High-frequency wind farm turbulence prediction method based on parallel liquid state space and flow matching

Abstract

The application discloses a high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching, and belongs to the field of deep learning and fluid mechanics prediction. In view of the serial calculation bottleneck, multi-objective optimization water bed effect and high-frequency information dissipation problems existing in the prior art, the application firstly constructs a prediction neural network connected by a Chebyshev KAN feature extraction module, an adaptive conditional normalization module and a parallel liquid state space module in sequence, outputs hidden state features, generates a preliminary prediction sequence through optimal transport flow matching, adopts a joint loss function to guide network training, and performs exponential moving average processing on parameters, and finally performs statistical re-calibration on the prediction result based on regional meteorological priori. The application effectively breaks the serial calculation bottleneck, overcomes the multi-objective optimization contradiction, significantly improves the prediction accuracy and high-frequency detail fidelity, and realizes high-fidelity real-time prediction of high-frequency wind fields.

Description

Technical Field

[0001] This application relates to the field of fluid dynamics prediction, and more specifically, to a high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching. Background Technology

[0002] Wind energy, as a clean and renewable energy source, has been widely applied in grid-connected wind power generation. However, the operation of wind turbines is significantly affected by complex flow fields on-site. Wind speed time-series data under complex flow field environments exhibit obvious non-stationary characteristics, compounded by random pulsations caused by turbulence, resulting in drastic changes in the high-frequency turbulent components of wind speed. Accurately reproducing and predicting the high-frequency turbulent characteristics of wind speed is a key prerequisite for achieving precise load control of wind turbines, quantitative assessment of turbine fatigue life, and safe operation and maintenance of wind farm projects.

[0003] Existing wind speed forecasting solutions mainly fall into three categories: numerical simulation forecasting, conventional deep learning time series forecasting, and general generative model forecasting. However, each solution has inherent shortcomings in terms of high-frequency turbulence detail reproduction and real-time inference efficiency. First, traditional numerical weather prediction combined with fluid dynamics simulation technology relies on iteratively solving the wind field evolution law based on fluid control equations. The solution process involves cumbersome grid partitioning and large iterative computation, resulting in high overall computational costs, making it difficult to meet the application requirements of millisecond-level online real-time forecasting at wind power sites. Second, conventional deep learning time series forecasting methods, represented by LSTM and Transformer models, generally use mean squared error as a loss constraint to complete model training. Under this optimization objective, the model converges to fit the expected sample distribution, which easily leads to mean regression problems during multi-step time series extrapolation. The predicted wind speed curve is over-smoothed, resulting in high-frequency information dissipation defects. It cannot effectively retain the key high-frequency turbulent kinetic energy and pulsation extreme value information of the wind field, and cannot reproduce the random pulsation characteristics of the real wind field. Third, while standard generative models such as basic flow matching and diffusion models have the ability to generate random samples, the conventional polynomial expansion-type deep networks that accompany these models are equivalent to low-pass filter structures, which naturally filter out high-frequency detail components. In addition, traditional time series networks use serial operation logic for long sequence inference, resulting in long computation links and technical drawbacks such as high computing power consumption and large online inference latency. They also cannot meet the requirements of high-frequency wind speed prediction accuracy and real-time performance. Summary of the Invention

[0004] The purpose of this invention is to provide a high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching, so as to solve the problems of high-frequency dissipation and low computational efficiency in the existing deep learning prediction models.

[0005] To achieve the above objectives, the present invention provides the following technical solution: A prediction neural network is constructed based on the KAN model and parallel liquid state space. The historical wind speed sequence is input into the prediction neural network, and the hidden state features are output. The latent state features are input into a vector field prediction network based on optimal transport flow matching to obtain a predicted wind speed vector field. The first predicted wind speed sequence is generated by solving the predicted wind speed vector field. The loss value between the first predicted wind speed sequence and the actual wind speed sequence is calculated using a joint loss function, and the training parameters of the prediction neural network are updated through backpropagation based on the loss value. The updated training parameters are processed by an exponential moving average to obtain the inference parameters; The inference parameters are loaded into the prediction neural network to obtain a trained prediction neural network. The historical wind speed sequence of the wind field to be predicted is input into the trained prediction neural network, and the second predicted wind speed sequence is output. The second predicted wind speed sequence is statistically recalibrated based on regional meteorological priors to obtain the final high-frequency wind field prediction sequence.

[0006] In one alternative implementation, the predictive neural network includes a feature extraction module, an adaptive conditional normalization module, and a parallel liquid state space module connected in sequence.

[0007] In one optional implementation, the feature extraction module performs feature extraction using a KAN model, specifically including: The input historical wind speed sequence is mapped to the interval [-1,1] by using the hyperbolic tangent function, and the Chebyshev polynomial is used as the basis function to perform a nonlinear transformation on the mapped sequence. The nonlinear features are extracted based on the nonlinear transformation results, expressed as follows:

[0008] In the formula, It is a nonlinear characteristic. The weight matrix is ​​the linear transformation matrix. This is the mapped historical wind speed sequence. Let be the highest order of the Chebyshev polynomial. For the first Learnable coefficients of the Chebyshev polynomial. For the first Chebyshev polynomials It is the hyperbolic tangent activation function.

[0009] In one alternative implementation, the adaptive conditional normalization module performs the following steps: Historical meteorological condition features and time step coding features are obtained, and then added element by element to obtain modulation features; The modulation features are sequentially input into a linear transformation layer and a nonlinear activation layer, and the output is a scaling factor and a translation factor. After performing layer normalization on the nonlinear features, and combining the scaling factor and translation factor, the nonlinear features fused with meteorological time series information are output, expressed as follows:

[0010] In the formula, To integrate the nonlinear characteristics of meteorological time series information, Presentation layer normalization operation, It is a nonlinear characteristic. As a scale factor, This is the translation factor.

[0011] In one alternative implementation, the parallel liquid state-space module performs the following steps: The time-varying step size is dynamically generated based on the nonlinear characteristics of fused meteorological time-series information, and the expression is:

[0012] In the formula, For dynamically generated time steps, It is the Softplus non-linear activation function. For linear mapping layer, To integrate the nonlinear characteristics of meteorological time series information, It is a Sigmoid non-linear activation function. This represents a multi-layer perceptron. This represents a one-dimensional convolution operation; By discretizing the system matrix and input matrix in the logarithmic domain using a time-varying step size, the expression is as follows:

[0013]

[0014] In the formula, The system matrix is ​​the result of discretization mapping. For the system matrix, The input matrix is ​​the result of discretization mapping. The input matrix; Based on the previous hidden state, the discretized system matrix, and the input matrix, the current hidden state is calculated as follows:

[0015] In the formula, This is the current hidden state. This is the hidden state from the previous moment. To integrate the nonlinear characteristics of meteorological time series information.

[0016] In one optional implementation, the parallel liquid state-space module further includes: Static graph compilation enables the computation of time-series loops to be pushed down to the underlying C++ operator layer for parallel execution.

[0017] In one optional implementation, the hidden state features are input into a vector field prediction network based on optimal transport flow matching to obtain a predicted vector field. A first predicted wind speed sequence is generated by solving the predicted vector field, specifically including: A uniformly distributed original time step is generated using stratified sampling. The original time step is then subjected to cosine transformation to obtain the optimized time step, expressed as:

[0018] In the formula, The time step after deformation optimization. The original time steps are uniformly distributed; A vector field prediction network based on optimal transport flow matching is constructed. The hidden state features and the optimized time step are input into the vector field prediction network, and the predicted wind speed vector field is output. Solving the ordinary differential equations of the predicted wind speed vector field yields the first predicted wind speed sequence, expressed as:

[0019] In the formula, This is the first predicted wind speed sequence. It is a standard Gaussian noise sequence. To predict the vector field.

[0020] In one optional implementation, the joint loss function includes a vector field mean square error loss term, a first- and second-order derivative joint physical constraint loss term, and a multi-resolution short-time Fourier transform spectral constraint loss term. The expression for the mean square error loss term of the vector field is:

[0021] In the formula, This represents the mean square error loss value for the vector field. Let be the mean square error function. To predict the velocity field, For the target velocity field; The expression for the joint physical constraint loss term of the first and second derivatives is as follows:

[0022] In the formula, The loss value is the combined physical constraint loss of the first and second derivatives. For stream matching time steps, The mean absolute error function, This represents the first-order difference operation. and These are the first predicted wind speed sequence and the actual wind speed sequence, respectively. This represents a second-order difference operation; The expression for the multi-resolution short-time Fourier spectrum constraint loss term is:

[0023] In the formula, This represents the multi-resolution short-time Fourier spectrum constraint loss value. For stream matching time steps, For window size, The mean absolute error function, This indicates the short-time Fourier transform operation. and These are the first predicted wind speed sequence and the actual wind speed sequence, respectively. To prevent extremely small positive numbers from being set to zero.

[0024] In one optional implementation, the expression for applying an exponential moving average to the updated training parameters is:

[0025] In the formula, For inference parameters, For smoothing coefficients, These are the updated training parameters.

[0026] In one optional implementation, the second predicted wind speed sequence is statistically recalibrated based on regional meteorological priors to obtain the final high-frequency wind field prediction sequence, specifically including: Obtain the mean and standard deviation of the historical wind speed sequence of the wind field to be predicted, and obtain the mean and standard deviation of the second predicted wind speed sequence; Further standardization and reconstruction of the second predicted wind speed sequence yields the final high-frequency wind field prediction sequence, expressed as:

[0027] In the formula, This is the final high-frequency wind field prediction sequence. This is the second predicted wind speed sequence. and These are the mean and standard deviation of the second predicted wind speed series, respectively. and These are the mean and standard deviation of the historical wind speed sequence for the wind field to be predicted, respectively.

[0028] This application has at least the following advantages or beneficial effects: (1) This invention obtains high-frequency nonlinear mutation characteristics through Chebyshev and KAN models, and combines regional meteorological prior statistics recalibration to correct the inherent energy dissipation problem of generative models, so as to ensure that the statistical characteristics of the predicted wind field completely restore the real physical meteorological environment.

[0029] (2) This invention introduces a parallelized ultra-fast liquid state space operator, which completely sinks the time series loop to the underlying operator for parallel execution through static graph compilation, solving the problem of low computational efficiency of autoregressive prediction models under long-sequence high-frequency data and meeting the requirements of industrial-grade ultra-fast response.

[0030] (3) The present invention uses a joint loss function to guide network training, so that the prediction neural network avoids losing the long-range macro trend in the process of learning micro-nonlinear turbulent fluctuations, and the predicted trajectory is robust and does not diverge. Attached Figure Description

[0031] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments of this application will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart of the high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching provided by the present invention; Figure 2 This is a comparison chart of the power spectrum and energy spectrum of the first predicted wind speed sequence and the actual wind speed sequence in an embodiment of the present invention. The left sub-chart is a comparison chart of the power spectral density across the entire frequency band, and the right sub-chart is a comparison chart of the premultiplied energy spectrum. Figure 3 This is a multi-resolution short-time Fourier phase error heatmap of the first predicted wind speed sequence and the actual wind speed sequence in an embodiment of the present invention. Figure 4 This is a comparison chart of the final high-frequency wind field prediction sequence and the actual wind speed sequence in an embodiment of the present invention. The upper sub-chart is a full-scale wind speed tracking comparison chart, and the lower sub-chart is a magnified view of the transition area between the historical input and the prediction interval. Detailed Implementation

[0033] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0034] Please refer to Figure 1 , Figure 1 This is a flowchart of a high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching, proposed in an embodiment of this application. Figure 1 As shown, a high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching includes: Step S1: Construct a prediction neural network based on the KAN model and parallel liquid state space, input the historical wind speed sequence into the prediction neural network, and output the hidden state features; In one embodiment, the predictive neural network includes a feature extraction module, an adaptive conditional normalization module, and a parallel liquid state space module connected in sequence.

[0035] In one embodiment, the feature extraction module performs feature extraction using a KAN model, specifically including: The input historical wind speed sequence is mapped to the interval [-1,1] by using the hyperbolic tangent function, and the Chebyshev polynomial is used as the basis function to perform a nonlinear transformation on the mapped sequence. Nonlinear features are extracted based on the nonlinear transformation results, expressed as follows:

[0036] In the formula, It is a nonlinear characteristic. The weight matrix is ​​the linear transformation matrix. This is the mapped historical wind speed sequence. Let be the highest order of the Chebyshev polynomial. For the first Learnable coefficients of the Chebyshev polynomial. For the first Chebyshev polynomials It is the hyperbolic tangent activation function.

[0037] The recursive formulas for the Chebyshev polynomial basis functions are as follows:

[0038]

[0039]

[0040] In the formula, This is the mapped historical wind speed sequence. , These are Chebyshev polynomials of order 0 and order 1, respectively. , and The first Rank, First Rank and first Chebyshev polynomial of order 1.

[0041] The feature extraction module uses learnable Chebyshev polynomial basis functions, which can adaptively fit arbitrarily complex nonlinear functions and is particularly good at capturing high-frequency abrupt changes in wind speed sequences.

[0042] The low-order Chebyshev polynomial is mainly responsible for capturing the overall trend and low-frequency fluctuations of the wind speed sequence, while the high-order Chebyshev polynomial is mainly responsible for capturing the high-frequency abrupt changes and fine structure of turbulence, thereby achieving comprehensive extraction of multi-scale wind field characteristics.

[0043] In one embodiment, the adaptive conditional normalization module performs the following steps: Historical meteorological condition features and time step coding features are obtained, and then added element by element to obtain modulation features; The modulation features are sequentially input into a linear transformation layer and a nonlinear activation layer, and the output is a scaling factor and a translation factor. After performing layer normalization on the nonlinear features, and combining the scaling factor and translation factor, the nonlinear features fused with meteorological time series information are output, expressed as follows:

[0044] In the formula, To integrate the nonlinear characteristics of meteorological time series information, Presentation layer normalization operation, It is a nonlinear characteristic. As a scale factor, This is the translation factor.

[0045] The adaptive normalization module integrates macroscopic meteorological conditions and time-step information into the turbulence features, achieving multi-scale information fusion. Compared with traditional fixed-layer normalization, this enables the model to better adapt to wind field variations under different weather conditions, significantly improving prediction robustness.

[0046] In one embodiment, the parallel liquid state-space module performs the following steps: The time-varying step size is dynamically generated based on the nonlinear characteristics of fused meteorological time-series information, and the expression is:

[0047] In the formula, For dynamically generated time steps, It is the Softplus non-linear activation function. For linear mapping layer, To integrate the nonlinear characteristics of meteorological time series information, It is a Sigmoid non-linear activation function. This represents a multi-layer perceptron. This represents a one-dimensional convolution operation; By discretizing the system matrix and input matrix in the logarithmic domain using a time-varying step size, the expression is as follows:

[0048]

[0049] In the formula, The system matrix after discretization and mapping. For the system matrix, The input matrix is ​​the result of discretization mapping. The input matrix; Based on the previous hidden state, the discretized system matrix, and the input matrix, the current hidden state is calculated as follows:

[0050] In the formula, This is the current hidden state. This is the hidden state from the previous moment. To integrate the nonlinear characteristics of meteorological time series information.

[0051] In one embodiment, the parallel liquid state space module further includes: using static graph compilation to push the time-series loop computation down to the underlying C++ operator layer for parallel execution.

[0052] For example, a time-series loop computation process can be compiled into a static computation graph using PyTorch's torch.jit.script or TensorFlow's tf.function, and the loop logic can be completely buried and solidified so that it is only executed at the underlying C++ operator layer.

[0053] It's important to explain that traditional state-space models rely on time-step computation using a for loop in the Python layer, and the overhead of the Python interpreter increases linearly with the sequence length. This invention completely eliminates the Python interpreter overhead by sinking the loop down to the underlying operator layer, and simultaneously leverages the parallel optimization capabilities of the underlying operators to achieve fully parallel computation in the time dimension.

[0054] The parallel liquid state space module is the core innovation of this invention in solving the bottleneck of serial computing and realizing ultra-fast inference. Through dynamic time-varying step size and static graph compilation technology, the serial computing of traditional state space model is transformed into parallel computing, which greatly improves the efficiency of long time series prediction.

[0055] Step S2: Input the hidden state features into the vector field prediction network based on optimal transport flow matching to obtain the predicted wind speed vector field, and generate the first predicted wind speed sequence by solving the predicted wind speed vector field. In one embodiment, the latent state features are input into a vector field prediction network based on optimal transport flow matching to obtain a prediction vector field. A first predicted wind speed sequence is generated by solving the prediction vector field, specifically including: A layered sampling method is used to generate an original time step uniformly distributed in the [0,1] interval. The original time step is then subjected to cosine transformation to obtain the optimized time step, expressed as:

[0056] In the formula, The time step after deformation optimization. The original time steps are uniformly distributed; A vector field prediction network based on optimal transport flow matching is constructed. The hidden state features and the optimized time step are input into the vector field prediction network, and the predicted wind speed vector field is output. An explicit Euler integrator is used to solve the ordinary differential equations of the predicted wind speed vector field, yielding the first predicted wind speed sequence, expressed as:

[0057] In the formula, This is the first predicted wind speed sequence. It is a standard Gaussian noise sequence. To predict the vector field.

[0058] The expression for the MSE loss of the wind speed vector field is:

[0059] In the formula, The MSE loss value for the vector field. Let be the mean square error function. The target wind speed vector field.

[0060] Cosine deformation increases the sampling density at both ends of the time step (t approaches 0 and t approaches 1), which are precisely the regions where the vector field changes most drastically. In this way, the accuracy of ordinary differential equation (ODE) solutions can be significantly improved without increasing the total number of sampling steps.

[0061] As an optional implementation, in some scenarios where higher prediction accuracy is required, the Runge-Kutta method can also be used to solve the problem, thereby further improving the prediction accuracy by increasing the computational load slightly.

[0062] This step transforms the wind field prediction problem into an optimal transmission problem, and achieves the mapping from Gaussian noise to the actual wind speed by fitting a vector field, thus avoiding the high latency problem of multi-step sampling in traditional diffusion models.

[0063] Step S3: Calculate the loss value between the first predicted wind speed sequence and the actual wind speed sequence using a joint loss function, and update the training parameters of the prediction neural network based on the loss value through backpropagation. In one embodiment, the joint loss function includes a vector field mean square error loss term, a first-order and second-order derivative joint physical constraint loss term, and a multi-resolution short-time Fourier transform spectral constraint loss term. The expression for the mean square error loss term of the vector field is:

[0064] In the formula, This represents the mean square error loss value for the vector field. Let be the mean square error function. To predict the velocity field, For the target velocity field; The expression for the joint physical constraint loss term of the first and second derivatives is as follows:

[0065] In the formula, The loss value is the combined physical constraint loss of the first and second derivatives. For stream matching time steps, The mean absolute error function, This represents the first-order difference operation. and These are the first predicted wind speed sequence and the actual wind speed sequence, respectively. This represents a second-order difference operation; The expression for the multi-resolution short-time Fourier spectrum constraint loss term is:

[0066] In the formula, This represents the multi-resolution short-time Fourier spectrum constraint loss value. For stream matching time steps, For window size, The mean absolute error function, This indicates the short-time Fourier transform operation. and These are the first predicted wind speed sequence and the actual wind speed sequence, respectively. To prevent extremely small positive numbers from being set to zero; In this embodiment, the power spectral density (PSD) and premultiplied energy spectrum of the first predicted wind speed sequence and the actual wind speed sequence are calculated respectively. Figure 2 As shown in the left sub-figure, it can be seen that the power spectral density curve of the first predicted wind speed sequence highly coincides with the actual wind speed across the entire frequency band, and the attenuation slope of the predicted curve in the inertial sub-region matches the gray dashed line (K41 -5 / 3 scaling law). Figure 2 As shown in the right subfigure, it can be seen that the peak frequency and peak amplitude of the premultiplied energy spectrum of the first predicted wind speed sequence have a high degree of overlap with the actual wind speed sequence. This indicates that the model has successfully learned the physical laws of the turbulent energy level series: the energy transfer process from large-scale energetic eddies to small-scale dissipative eddies, achieving accurate anchoring of energy distribution across the entire frequency band, and the power spectral density of the predicted wind speed follows Kolmogorov's K41 theory.

[0067] The window size values ​​cover three characteristic regions of the turbulent energy spectrum: the small window (W=16, about 10Hz resolution) has the highest time domain resolution and is used to capture the pulsation details of the high-frequency dissipation region; the medium window (W=32, about 5Hz resolution) has a balanced time domain and corresponds to the inertial sub-region in the turbulent energy spectrum; and the large window (W=64, about 2.5Hz) has the highest frequency domain resolution and ensures the spectral consistency of low-frequency large-scale structures.

[0068] like Figure 3 As shown, the multi-resolution short-time Fourier spectrum values ​​between the first predicted wind speed sequence and the actual wind speed sequence are calculated respectively, and the distribution of phase error on the time-frequency plane is displayed in the form of a heatmap. It can be seen that the three windows cover the complete frequency band from the high-frequency dissipation region to the low-frequency energy-containing region, with the color scale ranging from blue (zero error) to red (π-phase out of phase). The entire frequency band shows a large area of ​​blue, and the phase error is maintained at a low level, proving that the model output does not have phase lag or lead deviation in the temporal flow direction. This ensures the complete physical characteristic restoration of the predicted sequence from the large-scale macroscopic trend to the small-scale high-frequency fluctuations.

[0069] This step completely solves the "waterbed effect" problem in multi-objective optimization of traditional generative models by decoupling and constraining features at different scales in the frequency domain.

[0070] It's important to clarify that the "waterbed effect" refers to a technical problem in the field of multi-objective optimization in deep learning. Its physical principle is similar to a waterbed: when one area of ​​the waterbed is pressed, another area will inevitably bulge, making it impossible to keep all areas flat simultaneously. In wind field turbulence prediction tasks, the waterbed effect specifically refers to the situation where, when a model simultaneously optimizes the inherently competing objectives of "macro-trend accuracy" and "high-frequency detail fidelity," an improvement in the performance of one objective inevitably leads to a decline in the performance of the other.

[0071] Step S4: Perform an exponential moving average on the updated training parameters to obtain the inference parameters; In one embodiment, the expression for applying an exponential moving average to the updated training parameters is:

[0072] In the formula, For inference parameters, For smoothing coefficients, These are the updated training parameters.

[0073] Step S5: Load the inference parameters into the prediction neural network to obtain a trained prediction neural network. Input the historical wind speed sequence of the wind field to be predicted into the trained prediction neural network and output the second predicted wind speed sequence. Step S6: Statistically recalibrate the second predicted wind speed sequence based on regional meteorological priors to obtain the final high-frequency wind field prediction sequence.

[0074] In one embodiment, the second predicted wind speed sequence is statistically recalibrated based on regional meteorological priors to obtain the final high-frequency wind field prediction sequence, specifically including: Obtain the mean and standard deviation of the historical wind speed sequence of the wind field to be predicted, and obtain the mean and standard deviation of the second predicted wind speed sequence; Further standardization and reconstruction of the second predicted wind speed sequence yields the final high-frequency wind field prediction sequence, expressed as:

[0075] In the formula, This is the final high-frequency wind field prediction sequence. This is the second predicted wind speed sequence. and These are the mean and standard deviation of the second predicted wind speed series, respectively. and These are the mean and standard deviation of the historical wind speed sequence for the wind field to be predicted, respectively.

[0076] Through the above transformation, the mean and standard deviation of the predicted sequence can be forcibly aligned with the statistical characteristics of the real physical environment, thus completely correcting the inherent energy dissipation problem of the generative model.

[0077] To demonstrate the effectiveness of this invention, an experimental dataset was constructed using 50 Hz ultrasonic anemometer data collected at multiple altitudes (0.64 m, 1.16 m, 2.07 m, 3.05 m, 6.00 m, and 8.95 m) from a meteorological station. The dataset from the first 150 time steps was selected as the historical wind speed sequence. This dataset effectively reflects the non-stationary temporal characteristics and turbulent random fluctuations of wind speed under complex flow field conditions. Implementing the method of this invention yields the following results: Figure 4 The image shows a comparison between the final high-frequency wind field prediction sequence and the actual wind speed sequence.

[0078] like Figure 4 As shown in the subplot above, the predicted curve of this invention closely overlaps with the actual trajectory after 150 time steps, without any divergence.

[0079] Furthermore, a local magnification of the transition region between the input conditions and the prediction interval (time step 100 to 160) reveals the absence of boundary failure phenomena such as phase jumps, trend breaks, or abrupt variance changes. This demonstrates that the final model captures the continuous evolution dynamics of the wind field over time. The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0080] Although preferred embodiments of the present application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present application.

[0081] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0082] The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching provided in this application has been described in detail above. Specific examples have been used to illustrate the principle and implementation of this application. The description of the above embodiments is only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching, characterized in that, include: A prediction neural network is constructed based on the KAN model and parallel liquid state space. The historical wind speed sequence is input into the prediction neural network, and the hidden state features are output. The latent state features are input into a vector field prediction network based on optimal transport flow matching to obtain a predicted wind speed vector field. The first predicted wind speed sequence is generated by solving the predicted wind speed vector field. The loss value between the first predicted wind speed sequence and the actual wind speed sequence is calculated using a joint loss function, and the training parameters of the prediction neural network are updated through backpropagation based on the loss value. The updated training parameters are processed by an exponential moving average to obtain the inference parameters; The inference parameters are loaded into the prediction neural network to obtain a trained prediction neural network. The historical wind speed sequence of the wind field to be predicted is input into the trained prediction neural network, and the second predicted wind speed sequence is output. The second predicted wind speed sequence is statistically recalibrated based on regional meteorological priors to obtain the final high-frequency wind field prediction sequence.

2. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 1, characterized in that, The predictive neural network includes a feature extraction module, an adaptive conditional normalization module, and a parallel liquid state space module connected in sequence.

3. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 2, characterized in that, The feature extraction module performs feature extraction using the KAN model, specifically including: The input historical wind speed sequence is mapped to the interval [-1,1] by using the hyperbolic tangent function, and the Chebyshev polynomial is used as the basis function to perform a nonlinear transformation on the mapped sequence. The nonlinear features are extracted based on the nonlinear transformation results, expressed as follows: In the formula, It is a nonlinear characteristic. The weight matrix is ​​the linear transformation matrix. This is the mapped historical wind speed sequence. Let be the highest order of the Chebyshev polynomial. For the first Learnable coefficients of the Chebyshev polynomial. For the first Chebyshev polynomials It is the hyperbolic tangent activation function.

4. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 2, characterized in that, The adaptive conditional normalization module performs the following steps: Historical meteorological condition features and time step coding features are obtained, and then added element by element to obtain modulation features; The modulation features are sequentially input into a linear transformation layer and a nonlinear activation layer, and the output is a scaling factor and a translation factor. After performing layer normalization on the nonlinear features, and combining the scaling factor and translation factor, the nonlinear features fused with meteorological time series information are output, expressed as follows: In the formula, To integrate the nonlinear characteristics of meteorological time series information, Presentation layer normalization operation, It is a nonlinear characteristic. As a scale factor, This is the translation factor.

5. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 2, characterized in that, The parallel liquid state-space module performs the following steps: The time-varying step size is dynamically generated based on the nonlinear characteristics of fused meteorological time-series information, and the expression is: In the formula, For dynamically generated time steps, It is the Softplus non-linear activation function. For linear mapping layer, To integrate the nonlinear characteristics of meteorological time series information, It is a Sigmoid non-linear activation function. This represents a multilayer perceptron. This represents a one-dimensional convolution operation; By discretizing the system matrix and input matrix in the logarithmic domain using a time-varying step size, the expression is as follows: In the formula, The system matrix is ​​the result of discretization mapping. For the system matrix, The input matrix is ​​the result of discretization mapping. The input matrix; Based on the previous hidden state, the discretized system matrix, and the input matrix, the current hidden state is calculated as follows: In the formula, This is the current hidden state. This is the hidden state from the previous moment. To integrate the nonlinear characteristics of meteorological time series information.

6. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 5, characterized in that, The parallel liquid state space module further includes: Static graph compilation enables the computation of time-series loops to be pushed down to the underlying C++ operator layer for parallel execution.

7. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 1, characterized in that, The latent state features are input into a vector field prediction network based on optimal transport flow matching to obtain a prediction vector field. The first predicted wind speed sequence is generated by solving the prediction vector field, specifically including: A uniformly distributed original time step is generated using stratified sampling. The original time step is then subjected to cosine transformation to obtain the optimized time step, expressed as: In the formula, The time step after deformation optimization. The original time steps are uniformly distributed; A vector field prediction network based on optimal transport flow matching is constructed. The hidden state features and the optimized time step are input into the vector field prediction network, and the predicted wind speed vector field is output. Solving the ordinary differential equations of the predicted wind speed vector field yields the first predicted wind speed sequence, expressed as: In the formula, This is the first predicted wind speed sequence. It is a standard Gaussian noise sequence. To predict the vector field.

8. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 1, characterized in that, The joint loss function includes a vector field mean square error loss term, a first-order and second-order derivative joint physical constraint loss term, and a multi-resolution short-time Fourier transform spectral constraint loss term. The expression for the mean square error loss term of the vector field is: In the formula, This represents the mean square error loss value for the vector field. Let be the mean square error function. To predict the velocity field, For the target velocity field; The expression for the joint physical constraint loss term of the first and second derivatives is as follows: In the formula, The loss value is the combined physical constraint loss of the first and second derivatives. For stream matching time steps, The mean absolute error function, This represents the first-order difference operation. and These are the first predicted wind speed sequence and the actual wind speed sequence, respectively. This represents a second-order difference operation; The expression for the multi-resolution short-time Fourier spectrum constraint loss term is: In the formula, This represents the multi-resolution short-time Fourier spectrum constraint loss value. For stream matching time steps, For window size, The mean absolute error function, This indicates the short-time Fourier transform operation. and These are the first predicted wind speed sequence and the actual wind speed sequence, respectively. To prevent extremely small positive numbers from being set to zero.

9. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 1, characterized in that, The expression for applying the exponential moving average to the updated training parameters is: In the formula, For inference parameters, For smoothing coefficients, These are the updated training parameters.

10. The high-frequency wind field turbulence prediction method based on parallel liquid state space and flow matching according to claim 1, characterized in that, Based on regional meteorological priors, the second predicted wind speed sequence is statistically recalibrated to obtain the final high-frequency wind field prediction sequence, which specifically includes: Obtain the mean and standard deviation of the historical wind speed sequence of the wind field to be predicted, and obtain the mean and standard deviation of the second predicted wind speed sequence; The second predicted wind speed sequence is standardized and reconstructed to obtain the final high-frequency wind field prediction sequence, expressed as: In the formula, This is the final high-frequency wind field prediction sequence. This is the second predicted wind speed sequence. and These are the mean and standard deviation of the second predicted wind speed series, respectively. and These are the mean and standard deviation of the historical wind speed sequence for the wind field to be predicted, respectively.

Images