Wind power plant wind speed prediction method and system based on double-domain space-time dynamic local sensing

Abstract

The invention discloses a wind power plant wind speed prediction method and system based on double-domain space-time dynamic local sensing. The method comprises the following steps: integrating fans in a wind power plant, a connection relationship and time sequence characteristics of the fans to form a space-time diagram; constructing a depth model comprising a plurality of DDL layers, wherein each DDL layer comprises a cross-domain filter, a dual-domain encoder and a dynamic local sensing recursive predictor; and projecting time-space diagram data, performing depth feature extraction on projected signal features through a plurality of DDL layers in the model, and averaging outputs of the plurality of DDL layers to form a final prediction result. The wind power plant wind speed prediction precision can be significantly improved, and the dynamic change rule of the wind speed can be effectively captured.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of wind speed prediction, and particularly relates to a wind speed prediction method for a wind farm based on double-domain spatio-temporal dynamic local perception. BACKGROUND

[0002] With the continuous rise of the proportion of new energy power generation, the uncertainty of wind farm output has become one of the key factors affecting the safe and economic operation of the power grid and the new energy consumption capacity. Among them, wind speed as the most direct meteorological element to determine the wind power output, its high-precision, ultra-short-term and short-term prediction results are not only the core basis for the power grid dispatching department to formulate rolling generation plan and optimize rotating reserve capacity, but also the prerequisite for the station side to participate in the spot and auxiliary service market. Therefore, to build a prediction model that can fully tap the spatio-temporal dynamic characteristics of wind speed data, has robust denoising ability and can self-adaptively depict the local propagation mechanism has become a technical bottleneck that needs to be broken through in the field of wind power.

[0003] However, the existing wind speed prediction technology still has the following significant defects when facing the real operating environment of the wind farm:

[0004] 1. Insufficient extraction of spatio-temporal coupling characteristics

[0005] The wind speed sequence has obvious double coupling properties: in the time dimension, the wind speed evolution presents periodic fluctuations; in the space dimension, influenced by various environmental factors, different machine sites show complex nonlinear correlation. Traditional statistical models (such as ARIMA, Kalman filter) or shallow machine learning methods (such as SVR, shallow ANN) often process time and space features separately, making it difficult to simultaneously depict the spatio-temporal coupling relationship.

[0006] In recent years, graph neural network (GNN) has shown significant advantages in representing the global topology of the wind farm by abstracting wind machines as nodes, spatial relationships between units as edges, and constructing a spatio-temporal graph with node features (wind speed, wind direction, power, temperature, etc.) and edge features (geographical distance, wake angle, elevation difference). However, most existing GNN methods simply concatenate time series as node attributes, lack explicit modeling of wind speed frequency domain characteristics, and thus limit the model's expression ability under complex weather conditions.

[0007] 2. Limited ability to suppress data noise and abnormal interference

[0008] The wind farm monitoring data inevitably introduces various noise sources in the process of collection, transmission and storage: sensor drift, communication packet loss, abnormal spikes caused by extreme weather, etc. Such noise is highly overlapped with the real wind speed signal in amplitude and frequency domain characteristics, and the traditional filtering or wavelet threshold method is easy to lose effective high frequency details while denoising.

[0009] Under the GNN framework, since the features of the space-time graph nodes are usually multivariate time series, their anomalies or missing values will spread along the edges during message passing, leading to the phenomenon of "abnormal pollution". Although some studies propose anomaly detection strategies based on graph contrast learning, by constructing positive and negative sample pairs to learn robust representations, they are still limited to node-level or graph-level coarse-grained discrimination and fail to suppress fine-grained anomalies in the time-frequency domain of wind speed sequences, making it difficult to accurately separate noise and effective signals in the prediction task.

[0010] 3. Dynamic local dependence modeling is missing

[0011] The wind speed propagation within the wind farm has obvious local space-time diffusion characteristics: on the one hand, the influence range of the upstream wind turbine wake on the downstream machine site varies dynamically with wind direction, wind speed and atmospheric stability; on the other hand, local micro-meteorological processes (such as canyon wind, thermal internal boundary layer) can cause adjacent units to exhibit different coupling strengths at different times.

[0012] Existing technologies mostly use global information for prediction, ignoring the dynamic changes caused by the influence of external factors on local units. Although some GNN variants attempt to dynamically weight the adjacency relationship with an attention mechanism, the weight calculation is still limited to the spatial domain and fails to process local information transmission jointly in the time-frequency domain, resulting in large prediction bias in the presence of local environmental differences. In addition, the lack of dual-domain (time-frequency) collaborative modeling methods makes it difficult for existing methods to accurately capture the local action radius and propagation speed corresponding to different frequency components.

[0013] In summary, there is an urgent need for a wind speed prediction method to overcome the systematic defects of existing GNN and space-time graph technologies in complex wind farm scenarios. SUMMARY

[0014] The present application is directed to the above problems, and provides a wind speed prediction method for wind farms based on dual-domain space-time dynamic local perception.

[0015] The technical solution of the present application is: a wind speed prediction method for wind farms based on dual-domain space-time dynamic local perception, comprising the following steps:

[0016] Step 1: Integrate the wind turbines in the wind farm, the connection relationship and the time sequence characteristics of the wind turbines to form a space-time graph;

[0017] Step two, constructing a deep model containing multiple DDL layers, each DDL layer containing a cross-domain filter, a dual-domain encoder and a dynamic local perception recursive predictor;

[0018] Step three, projecting the spatio-temporal graph data, and the signal features after projection are subjected to deep feature extraction through multiple DDL layers in the model, and the outputs of multiple DDL layers are averaged to form the final prediction result.

[0019] The spatio-temporal graph is constructed in the following manner:

[0020]

[0021] Among them, is a spatio-temporal graph, V is a node set, A is an adjacency matrix, is a time series feature matrix of the fan node, T x is the historical time step, and N is the number of nodes.

[0022] The spatio-temporal graph data is mapped to a hidden d-dimensional feature space through a multi-layer perception.

[0023] The cross-domain filter performs multi-scale decomposition on the signal through wavelet transform, frequency denoising through soft thresholding, and signal reconstruction through inverse transform.

[0024] In multi-scale decomposition, multiple different wavelet functions are used, including Daubechies wavelet function and Symlets wavelet function.

[0025] The dual-domain encoder encodes the denoised signal in time domain and frequency domain, and adopts a combination of multi-head self-attention mechanism and graph convolution kernel to capture the time series and spatial features of the wind farm wind speed data.

[0026] Through the dynamic local perception recursive predictor, based on the local state decoding and dynamic local update mechanism, the future fan state is iteratively predicted.

[0027] By decoding the historical features, the dynamic local dependence is modeled as an edge weight, which is used to weigh the connection between nodes, and based on the local state, the predicted future state and the past context state are iteratively updated.

[0028] A wind farm wind speed prediction system based on dual-domain spatio-temporal dynamic local perception, comprising:

[0029] The integration module is used to integrate the fans in the wind farm, the connection relationship and the time series features of the fans to form a spatio-temporal graph;

[0030] A construction module is configured to construct a deep model comprising a plurality of DDL layers, each DDL layer comprising a cross-domain filter, a dual-domain encoder, and a dynamic local perception recursive predictor.

[0031] A prediction module is configured to project the spatio-temporal graph data, and the signal features after the projection are subjected to deep feature extraction through the plurality of DDL layers in the model, and the outputs of the plurality of DDL layers are averaged to form a final prediction result.

[0032] The present application has the following effects:

[0033] 1. Improved prediction accuracy: By fine modeling of spatio-temporal correlation and dynamic local characteristics, the present application can more accurately predict wind speed changes, and has a significant improvement in prediction accuracy compared with traditional methods.

[0034] 2. Strong noise removal capability: The cross-domain filter effectively removes noise in wind power data, so that the model can still maintain good prediction performance when facing low-quality data, and enhances the robustness of the model.

[0035] 3. Efficient feature extraction: The dual-domain encoder combines the advantages of time domain and frequency domain features, fully captures the internal rules of wind power data, and improves the feature expression capability.

[0036] 4. Dynamic adaptability: The dynamic local perception recursive predictor can update the local dependency relationship in real time, adapt to the dynamic changes of wind speed, and improve the real-time prediction capability of the model.

[0037] Compared with traditional methods, the present application can significantly improve the wind speed prediction accuracy of the wind farm, and effectively capture the dynamic change rule of the wind speed. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1 is a principle block diagram of the present application,

[0039] Figure 2 is a working principle diagram of the cross-domain filter,

[0040] Figure 3 is a structure diagram of the dual-domain encoder,

[0041] Figure 4 is a schematic diagram of the dynamic local perception recursive predictor. DETAILED DESCRIPTION

[0042] The present application will be described in detail below in combination with the drawings and specific embodiments.

[0043] As Figures 1-4 shown, the present application provides a wind farm wind speed prediction method based on dual-domain spatio-temporal dynamic local perception, comprising the following steps:

[0044] Step one, integrate the wind turbines in the wind farm, the connection relationship and the time sequence characteristics of the wind turbines to form a space-time graph; used for capturing the space-time characteristics of the wind speed data of the wind farm.

[0045] Step two, build a deep model containing multiple DDL layers, each DDL layer contains a cross-domain filter, a dual-domain encoder and a dynamic local perception recursive predictor; to realize the deep feature extraction and prediction of the wind speed data of the wind farm.

[0046] Step three, project the space-time graph data, and the signal characteristics after projection are extracted through multiple DDL layers in the model, and the output of multiple DDL layers is averaged to form the final prediction result.

[0047] In view of the complex space-time correlation and noise interference problem in wind power data, a model containing DDL layer is constructed, and through the synergistic effect of cross-domain filter, dual-domain encoder and dynamic local perception recursive predictor, the wind speed is efficiently predicted.

[0048] Specifically as follows:

[0049] Step one: space-time graph structure design

[0050] In the present application, the wind farm has both time and space information, and a suitable organization structure needs to be used to describe it. This step integrates the wind space-time characteristics into a graph structure, providing an efficient data organization structure with rich information for subsequent processing and prediction.

[0051] Firstly, each wind turbine in the wind power system is regarded as a node, an adjacency matrix A is constructed according to the device connection relationship, and the historical wind speed data of each node is collected as a time sequence characteristic matrix Forming a space-time graph Wherein, V is the node set, T x is the historical time step, and N is the number of nodes.

[0052] Take this as input, the goal is to predict the wind speed of the future T y time step And the model parameters are optimized by minimizing the prediction error, and the optimization objective of the model is expressed as:

[0053]

[0054] Wherein, f(·) is a model with parameters θ, is the prediction result, T y is the predicted future time step.

[0055] Step two: multi-layer DDL neural network architecture building

[0056] In order to fully capture the time dependence and spatial correlation of the data of each wind turbine in the wind farm, the present application proposes a neural network architecture composed of multiple layers of DDL to model the features of the data and extract information, thereby improving the understanding of the dynamic changes of wind speed.

[0057] This step constructs a model containing L DDL layers (Dual-domain Dynamic Locality-aware Layer, hereinafter referred to as DDL). The input spatiotemporal graph data of the wind farm is first projected into a d-dimensional hidden feature space through a multi-layer perceptron (Multi-Layer Perceptron, hereinafter referred to as MLP), thereby obtaining a compact representation of information about the wind speed data of the wind turbine and possibly the operating state of the wind turbine. The MLP mapping process can be represented by the following formula:

[0058] H 0 =MLP(X)

[0059] where H 0 is the hidden state obtained after MLP mapping, which can be regarded as the output of the 0th layer of the model, MLP = X·W0+b0, W0 and b0 are the learnable network weights and bias terms, respectively. Subsequently, the projected signal features will be sent to multiple stacked DDL layers for deep feature extraction. The outputs of the multiple DDL layers are averaged and used as the final prediction result.

[0060] Cross-domain filtering denoising:

[0061] During the operation of the wind power system, wind power data may change in the short term. This change not only contains noise, but also coexists with long-term periodicity. This means that directly inputting the time series into the neural network not only introduces noise, but also requires a complex decomposition mechanism to decouple short-term changes and long-term patterns. In applications, first, a cross-time domain and frequency domain filtering strategy is used to suppress the noise contained in the data, so as to retain the periodicity characteristics that are more consistent with the actual operating conditions.

[0062] The cross-domain filter contained in the DDL layer includes three specific processing strategies, which are multi-scale decomposition of the signal through wavelet transform, frequency denoising through soft thresholding, and signal reconstruction through inverse transform. In the multi-scale decomposition strategy, the present application uses multiple different wavelet functions, including Daubechies wavelet function (abbreviated as db) and Symlets wavelet function (abbreviated as sym). For the input signal, i.e. the time series feature matrix X of the wind turbine node, the final denoised signal obtained after filtering can be represented as

[0063] First, the signal mapped by the MLP is decomposed into multiple scales through discrete wavelet transform:

[0064]

[0065] where j is the scale (decomposition level), representing the level of wavelet decomposition. j = 1 corresponds to the finest scale (high frequency details), and j = J is the coarsest scale (low frequency approximation). The resolution of the signal is halved with each level increase. n is the discrete time index, representing the sample point position at scale j, and m is the filter coefficient index. h[·] and g[·] are the low-pass and high-pass filters, respectively, which decompose the signal x into approximation coefficients a j and detail coefficients d j Here, the detail coefficients d j will be subjected to soft thresholding processing:

[0066]

[0067] where is a learnable threshold parameter. After the above process, the soft thresholding processed detail coefficients are denoted as Soft thresholding processing can suppress high-frequency noise while preserving the inherent patterns contained in the state information of each wind turbine node during wind power system operation.

[0068] Subsequently, the denoised signal needs to be reconstructed through inverse wavelet transform:

[0069]

[0070] where k is the discrete position index, representing the sample point number of the wavelet coefficients (or approximation coefficients) at scale j, and t is the discrete time index of the reconstructed signal, representing the time point of the original signal recovered . is the reconstructed low-pass filter, is the reconstructed high-pass filter, used to reconstruct the denoised signal back to the time domain. The reconstructed signal is denoted as

[0071] Dual-domain feature encoding:

[0072] After obtaining the denoised signal and hidden state features, it is necessary to model them in the time domain and frequency domain, respectively, to better utilize the complementarity of time-frequency signals, capturing both short-term patterns and long-term regularity changes in wind speed state.

[0073] The dual-domain feature encoder in the DDL layer will process the denoised signal and hidden state features

[0074] where represents the hidden state output by the dual-domain feature encoder in the lth DDL layer, Hl-1 represents the output result of the l-1th DDL layer dual-domain encoder. In the following, unless otherwise specified, the subscript l in the variable represents that the variable is the intermediate result calculated by the lth layer in the DDL model.

[0075] The time-domain encoder of the model first captures the historical dependency of the denoised signal by using a multi-layer stacked multi-head self-attention mechanism:

[0076]

[0077] where pe is the position encoding used to inject the order information of the sequence. SA(·) represents the self-attention mechanism operation, which enables the model to focus on the internal association of the input sequence by using the same matrix as q, k and v in MHA, i.e. SA(h) = MHA(h, h, h). represents the historical dependency result output by the time-domain encoder, and the dimension of the output result is N x T x d. x

[0078] The process of capturing sequence dependency by multi-head attention is represented by the following formula:

[0079] MHA(q, k, v) = Concat(head1,..., head h )·W o

[0080]

[0081] where Multi-head Attention is represented as MHA, and head represents the output result of each head in the multi-head attention, and there are h heads in total. q, k and v are query, key and value, respectively, and are learnable parameters in the neural network, and d is the dimension of the sequence feature hidden state.

[0082] Subsequently, a graph convolution kernel is applied to capture the spatial correlation of the last step:

[0083]

[0084] where A is the adjacency matrix of the wind farm network, represents the last step learned feature. GC(·) represents the graph convolution kernel operation. represents the spatial correlation result output by the time-domain encoder, and the dimension of the output result is N x d. Finally, the output of the time-domain encoder is obtained by MLP:

[0085] ​

[0086] wherein, denotes the final output of the time-domain encoder at time step t.

[0087] The frequency-domain encoder of the model first performs a fast Fourier transform on the denoised signal:

[0088]

[0089] wherein, FFT stands for the computation of the fast Fourier transform, denotes the complex domain, F l is the result of the transformation. Then, the real and imaginary parts are concatenated to form a new feature matrix, which has a 2d dimension:

[0090]

[0091] wherein, denotes the feature matrix formed by concatenating the real and imaginary parts, rm is the abbreviation for real part and imaginary part.

[0092] Then, the self-attention mechanism SA and the graph convolution kernel GC are applied respectively to capture the temporal dependencies and spatial correlations existing in the features:

[0093]

[0094] wherein, is the result obtained through the self-attention mechanism, is the result obtained through the graph convolution kernel operation. Finally, is split into real and imaginary parts, and MLP is applied respectively to obtain the output of the frequency-domain encoder:

[0095]

[0096] wherein, is the real part output obtained through the MLP, is the imaginary part output obtained through the MLP.

[0097] The dual-domain fusion aggregates the output of the dual-domain encoder using the inverse fast Fourier transform (iFFT) and the MLP:

[0098]

[0099] wherein, iFFT(·) converts the complex features in the frequency domain back to the time domain, ensuring that its dimension is aligned with the original input, and finally obtains the output result H l of the dual-domain encoder.

[0100] Dynamic local-aware prediction:

[0101] In the process of wind power system operation, the spatial dependence range of each wind turbine may change over time. In order to capture the dynamic dependence relationship changes of the local area over time, the present application adds a dynamic local-aware recursive predictor in the DDL layer.

[0102] By decoding the historical features, the dynamic local dependence is modeled as edge weights, which are used to weigh the connection between nodes, and the predicted future state Z and the context state of the past k steps (corresponding to λ l [i] in the formula below) are iteratively updated. l

[0103] The dynamic local-aware predictor contained in the DDL layer models the locality dependence relationship as the weight on the edge in the space-time graph, which is then modeled by the graph neural network (GNN). Specifically, it includes:

[0104] 1. The dynamic local-aware recursive predictor first obtains the state of the future k time steps through the MLP to realize the decoding of the local state:

[0105]

[0106] Wherein, represents the state of the future k time steps.

[0107] Then, the long-term future state is decoded through the self-attention and multi-head attention mechanism in turn:

[0108]

[0109] Wherein, is the intermediate result obtained by the self-attention mechanism acting on , and Z l is the long-term future state obtained by the multi-head attention mechanism.

[0110] 2. The locality is decomposed into the internal dynamics of the node itself and the cross dynamics caused by the connection of the nodes to realize the update of the dynamic locality. At each step t (0≤t<T y , first, the context κ l is accumulated through a sliding window of size k to obtain the context state λ l [t]:

[0111]

[0112] Wherein, κ l ​[t] is the accumulated context state of the past k time steps, λ l [i] is the context of the i-th historical time step.

[0113] Then, the locality of node u at time step t is calculated as:

[0114]

[0115] where, is the neighbor set of node u, denotes the state features collected from node u itself, denotes the state features collected from node u itself and its neighbors, v denotes a node in the neighbor set, is the locality state of node u at time step t.

[0116] 3. In the output stage, first, the edge weight needs to be calculated, which is mapped by the MLP to the sum of the locality of node u and node v respectively:

[0117]

[0118] where, denotes the weight of edge e connecting node u and node v at time step t.

[0119] After the edge weight calculation for each pair of nodes in the graph, the graph convolution operation is performed:

[0120]

[0121] where, is the locality representation of the full graph at time step t. Similar to the dual-domain encoder, the dynamic locality is decoded and modeled in the frequency domain by FFT and iFFT, obtaining the output in time and frequency domains and

[0122] Finally, the prediction of wind speed is aggregated in a gated manner and

[0123]

[0124] where δ is a learnable gating parameter that balances the contributions of time and frequency domains. and are the real and imaginary parts of respectively.

[0125] A wind farm wind speed prediction system based on dual-domain spatio-temporal dynamic locality perception, comprising:

[0126] An integration module is configured to integrate wind turbines, connection relationships and time sequence characteristics of the wind turbines in a wind farm to form a space-time graph;

[0127] A construction module is configured to construct a deep model comprising a plurality of DDL layers, each DDL layer comprising a cross-domain filter, a dual-domain encoder and a dynamic local perception recursive predictor.

[0128] A prediction module is configured to project the space-time graph data, and the signal characteristics after the projection are subjected to deep feature extraction through a plurality of DDL layers in the model, and the outputs of the plurality of DDL layers are averaged to form a final prediction result.

[0129] The present application has the following innovations:

[0130] 1. Space-time graph structure design: The wind farm network is modeled as a space-time graph, and the nodes, connection relationships and time sequence characteristics are integrated to provide a unified data structure for subsequent space-time feature extraction.

[0131] 2. Multi-layer DDL neural network architecture: A deep model comprising a plurality of DDL layers is constructed, and each DDL layer cooperates through a cross-domain filter, a dual-domain encoder and a dynamic local perception recursive predictor to gradually extract deep features of the data.

[0132] 3. Cross-domain filtering denoising: The cross-domain filter is used to perform frequency domain filtering processing on the input signal, the signal is decomposed through wavelet transform, soft threshold denoising is performed and the signal is reconstructed through inverse transform, and the noise components in the wind farm wind speed data are removed while the key patterns matching the real scene wind speed change rule are retained.

[0133] 4. Dual-domain feature encoding: The denoised signal is encoded in the time domain and the frequency domain respectively, and a multi-head self-attention mechanism and a graph convolution kernel are combined to capture the time sequence dependency and spatial correlation features of the wind farm wind speed data, and realize complementary feature extraction in time and frequency.

[0134] 5. Dynamic local perception prediction: The dynamic local perception recursive predictor is used to iteratively predict the future wind turbine state based on local state decoding and dynamic local updating mechanism, and dynamically model the local dependency relationship, and the local characteristics are taken as edge weights in the modeling process of the graph neural network to enhance the model's ability to capture the dynamic changes of the wind farm wind speed.

[0135] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and not to limit it, although the present application has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present application can be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present application, and they should be covered in the scope of the claims of the present application.

Claims

1. A wind farm wind speed prediction method based on dual-domain spatiotemporal dynamic local perception, characterized in that, The method comprises the following steps: Step one, integrating wind turbines, connection relationships and time sequence characteristics of the wind turbines in a wind farm to form a space-time graph; Step two, constructing a deep model comprising multiple DDL layers, each DDL layer comprising a cross-domain filter, a dual-domain encoder and a dynamic local perception recursive predictor; Step three, projecting the space-time graph data, and performing deep feature extraction on the projected signal features through multiple DDL layers in the model, and averaging the outputs of the multiple DDL layers to form a final prediction result.

2. The method of claim 1, wherein, The space-time graph is constructed in the following manner: wherein, is a space-time graph, V is a set of nodes, A is an adjacency matrix, is a time-series feature matrix of the fan nodes, T x is a historical time step, N is the number of nodes.

3. The method of claim 1, wherein, The space-time graph data is mapped to a hidden d-dimensional feature space through a multi-layer perception.

4. The method of claim 1, wherein, The cross-domain filter performs multi-scale decomposition on the signal through wavelet transform, frequency denoising through a soft threshold and signal reconstruction through inverse transform.

5. The method of claim 1, wherein, In the multi-scale decomposition, multiple different wavelet functions are used, including Daubechies wavelet functions and Symlets wavelet functions.

6. The method of claim 1, wherein, The dual-domain encoder performs time-domain and frequency-domain encoding processing on the denoised signal, and adopts a combination of a multi-head self-attention mechanism and a graph convolution kernel to capture the time sequence and spatial characteristics of the wind speed data of the wind farm.

7. The method of claim 1, wherein, Through the dynamic local perception recursive predictor, the future state of the wind turbine is iteratively predicted based on local state decoding and dynamic local update mechanism.

8. The method of claim 1, wherein, By decoding the historical features, the dynamic local dependence is modeled as an edge weight, which is used to weigh the connection between nodes, and based on the local state, the predicted future state and the past context state, the future state is iteratively updated.

9. A wind farm wind speed prediction system based on dual-domain spatiotemporal dynamic local perception, characterized in that, The method comprises: an integration module for integrating wind turbines, connection relationships and time sequence characteristics of the wind turbines in a wind farm to form a space-time graph; a construction module for constructing a deep model comprising multiple DDL layers, each DDL layer comprising a cross-domain filter, a dual-domain encoder and a dynamic local perception recursive predictor; a prediction module for projecting the space-time graph data, and performing deep feature extraction on the projected signal features through multiple DDL layers in the model, and averaging the outputs of the multiple DDL layers to form a final prediction result.

Images