Wind power prediction system based on multi-source data meteorological factors
Through multi-source data preprocessing and dual-channel DCCGA-ACNN model, combined with gray correlation analysis and variational modal decomposition, the problem of insufficient modal aliasing and characteristic regularity in wind power prediction is solved, and the accuracy of wind power prediction is improved. Stability, achieving high-quality wind power power prediction.
Patent Information
- Application Number
- CN202411088854.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-08
- Publication Date
- 2024-11-01
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
The existing wind power prediction model has modal aliasing and insufficient feature regularity when processing multivariate inputs, resulting in limited prediction accuracy. The component reconstruction after separate decomposition ignores the correlation between reconstruction vectors, increasing error and time-consuming .
Adopt a wind power prediction system based on multi-source data, including multi-source data acquisition, data preprocessing, gray correlation analysis, improved singular spectrum analysis and variational modal decomposition modules, combined with the dual-channel DCCGA-ACNN neural network model, through attention Mechanism and Bayesian optimization methods integrate spatial and temporal features to provide high-quality prediction results.
By removing noise components and extracting characteristics of stationary and periodic trends, the model's ability to capture space-time features is enhanced, the accuracy and interpretation ability of short-term wind field power prediction is improved, and high-quality prediction results are provided.
Smart Images

Figure CN118889409A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of wind power prediction, and in particular to a wind power prediction system based on multi-source data meteorological factors. Background Art
[0002] With the continuous development of the global economy, the problem of energy shortage has become increasingly prominent. At the same time, the widespread use of fossil energy has caused large-scale environmental problems. In this case, it is imperative to develop emerging clean energy to replace traditional energy. As an important part of new energy generation, wind power generation has played a positive role in energy conservation, emission reduction and sustainable development around the world. However, the output of wind power generation has a certain degree of randomness and instability, which has brought a series of negative impacts on the collection, conversion and dispatch of power Internet electricity. Therefore, it is of practical significance to improve the accuracy of short-term wind power prediction. In order to improve the accuracy of wind power prediction, scholars at home and abroad have proposed a series of prediction models. These methods can be roughly summarized into four categories: physical model, statistical model, artificial intelligence model and hybrid model.
[0003] However, in the current research on hybrid models for wind power prediction, most of them simply combine signal processing algorithms with neural models. Even if the relationship between multiple inputs and outputs can be considered, the use of a specific decomposition algorithm usually leads to modal aliasing of the reconstructed components, resulting in limited final prediction accuracy. Secondly, in existing studies, serial network structures are often used to extract spatiotemporal features of multiple inputs. The features extracted from one module are often passed to the next module, resulting in weak regularity of features extracted from the original data, which ultimately affects the prediction accuracy. In addition, predicting each component after decomposition and then adding them together not only ignores the correlation between the reconstructed vectors, but also leads to increased errors and time consumption. Similarly, if an inappropriate comprehensive prediction main network is selected, it may not only fail to correctly handle the correlation between the feature quantities, but also make the comprehensive prediction accuracy less accurate than the group prediction accuracy. To solve the above problems, we propose a wind power prediction system based on meteorological factors of multi-source data. Summary of the invention
[0004] In view of the deficiencies in the prior art, the present invention provides a wind power prediction system based on multi-source data meteorological factors to solve the problems raised in the background technology.
[0005] The above technical objectives of the present invention are achieved through the following technical solutions:
[0006] A wind power prediction system based on multi-source data meteorological factors, comprising:
[0007] 1.1 Multi-source data acquisition module, used to collect various types of data, including meteorological data (such as temperature, humidity, wind speed, wind direction, air pressure), historical wind power data, geographic information data and environmental data;
[0008] 1.2 Data preprocessing module, including:
[0009] A data cleaning unit, used to remove noise and erroneous data from the collected data, including a noise filtering algorithm;
[0010] Missing value processing unit, used to fill missing values;
[0011] An outlier detection and processing unit, used to detect and process outliers;
[0012] Data standardization unit, used to normalize data;
[0013] 1.3 Grey correlation analysis module, used to calculate the correlation between each meteorological factor and wind power, and screen out the meteorological factors with greater impact on wind power based on grey correlation analysis;
[0014] 1.4 Improve the singular spectrum analysis module, use singular value decomposition (SVD) technology to extract the main trends and change patterns in the data, and remove noise components;
[0015] 1.5 Variational Mode Decomposition (VMD) module uses an adaptive decomposition algorithm to perform secondary decomposition on complex high singular value data, extract modal signals with physical significance, and reduce the modal stacking phenomenon of high-dimensional components;
[0016] In the secondary decomposition process, different decomposition values k are tested to ensure that the sub-columns finally decomposed and reconstructed show a stable and periodic trend;
[0017] 1.6 Wind power prediction model module, including a DCCGA-ACNN neural network combination model with dual-channel input, where:
[0018] DCCGA (Dual Channel CNN and GRU with Attention Mechanism) submodule, including:
[0019] A two-channel convolutional neural network (CNN) is used to extract spatial features between multi-source data;
[0020] The gated recurrent unit (GRU) is used to extract the time-related features of the corresponding time series data;
[0021] The attention mechanism is used to enhance the model’s attention to important temporal and spatial features;
[0022] ACNN (Attention Convolutional Neural Network) submodule, including:
[0023] Through the global attention mechanism, each mixed spatiotemporal feature is deeply mined;
[0024] Horizontal fusion and vertical fusion operations are used to integrate spatial and temporal features;
[0025] The fully connected layer is used to generate the final wind power prediction results;
[0026] 1.8 Model integration and optimization module, including Bayesian optimization methods, for automatically adjusting model hyperparameters to improve the performance of predictive models;
[0027] 1.9 Prediction result output module, which provides prediction results including confidence intervals and can generate relevant statistical analysis reports;
[0028] 1.10 The system self-learning and updating module can regularly update the prediction model, retrain the model based on the latest collected data, and update the parameters to adapt to the impact of environmental changes on wind power.
[0029] Preferably, the multi-source data acquisition module further comprises a weather station data acquisition unit for acquiring weather data around the wind farm.
[0030] Preferably, the multi-source data acquisition module includes a remote sensing data acquisition unit for acquiring satellite remote sensing data, including cloud distribution and surface temperature.
[0031] Preferably, the multi-source data acquisition module includes an online monitoring data acquisition unit for acquiring real-time wind farm equipment operation data.
[0032] Preferably, the data cleaning unit in the data preprocessing module includes noise filtering and data cleaning algorithms for removing abnormal and erroneous data in the collected data.
[0033] Preferably, the grey correlation analysis module screens out the meteorological factors having a greater impact on wind power by calculating the grey correlation between each meteorological factor and wind power.
[0034] Preferably, the singular spectrum analysis module utilizes singular value decomposition technology to extract the main trends and change patterns in the data while removing noise components.
[0035] Preferably, the variational mode decomposition (VMD) module adopts an adaptive decomposition algorithm to perform secondary decomposition on complex data to extract modal signals with physical significance and reduce the modal stacking phenomenon.
[0036] Preferably, the processing corresponding advantage neural network submodule (CNN, GRU) in the wind power prediction model module is used to capture the spatial connection between multi-source data and the long-term dependence and nonlinear relationship in the time series data.
[0037] Preferably, the model integration and optimization module includes a Bayesian optimization method for automatically adjusting model hyperparameters to improve model performance.
[0038] Preferably, the prediction result provided by the prediction result output module includes a confidence interval and can generate a related statistical analysis report.
[0039] Preferably, the system self-learning and updating module can regularly update the prediction model, retrain the model based on the latest collected data, and update parameters to cope with the impact of environmental changes on wind power.
[0040] Preferably, the wind power prediction model module comprises the following steps:
[0041] 2.1 The geometric similarity between each meteorological factor and wind power is calculated through grey correlation analysis, and multi-dimensional input data with high correlation is screened out;
[0042] 2.2 Use the improved CAO-SSA algorithm to determine the number of decompositions, decompose the data, and discard the components with low contribution to remove the noise components;
[0043] 2.3 Perform VMD secondary decomposition on the decomposed data to ensure that each component has a stable and periodic trend, and remove irrelevant noise on high-dimensional components to reduce mode stacking;
[0044] 2.4 The multi-source data after secondary decomposition are mixed and input into the dual-channel DCCGA model according to the frequency, where:
[0045] Two-channel CNN is used to extract spatial features between multi-source data;
[0046] The GRU network is used to extract the time-related features of the corresponding data;
[0047] Attention mechanism is used to emphasize important temporal and spatial features;
[0048] 2.5 Perform vertical fusion through the ACNN network to explore the correlation between mixed spatiotemporal features, and output the final prediction results through the fully connected layer.
[0049] In summary, the present invention mainly has the following beneficial effects:
[0050] 1. By analyzing the geometric similarity between meteorological factors and wind power, factors that have a significant impact on wind power are screened out, providing high-quality features for model input, determining the optimal number of decompositions, eliminating components with low contribution, effectively removing noise, and improving data quality. By performing a secondary decomposition of the data, it is ensured that the reconstructed sub-columns present a stable and periodic trend, thereby improving data stability and prediction accuracy.
[0051] 2. The two-layer CNN is used to extract spatial features, and the GRU is used to extract temporal features. Parallel processing can enhance the capture of spatial and temporal features, help focus on important spatiotemporal features, improve sensitivity to key features, and enhance the model's ability to learn important information.
[0052] 3. Through the feature attention mechanism, the extraction process of spatiotemporal features is optimized, and the prediction accuracy of short-term wind farm power is improved. The global attention mechanism further explores the correlation characteristics between various mixed spatiotemporal features, and improves the prediction and explanation capabilities of the model. The prediction results after processing by the fully connected layer provide accurate wind power prediction values. The system can provide explainable results to help users understand the basis and process of the prediction.
[0053] In summary, the system can significantly improve the accuracy of short-term wind farm power forecasting through advanced model design and data processing technology. Through precise data decomposition, feature extraction and fusion, and the application of attention mechanism, the system can process complex time series data and provide high-quality forecast results. This comprehensive forecasting system is particularly suitable for wind farm operation and management that requires high-precision wind power forecasting. BRIEF DESCRIPTION OF THE DRAWINGS
[0054] Figure 1 It is a model diagram of the feature attention mechanism of the present invention;
[0055] Figure 2 It is a network structure diagram of the DCCGA-ACNN prediction model of the present invention;
[0056] Figure 3 It is the prediction framework and flow chart of the present invention;
[0057] Figure 4 is the spatial distance fluctuation diagram of the present invention;
[0058] Figure 5 is a singular value sequence contribution graph of the present invention;
[0059] Figure 6 It is the constructed mixed vector input graph of the present invention;
[0060] Figure 7 It is the grey relational graph of the present invention. DETAILED DESCRIPTION
[0061] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution of the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings of the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all the embodiments. Based on the described embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work belong to the scope of protection of the present invention.
[0062] The following examples are used to illustrate the present invention, but they cannot be used to limit the scope of protection of the present invention. The conditions in the examples can be further adjusted according to specific conditions. Simple improvements to the method of the present invention under the premise of the concept of the present invention belong to the scope of protection claimed in the present invention.
[0063] refer to Figure 1-Figure 7 , a wind power prediction system based on multi-source data meteorological factors, comprising:
[0064] 1.1 Multi-source data acquisition module, used to collect various types of data, including meteorological data (such as temperature, humidity, wind speed, wind direction, air pressure), historical wind power data, geographic information data and environmental data;
[0065] 1.2 Data preprocessing module, including:
[0066] A data cleaning unit, used to remove noise and erroneous data from the collected data, including a noise filtering algorithm;
[0067] Missing value processing unit, used to fill missing values;
[0068] An outlier detection and processing unit, used to detect and process outliers;
[0069] Data standardization unit, used to normalize data;
[0070] 1.3 Grey correlation analysis module, used to calculate the correlation between each meteorological factor and wind power, and screen out the meteorological factors with greater impact on wind power based on grey correlation analysis;
[0071] 1.4 Improve the singular spectrum analysis module, use singular value decomposition (SVD) technology to extract the main trends and change patterns in the data, and remove noise components;
[0072] 1.5 Variational Mode Decomposition (VMD) module, which uses an adaptive decomposition algorithm to perform secondary decomposition on complex data, extract modal signals with physical significance, and remove irrelevant noise;
[0073] In the secondary decomposition process, different k values are tested to select the appropriate decomposition dimension to ensure that the sub-sequences finally decomposed and reconstructed show a stable and periodic trend;
[0074] 1.6 Wind power prediction model module, including a DCCGA-ACNN neural network combination model with dual-channel input, where:
[0075] DCCGA (Dual Channel CNN and GRU with Attention Mechanism) submodule, including:
[0076] A two-channel convolutional neural network (CNN) is used to extract spatial features of multi-source data;
[0077] The gated recurrent unit (GRU) is used to extract the time-related features of time series data;
[0078] The attention mechanism is used to enhance the model’s attention to important temporal and spatial features;
[0079] ACNN (Attention Convolutional Neural Network) submodule, including:
[0080] Through the global attention mechanism, each mixed spatiotemporal feature is deeply mined;
[0081] Horizontal fusion and vertical fusion operations are used to integrate spatial and temporal features;
[0082] The fully connected layer is used to generate the final wind power prediction results;
[0083] 1.7 Model integration and optimization module, including Bayesian optimization methods, for automatically adjusting model hyperparameters to improve the performance of prediction models;
[0084] 1.8 Prediction result output module, which provides prediction results including confidence intervals and can generate relevant statistical analysis reports;
[0085] 1.9 The system self-learning and updating module can regularly update the prediction model, retrain the model based on the latest collected data, and update the parameters to adapt to the impact of environmental changes on wind power.
[0086] Example 1
[0087] Secondary decomposition: The power of a wind farm is restricted by multiple external factors such as wind speed, wind direction, and temperature, and has obvious lag. Considering that neural networks can mine deep correlation features between data at multiple levels, the final prediction accuracy can be significantly improved by adding multidimensional input features with high correlation characteristics with wind power output power. The grey correlation analysis (GRA) method, which can judge the degree of geometric similarity between sequences, is used to select multivariate meteorological data with high correlation characteristics with wind power. This paper uses secondary decomposition to decompose and reconstruct the screened multi-source data to improve the accuracy of power prediction. First, the improved CAO-SSA is used to remove the noise sequence of the screened multi-source data; the reconstructed high singular value sub-sequences are decomposed again by VMD to reduce the nonlinearity of the high-frequency sequence and make the sequence after secondary decomposition present a periodic trend, which is convenient for the multi-branch network module to extract spatiotemporal features;
[0088] Multivariate meteorological data preprocessing: gray correlation analysis is performed on the meteorological data of wind farms, such as wind speed, wind direction, temperature, humidity, air pressure, etc. at the hub and at different heights, and the filtered data is normalized to the maximum and minimum values and used as the subsequent research object;
[0089]
[0090] Selection of multivariate meteorological data: Changes in meteorological and geographical environments are important factors affecting the output of wind turbines. Randomly changing meteorological conditions have an important impact on the accuracy of predicting wind speed and wind power. In order to give full play to the ability of neural networks to extract multiple mappings between input and output, it is important to improve the prediction accuracy by supplementing the coupling relationship between multidimensional variables and wind power output. Text selection uses the grey correlation analysis method to obtain multivariate input data sets.
[0091] The basic principle of grey correlation analysis is to judge whether the connection between different sequences is close based on the geometric shape of the sequence curve. The sequence with the greatest impact factor is selected by comparing the similarity between the external multivariate factor data and wind power. i ={x i (1), x i (2), x i (3), …, x i (n)}, where X0(n), X i (n) is the comparison sequence, γ(x0(n), x i (n)) The relevant relationship is:
[0092]
[0093] Where: x 0i =x0(k)-x i (k); is the k point x0(k) and x i (k) minimum absolute difference;
[0094]
[0095] is the maximum absolute difference of k points; ρ∈[0,1] is the resolution coefficient, which is generally taken as 0.5.
[0096] X0,X i The grey relational degree γ(X0, X i )for:
[0097]
[0098] Improved singular spectrum analysis denoising: Excessive noise components in the original data will cause distortion in neural network feature extraction and reduce prediction accuracy. The noise generated by external factors in nature and the limitations of the sensor's own conditions is indispensable. Singular spectrum analysis is a method of principal component analysis of time series based on singular value decomposition (SVD). As an adaptive decomposition algorithm, SSA plays a good role in non-stationary signal processing. The decomposition process includes four steps: trajectory matrix nesting, SVD decomposition, grouping, and reconstruction. However, there are currently two major problems: one is the selection of embedding dimensions for trajectory matrix construction, and the other is to determine the main sequence after grouping and reconstruction. This paper uses the phase space reconstruction CAO method to determine the embedding dimension and retain a certain degree of sequence (cumulative singular value contribution rate ≥ 95%) as the main component.
[0099] Maps the F time series with N samples to a multidimensional lag vector sequence, integer As the embedding window width, let the sequence {f i , f i+1 , ..., f i+L-1}where i=0,...,NL constitutes the trajectory matrix X:
[0100]
[0101] Perform SVD decomposition on the trajectory matrix X:
[0102]
[0103] The singular value σ i It is XX T The square root of the eigenvalue λ is greater than 0. u i and v i are the left and right singular vectors corresponding to the i-th singular value of the trajectory matrix X; the rank of the matrix X is given.
[0104] In the grouping stage, d sub-matrices X iDivided into several non-overlapping subsets, the original sequence F can be regarded as a collection of trend signals and noise signals. The purpose of grouping is mainly to select appropriate principal components and filter out the influence of noise on the original signal:
[0105]
[0106] The basic matrix X i The diagonal average is used to obtain a new time series sub-series with a length of N. Known X i yes
[0107] L×K (K=N-L+1) matrix, where the elements are x m,n ,
[0108] 1≤m≤L, 1≤n≤K,K*=MAX(L,K),L * =MIN(L, K). Use the following formula to convert the matrix X i Reconstructing into time series
[0109]
[0110] SSA is highly non-parametric. Too large or too small an embedding dimension L will affect the size of the feature differences contained in different components after decomposition. This paper uses the phase space reconstruction CAO method to determine the embedding dimension L, and determines the parameter L by calculating the geometric distance variables related to the attractor. When the embedding dimension increases to a certain stage, the mean of these spatial distance variables will enter a stable state, and the appropriate embedding dimension L will be obtained.
[0111] Calculate the change in distance between nearest neighbors under different embedding dimensions L:
[0112]
[0113] Where: i = 1, 2, ..., NL; ||·|| is the ∞-norm; X i (L+1) is the i-th reconstructed phase space vector of L+1, X n(i,L) (L+1) is the ∞-norm and X i (L+1) The vector closest to the vector, where 1≤n(i, L)≤(NL), and the value of n(i, L) remains unchanged under different embedding dimensions.
[0114] Calculate the mean of the spatial distance variable under the same embedding dimension L:
[0115]
[0116] Use E1(L) to monitor the fluctuation of E(L) under different dimensions L: When the corresponding E1(L) value stops changing or fluctuates less than a certain threshold, the minimum L is l0+1, and the optimal embedding dimension is l0.
[0117] After decomposing and reconstructing the multivariate data, the influence of noise is reduced by removing the low singular value contribution rate sequence. At this time, the few remaining main components account for most of the original signal information. To this end, this paper uses the VMD method to perform secondary decomposition on the high singular value sequence to further reduce its complexity.
[0118] The VMD method is a completely non-recursive modal decomposition method. The specific approach is to convert the sequence into a variational problem during the decomposition process, and adaptively match the optimal center frequency and corresponding bandwidth of each mode by iterative search for the optimal solution, so as to achieve effective separation of the intrinsic mode components. The decomposition process is as follows:
[0119] 1). Decompose the original signal X into K finite bandwidth modal functions, and obtain each modal function u through Hibert transformation j (t) One-sided spectrum.
[0120] 2). For u j (t) The center frequency of the spectrum ξ j (t) Add item and adjust it to the baseband area.
[0121] 3) Estimate u by calculating the L2 norm of the modulated signal j The bandwidth of (t) is used to construct a variational model with the bandwidth of each component and the minimum constraint:
[0122]
[0123] In the formula {u j},{ξ j} are respectively the set of modal functions and the set of their corresponding spectrum center frequencies; is the partial derivative operator; * represents the convolution operation; δ(t) is the unit pulse function.
[0124] Feature attention mechanism: Although deep learning algorithms can fit data through network parameters and iterative loss functions, complex network structures usually lead to overfitting. There are many reasons for abnormalities, but they are often difficult to eliminate intuitively due to the unique operating mechanism within the network. The attention mechanism adaptively allocates input variable weights, allowing the model to focus on important input information, thereby reducing computational loss and improving computational stability and generalization capabilities.
[0125] For a variable X with an input length of M, X = [x1, x2, ..., x M ] as an example, a linear layer is used to calculate the attention weight vector Y as shown below.
[0126] Y=σ(WX+white)#(1)
[0127] In the formula: Y=[y1, y2,...,y M ] represents the combination of attention weight vectors corresponding to each input feature at the current moment.
[0128] In order to stabilize the training process and normalize the weights, the output Y is processed using the Sigmoid function to obtain a series of continuous smooth values, and then the Softmax function is used to convert these real numbers into α so that the sum of their probability distribution is 1. Considering the influence of the attention weight, the input vector X is re-weighted to obtain X′, as shown in (2):
[0129] X′=αX=[α1x1, α2x2,…,α M x M ]#(2)
[0130] Feature attention mechanism: With the increase of sensors, the scale of meteorological features that can be captured is gradually increasing. However, excessive meteorological features are not only unhelpful to the prediction results of the model, but may even interfere with the prediction results. Although the appropriate input vector can be screened out through correlation analysis, the impact of the screened vectors on the final power prediction varies. The feature attention mechanism is used to extract features from the model input to focus on key information.
[0131] Feature attention mechanism model such as Figure 1 As shown in the figure, the difference is that the input at this time is different feature data dimensions in the wind farm. The weight of each feature at each moment is calculated through a multi-layer perceptron, and the contribution of many input features of the model to the wind farm output is deeply mined, the expression of feature information that has a greater impact on the output is enhanced, and the impact of features with low correlation on the prediction model is reduced.
[0132] Spatial attention mechanism: When comprehensively predicting the spatiotemporal features of each mixed vector, the CNN network is used to extract features from the coupling connections of each component. Since the convolution kernel of the convolutional network is limited by size and cannot capture global information, in order to avoid the final prediction accuracy being reduced due to local information, the spatial attention mechanism is introduced to weight the input features with global information, so that it can dynamically focus on the importance of different positions in the input, thereby better extracting key information without increasing too much computational complexity.
[0133] This paper introduces the spatial attention mechanism based on the connection between the mixed vectors extracted by the convolutional network. By processing the input feature map, a weight map of the same size as the input feature map is generated, and it is multiplied with the original input to obtain a weighted feature map. The attention area is dynamically adjusted according to the input features so that the model can capture important features more flexibly and enhance the model's representation ability. At this time, Figure 1 The input is the mixed spatiotemporal features and the output is the spatial attention weight. This process can effectively solve the field of view problem of traditional convolution kernels while avoiding the feature accuracy defects caused by using dilated convolution kernels to extract discontinuous features.
[0134] Multi-branch input DCCGA-ACNN deep learning model with attention mechanism: In order to make up for the problem that some details and regularities between components are ignored when extracting serial spatiotemporal features, a DCCGA-ACNN prediction model is proposed, and its structure is as follows Figure 2 As shown. A multi-branch network composed of parallel multi-layer gated GRU and convolutional CNN extracts spatiotemporal features from each mixed component after decomposition, and at the same time, the rules of multivariate components and wind power data are complemented, thereby enhancing effective feature input and improving prediction accuracy. The model proposed in this paper consists of two parts: one is the spatiotemporal feature extraction model, and the other is the feature mixed prediction module. Feature attention is added to each mixed component composed of multivariate factors and wind power data to highlight key information and use it as the input of the multi-branch parallel spatiotemporal feature extraction module, thereby obtaining the spatiotemporal fusion features output by a single branch network. The integrated multiple two-dimensional spatiotemporal features are then vertically spliced and input into the convolutional network through spatial attention to comprehensively extract the internal connections of multiple spatiotemporal features, which alleviates the local field of view problem of the convolution kernel and improves the prediction performance of the overall network model.
[0135] Spatiotemporal feature extraction module: The spatiotemporal feature extraction module proposed in this paper consists of two parts. One is to use the feature attention mechanism for a given multivariate time series, and the other is to connect two layers of convolutional CNN and GRU networks in parallel to the weighted input sequence to fully extract the spatiotemporal mixed vector. For the input, there are m multivariate time series as The weight of the i-th feature at time t is calculated through a multi-layer perceptron
[0136] λ t =o(W λ x t +b λ )#(9)
[0137] Where: W λ is the training weight matrix; b λ is the bias vector. The Softmax function is used to adjust the weight scores λ on the output results. tNormalize it so that the sum of the weights is 1. t With the input feature vector x t Dot multiplication to get the weight vector
[0138] Vector As the input of the single-branch spatiotemporal feature extraction module, the temporal features of the multivariate mixed components are fully extracted through a two-layer GRU network. Extract spatial features through a two-layer convolutional CNN network In order to simplify the calculation and avoid overfitting, the activation function RELU is added after each convolution layer. In order to improve the accuracy of the extracted spatiotemporal features, the number of hidden layers of the GRU network and the number of convolution kernels of the CNN network are twice the number of the previous layer. The spatiotemporal vectors obtained from the two channels are horizontally spliced to obtain the fused spatiotemporal vector
[0139] Feature integration prediction module: The feature integration prediction module is composed of the Aadding feature fusion module connected in series with the convolutional CNN network with spatial attention and the fully connected layer to obtain the final prediction result. First, the spatiotemporal vectors output by multiple branches are Perform vertical fusion to obtain the spatiotemporal signal F of the entire multivariate input st , then the spatiotemporal vectors are serially input into the convolutional network with spatial attention, and the coupling relationship between the mixed components of the multivariate input is fully extracted while alleviating the local field of view problem of the convolution kernel. Finally, the prediction result Y is obtained through the fully connected layer. t .
[0140] Example 2
[0141] In order to improve the wind power prediction accuracy of wind farms, this paper proposes a power prediction method based on quadratic decomposition and multi-branch input (DCCGA-ACNN). The specific process of this method is as follows Figure 3 It mainly consists of three steps: multivariate factor selection and data preprocessing, secondary decomposition to construct mixed vectors, and integrated spatiotemporal feature prediction. The corresponding data is supported by the data of a wind farm in Ningxia in 2017.
[0142] 1) In the neural network prediction research on wind farm power generation, it was found that the prediction accuracy of a single power input was not high. Providing multiple related input data can complement the characteristics of the power component and effectively improve the prediction accuracy. Therefore, the multi-dimensional data collected in the wind farm are classified for power correlation, and the multi-dimensional signals in the wind farm, such as the hub and meteorological data such as wind speed, wind direction, temperature, humidity, air pressure at different heights, are analyzed with the power data for gray correlation. The screened data are normalized and used as subsequent research objects.
[0143] 2) The original multivariate signal contains too many disordered and noisy factors, which need to be further decomposed to improve the prediction accuracy. n The CAO method is used to determine the embedding dimension of the singular value decomposition, and the original signal is decomposed into several mixed subsequences according to the frequency. By calculating the cumulative contribution of the singular values of each mixed subsequence, each sequence is classified into high singular value, periodic trend and noise mixed components. At this time, the improved singular value decomposition method can well remove the noise component in the original mixed signal, but the high singular value mixed components left by the decomposition still have a large number of signal stacking phenomena. In order to reduce its complexity, considering that low-dimensional replacement of high-dimensional can reduce model calculation, the VMD decomposition algorithm is used to perform secondary decomposition of the high singular value mixed sequence. Afterwards, the mixed components after the secondary decomposition and the periodic trend components after the initial decomposition are used as the input signals of each branch network to predict the wind power value.
[0144] 3) For each mixed vector X′ constructed after the secondary decomposition input ={X′1,X′2,…,X′ n} is used as network input, and the key features between the mixed components are highlighted through feature attention. A parallel spatiotemporal network is used to perform spatiotemporal feature fusion. At the same time, the spatiotemporal features of each branch network are vertically spliced and sent to the convolutional network with spatial attention to integrate and predict the final wind power regression value.
[0145] In the study of neural network prediction of wind farm power generation, it is found that providing multiple related input data can complement the power component to effectively improve the prediction accuracy, but too many input variables will increase the amount of network calculation and make convergence difficult. Therefore, gray correlation analysis is used to show the correlation between power and multiple meteorological factors to reduce the input dimension and improve the prediction accuracy. The gray correlation between each variable and wind power is calculated by formula (1) (2), and the results are shown in Table 1:
[0146] Table 1 shows the correlation coefficients between various multivariate meteorological data and wind power, where the wind direction angle is uniformly converted into a cosine function value and then correlated with the power. Table 1 shows that the wind speed at the hub height has the largest correlation coefficient with the power.
[0147] Among the temperature, air pressure and humidity variables, the temperature environmental variable has the greatest correlation with power. Considering that the hub wind speed has the greatest impact on power and the correlation of wind directions at different heights is not much different, the hub wind direction data is selected here. Finally, the hub wind speed, wind direction, temperature and wind power are selected as the subsequent research objects.
[0148] Table 1. Correlation coefficients between multiple meteorological variables and wind power
[0149]
[0150] In order to reduce the randomness and instability in multivariate data and wind power time series, this paper uses the improved singular value analysis method to decompose and reconstruct the original mixed vector after screening. In order to determine the embedding dimension of the SSA method, the CAO method based on the spatial reconstruction principle is used to calculate the spatial distance fluctuation of different vectors. Figure 4 As shown;
[0151] from Figure 4 It can be seen that when L is selected as 22, power, temperature and wind speed stop fluctuating first, and the wind direction sequence is still fluctuating. Considering that the correlation between wind direction and power is the weakest and to avoid over-decomposition, the wind direction embedding dimension is selected to be the same as the above sequence.
[0152] Singular value decomposition achieves the purpose of denoising by truncating singular values, and can effectively retain the main information of the trajectory matrix by retaining a larger part of the singular values. Figure 5 As shown in the figure, each component 1 after decomposition contains most of the features in the original signal, each component 2 occupies secondary information, the first 7 components of power, wind speed, and temperature vectors occupy 99.5% of the information, and the subsequent components are considered to be noise components, while the wind direction from components 3 to 10 steadily contributes about 0.5%-1%, and the cumulative contribution reaches 95% in the first sequence 10, and is considered to be a noise component afterwards. Therefore, each component 1 is considered to be a high singular value signal, and the first 7 components of power, wind speed, and temperature vectors are reconstructed into periodic trend signals, and the rest are discarded as noise signals. Wind direction components 2 to 10 constitute trend signals, and the rest are discarded, as shown in Table 2.
[0153] Table 2. SSA decomposition and reconstruction
[0154]
[0155]
[0156] Since component 1 generated by SSA decomposition occupies most of the original signal features, it will have problems such as high frequency and mode stacking, which is difficult to predict. Therefore, based on this, this paper uses VMD secondary decomposition to further decompose each component 1 into more periodic component signals to improve the prediction effect. Compare whether the central frequencies of each decomposition component are close under different k values to determine the number of modes decomposed by VMD. The results are shown in Table 3.
[0157] Table 3. Center frequencies for different K values under different input vectors (before normalization)
[0158]
[0159] As shown in Table 3, as the K value increases, the center frequency difference of different components under each vector decreases significantly. When K = 5, the center frequencies of power, wind speed and wind direction vectors v4 and v5 are similar, and the center frequencies of temperature vectors v3 and v5 are similar, resulting in excessive decomposition. Therefore, the decomposition number K value is selected as 4.
[0160] 3.2.3 Constructing multi-source data mixed components
[0161] In order to avoid the defects of the single input decomposition method, the power, wind speed, temperature, wind direction components after secondary decomposition and the reconstructed trend components are used to construct a mixed vector X' according to the frequency. input , in order to achieve complementary advantages of multivariate components in similar frequency bands. It should be noted that after the temperature vector is decomposed by VMD, the center frequencies of IMF3 and IMF4 components are opposite to those of other components and need to be reconstructed. The specific results are as follows Figure 6 shown.
[0162] In summary, this paper proposes a dual-channel input DCCGA-ACNN (Dual channel CNN and GRU with Attention Mechanism-Attention MechanismviaCNN) neural network combination model based on secondary decomposition and with attention mechanism for short-term wind farm power prediction. Firstly, the grey correlation method is used to analyze the geometric similarity between multi-source meteorological data and wind power data. The improved CAO-SSA algorithm is used to determine the number of decompositions for multi-dimensional inputs with high correlation, and the low-contribution components are discarded to remove the noise components. At the same time, the highest singular value sub-columns are subjected to VMD secondary decomposition so that the sub-columns after decomposition and reconstruction show a stable and periodic trend. Secondly, the multi-source components after secondary decomposition are mixed according to the frequency and input into the DCCGA dual-channel spatiotemporal feature extraction module with feature attention. The spatial similarity features and temporal correlation features between the mixed components are extracted in parallel using the double-layer CNN and GRU networks and then horizontally fused. Finally, the spatiotemporal features of the fused multivariate mixed components are vertically fused and sent to the ACNN network with a global attention mechanism to mine the correlation features between the mixed spatiotemporal features. The fully connected layers are connected to obtain the final prediction results, which can significantly improve the accuracy of short-term wind farm power prediction. Through precise data decomposition, feature extraction and fusion, and the application of attention mechanism, the system can process complex time series data and provide high-quality prediction results. This comprehensive prediction system is particularly suitable for wind farm operation and management that requires high-precision wind power prediction.
[0163] Although the embodiments of the present invention have been shown and described, it is understandable to those skilled in the art that, unless otherwise defined, the technical terms or scientific terms used in the present invention shall have the usual meanings understood by persons with ordinary skills in the field to which the present invention belongs. The words "including" or "comprising" and the like used in the present invention mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. The words "connect" or "connected" and the like are not limited to physical or mechanical connections, but may also include electrical connections, whether direct or indirect. "Up", "down", "left", "right", etc. are only used to indicate relative position relationships. When the absolute position of the described object changes, the relative position relationship may also change accordingly.
[0164] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.
Claims
1. A wind power prediction system based on multi-source data meteorological factors, characterized by comprising: 1.1 Multi-source data acquisition module, used to collect various types of data, including meteorological data (such as temperature, humidity, wind speed, wind direction, air pressure), historical wind power data, geographic information data and environmental data; 1.2 Data preprocessing module, including: A data cleaning unit, used to remove noise and erroneous data from the collected data, including a noise filtering algorithm; Missing value processing unit, used to fill missing values; An outlier detection and processing unit, used to detect and process outliers; Data standardization unit, used to normalize data; 1.3 Grey correlation analysis module, used to calculate the correlation between each meteorological factor and wind power, and screen out the meteorological factors with greater impact on wind power based on grey correlation analysis; 1.4 Improve the singular spectrum analysis module, use singular value decomposition (SVD) technology to extract the main trends and change patterns in the data, and remove noise components; 1.5 Variational Mode Decomposition (VMD) module uses an adaptive decomposition algorithm to perform secondary decomposition on complex data, extract modal signals with physical significance, and reduce the phenomenon of modal stacking; In the secondary decomposition process, different k values are tested to select the appropriate decomposition dimension to ensure that the sub-sequences finally decomposed and reconstructed show a stable and periodic trend; 1.6 Wind power prediction model module, including a DCCGA-ACNN neural network combination model with dual-channel input, where: DCCGA (Dual Channel CNN and GRU with Attention Mechanism) submodule, including: A two-channel convolutional neural network (CNN) is used to extract spatial features between multi-source data; It is used with the gated recurrent unit (GRU) to extract temporal correlation features of time series data; The attention mechanism is used to enhance the model’s attention to important temporal and spatial features; ACNN (Attention Convolutional Neural Network) submodule, including: Through the global attention mechanism, each mixed spatiotemporal feature is deeply mined; Horizontal fusion and vertical fusion operations are used to integrate spatial and temporal features; The fully connected layer is used to generate the final wind power prediction results; 1.7 Model integration and optimization module, including Bayesian optimization methods, for automatically adjusting model hyperparameters to improve the performance of prediction models; 1.8 Prediction result output module, which provides prediction results including confidence intervals and can generate relevant statistical analysis reports; 1.9 The system self-learning and updating module can regularly update the prediction model, retrain the model based on the latest collected data, and update the parameters to adapt to the impact of environmental changes on wind power.
2. A wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The multi-source data acquisition module further includes a weather station data acquisition unit for acquiring weather data around the wind farm.
3. A wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The multi-source data acquisition module includes a remote sensing data acquisition unit for acquiring satellite remote sensing data, including cloud distribution and surface temperature.
4. A wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The multi-source data acquisition module includes an online monitoring data acquisition unit for acquiring real-time wind farm equipment operation data.
5. The wind power prediction system based on multi-source data meteorological factors according to claim 1 is characterized in that: The data cleaning unit in the data preprocessing module includes noise filtering and data cleaning algorithms, which are used to remove abnormal and erroneous data in the collected data.
6. A wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The grey correlation analysis module screens out the meteorological factors that have a greater impact on wind power by calculating the grey correlation between each meteorological factor and wind power.
7. The wind power prediction system based on multi-source data meteorological factors according to claim 1 is characterized in that: The improved singular spectrum analysis module utilizes singular value decomposition technology to extract the main trends and change patterns in the data while removing noise components.
8. The wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The variational mode decomposition (VMD) module adopts an adaptive decomposition algorithm to perform secondary decomposition on complex high singular value data to extract modal signals with physical significance and reduce the modal stacking phenomenon of high singular value components.
9. The wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The wind power prediction model module uses a neural network submodule with corresponding advantages to capture the spatial connection between multi-source data and the long-term dependence and nonlinear relationship in the corresponding time series data.
10. A wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The model integration and optimization module includes a Bayesian optimization method for automatically adjusting network model hyperparameters to improve model performance and enhance prediction fit.
11. The wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The prediction results provided by the prediction result output module include confidence intervals and can generate relevant statistical analysis reports.
12. A wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The system self-learning and updating module can regularly update the prediction model, retrain the model based on the latest collected data, and update parameters to cope with the impact of environmental changes on wind power.
13. The wind power prediction system based on multi-source data meteorological factors according to claim 1, characterized in that: The wind power prediction model module The following steps are involved: 2.1 The geometric similarity between each meteorological factor and wind power is calculated through grey correlation analysis, and multi-dimensional input data with high correlation is screened out; 2.2 Use the improved CAO-SSA algorithm to determine the number of decompositions, decompose the data, and discard the components with low contribution to remove the noise components; 2.3 Perform VMD secondary decomposition on the decomposed high singular value data to ensure that each component has a stable and periodic trend and remove irrelevant noise; 2.4 The multi-source data after secondary decomposition are mixed and input into the dual-channel DCCGA model according to the frequency, where: The CNN network is used to extract spatial features between multi-source data; The GRU network is used to extract the time-related features of the corresponding data; Attention mechanism is used to emphasize important temporal and spatial features; 2.5 Perform vertical fusion through the ACNN network to explore the correlation between mixed spatiotemporal features, and output the final prediction results through the fully connected layer.
Citation Information
Cited By
Virtual power plant productivity prediction method and system
CN119417071A
Radar working mode single-step prediction method based on time sequence fusion Transform
CN119885067A
Rainfall prediction method and system based on multi-source data fusion and dynamic and static space-time network
CN119960088A
Optical storage direct flexible system short-term load prediction method based on multilevel data association fusion
CN120127625A
Multi-level data correlation fusion-based short-term load forecasting method for optical storage and direct flexible system
CN120127625B
