New energy power prediction method fusing typhoon meteorological information and micrometeorological prediction result

Abstract

The invention relates to the technical field of new energy power generation prediction, in particular to a new energy power prediction method fusing typhoon meteorological information and a micrometeorological prediction result, which comprises the following steps: collecting multi-source observations such as a satellite scatterometer, a radar wind profile and laser wind measurement, and unifying coordinates; constructing a mesoscale wind field by adopting four-dimensional variational assimilation, inferring a micrometeorological field by utilizing a spectrum embedding diffusion network, and generating a multi-scale meteorological field by frequency domain phase consistency fusion; probabilistic wind speed is sampled in the countercurrent model meeting the condition of mass and momentum conservation, an energy conservation graph neural network is input, the wake effect is coupled, and a unit power quantile value is obtained; the power probability is sent to a risk weighting model, the weight is adjusted in real time according to the peak load, the reserve capacity and the climbing rate risk, a dispatching power curve and uncertainty are output, and grid-connected power errors are used for periodically updating the weight and the parameters of the last layer of the reversible model. According to the invention, the safety acceptance margin of the power grid to new energy is obviously improved.

Description

Technical Field

[0001] The present invention relates to the technical field of new energy power generation prediction, and particularly to a new energy power prediction method that integrates typhoon meteorological information and micro-meteorological prediction results. Background Art

[0002] With the grid connection of high-penetration wind power in coastal and mountainous areas, the strong wind shear, gusts, and wake coupling caused by the rapid passage of typhoons have become the main reasons for significant power fluctuations. If only relying on conventional numerical weather prediction (NWP) data, it is difficult for the prediction model to capture mesoscale processes such as eyewall replacement and spiral rainbands on the scale of dozens of minutes, thereby weakening the dispatching safety margin. Therefore, a power prediction method that integrates typhoon macro-circulation and micro-meteorology at the unit scale is of crucial significance for improving power dispatchability under extreme weather conditions, reducing the curtailment rate of wind power, and standby costs.

[0003] The prior art adopts the idea of "wind speed offset identification + dual TCN power model" (Chinese invention patent, publication number: CN119419801B): first, use a directed acyclic graph network to determine whether the NWP wind speed is a "positive offset" or a "negative offset", and then call two sets of TCNs respectively and embed directional penalties in the loss function for correction. The defects of the prior art are as follows: only dealing with NWP grid wind speeds, ignoring the three-dimensional structure of typhoons and micro-meteorological disturbances, it is difficult to depict the wind power peaks during the rapid intensification stage of typhoons. The dual model only gives a single-value power, unable to quantify the tail risk, and the dispatching margin is low. Once the model weights are trained, they are fixed and cannot adapt to the real-time grid connection error; the accuracy drops sharply in the scenario of typhoon path offset. The mutual shielding between units and the energy conservation relationship are not explicitly modeled, and the power array effect estimation deviation is large. Summary of the Invention

[0004] In view of the many problems existing in the above-mentioned prior art, the present invention provides a new energy power prediction method that integrates typhoon meteorological information and micro-meteorological prediction results. The present invention first uses multi-source observations and four-dimensional variational assimilation to reconstruct the typhoon mesoscale field, and obtains a micro-meteorological resolution wind field through a spectral embedding diffusion network; then, taking the fusion field as a condition, samples the probability wind speed that satisfies the conservation constraint in a reversible flow network, and calculates the unit power by coupling the wake through an energy conservation graph neural network; finally, uses a risk-weighted model to convert the probability power into a grid-side dispatchable curve and adapt online with the grid connection error, significantly reducing the power peak error and dispatching risk during typhoons.

[0005] A new energy power prediction method that integrates typhoon meteorological information and micro-meteorological prediction results includes: Obtain satellite observations, radar wind profiles, surface wind measurements, and unit operation data of the wind farm, and generate multi-source meteorological data through time correction, coordinate unification, quality control, and normalization; Implement numerical assimilation on the multi-source meteorological data to generate a mesoscale typhoon field, infer the mesoscale typhoon field to a micro-meteorological field through super-resolution, and use the frequency-domain phase consistency method for fusion to output a multi-scale fused meteorological field; Taking the multi-scale fused meteorological field as a condition, sample probability wind field data in a reversible generation model based on a flow-based reversible network that satisfies the constraints of mass conservation and momentum conservation, and combine wake interaction in an energy conservation graph model based on a graph neural network to calculate the unit power prediction data; Input the unit power prediction data into a risk-weighted decision model, generate a scheduling power curve and power uncertainty based on the conditional risk value of the peak load, the conditional risk value of the reserve capacity, and the conditional risk value of the ramp rate, and online update the parameters of the risk-weighted decision model within a preset period according to the grid-connected power feedback, and output the grid-side schedulable power prediction data.

[0006] Preferably, the multi-source meteorological data consists of satellite scatterometer observation data, airborne sounding profile data, three-dimensional Doppler radar wind profile data, acoustic wind profiler observation data, lidar wind observation data, unmanned aerial vehicle micro-layer flow observation data, and unit operation monitoring data, and enters the subsequent processing process after constructing a unified data structure under a unified time reference and a unified geographic coordinate system.

[0007] Preferably, for the quality control, first use statistical threshold filtering to eliminate the observed values that exceed the abnormal discrimination boundary, then use smooth interpolation to complete the short-term continuous missing data, and use geostatistical interpolation to complete the long-term continuous missing data. After completion, set a quality identification code for each observation record.

[0008] Preferably, the numerical assimilation uses a four-dimensional variational assimilation algorithm, takes wind speed, temperature, humidity, and pressure as control vectors, and generates a mesoscale typhoon field by minimizing the weighted error function between the background field and the observation field.

[0009] Preferably, for the resolution improvement, use a spectral embedding diffusion network with a frequency-domain convolutional architecture, and convert the mesoscale typhoon field into a micro-meteorological resolution meteorological field through forward noise perturbation and reverse denoising reconstruction sequence mapping.

[0010] Preferably, the frequency-domain phase consistency method performs wavelet transform on the mesoscale typhoon field and the micro-meteorological resolution meteorological field, performs phase linear combination according to the set weights in the corresponding frequency band range, and then performs inverse transform to generate a multi-scale fused meteorological field.

[0011] Preferably, the reversible generation model uses a conditional reversible network with a residual coupling structure, records the logarithmic Jacobian determinant in the forward mapping, and simultaneously minimizes the mass conservation residual term and the momentum conservation residual term during the reverse sampling process to obtain probability wind field data.

[0012] Preferably, the energy conservation graph model uses a wind turbine and a meteorological tower as nodes and wake interaction as edges, calculates node hidden vectors through multi-layer attention message passing, applies power conservation projection at the output layer, and maps wind speed information to unit power prediction data.

[0013] Preferably, the risk-weighted decision model calculates risk weight coefficients based on the conditional risk value of peak load, the conditional risk value of reserve capacity, and the conditional risk value of ramp rate respectively, applies the risk weight coefficients to the corresponding risk components in the power prediction data and performs weighted synthesis to generate a dispatch power curve and synthesize power uncertainty according to the variances of the risk components.

[0014] Preferably, grid-connected power feedback triggers an online update process at a preset time interval, and in this process, the parameters of the risk-weighted decision model and the parameters of the last stage of the reversible generation model are synchronously adjusted according to the error gradient between the predicted power and the grid-connected power.

[0015] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: The present invention realizes the multi-scale seamless fusion of the typhoon macro field and the micro-meteorological field and captures the rapidly evolving local wind speed gradient through the technical means of four-dimensional variational assimilation + spectral embedding diffusion super-resolution; The present invention realizes the probability wind field sampling that simultaneously satisfies mass-momentum conservation and outputs a confidence interval rather than a single point value through the technical means of the residual coupling conditional reversible flow model; The present invention realizes array-level power mapping and unit coupling loss compensation through the technical means of energy conservation graph neural network + wake interaction constraint; The present invention realizes the three-dimensional risk collaborative minimization of peak load, reserve capacity, and ramp rate through the technical means of risk-weighted decision + gradient closed-loop online update. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic flowchart of the method of the present invention; Figure 2 is a data structure association diagram of the new energy power prediction method of the present invention; Figure 3 is a schematic diagram of the conserved reversible network structure and the residual feedback mechanism of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0017] Hereinafter, embodiments of the present disclosure will be described with reference to the accompanying drawings. However, it should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present disclosure. In the following detailed description, for the sake of explanation, numerous specific details are set forth in order to provide a comprehensive understanding of the embodiments of the present disclosure. However, it is obvious that one or more embodiments can also be implemented without these specific details. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessarily obscuring the concepts of the present disclosure.

[0018] The terms used herein are merely for describing specific embodiments and are not intended to limit the present disclosure. The terms "including", "comprising", etc. used herein indicate the presence of the described features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0019] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein should be interpreted as having a meaning consistent with the context of this specification and should not be interpreted in an idealized or overly rigid manner.

[0020] As Figure 1 shown, a new energy power prediction method that fuses typhoon meteorological information and micro-meteorological prediction results includes: Obtaining satellite observations, radar wind profiles, surface wind measurements, and unit operation data of a wind farm, and generating multi-source meteorological data through time correction, coordinate unification, quality control, and normalization. This embodiment elaborates on the technical principles, key algorithms, and application effects of obtaining satellite observations, radar wind profiles, surface wind measurements, and unit operation data of a wind farm and generating multi-source meteorological data. The overall process sequentially completes spatio-temporal unification, data correction, quality control, and normalization, providing highly consistent inputs that can be directly called by subsequent typhoon-micro-meteorological fusion and power prediction modules.

[0021] First, the along-track observations of the satellite scatterometer, the airborne sounding profiles, the three-dimensional Doppler radar volume scan voxels, the acoustic wind profiler echoes, the lidar volume scan, the UAV micro-layer flow profiles, and the unit operation records are respectively transmitted to the data receiving node through dedicated communication links. Each observation source carries an independent timestamp and local coordinates and must be synchronized under a unified clock and unified coordinate framework.

[0022] The system constructs an absolute time reference using the global positioning system timing signal and converts the original time to Coordinated Universal Time :

[0023] Wherein Indicates the deviation between the satellite timing pulse and the internal clock of the device. The corrected timestamp error is kept at the microsecond level, which can meet the requirements of splicing high-frequency unit data and radar volume scan data. Spatial unification is achieved through a strict conversion of ellipsoidal geodetic coordinates to projected plane coordinates: the longitude and latitude triplet is converted to the Map to horizontal axis , vertical axis and senior After the coordinates are unified, the system aggregates multi-source observations at the same time for the same grid unit to form a multi-dimensional vector Quality control is performed in two levels. First, the sliding median-absolute deviation test is used in the time dimension to remove outliers: Calculate the sliding window median The median absolute deviation (MAD) is the same as the median absolute deviation (MAD), if:

[0024] Then Marked as an exception, parameter According to the instrument stability setting, the spatial missing data are supplemented by three-dimensional kriging interpolation, and its semivariogram function Fitting based on neighborhood sample empirical statistics, interpolation weights Obtained by solving the Kriging linear equations.

[0025] All observed quantities are uniformly converted into dimensionless normalized form. Taking wind speed as an example, the normalization formula is:

[0026] in is the sliding mean, Normalization maps data of different dimensions to a zero-mean unit variance space, which facilitates the subsequent reversible generation of flow models to process multiple meteorological elements simultaneously.

[0027] To verify the feasibility of this process, a typhoon-affected event was used as an example: the system installed two laser wind radars and three acoustic wind profilers at the wind farm, and enabled satellite scatterometer track coverage. After the observation data was processed through the above steps, a multi-source meteorological data matrix containing more than 50 dimensions of wind speed, wind direction, temperature, humidity, and unit power was generated. This matrix is used as a unified input in the subsequent assimilation-diffusion-phase locking module. The measured splicing error is lower than the calibration accuracy of the wind tower, meeting the requirements of microstructure analysis under typhoon strong wind shear field.

[0028] Through the above technical solution, under the complex wind field conditions of a rapidly approaching typhoon, multi-source meteorological data with consistent time and space and controllable quality can be stably output, laying a solid foundation for subsequent probabilistic wind field sampling and power mapping.

[0029] Preferably, as Figure 2 shown, the multi-source meteorological data consists of satellite scatterometer observation data, airborne sounding profile data, three-dimensional Doppler radar wind profile data, acoustic wind profiler observation data, lidar wind observation data, unmanned aerial vehicle micro-layer flow observation data, and unit operation monitoring data, and enters the subsequent processing process after constructing a unified data structure under a unified time reference and a unified geographic coordinate system.

[0030] This embodiment elaborates on the construction principle of multi-source meteorological data, key processing algorithms, and application effects in the new energy power prediction scenario of fusing typhoon meteorological information and micro-meteorological prediction results. The multi-source meteorological data consists of satellite scatterometer observation data, airborne sounding profile data, three-dimensional Doppler radar wind profile data, acoustic wind profiler observation data, lidar wind observation data, unmanned aerial vehicle micro-layer flow observation data, and unit operation monitoring data. By constructing a unified data structure under a unified time reference and a unified geographic coordinate system, spatial misalignment and time drift caused by source differences can be eliminated in the subsequent assimilation-diffusion-phase locking link, thereby improving the power prediction accuracy under typhoon extreme wind fields.

[0031] There are deviations between the built-in clocks of each data source and the timing signals of the Global Navigation Satellite System. The time is unified through the following formula:

[0032] where represents the original timestamp; represents the difference between the timing pulse and the device clock; represents the Coordinated Universal Time after unification. After correction, all observed values are aligned within a time error of the microsecond order, which can ensure the synchronization of the radar volume scan voxels and the unit ten-hertz power curve.

[0033] Satellite and radar products are given in geocentric ellipsoidal coordinates and projected onto plane coordinates through the following formula:

[0034]

[0035] where is the first radius of curvature in the prime vertical; is the radius of curvature of the meridian; and are the differences in longitude and latitude; the elevation is unified with reference to the geoid, and the relative height returned by the acoustic wind profiler and the lidar wind observation is converted to orthometric height . All observation points are then mapped to a unified three-dimensional grid index .

[0036] Perform two-level quality inspections on each grid time series : Primary inspection: Sliding median And the median absolute deviation MAD:

[0037] Where is a constant factor.

[0038] Secondary inspection: Re-screen the abnormal marked points according to the neighborhood space gradient, and use linear interpolation or Kriging interpolation to complete the missing segments. The Kriging weights , by solving:

[0039] represents the empirical semi-variogram function; is the Lagrange multiplier, and the quality identification code QF is generated after completion as the subsequent weighting basis.

[0040] For each meteorological quantity Perform zero-mean unit-variance mapping:

[0041] represents the sliding mean; represents the sliding standard deviation; After normalization, construct a unified vector:

[0042] Where is the feature dimension. This vector is directly input into the mass conservation - momentum conservation reversible generative flow model without further dimensionality transformation.

[0043] If the Kriging interpolation residual exceeds the threshold , the system immediately triggers redundant switching, and uses another acoustic wind profiler or lidar at the same height as the compensation source. If the continuous missing measurement time exceeds the threshold , then the historical statistical field is applied to perform playback filling to ensure the integrity of the input tensor.

[0044] Embedding this embodiment into the case of typhoon "around Bohai Sea", the mean square error of the wind speed of the fusion grid is significantly reduced compared with the traditional site interpolation scheme, and the probability power quantile curve remains continuous near the cut-out wind speed of the unit, which can support the fine description of wake coupling by the energy conservation diagram model. Experimental verification shows that the RMS error of the wind vector output by the assimilation - diffusion - phase locking module near the extreme typhoon center is significantly reduced, and the absolute error curve of power prediction drops synchronously.

[0045] In the embodiment, two lidar wind profilers, an airborne sounding system and an acoustic wind profiler array are deployed for a coastal wind farm. The system aggregates each observation source with a time granularity of one second, and executes the above-mentioned spatio-temporal unification and quality control processes to generate vectors . Subsequently, the reversible generative flow is applied to sample the wind field and map the power curve. The results show that the active power prediction deviation within six hours before the typhoon makes landfall remains within 10% of the rated power of the unit, verifying the effective support of multi-source meteorological data for the new energy power prediction scenario.

[0046] Preferably, for the quality control, first, the statistical threshold filtering is used to eliminate the observed values that exceed the abnormal discrimination boundary, then the short-term continuous missing data is complemented by smooth interpolation, and the long-term continuous missing data is complemented by geostatistical interpolation. After the complementation is completed, a quality identification code is set for each observation record.

[0047] To ensure that the subsequent typhoon-micro-meteorological fusion model obtains credible inputs, in this embodiment, a set of quality control links is constructed for multi-source meteorological data. The core consists of three types of operations: anomaly rejection, missing data complementation, and quality identification, which not only make full use of high-frequency observations but also avoid the amplification of abnormal noise in the probability wind field sampling stage.

[0048] First, the system builds a time series for each type of observed variable on a unified spatio-temporal grid . The sliding median and the median absolute deviation are calculated for this series. If there is an observed value that satisfies:

[0049] then it is determined that it exceeds the abnormal discrimination boundary, represents the multiple factor set for the sensor stability, represents the median, represents the median absolute deviation. This statistical threshold filtering can be completed within thirty milliseconds of data access and output a one-dimensional boolean mask. After anomaly rejection, the system detects the length of the continuous missing data segment . If:

[0050] where is a preset threshold, then smooth interpolation is used for complementation. The smooth interpolation is described by the cubic spline function , and its natural boundary conditions make to ensure smooth curvature. The reconstructed observed value after interpolation is denoted as . If , then geostatistical interpolation is called. The system constructs an empirical semi-variance function based on the surrounding spatial samples :

[0051] where represents the sample spacing, represents the number of sample pairs at this spacing. By fitting the theoretical range and the sill are obtained, and then the Kriging equations are solved:

[0052] The weights are obtained, and based on this, the estimated value of the missing grid is calculated:

[0053] where represents the weight coefficient, represents the distance between the point to be estimated and the known sample point, represents the Lagrange multiplier.

[0054] After interpolation, the system writes the quality identification code QF for each observation record. QF = 0 indicates that the original observation passes the threshold test; QF = 1 indicates that it is completed by smooth interpolation; QF = 2 indicates that it is completed by geostatistical interpolation; QF = 3 indicates that it is masked due to severe missing but still occupies a position. The quality identification code is reflected in the form of weights in the loss function of the reversible generation flow model, and the contribution of low-quality samples is automatically reduced to ensure that the training process of the conservation constraint is robust to the error distribution.

[0055] Example: During a typhoon's rapid landing, the acoustic wind profiler experienced signal attenuation in a high-humidity environment and had continuous missing measurements for two minutes. The system detected greater than the threshold, initiated Kriging interpolation, constructed spatial samples using the lidar wind measurement radar and the airborne sounding profile at the same height, and reconstructed the wind speed field through semi-variance fitting and weight solution. The interpolation residual variance was lower than 20% of the mean square error of the original wind measurement tower. This interpolation result was assigned QF = 2 and had a lower weight in the subsequent probability wind field sampling stage, reducing the data uncertainty diffusion of the abnormal segment.

[0056] Through the above multi-level quality control, the scheme still maintains the integrity of the observation vector under extreme typhoon wind speed shear, provides noise-controlled input for the subsequent typhoon - micrometeorological fusion algorithm, and finally significantly reduces the mean square error of the unit power prediction compared to the baseline without quality control.

[0057] Numerical assimilation is performed on the multi-source meteorological data to generate a mesoscale typhoon field, the mesoscale typhoon field is inferred to a micrometeorological field through super-resolution, and the frequency domain phase consistency method is used for fusion to output a multi-scale fused meteorological field; The original multi-source meteorological data is used as the observation vector after being unified into a grid 。The assimilation system uses the three-dimensional state vector of the finite-difference non-hydrostatic model as the control variable, and by minimizing the four-dimensional variational objective function:

[0058] where represents the background field, represents the background error covariance matrix, represents the observation operator at time and represents the observation error covariance matrix.

[0059] Solving yields a mesoscale typhoon field with a horizontal resolution of three kilometers , capturing the maximum radial wind shear in the typhoon eyewall and providing a typhoon-scale forcing boundary for wind power prediction.

[0060] Sending slices into the spectral embedding diffusion network. The network gradually injects Gaussian noise into the state tensor during the forward process and performs frequency-domain convolution denoising during the reverse process to achieve reconstruction with a horizontal resolution of thirty meters and a vertical resolution of five meters, and outputs the micrometeorological field . This process shows that it maintains the relative phase structure of the typhoon vortex and avoids the small-scale energy decay introduced by traditional interpolation.

[0061] To eliminate the phase mismatch at the meso-micro two-scale boundary, the system first performs three-dimensional discrete wavelet transform on and to obtain the corresponding frequency band phases and . A fused phase is constructed within each wavelet frequency band:

[0062] where represents the weight coefficient. The amplitude is retained as the micrometeorological energy spectrum amplitude , and after combining with , inverse wavelet transform is performed to output the multi-scale fused meteorological field . This fusion simultaneously maintains mesoscale energy and microscale details, and the phase continuity ensures the differentiability of the wind speed field gradient, meeting the conservation constraint requirements of the subsequent reversible generative flow model. It should be noted that in the present invention, represents the multi-scale fused meteorological field (i.e., a three-dimensional meteorological tensor that has completed typhoon-micrometeorological phase fusion and has a spatial resolution reaching the unit scale). If a local window needs to be extracted from this tensor or a low-dimensional conditional vector is obtained through linear embedding, it is written as , where Indication: For any condition vector representing its derivative, must be used

[0063] The experiment was verified on the case of the landing typhoon "SA Outpost": the mean square error of the wind vector of the traditional pure assimilation - interpolation scheme is , and after introducing super - resolution - phase fusion, the error drops to . The 95 - quantile prediction error of the mapped wind farm power is reduced by more than 30%, and the cut - out judgment of the unit exceeding the rated wind speed is issued 20 minutes in advance, effectively supporting grid connection dispatching.

[0064] In an embodiment, two lidar wind sensors, a set of coastal radar arrays and an airborne sounding system are arranged in a coastal wind farm of one hundred megawatts. Two hours after the outer circulation of the typhoon enters the observation domain, the system completes the integrated calculation of assimilation - inference - fusion, generates and pushes it to the reversible generation flow module. Monitoring shows that during the passage of the typhoon eyewall, the capture delay of the fusion field for gust pulses is less than 15 seconds, and finally the smooth downward adjustment of the power dispatch curve is realized, avoiding frequent tripping of the unit.

[0065] Preferably, the numerical assimilation adopts the four - dimensional variational assimilation algorithm, taking wind speed, temperature, humidity and pressure as control vectors, and generating a mesoscale typhoon field by minimizing the weighted error function of the background field and the observation field.

[0066] The present invention uses the four - dimensional variational assimilation algorithm to invert the mesoscale typhoon field from multi - source observations, laying a dynamic boundary for subsequent micro - meteorological super - resolution inference and new - energy power prediction. Its core principle is to jointly regard wind speed, temperature, humidity and static pressure as state variables, and within the differentiable framework of the finite - difference non - hydrostatic numerical model, approximate the background field and the observation field simultaneously by minimizing the weighted error function.

[0067] Let represent the three - dimensional state vector to be estimated, represent the background field of model extrapolation, represent the observation vector at the th moment within the assimilation window, be the corresponding observation operator, then the objective function is:

[0068] where represents the background error covariance matrix, used to measure the spatial correlation of background deviation; represents the observation error covariance matrix, used to characterize the uncertainty of each observation source.

[0069] The algorithm calculates through the adjoint model, and uses conjugate gradient or quasi - Newton iteration to search for . ​It is a mesoscale three-dimensional wind field that includes the typhoon circulation, eyewall, and outer rainband structures. The analytical horizontal grid of this wind field is much smaller than that of conventional global reanalysis products, and it can resolve the radial wind shear at the tip height of wind turbine blades; in the vertical direction, it maintains the same resolution as the sounding profile, enabling synchronous constraints on the temperature and humidity gradients and momentum coupling.

[0070] To enable the grid dispatcher to obtain a reliable power reduction plan before the typhoon arrives, the present invention selects the assimilation window in the forecast period before the rapid evolution of the wind field; if there are missing measurements in the satellite scatterometer strip, the observation operator automatically excludes the corresponding observation channels and adjusts to maintain the stability of the condition number. The abnormal wind speed observations have been assigned a quality identification code QF in the above-mentioned quality control steps; during assimilation, by increasing the corresponding diagonal element to reduce the weight, it is ensured that the iterative convergence direction is not restricted by local anomalies.

[0071] Example: For the landing process of a medium-strength typhoon, airborne sounding, coastal three-dimensional radar, and two unmanned aerial vehicle profile data are incorporated into the assimilation window. The deviation between the maximum near-surface radial wind speed of the finally obtained mesoscale typhoon field and the instantaneous wind speed at the center of the measured turbine impeller is less than half of the output of the reference model. Subsequently, the micro-meteorological super-resolution inference inherits the phase of this field to generate a multi-scale fused meteorological field. After power mapping, the cut-out wind speed moment of the turbine is accurately given three hours in advance, providing a basis for the power station to implement active power reduction and slow down the blade root load.

[0072] Preferably, the resolution improvement uses a spectral embedding diffusion network with a frequency-domain convolution architecture. Through forward noise perturbation and reverse denoising reconstruction sequence mapping, the mesoscale typhoon field is converted into a micro-meteorological resolution meteorological field.

[0073] The present invention refines the mesoscale typhoon wind field with a horizontal grid of about 3 km into a micro-meteorological wind field with a horizontal resolution of about 30 m and a vertical resolution of about 5 m through a spectral embedding diffusion network with a frequency-domain convolution architecture, providing wind vectors at the turbine scale for new energy power prediction. The core process and mathematical expressions are as follows.

[0074] Let the assimilated mesoscale three-dimensional wind field tensor be . In the diffusion forward process, noise is injected in discrete steps to obtain:

[0075] where is the retention factor at the step, . The noise is first subjected to a three-dimensional fast Fourier transform, and after being re-distributed according to the observed energy spectrum weight in the wave number dimension, it is inversely transformed and injected into the above formula to ensure that the noise energy matches the actual energy spectrum of the typhoon.

[0076] Frequency-domain convolution for inverse denoising. The inverse network uses a multi-level UNet, but all convolutional kernels are implemented in the frequency domain: perform a fast Fourier transform on the input tensor, multiply it with the convolutional kernel spectrum and then perform an inverse transform back to the spatio-temporal domain. Frequency-domain convolution expands the receptive field and has periodic boundaries, effectively eliminating eye wall edge artifacts.

[0077] Spectral embedding phase correction. The method for maintaining the phase continuity of the typhoon is as follows: introduce the phase of the same layer of the encoder during the decoding stage , and according to:

[0078] generate the fused phase , where is the current decoder phase, is a fixed weight. The amplitude takes the decoder amplitude spectrum . After inverse wavelet transform, the denoised tensor is obtained.

[0079] According to the height-correlated mean and the standard deviation for denormalization:

[0080] where the symbol " " represents element-wise multiplication, is the three-dimensional wind field with micrometeorological resolution and can be directly input into the subsequent conservation reversible generation flow model and power mapping model. Comparing the relative error between the measured energy spectrum of the lidar and the output energy spectrum :

[0081] The present invention is about lower than the traditional bicubic interpolation scheme. The power ninety-fifth percentile prediction deviation is significantly reduced, enabling the dispatching side to issue a power reduction instruction ten minutes in advance and reducing the risk of unit overload and tripping.

[0082] In an embodiment, two lidar, a coastal Doppler radar, and an airborne sounding system are deployed in a coastal multi-megawatt wind farm. After the outer circulation of the typhoon enters the observation domain, the system uses the diffusion network to generate . The measured results show that during the passage of the typhoon eye wall, the response delay of the model to gust pulses is less than 15 s, and the peak load of the unit is controlled below 90% of the rated load, verifying the practical value of the resolution improvement link under extreme wind conditions.

[0083] Preferably, the frequency-domain phase consistency method performs wavelet transform on the mesoscale typhoon field and the micrometeorological resolution meteorological field, performs inverse transform after linearly combining phases within the corresponding frequency band range according to a set weight, and generates a multi-scale fused meteorological field.

[0084] To ensure the continuous presentation of typhoon-scale dynamic information and turbine-scale details in the same wind field, the present invention introduces the frequency-domain phase consistency method after resolution improvement, and combines the mesoscale typhoon field with the micrometeorological field into a multi-scale meteorological field . This method retains the mesoscale energy and embeds the microscale texture through linear phase combination in the wavelet domain. The specific principle and implementation steps are as follows.

[0085] Wavelet transform and frequency band mapping: Perform three-dimensional discrete wavelet transform on and respectively to obtain the complex coefficients at the scale-frequency index :

[0086]

[0087] and represent the amplitude spectrum; and represent the phase spectrum; is the direction sub-band, is the scale level.

[0088] Phase linear combination: Let the phase weight coefficient be , and perform linear interpolation within the same sub-band :

[0089] , with the microscale level taken as as the scale increases, the microscale phase is fully adopted, and the mesoscale main energy layer is taken as , the typhoon phase amplitude selection is maintained, and the inverse transform amplitude takes the micrometeorological amplitude to retain the high-resolution energy spectrum:

[0090] Perform inverse wavelet transform on to obtain:

[0091] Variable description: —Mesoscale typhoon wind field tensor; — The micro-meteorological wind field tensor obtained through the spectral embedding diffusion network; — Output multi-scale fused meteorological wind field; — Phase weight coefficients at different scale levels.

[0092] After phase consistency processing, while maintaining the circulation phase of the typhoon eyewall and the details of the horizontal vortex of the unit. Compared with the non-fused scheme, the root mean square error of the wind speed in the unobstructed area of the radar in the fused field is reduced by about 35%. In the power prediction link, the 95th percentile power deviation is reduced and the lead time is increased, enabling the dispatching system to issue power reduction instructions earlier and reducing the tripping events when the typhoon center passes.

[0093] Example, in the case of Typhoon Kalmaegi, take the three-dimensional wavelet Daubechies-4 basis and set the phase weight to linearly decrease from 1 to 0. The fusion result is verified by a lidar wind sensor. The mean value of the wind direction error at a height of 200 meters is reduced from to . Based on this, the station issues a cluster load reduction instruction ten minutes in advance, and the maximum torque of the unit is reduced by 11% compared with the non-fused baseline, proving the practical value of this frequency-domain phase consistency method in typhoon-driven new energy power prediction.

[0094] As Figure 3 shown, conditional on the multi-scale fused meteorological field, probability wind field data is sampled in a reversible generative model based on a flow-based reversible network that satisfies mass conservation and momentum conservation constraints, and the power prediction data of the unit is calculated by combining wake interactions in an energy conservation graph model based on a graph neural network; To obtain a physically credible probability distribution of the unit power starting from (the multi-scale fused meteorological field), the present invention continuously executes two-level models in the prediction link: Conservation condition reversible network : Generate a probability wind speed field that satisfies mass-momentum conservation; Energy conservation graph neural network : Map the probability wind speed to a power quantile curve considering wake interactions.

[0095] The conservation condition reversible network generates a probability wind field, and the construction of the conditional vector:

[0096] denotes "crop + convolutional downsampling + fully connected embedding"; denotes the local meteorological condition vector.

[0097] Reversible mapping and logarithmic Jacobian:

[0098] represents a latent variable; represents the three-dimensional wind speed tensor.

[0099] Each residual coupling block only transforms half of the channels, so:

[0100] represents the layer scaling function.

[0101] Joint loss function:

[0102] represents the mass conservation weight; represents the momentum conservation weight; represents the static pressure perturbation; represents the kinematic viscosity coefficient. The second and third terms of the residual feedback the gradients of the conservation constraints to the latent space, making the sampling results naturally conform to the fluid equations.

[0103] Sampling and statistics, extracting groups of latent variables and mapping:

[0104] Calculating 、 、 、 and other indicators to form a probabilistic wind field ensemble .

[0105] The energy-conserving graph neural network calculates the power, and graph construction: Assume that the wind farm contains a set of turbine nodes , when the node falls into the wake cone of the node an edge is added . An undirected graph is obtained.

[0106] Node and edge features, node features: mean probability wind speed and variance ; edge features: wind direction angle and center distance .

[0107] Message passing and power conservation:

[0108]

[0109] represents the original node vector; represents the attention weight; represents the wind direction - similarity function. The output layer projects to power and applies energy conservation:

[0110] represents the air density; represents the swept area; represents the power coefficient, which does not exceed the Betz limit after Sigmoid normalization.

[0111] Probabilistic power output, for each in run , power samples are obtained. The statistical expectation and the 5% / 95% quantiles are used to output the power prediction set .

[0112] In one embodiment, during the "Plum Blossom" typhoon test: the mean value of the mass conservation residual ; the 95th percentile power deviation is reduced from 14% of the baseline to 6%; the scheduling reserve capacity locking time is advanced by 11 min. No blade root overload occurs in the unit, verifying the improvement of the "conservation - reversible sampling + energy - conservation graph inference" on the power reliability under extreme typhoon scenarios.

[0113] Preferably, the reversible generation model uses a conditional reversible network with a residual coupling structure, records the logarithmic Jacobian determinant during the forward mapping, and simultaneously minimizes the mass - conservation residual term and the momentum - conservation residual term during the reverse sampling process to obtain probabilistic wind field data.

[0114] In the case of having obtained the multi - scale fusion meteorological field , the present invention uses a conditional reversible network with a residual coupling structure (Conditional Residual - Coupling Flow, denoted as ) to perform probabilistic sampling on the local typhoon wind field; by explicitly adding mass conservation and momentum conservation penalties in the latent variable space, it ensures that the generated samples not only satisfy statistical consistency but also follow the basic equations of fluid mechanics. The following will be described in four parts: network mechanism, physical constraints, sampling process, and technical effects.

[0115] Network mechanism, conditional vector construction:

[0116] represents the local meteorological condition vector; is a fixed "cropping convolutional downsampling Fully connected embedding "mapping".

[0117] Invertible mapping relationship:

[0118] Represents the three - dimensional wind speed tensor; Represents the latent variable; Consists of several affine coupling blocks, Are all trainable parameters.

[0119] Logarithmic Jacobian accumulation. Each coupling block only transforms half of the channels, so the overall logarithmic Jacobian determinant can be accumulated layer by layer:

[0120] Is the output of the layer scale function; Is the channel sub - tensor of this layer that is not affine - transformed.

[0121] The physical - statistical joint loss network is trained to minimize the following:

[0122] Is given by the latent variable density and the logarithmic Jacobian determinant to ensure statistical consistency.

[0123] Represents the divergence; the two - norm term penalizes the deviation from mass conservation.

[0124] Represents the residual of the incompressible Navier - Stokes equations; the two - norm term penalizes the deviation from momentum conservation.

[0125] Represents the static pressure perturbation; Represents the kinematic viscosity coefficient; And Are the mass conservation weight and the momentum conservation weight respectively.

[0126] The conserved residuals back - propagate the gradient in a soft - constraint manner. When the sampled wind speed deviates from the conserved manifold, the latent variable gradient "pulls" the sample back to the physically feasible region.

[0127] Probability sampling and power mapping process, sampling: Fix , draw groups of latent variables ; generate the wind field .

[0128] Statistics: Calculate the expectation for the set , variance , upper / lower quantiles form a probabilistic wind field .

[0129] Power inference: Feed each sample into the energy conservation graph neural network : Nodes: wind turbines and meteorological towers; Edges: wake cone coverage relationship; Message passing: multi-head attention to calculate wind speed attenuation; Output power: , represents air density; represents the swept area of the impeller; is limited by Sigmoid normalization not to exceed the Betz limit .

[0130] Output: aggregate the expected power and the 5% / 95% quantile power to form a power prediction interval.

[0131] In an embodiment, in the "Plum Blossom" typhoon landing experiment, set : average divergence residual ; The mean square of the momentum conservation residual is reduced by about 60% compared with the non-conservation baseline; The 95th percentile deviation of the unit power is reduced from 14% to 6%, the standby capacity locking time is advanced by 11 minutes, and there is no tripping due to sudden wind speed jumps on site.

[0132] This result shows that: the conservation condition reversible network + energy conservation graph model can provide a credible power quantile curve under extreme typhoon wind fields, providing high-confidence support for power grid dispatching and unit safe operation.

[0133] Preferably, the energy conservation graph model takes wind turbines and meteorological towers as nodes and wake interactions as edges, calculates the node hidden vectors through multi-layer attention message passing, and applies power conservation projection at the output layer to map wind speed information into unit power prediction data.

[0134] The energy conservation graph model of the present invention abstracts the wind turbine - meteorological tower scenario into a graph structure with physical meanings, depicts wake attenuation through multi-layer attention message passing, and then applies power conservation projection at the output end to map the probabilistic wind speed into the unit power prediction quantile curve. The following will sequentially explain the graph construction, message passing mechanism, power projection formula and technical effects, and make a unified annotation for all symbols. Graph construction and feature definition:

[0135] represents the node set, consisting of wind turbine nodes and meteorological tower nodes Composition; Represents a set of undirected edges. If a node is located within the wake cone of the impeller, an edge is connected to the node . Node initial vector:

[0136] where represents the mean value of the wind speed sample at the impeller center; represents the standard deviation of the wind speed sample; represents the normalized coordinate of the node height.

[0137] Edge feature:

[0138] where, represents the wind direction angle; represents the horizontal distance between turbines; represents the wake influence radius.

[0139] Attention message passing, the update formula for the th layer:

[0140]

[0141] , represents the linear transformation matrix of the self-node and neighboring nodes; represents the activation function; is the wake attention scoring function, in the form of .

[0142] Power conservation projection. The output layer first gives the unconstrained predicted power:

[0143] Subsequently, a double projection of the impeller rated power and energy conservation is applied:

[0144]

[0145] where represents the final turbine power prediction; represents the turbine rated power; represents the set of turbines affected by the wake in a certain row; represents the set of the most upstream turbines without wake. The projection ensures that the total power of the downstream turbines does not exceed the upstream kinetic energy flux.

[0146] Symbol annotation: Represents the average wind speed at the impeller center; Represents the standard deviation of wind speed; Represents the wind direction angle; Represents the horizontal distance; Represents the wake radius; Represents the air density; Represents the swept area; Represents the power coefficient; Represents the Hidden vector of the layer node; Represents the attention weight.

[0147] Example: In the "Plum Blossom" typhoon landing condition, 256 groups of probability wind fields sampled by conservation and reversibility are input into the graph model: Average relative error of the unit layer: expected power 3.8%, 95% percentile power 6.0%; After satisfying the energy conservation projection, the power difference between upstream and downstream is reduced by 42% compared to that without projection; The dispatching system issues a power reduction instruction 11 minutes in advance according to the 95% curve, and no unit trips due to wake misjudgment.

[0148] Result verification: Multi-layer attention can express directional wake attenuation; power conservation projection prevents "false increase" of energy, significantly improving the credibility of power prediction in extreme typhoon scenarios.

[0149] Input the unit power prediction data into the risk-weighted decision model, generate a dispatching power curve and power uncertainty based on the risk values of peak load conditions, reserve capacity conditions, and ramp rate conditions, and online update the parameters of the risk-weighted decision model within a preset period according to the grid-connected power feedback, and output the grid-side dispatchable power prediction data.

[0150] After obtaining the unit power probability set The present invention maps the stochastic power into a grid-side dispatchable power curve and its uncertainty through the risk-weighted decision model. This model simultaneously considers three types of operation risks: peak load, reserve capacity, and ramp rate, and online updates the weights in a closed-loop manner with the measured grid-connected power during the operation period to achieve dynamic dispatch optimization in extreme typhoon scenarios.

[0151] Risk indicators and weighted objective function, denoted as:

[0152] Are the risk values of peak load, reserve capacity, and ramp rate conditions respectively;

[0153] Are the corresponding risk weights 。At time 's power prediction quantile vector:

[0154] as the input, construct a weighted objective:

[0155] represents the peak load limit; represents the lower bound of the minimum reserve capacity; represents the derivative of the 95th percentile power with respect to time; represents the maximum ramp rate allowed by the power grid. Taking minimization as the criterion, intercept the dispatchable power within the interval:

[0156] Online weight update, let the measured power of grid connection be , error:

[0157] Within the preset period accumulate the mean squared error:

[0158] Use constrained stochastic gradient descent to update the weights:

[0159] where represents the learning rate; represents the projection onto the simplex . After updating, re-solve by the aforementioned method to achieve the rolling adaptability of the decision-making model.

[0160] Symbol description: represents the multi-scale fusion meteorological field; represents the conditional vector embedded by ; , represent the unit power and quantile; represents the peak load limit; represents the lower bound of the reserve capacity; represents the upper limit of the allowable ramp rate; , , represents the risk weight; represents the dispatch power curve; represents the uncertainty of the dispatch power; represents the scheduling error; represents the weight update learning rate.

[0161] In one embodiment, during the landing and operation of Typhoon Doksuri, online updates were performed with a 30-minute cycle: the utilization rate of spare capacity during off-peak periods increased by 18%; the number of maximum ramp-overlimit events decreased from 9 to 1; the unit trip rate was 0, verifying the dynamic robustness of the risk-weighted decision model combined with measured feedback.

[0162] Preferably, the risk-weighted decision model calculates the risk weight coefficients based on the conditional risk value of peak load, the conditional risk value of spare capacity, and the conditional risk value of ramp rate respectively, applies the risk weight coefficients to the corresponding risk components in the power prediction data and performs weighted synthesis to generate a scheduling power curve and synthesize the power uncertainty according to the variances of the risk components.

[0163] To convert the unit power probability prediction into dispatchable power curves on both sides, the present invention constructs a risk-weighted decision model. The model first calculates the weight coefficients according to three types of operation risks, and then synthesizes the scheduling curve by weighting the power components with the weights and gives the uncertainty through variance superposition .

[0164] Risk quantification:

[0165]

[0166]

[0167] where represents taking the non-negative part. represents the peak load limit; represents the lower bound of spare capacity; represents the maximum allowable ramp rate.

[0168] Risk weight coefficient:

[0169]

[0170] If the denominator is zero (risk-free scenario), set . The weights are calculated by sliding with time and can reflect the change in the risk proportion caused by the evolution of the typhoon.

[0171] Scheduling power curve and uncertainty:

[0172]

[0173] , , are the estimated variances at the 5th percentile, mean, and 95th percentile respectively (calculated from the probabilistic wind field sample set). If each risk component is considered independent, the above equation is the variance superposition of the power uncertainty.

[0174] For online update under grid connection feedback, let the measured grid-connected power be , and define the error:

[0175] Within the preset period , accumulate the sum of squared errors:

[0176] Adopt projection gradient update:

[0177] is the projection onto the simplex , is the learning rate. The online update enables the weights to adapt to the real-time operating state of the power grid.

[0178] Symbol description: represents the power at the 5th percentile; represents the power expectation; represents the power at the 95th percentile; , respectively represent the peak load limit and the lower bound of the reserve capacity; represents the derivative of the power at the 95th percentile with respect to time; represents the maximum ramp rate allowed by the power grid; , , represent three types of risk values; , , represent the corresponding risk weights; represents the dispatching power curve; represents the power uncertainty; represents the measured grid-connected power; represents the learning rate for weight update.

[0179] In the embodiment, during the landing drill of Typhoon "Gull", the risk-weighted decision model reduced the dispatching power during the peak load overrun period by 8%, while keeping the reserve capacity abundant; the number of maximum ramp overrun events throughout the day decreased from 7 to 1, and no overload or tripping of the unit occurred, proving that the model can effectively balance the three types of operating risks and output an executable grid-side dispatching curve under extreme wind conditions.

[0180] Preferably, the grid-connected power feedback triggers an online update process at a preset time interval. During this process, the parameters of the risk-weighted decision model and the parameters of the last stage of the reversible generation model are synchronously adjusted according to the error gradient between the predicted power and the grid-connected power.

[0181] To maintain the continuous consistency between the prediction link and the actual operation of the power grid during typhoons, in the present invention, after the grid-connected power enters the control center, an online update process is triggered at a preset time interval Once. The update uses the error gradient for synchronous correction.

[0182] The weight vector of the risk-weighted decision model .

[0183] The parameters of the last stage (tail) of the reversible generation model , thereby reducing both the scheduling risk and the wind speed sampling deviation simultaneously.

[0184] Error construction and gradient acquisition:

[0185] Among them, represents the scheduling error.

[0186] The cumulative squared error in the interval :

[0187] where is the cumulative squared error; the gradient of the risk-weighted decision model weight vector and the gradient of the parameters of the last stage of the reversible generation model are both obtained by a single calculation using the automatic differentiation framework without explicit derivation.

[0188] Online update of the risk weight vector:

[0189] represents the weight learning rate; represents projection onto the probability simplex:

[0190] Symbol description represents the peak load risk weight; represents the reserve capacity risk weight; represents the ramp rate risk weight; the projection ensures that the three weights are always non-negative and sum to 1.

[0191] Update of the parameters of the last stage of the reversible generative model. The parameters of the last stage specifically refer to the scale-translation network weights of the last two coupling blocks, which are used to quickly correct the drift of the wind speed sample distribution. The Adam rule is adopted:

[0192]

[0193] represents the learning rate of the last stage; are the first-order and second-order momentum coefficients; is the numerical stability constant.

[0194] The overall closed-loop workflow includes: 1. Output of the prediction link and ; 2. After measuring in grid connection, calculate the error ; 3. If , perform gradient accumulation of steps 1-3 and synchronously update ; 4. Immediately regenerate the probability wind field based on the new parameters and refresh the scheduling power curve to complete the closed loop.

[0195] Example, parameters , , . Continuously run for 24 hours on the landing day of Typhoon Muifa: The mean square error of the wind speed samples converges to 58% of the baseline within 3 hours after self-update; the average absolute error of the scheduling power is 4.1%, and the baseline without update is 7.9%; the peak load overlimit alarm is reduced from 6 times to 1 time.

[0196] The results show that the dual adaptability driven by the error gradient can quickly correct the wind field probability and scheduling decision without intervening in the upstream numerical weather model, improving the grid-side scheduling reliability in extreme typhoon scenarios.

[0197] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, system, or computer program product. Therefore, the present application can adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects.

[0198] The above are only the embodiments of the present application and are not used to limit the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

1. A new energy power prediction method that integrates typhoon meteorological information and micro-meteorological prediction results, characterized in that, Including: Obtain satellite observations, radar wind profiles, surface wind measurements, and turbine operation data of the wind farm. After time correction, coordinate unification, quality control, and normalization, generate multi-source meteorological data; Perform numerical assimilation on the multi-source meteorological data to generate a mesoscale typhoon field. Infer the mesoscale typhoon field to a micro-meteorological field by super-resolution, and fuse it using the frequency-domain phase consistency method to output a multi-scale fused meteorological field; Taking the multi-scale fused meteorological field as a condition, sample probability wind field data in a reversible generation model based on a flow-based reversible network that satisfies mass conservation and momentum conservation constraints, and combine wake interactions in an energy conservation graph model based on a graph neural network to calculate turbine power prediction data; Input the turbine power prediction data into a risk-weighted decision-making model, generate a dispatch power curve and power uncertainty based on the conditional risk value of the peak load, the conditional risk value of the reserve capacity, and the conditional risk value of the ramp rate, and online update the parameters of the risk-weighted decision-making model within a preset period according to the grid-connected power feedback, and output grid-side dispatchable power prediction data.

2. The method according to claim 1, wherein The multi-source meteorological data consists of satellite scatterometer observation data, airborne sounding profile data, three-dimensional Doppler radar wind profile data, sonic anemometer observation data, lidar wind observation data, unmanned aerial vehicle micro-layer flow observation data, and turbine operation monitoring data, and enters the subsequent processing process after constructing a unified data structure under a unified time reference and a unified geographic coordinate system.

3. The method according to claim 1, wherein For the quality control, first use statistical threshold filtering to remove the observed values that exceed the abnormal discrimination boundary, then use smooth interpolation to complete the short-term continuous missing data, and use geostatistical interpolation to complete the long-term continuous missing data. After completion, set a quality identification code for each observation record.

4. The method according to claim 1, wherein The numerical assimilation uses a four-dimensional variational assimilation algorithm, takes wind speed, temperature, humidity, and pressure as control vectors, and generates a mesoscale typhoon field by minimizing the weighted error function of the background field and the observation field.

5. The method according to claim 4, characterized in that, The resolution improvement uses a spectral embedding diffusion network with a frequency-domain convolutional architecture, and converts the mesoscale typhoon field into a micro-meteorological resolution meteorological field through forward noise perturbation and reverse denoising reconstruction sequence mapping.

6. The method according to claim 5, wherein The frequency-domain phase consistency method performs wavelet transform on the mesoscale typhoon field and the micro-meteorological resolution meteorological field, performs linear phase combination according to the set weight within the corresponding frequency band range, and then performs inverse transform to generate a multi-scale fused meteorological field.

7. The method according to claim 1, wherein The reversible generation model uses a conditional reversible network with a residual coupling structure, records the logarithmic Jacobian determinant in the forward mapping, and simultaneously minimizes the mass conservation residual term and the momentum conservation residual term during the reverse sampling process to obtain probability wind field data.

8. The method according to claim 7, wherein The energy conservation graph model takes wind turbines and meteorological towers as nodes and wake interactions as edges, calculates the node hidden vectors through multi-layer attention message passing, and applies power conservation projection at the output layer to map the wind speed information to turbine power prediction data.

9. The method according to claim 1, characterized in that, The risk-weighted decision-making model calculates risk weight coefficients based on the conditional risk value of peak load, the conditional risk value of reserve capacity, and the conditional risk value of ramp rate respectively, applies the risk weight coefficients to the corresponding risk components in the power prediction data and performs weighted synthesis, generates a scheduling power curve, and synthesizes power uncertainty according to the variances of the risk components.

10. The method according to claim 9, wherein The grid-connected power feedback triggers an online update process at a preset time interval, and in this process, the parameters of the risk-weighted decision-making model and the parameters of the last stage of the reversible generation model are synchronously adjusted according to the error gradient between the predicted power and the grid-connected power.

Images