemporal data features through the graph convolutional neural network, adaptively optimizes feature extraction and fusion strategies through joint training, and sends the processed data to the data storage module.

[0089] A data storage module, which receives data and stores the data in a data warehouse according to classification;

[0090] The large model training module builds a 3D transformer wind speed prediction model based on the visual transformer, sets model training hyperparameters, uses self-supervised training methods to train the model, inputs annotated wind history data sets to train the basic parameters of the large model, and then uses Fine-Tuning technology to further train the basic parameters of the large model until the average error and mean square error of the large model for real-time wind prediction meet expectations;

[0091] Display module: Based on the development of an interactive Web platform, the system development and visualization technology are integrated to visualize the prediction results of the large wind prediction model.

[0092] In this embodiment, the collected data include: aster terrain data with a resolution of 30m; surface air temperature related data from the latest released ERA6 reanalysis data set, including 2m air temperature, 2m dew point temperature, daily minimum air temperature at 2m, and daily maximum air temperature at 2m; surface air pressure related data from the latest released ERA6 reanalysis data set, including daily average sea level pressure and surface air pressure; precipitation related data from the latest released ERA6 reanalysis data set, including daily total precipitation; observable large-scale wind data from the latest released ERA6 reanalysis data set, including 10m u wind component and 10m v wind component; high-altitude wind speed data from the latest released ERA6 reanalysis data set, etc.

[0093] Among them, aster terrain data: standardize the terrain data with a resolution of 30 meters and align it with other data through interpolation and resampling operations. Extract terrain height, slope and other features to capture the impact of terrain on wind speed;

[0094] Surface air temperature related data: including 2m air temperature, 2m dew point temperature, daily minimum and maximum air temperature at 2m. These data are used to extract the impact of thermal conditions on wind speed;

[0095] Surface pressure related data: including daily average sea level pressure and surface pressure, mainly used to analyze the impact of pressure gradient on wind force;

[0096] Rainfall data: Daily totals are used to account for the changing effects of humidity and weather conditions on wind speed;

[0097] Large-scale wind data: including 10m u wind component and v wind component, as well as high-altitude wind speed data, are directly used to establish wind characteristics and correlate with surface microscale wind speed.

[0098] In this embodiment, the above-mentioned various types of data are feature extracted and fused through the dual attention mechanism and graph convolutional neural network (GCN) technology. The dual attention mechanism helps the model focus on important spatiotemporal features, and GCN is used to capture the spatial relationship between different data sources.

[0099] Furthermore, preprocessing includes designing a unified spatiotemporal representation framework for meteorological, geographic, remote sensing and other data from different sources and in different formats, achieving data standardization and alignment, and eliminating scale differences and inconsistencies between different data sources through spatiotemporal interpolation and resampling.

[0100] Spatiotemporal interpolation methods estimate and interpolate missing or inconsistent data by considering the correlation of data in time and space, thereby improving the consistency and accuracy of data; resampling techniques eliminate inconsistencies in data sets by changing the time interval or spatial resolution of data. Resampling can convert time series data from one frequency to another, increase granularity through upsampling or reduce granularity through downsampling, thereby achieving data consistency.

[0101] Furthermore, multi-dimensional data features are extracted through the dual attention mechanism, including:

[0102] Step 1: Introduce the dual attention mechanism to adaptively learn the importance weights of different data sources and different spatiotemporal locations;

[0103] The dual attention mechanism can effectively capture the differences and correlations between feature maps by calculating the attention weights in both spatial and channel dimensions.

[0104] Step 2: Automatically adjust the weight distribution between different data sources through the channel attention module;

[0105] Step 3: Dynamically focus on key information in different spatiotemporal regions through the spatiotemporal attention module;

[0106] The dual attention mechanism can also dynamically adjust the importance weights of different data sources and spatiotemporal locations. It learns the internal correlation between channels and the spatiotemporal dependency of features through the channel attention module and the spatiotemporal attention module, respectively, thereby achieving adaptive weight allocation for different data sources and spatiotemporal locations.

[0107] Step 4: The above two attention modules cooperate with each other to achieve adaptive optimization of feature extraction.

[0108] Furthermore, the spatiotemporal data feature fusion is performed through the graph convolutional neural network, specifically including:

[0109] Each spatiotemporal location is regarded as a node of the graph, and the associations between different data sources are regarded as edges of the graph. GCN is used to propagate and aggregate information on the graph structure to achieve feature interaction and fusion between different data sources and different spatiotemporal locations, and obtain a comprehensive feature representation that integrates time and space.

[0110] Furthermore, during joint training, reconstruction loss and adversarial loss are introduced simultaneously to ensure that the preprocessed features have good distinguishability and generalization while retaining key information.

[0111] Reconstruction loss is mainly used to measure the difference between the generated image and the original image. Common reconstruction loss functions include L1 loss and L2 loss. L1 loss is an absolute error loss, which is usually clearer and converges faster than L2 loss, i.e., mean square error loss, and is suitable for image reconstruction tasks. In some cases, L2 loss is also used for reconstruction loss. For example, in image super-resolution reconstruction, L2 reconstruction loss is used to ensure that the pixel difference between the reconstructed image and the original image is minimized.

[0112] Adversarial loss is used to train the generator and discriminator. The goal of the generator is to generate realistic samples, while the goal of the discriminator is to distinguish between real samples and generated samples. The adversarial loss function trains the generator by maximizing the error rate of the discriminator, so that it generates more realistic samples.

[0113] Furthermore, the design points of the 3D transformer wind speed prediction model specifically include: reconstructing the encoder and decoder to meet the model's ability to process highly characteristic historical wind-related data, using the Swin-Transformer sliding window attention mechanism to reduce the model's calculation amount, and using a hierarchical time domain aggregation strategy to reduce the number of model prediction iterations.

[0114] The sliding window attention mechanism in Swin-Transformer is used to limit the attention calculation to a local window, which helps reduce the computational complexity. Each local window only focuses on a limited adjacent space instead of all global positions, which significantly reduces the computational overhead. In the 3D variant transformer, this mechanism is extended to a three-dimensional window, allowing the model to efficiently learn local features at different time segments, spatial positions, and height levels.

[0115] Ordinary Transformers are mainly used for two-dimensional data (such as natural language processing, image recognition, etc.). The encoder and decoder modules of Transformers have limited processing of temporal or spatial information. In this application, the encoder and decoder are reconstructed to adapt to the data input of three-dimensional space, including:

[0116] First, the input layer is improved: the input layer is expanded to enable it to receive and process three-dimensional tensors. Specifically, the convolution layer of the input data is adjusted from 2D convolution to 3D convolution to capture the three-dimensional spatial characteristics of meteorological data.

[0117] Then, the three-dimensional extension of the self-attention mechanism: the self-attention mechanism is extended from 2D to 3D, so that the model can simultaneously pay attention to the correlation in time, space and height. Both the multi-head self-attention mechanism in the encoder and the cross-attention mechanism in the decoder are adjusted to adapt to the modeling of three-dimensional spatiotemporal relationships.

[0118] The main task of the encoder is to map the input data to a low-dimensional space. The feature extraction capability of the encoder can be improved by adding more hidden layers and adjusting the number of neurons in the hidden layers. For example, a convolutional layer can be used to extract features of image data. The improved encoder of the present application introduces a three-dimensional position encoding method, which enables the model to distinguish data points at different spatial positions and altitude levels. This encoding method combines spatial information such as altitude, latitude, and longitude to ensure that the model does not lose spatial structure information during the extraction of high-dimensional features. It can process three-dimensional features more efficiently and extract complex feature relationships of input data in space, time, and meteorological elements by using 3D convolution and 3D pooling layers.

[0119] The task of the decoder is to reconstruct the data in the low-dimensional space into the original data. The reconstruction effect can be improved by adding more decoding layers and adjusting the number of neurons in the decoding layer. This application enhances the feature back propagation capability of the decoder so that it can restore accurate wind speed prediction values ​​from high-dimensional features. The design of the decoder focuses on the output dimension of the prediction to improve the prediction accuracy of the wind speed in the target area.

[0120] The present application also uses a layered, fully connected layer with variable width (or a hybrid multi-layer perceptron MLP) to adapt to diverse three-dimensional feature inputs. This design can flexibly adjust the number of layers and neurons to improve the adaptability and processing capabilities of complex wind speed data.

[0121] Swin-Transformer introduces a sliding window operation to limit the attention calculation to a local window, thereby significantly reducing the computational complexity. Specifically, Swin-Transformer uses two sliding window operations: non-overlapping local windows and overlapping cross windows.

[0122] The core idea of ​​the hierarchical time-domain aggregation strategy is to take time as an important dimension and organize and display relevant information according to the chronological order. Through the hierarchical time-domain aggregation strategy, the model extracts and aggregates features at multiple time scales, thereby reducing the number of iterations. This hierarchical aggregation can integrate historical wind speed data hierarchically through a recursive window aggregation method, aggregate step by step from the short-term to the medium-term, and finally form a stable forecast output.

[0123] Furthermore, model training hyperparameters include learning rate, batch size, optimizer, and number of iterations.

[0124] Hyperparameters are parameters that need to be manually set before training a model. These parameters are not automatically learned during the training process, but need to be manually adjusted to optimize model performance. The adjustment of hyperparameters has an important impact on the performance of the model and the training process.

[0125] The learning rate is one of the most important parameters in the optimizer, which determines the step size of each parameter update. A reasonable learning rate can make the optimizer converge quickly, while a too large learning rate may cause instability in the training process, and a too small learning rate may lead to slow convergence. In the early stages of training, a larger learning rate is usually used. As the training progresses, the learning rate is gradually reduced to improve the stability of the model. Common learning rate adjustment strategies include learning rate decay and warm-up strategies.

[0126] Batch size refers to the number of samples used each time the model is trained. Larger batch sizes can improve training stability and memory utilization, but may require more computing resources and may reduce the generalization ability of the model. Small batch sizes can make the model more stable, but the training time is longer. The choice of batch size depends on many factors, including the complexity of the model, the available memory (GPU or CPU), the size of the training data, and the requirements of the specific task.

[0127] An optimizer is an algorithm used to adjust model parameters to minimize the loss function. Common optimizers include SGD, Adam, RMSProp, etc. Different optimizers have different characteristics and applicable scenarios. For example, the Adam optimizer improves training efficiency and stability through adaptive learning rate.

[0128] The number of iterations refers to the number of times the model completely traverses the training data. A reasonable number of iterations can improve the convergence speed and generalization ability of the model and avoid overfitting and underfitting. Usually you can start with 10 and decide whether to increase or decrease it based on the performance of the validation set. Too many iterations will increase the computational complexity and convergence speed.

[0129] Furthermore, the basic parameters of the large model are the learnable parts within the model, including weights and biases.

[0130] The weight class is responsible for connecting each neuron and adjusting the importance of signal transmission. The bias is responsible for adding a fixed value to the input signal of the neuron. These parameters are automatically updated during the model training process to optimize the performance of the model.

[0131] Furthermore, the Fine-Tuning technology is used to further train the basic parameters of the large model, including:

[0132] Step 1: Collect wind-related data in a specific area and form a specific fine-tuning dataset after preprocessing steps such as data cleaning;

[0133] Specific fine-tuning datasets refer to datasets used for specific tasks or fields during the fine-tuning of large models. These datasets usually contain input data and target data. The input data is the data that the model needs to process, and the target data is the result that the model needs to predict.

[0134] Step 2: Use Fine-Tuning technology to fine-tune and optimize the wind prediction model based on the constructed specific fine-tuning dataset;

[0135] Fine-Tuning technology involves updating all parameters of the model to adapt to the data of specific tasks. This method is effective and has strong fine-tuning capabilities for the model. Wind forecasting tasks require processing a large amount of time series data and complex meteorological information. Through Fine-Tuning technology, the performance of the model in wind forecasting tasks can be significantly improved.

[0136] Step 3: Establish relevant fine-tuning technical evaluation indicators to comprehensively evaluate the accuracy and generalization performance of wind forecasting in different actual regions and areas.

[0137] Furthermore, relevant fine-tuning technology evaluation indicators include model accuracy, precision and recall.

[0138] Accuracy refers to the ratio of the number of samples correctly predicted by the model to the total number of samples. Precision refers to the ratio of all samples predicted to be positive that are actually positive. Recall refers to the ratio of all samples actually positive that are correctly predicted to be positive. Accuracy, precision and recall are important indicators for evaluating the performance of classification models, and they reflect the quality of the model from different perspectives.

[0139] like Figure 2 As shown, in this embodiment, the comparison between the wind force prediction result and the actual wind force is shown in the figure, the prediction of the peak change and change range of the wind force is more accurate, and the error of the long-term prediction is smaller.

[0140] It should be noted that the present application is not limited to the above-mentioned embodiments. The above-mentioned embodiments are only examples, and the embodiments having the same structure as the technical idea and exerting the same effect within the scope of the technical solution of the present application are all included in the technical scope of the present application. In addition, without departing from the scope of the main purpose of the present application, various modifications that can be thought of by those skilled in the art to the embodiments and other methods of combining some of the constituent elements in the embodiments are also included in the scope of the present application.

Claims

1. A wind power prediction system for wind power generation, characterized in that: include: The data acquisition module collects real-time wind-related data, including geographical environment information, seasonal changes, and meteorological element change information, and sends the data to the data processing module; A wind power historical data collection module is used to collect historical data of wind power related data, mark the historical data, and send the data to the data processing module; The data processing module pre-processes the received data and extracts multi-dimensional data features through the dual attention mechanism. It then fuses spatiotemporal data features through the graph convolutional neural network, adaptively optimizes feature extraction and fusion strategies through joint training, and sends the processed data to the data storage module. A data storage module, which receives data and stores the data in a data warehouse according to classification; The large model training module builds a 3D transformer wind speed prediction model based on the visual transformer, sets model training hyperparameters, uses self-supervised training methods to train the model, inputs annotated wind history data sets to train the basic parameters of the large model, and then uses Fine-Tuning technology to further train the basic parameters of the large model until the average error and mean square error of the large model for real-time wind prediction meet expectations; Display module: Based on the development of an interactive Web platform, the system development and visualization technology are integrated to visualize the prediction results of the large wind prediction model.

2. The wind power prediction system for wind power generation according to claim 1, characterized in that: The preprocessing includes designing a unified spatiotemporal representation framework for meteorological, geographical, remote sensing and other data from different sources and in different formats, achieving data standardization and alignment, and eliminating scale differences and inconsistencies between different data sources through spatiotemporal interpolation and resampling.

3. The wind power prediction system for wind power generation according to claim 1, characterized in that: The method of extracting multidimensional data features through the dual attention mechanism specifically includes: Step 1: Introduce the dual attention mechanism to adaptively learn the importance weights of different data sources and different spatiotemporal locations; Step 2: Automatically adjust the weight distribution between different data sources through the channel attention module; Step 3: Dynamically focus on key information in different spatiotemporal regions through the spatiotemporal attention module; Step 4: The above two attention modules cooperate with each other to achieve adaptive optimization of feature extraction.

4. The wind power prediction system for wind power generation according to claim 1, characterized in that: The spatiotemporal data feature fusion by using a graph convolutional neural network specifically includes: Each spatiotemporal location is regarded as a node of the graph, and the associations between different data sources are regarded as edges of the graph. GCN is used to propagate and aggregate information on the graph structure to achieve feature interaction and fusion between different data sources and different spatiotemporal locations, and obtain a comprehensive feature representation that integrates time and space.

5. The wind power prediction system for wind power generation according to claim 1, characterized in that: During the joint training, reconstruction loss and adversarial loss are also introduced to ensure that the preprocessed features have good distinguishability and generalization while retaining key information.

6. The wind power prediction system for wind power generation according to claim 1, characterized in that: The design points of the 3D transformer wind speed prediction model specifically include: reconstructing the encoder and decoder to meet the model's ability to process highly characteristic historical wind-related data, using the Swin-Transformer sliding window attention mechanism to reduce the model's calculation amount, and using a hierarchical time domain aggregation strategy to reduce the number of model prediction iterations.

7. The wind power prediction system for wind power generation according to claim 1, characterized in that: The model training hyperparameters include learning rate, batch size, optimizer, and number of iterations.

8. The wind power prediction system for wind power generation according to claim 1, characterized in that: The basic parameters of the large model are the learnable parts within the model, including weights and biases.

9. The wind power prediction system for wind power generation according to claim 1, characterized in that: The use of Fine-Tuning technology to further train the basic parameters of the large model specifically includes: Step 1: Collect wind-related data in a specific area and form a specific fine-tuning dataset after preprocessing steps such as data cleaning; Step 2: Use Fine-Tuning technology to fine-tune and optimize the wind prediction model based on the constructed specific fine-tuning dataset; Step 3: Establish relevant fine-tuning technical evaluation indicators to comprehensively evaluate the accuracy and generalization performance of wind forecasting in different actual regions and areas.

10. The wind power prediction system for wind power generation according to claim 9, characterized in that: The related fine-tuning technology evaluation indicators include the accuracy, precision and recall rate of the model.

Images