A method and device for predicting wind and solar power of new energy based on large meteorological model

Abstract

The present invention proposes a method and device for predicting wind and solar power of new energy based on a large meteorological model, including: making a large meteorological model ensemble forecast data set ENS0, performing ensemble assimilation and forecasting to obtain a deterministic forecast data set F0; constructing a fine-tuning localization matrix, based on the fine-tuning localization matrix and the large meteorological model ensemble forecast data set ENS0, traversing and optimizing to obtain a deterministic grid forecast data set F100; performing spatial scale separation on the deterministic grid forecast data set F100, traversing and optimizing to obtain a large meteorological model deterministic grid forecast data set D100; using the historical gridding data of the large meteorological model deterministic grid forecast data set D100 to calculate weights, performing weighted averaging according to the weights, and obtaining a fused power forecast time series. The present invention comprehensively considers the influence of the localization matrix, distinguishes the influences at different spatial scales, and can significantly improve the stability and accuracy of new energy power prediction.

Description

Technical Field

[0001] The present invention belongs to the technical field of new energy power, and in particular relates to a method and device for predicting new energy wind and solar power based on a large meteorological model. Background Art

[0002] With the growth of wind and solar energy, more than half of the newly added installed capacity in recent years is renewable energy. However, due to the intermittent and volatile characteristics of renewable energy, how to accurately predict the power of wind and solar power and reduce the impact of its volatility on the power grid is a major problem facing the current renewable energy power prediction.

[0003] The optimization goals of traditional data assimilation and numerical weather forecasting technology often do not take into account new energy meteorological elements and new energy power prediction, and only focus on the optimization of traditional meteorological elements. When using the forecast field, various spatial scales are mixed. Neither new energy meteorological elements are used as optimization goals when making the initial field, nor are the signals of various spatial scales of the forecast field optimized or corrected. The prediction results of new energy meteorological elements and power prediction have great risks in the initial field production and the application of the forecast field, so there are certain deviations in the prediction of new energy meteorological elements. For example, when performing data assimilation based on ensemble forecasts, key technical links such as the calculation and optimization of localized matrices often do not pay attention to the accuracy of new energy meteorological element predictions. In addition, the dynamic framework and physical schemes used in existing numerical weather forecasts are not perfect, and the error levels of prediction signals at different spatial scales are often different. The existing prediction of new energy meteorological elements often mixes various spatial scales without distinguishing them, which introduces large errors in power prediction. Summary of the invention

[0004] The present invention proposes a method and device for predicting new energy wind and solar power based on a large meteorological model, which comprehensively considers the impact of new energy meteorological element prediction and power prediction on the localized matrix in data assimilation, distinguishes the deviation level of gridded forecast data of large meteorological models at different spatial scales and its impact as input on new energy power prediction, and can significantly improve the stability and accuracy of new energy power prediction.

[0005] To achieve the above object, the technical solution of the present invention is achieved as follows:

[0006] A method for predicting wind and solar power of renewable energy based on a large meteorological model, comprising:

[0007] S1: Create a large meteorological model ensemble forecast dataset ENS0, perform ensemble normalization and forecast to obtain a deterministic forecast dataset F0;

[0008] S2: Construct a fine-tuning localization matrix, and based on the fine-tuning localization matrix and the meteorological large model ensemble forecast dataset ENS0, traverse and optimize to obtain the meteorological large model deterministic grid forecast dataset F100;

[0009] S3, performing spatial scale separation on the meteorological large model deterministic grid forecast dataset F100, and traversing and optimizing to obtain the meteorological large model deterministic grid forecast dataset D100;

[0010] S4. Calculate weights using historical gridded data of the meteorological large model deterministic grid forecast data set D100, calculate power forecast time series using future forecast gridded data of the meteorological large model deterministic grid forecast data set D100, and perform weighted averaging based on the weights to obtain a fused power forecast time series.

[0011] Furthermore, step S1 specifically includes:

[0012] S11, making the ensemble initial field required for the meteorological large model ensemble forecast, adding the initial disturbance on the basis of the deterministic forecast of the meteorological center, using the meteorological large model to make forecasts with the initial field of each ensemble member as input, and obtaining a set of meteorological large model ensemble forecast data sets ENS0;

[0013] S12. Obtain the available meteorological observation dataset OBS0 of the global assimilation system JEDI; calculate the localization matrix L0 based on the meteorological large model ensemble forecast dataset ENS0, assimilate the meteorological observation dataset OBS0; update the forecast field of the meteorological large model to obtain the analysis field; use the analysis field as the input of the numerical weather forecast or the meteorological large model to obtain the deterministic forecast dataset F0.

[0014] Furthermore, step S2 specifically includes:

[0015] S21. For a submatrix corresponding to a certain meteorological element variable in the localized matrix L0, a fine-tuning coefficient matrix is ​​constructed, whose value is a Gaussian weight centered on the new energy station and decaying toward the periphery; the amplitude and influence range of the traversal adjustment weight are combined to form a set of matrices corresponding to the Gaussian weights, forming a fine-tuning coefficient matrix C tuning ; When adjusting the amplitude and influence range of Gaussian weights by combination traversal, the fine-tuning coefficient matrix composed of each adjustment of Gaussian weights is multiplied element by element with the corresponding localization matrix, and the combination traversal obtains the fine-tuning localization matrix L tuning ;

[0016] S22, for the sub-matrices of all meteorological elements corresponding to a certain combination of amplitude and influence range in the fine-tuning localization matrix, based on the sub-matrices and the meteorological large model ensemble forecast data set ENS0, perform ensemble normalization and forecasting to obtain the corresponding meteorological large model deterministic forecast data set F1;

[0017] Traverse all fine-tuning localization matrices L tuning The amplitude and influence range are combined to obtain the fine-tuning localization matrix L tuning The corresponding set of meteorological large model deterministic grid forecast data sets F100 all ; Establish a set of meteorological large model deterministic grid forecast data sets F100 all The first mapping relationship between the historical prediction data set and the measured power corresponding to the same period in history is used to convert the set of meteorological large model deterministic grid forecast data set F100 all Substitute the future prediction data set of the first sub-data set into the first mapping relationship to obtain a set of future prediction power, and calculate the error between the future prediction power and the actual measurement during the same period; traverse the set of meteorological large model deterministic grid forecast data sets F100 all , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set F100 is obtained.

[0018] Furthermore, step S3 specifically includes:

[0019] S31. For each member and each meteorological element of the large-scale deterministic grid forecast data set F100 of the meteorological model, the correlation coefficient between each grid point and other grid points is calculated, and the relationship between the correlation coefficient and the grid point spacing is statistically analyzed to obtain the grid point spacing corresponding to the significant local extreme value of the correlation coefficient greater than the threshold value, which is used as the correlation spatial scale separation coefficient corresponding to the grid data;

[0020] For the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast dataset F100 of the meteorological model, power spectrum analysis is performed in the longitude and latitude directions respectively, and the spatial scale corresponding to the local extreme value greater than the test value is obtained as the power spectrum spatial scale separation coefficient corresponding to the gridded data;

[0021] The correlation spatial scale separation coefficient and the power spectrum spatial scale separation coefficient together constitute the spatial scale separation coefficient vector C of the meteorological large model deterministic grid forecast data set F100 filter ;

[0022] S32, gridded data of each member and each meteorological element of the large-scale deterministic grid forecast data set F100 of the meteorological model, and the spatial scale separation coefficient vector C corresponding to the gridded data filter , and take each element in the subvector from large to small as the cutoff spatial scale to perform two-dimensional local weighted regression on the gridded data, retain the low-pass field and recursively perform two-dimensional local weighted regression on the high-pass field in turn, and obtain a set of C filter Corresponding spatial scale-separated deterministic grid forecast dataset F100 filter ;

[0023] Establishing the deterministic grid forecast dataset F100 filter The second mapping relationship between the historical forecast dataset and the measured power corresponding to the same period in history. filter Substitute the future prediction data set of a sub-data set into the second mapping relationship to obtain a set of future prediction powers, and calculate the error between the obtained future prediction power and the actual measurement during the same period; traverse the deterministic grid forecast data set F100 filter , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set D100 is obtained.

[0024] On the other hand, the present invention also proposes a new energy wind and solar power prediction device based on a large meteorological model, comprising:

[0025] Dataset module: Create a large meteorological model ensemble forecast dataset ENS0, perform ensemble normalization and forecast to obtain a deterministic forecast dataset F0;

[0026] Fine-tuning localization module: construct a fine-tuning localization matrix, and based on the fine-tuning localization matrix and the meteorological large model ensemble forecast dataset ENS0, traverse and optimize to obtain the meteorological large model deterministic grid forecast dataset F100;

[0027] Scale separation module: performing spatial scale separation on the meteorological large model deterministic grid forecast data set F100, and traversing and optimizing to obtain the meteorological large model deterministic grid forecast data set D100;

[0028] Power prediction module: The weights are calculated using the historical gridded data of the meteorological large model deterministic grid forecast data set D100, and the power prediction time series is calculated using the future forecast gridded data of the meteorological large model deterministic grid forecast data set D100, and weighted averaging is performed according to the weights to obtain a fused power prediction time series.

[0029] Furthermore, the dataset module includes:

[0030] Ensemble forecast data set unit: produces the ensemble initial field required for the meteorological large model ensemble forecast, adds initial disturbances based on the deterministic forecast of the meteorological center, and uses the meteorological large model to make forecasts with the initial field of each ensemble member as input, thus obtaining a set of meteorological large model ensemble forecast data sets ENS0;

[0031] Deterministic forecast dataset unit: obtain the available meteorological observation dataset OBS0 of the global assimilation system JEDI; calculate the localization matrix L0 based on the meteorological large model ensemble forecast dataset ENS0, assimilate the meteorological observation dataset OBS0; update the forecast field of the meteorological large model to obtain the analysis field; use the analysis field as the input of the numerical weather forecast or the meteorological large model to obtain the deterministic forecast dataset F0.

[0032] Furthermore, fine-tuning the localization module includes:

[0033] Fine-tuning localized matrix unit: for a submatrix corresponding to a meteorological element variable in the localized matrix L0, a fine-tuning coefficient matrix is ​​constructed, whose value is a Gaussian weight centered on the new energy station and decaying toward the periphery; the amplitude and influence range of the weight are combined and traversed to form a set of matrices corresponding to the Gaussian weights, forming a fine-tuning coefficient matrix C tuning ; When adjusting the amplitude and influence range of Gaussian weights by combination traversal, the fine-tuning coefficient matrix composed of each adjustment of Gaussian weights is multiplied element by element with the corresponding localization matrix, and the combination traversal obtains the fine-tuning localization matrix L tuning ;

[0034] The first traversal optimization unit: for the sub-matrix of all meteorological elements corresponding to a certain combination of amplitude and influence range in the fine-tuning localization matrix, based on the sub-matrix and the meteorological large model ensemble forecast data set ENS0, ensemble assimilation and forecast are performed to obtain the corresponding meteorological large model deterministic forecast data set F1;

[0035] Traverse all fine-tuning localization matrices L tuning The amplitude and influence range are combined to obtain the fine-tuning localization matrix L tuning The corresponding set of meteorological large model deterministic grid forecast data sets F100 all ; Establish a set of meteorological large model deterministic grid forecast data sets F100 all The first mapping relationship between the historical prediction data set and the measured power corresponding to the same period in history is used to convert the set of meteorological large model deterministic grid forecast data set F100 all Substitute the future prediction data set of the first sub-data set into the first mapping relationship to obtain a set of future prediction power, and calculate the error between the future prediction power and the actual measurement during the same period; traverse the set of meteorological large model deterministic grid forecast data sets F100 all , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set F100 is obtained.

[0036] Furthermore, the scale separation module includes:

[0037] Separation coefficient vector unit: for the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast data set F100 of the meteorological model, the correlation coefficient between each grid point and other grid points is calculated, and the relationship between the correlation coefficient and the grid point spacing is statistically analyzed to obtain the grid point spacing corresponding to the significant local extreme value of the correlation coefficient greater than the threshold value, which is used as the correlation spatial scale separation coefficient corresponding to the gridded data;

[0038] For the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast dataset F100 of the meteorological model, power spectrum analysis is performed in the longitude and latitude directions respectively, and the spatial scale corresponding to the local extreme value greater than the test value is obtained as the power spectrum spatial scale separation coefficient corresponding to the gridded data;

[0039] The correlation spatial scale separation coefficient and the power spectrum spatial scale separation coefficient together constitute the spatial scale separation coefficient vector C of the meteorological large model deterministic grid forecast data set F100 filter ;

[0040] Second traversal optimization unit: gridded data of each member and each meteorological element of the large meteorological model deterministic grid forecast data set F100, and the spatial scale separation coefficient vector C corresponding to the gridded data filter , and take each element in the subvector from large to small as the cutoff spatial scale to perform two-dimensional local weighted regression on the gridded data, retain the low-pass field and recursively perform two-dimensional local weighted regression on the high-pass field in turn, and obtain a set of C filter Corresponding spatial scale-separated deterministic grid forecast dataset F100 filter ;

[0041] Establishing the deterministic grid forecast dataset F100 filter The second mapping relationship between the historical forecast dataset and the measured power corresponding to the same period in history. filter Substitute the future prediction data set of a sub-data set into the second mapping relationship to obtain a set of future prediction powers, and calculate the error between the obtained future prediction power and the actual measurement during the same period; traverse the deterministic grid forecast data set F100 filter , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set D100 is obtained.

[0042] Compared with the prior art, the present invention has the following beneficial effects:

[0043] 1) The present invention gives full play to the high efficiency and high precision technical characteristics of large meteorological models, optimizes the key technical links of centralized assimilation calculation of localized matrices in traditional numerical weather forecasting and data assimilation technology, and takes the new energy power prediction error as the optimization target, organically integrates data assimilation with new energy prediction, which can effectively improve the effects of numerical weather forecasting and data assimilation technology, and thus improve the accuracy of new energy meteorological element prediction and power prediction.

[0044] 2) The present invention screens out the spatial scales with smaller errors and better forecast quality and the corresponding gridded data in each spatial scale of the meteorological model by spatial scale separation and optimization of the gridded forecast data of the large meteorological model. This can effectively remove the noise signal of the prediction of new energy meteorological elements of the meteorological model, and further improve the accuracy of photovoltaic or wind power prediction with the meteorological source optimized by spatial scale separation as input. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 is a schematic diagram of a flow chart of an embodiment of the present invention;

[0046] Figure 2 is a schematic diagram of the detailed steps included in S101 in an embodiment of the present invention;

[0047] Figure 3 is a schematic diagram of the detailed steps included in S102 in an embodiment of the present invention;

[0048] Figure 4 It is a schematic diagram of the detailed steps included in S103 in the embodiment of the present invention. DETAILED DESCRIPTION

[0049] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0050] In order to make the purpose and features of the present invention more obvious and understandable, the specific implementation methods of the present invention are further described below in conjunction with the accompanying drawings.

[0051] First, the professional terms involved in this embodiment are explained as follows:

[0052] Numerical weather forecast: Based on the actual atmospheric conditions, under certain initial and boundary conditions, numerical calculations are performed on large computers to solve the fluid mechanics and thermodynamics equations that describe the weather evolution process, and to predict the atmospheric thermodynamic state and weather phenomena for a certain period of time in the future.

[0053] Data assimilation: The process of comparing and adjusting the latest observational data (such as temperature, humidity, wind speed, etc.) with the output of the numerical weather prediction model to correct and update the initial conditions of the model to provide the best estimate of the state of the atmosphere. Data assimilation often uses ensemble forecasts as input to evaluate the errors in the output of the numerical weather prediction model.

[0054] Meteorological big model: Meteorological big model uses artificial intelligence and big data technology, especially deep learning algorithms, to deeply integrate and mine multiple data sources such as meteorological observation data and numerical weather forecast model output to build a high-precision forecast model that can capture the complexity and nonlinear characteristics of atmospheric movement. These models can automatically extract the regularity and characteristics of atmospheric movement through training and learning of historical meteorological data, and then predict future weather conditions.

[0055] Ensemble forecast: A group of different forecast results for the same effective forecast time. The differences between the forecasts can provide information about the probability distribution of the forecasted quantity. Each forecast in the ensemble forecast can have different initial conditions, boundary conditions, parameter settings, and can even be generated using completely independent numerical weather forecast models.

[0056] Localization matrix: In the process of meteorological data assimilation, the localization matrix is ​​a tool used to reduce or eliminate the impact of long-distance observation data on the analysis variables. The main purpose of localization is to limit the scope of the influence of observation data on the analysis variables, thereby avoiding errors introduced by inaccurate or unreliable long-distance correlations. In data assimilation based on ensemble forecasts, if ensemble forecasts with a limited number of members are used, long-distance false correlations will often be introduced into the background covariance, thereby reducing the quality of data assimilation. Therefore, it is necessary to introduce a localization matrix so that the background covariance only considers the correlation within the local area.

[0057] Local weighted regression: Local weighted regression is a nonparametric regression method used to fit the regression model in the local area of ​​the data. Unlike the global regression method, local weighted regression performs weighted fitting in the local area near each predicted point, so that the model can better capture the local structure and changes of the data. The above characteristics of local weighted regression make this method often used in two-dimensional spatial filtering.

[0058] Planetary scale: Planetary scale refers to atmospheric phenomena and processes that affect the entire Earth or large parts of it. These phenomena usually have very large spatial scales, often thousands of kilometers or more, and last for long periods of time, from days to weeks or even longer.

[0059] Power spectrum: The power distribution of a signal at different frequency components. It shows how the signal power changes with frequency, that is, the distribution of signal power in the frequency domain. The power spectrum shows the relationship between the signal power and frequency.

[0060] Power spectral density: Power spectral density describes the distribution density of signal power over frequency. It provides a signal power distribution per unit frequency and is the frequency-normalized form of the power spectrum.

[0061] GraphCast Weather Model: A global medium-term weather forecast model based on graph neural network (GNN) developed by Google DeepMind team and others. This model has made significant progress in the field of weather forecasting, providing unprecedented accuracy and efficiency.

[0062] NeuralGCM Meteorological Model: A global medium-term weather forecast model that is a hybrid of physical models and machine learning models, developed by the Google Research team and others. It has demonstrated excellent performance in weather forecasting and climate simulation, and has also opened up new avenues for future climate research and applications.

[0063] In order to solve the problem of improving the accuracy of new energy meteorological element prediction and power prediction at multiple spatial scales, the present invention is based on the deterministic forecast and ensemble forecast produced by the world's mainstream meteorological large model, takes the new energy meteorological element prediction and power prediction as the optimization goal, fine-tunes the localized matrix of data assimilation, scales the gridded forecast data of the meteorological large model, and optimizes the meteorological large model data of different spatial scales as the new energy power prediction input, thereby obtaining the revised future numerical weather forecast data.

[0064] Based on the above design concept, the method of this embodiment is as follows Figure 1 As shown, specifically including:

[0065] S101: Create a large meteorological model forecast dataset.

[0066] First, prepare the meteorological model ensemble forecast dataset ENS0.

[0067] The input data of ensemble forecast can use the ensemble forecast data of major world meteorological centers, the deterministic forecast data of major world meteorological centers with random disturbances added, or the independently produced global ensemble assimilated updated analysis field data. Use the above input data to drive one or more large meteorological models to make forecasts, and obtain a set of meteorological large model ensemble forecast data sets ENS0.

[0068] Then, the aggregate assimilation and forecasting are performed to obtain the meteorological large model deterministic forecast data set F0.

[0069] The background field of data assimilation can use the deterministic forecast data of major world meteorological centers or the forecast data of large meteorological models. Use the empirical constant localization matrix or calculate the localization matrix L0 according to ENS0. Use the large meteorological model ensemble forecast data set ENS0 and the localization matrix L0 as input to perform ensemble-variation assimilation on the background field to obtain the analysis field after data assimilation. Use the analysis field as the input of numerical weather forecast or large meteorological model to obtain the forecast field.

[0070] Repeat the above steps to obtain the meteorological large model ensemble forecast dataset ENS0 and the deterministic forecast dataset F0 for historical forecasts and future forecasts.

[0071] S102: Fine-tune and traverse the optimal localization matrix L0 to obtain the optimal deterministic grid forecast data set F100.

[0072] First, construct the fine-tuning coefficient matrix C tuning and fine-tune the localization matrix L tuning .

[0073] For the submatrix corresponding to a certain meteorological element variable in the localization matrix L0, a fine-tuning coefficient matrix is ​​constructed, whose value is a Gaussian weight centered on the new energy site and decaying toward the periphery. The amplitude and influence range of the weight can be adjusted. The amplitude adjustment can make the value of the localization matrix of the new energy site location change from the grid spacing scale to the planetary scale, and the influence range adjustment can make the Gaussian weight influence range change from the grid spacing scale to the planetary scale. A set of matrices corresponding to Gaussian weights composed of the amplitude and influence range of the combined traversal adjustment weights constitute the fine-tuning coefficient matrix C tuning When the amplitude and influence range of the Gaussian weights are adjusted by combination traversal, the fine-tuning coefficient matrix composed of each adjustment of the Gaussian weights is multiplied element-wise with the corresponding localization matrix. In this way, the fine-tuning localization matrix L is obtained by combination traversal. tuning .

[0074] Then, the optimal large meteorological model deterministic grid forecast data set F100 is traversed.

[0075] For the submatrix of all meteorological elements corresponding to a certain combination of amplitude and influence range in the fine-tuning localization matrix, based on the submatrix and the meteorological large model ensemble forecast data set ENS0, repeat the ensemble assimilation and forecasting process of step S101 to obtain the corresponding meteorological large model deterministic forecast data set. Traverse the amplitude and influence range combinations of all fine-tuning localization matrices, repeat the above steps, and obtain a group of meteorological large model deterministic forecast data sets corresponding to the fine-tuning localization matrix. Establish a mapping relationship between the historical prediction data set and the measured power corresponding to the same period in a group of meteorological large model deterministic forecast data sets. Substitute the future prediction data set of the meteorological large model into the mapping relationship to obtain a group of future predicted power, and further calculate the error between the future predicted power and the measured power in the same period. Based on the error traversal, the optimal meteorological large model deterministic forecast is selected, and the ones with larger errors are eliminated to obtain the traversal preferred meteorological large model deterministic grid forecast data set F100.

[0076] S103: scale-separate the large-scale meteorological model deterministic grid forecast dataset F100 to obtain an optimal large-scale meteorological model deterministic grid forecast dataset D100.

[0077] First, construct the spatial scale separation coefficient vector C filter .

[0078] For the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast dataset F100 of the meteorological model, the correlation coefficient between each grid point and other grid points is calculated, and the relationship between the correlation coefficient and the grid point spacing is statistically analyzed to obtain the grid point spacing corresponding to the local extreme value with significant correlation coefficient, which is the correlation spatial scale separation coefficient corresponding to the gridded data. For the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast dataset F100 of the meteorological model, power spectrum analysis is performed in the longitude and latitude directions respectively to obtain the spatial scale corresponding to the local extreme value greater than the test value, which is the power spectrum spatial scale separation coefficient corresponding to the gridded data. The correlation spatial scale separation coefficient and the power spectrum spatial scale separation coefficient of the large-scale deterministic grid forecast dataset F100 of the meteorological model together constitute the spatial scale separation coefficient vector C of the large-scale deterministic grid forecast dataset F100 of the meteorological model filter .

[0079] Then, the scale-separated grid forecast data D100 are calculated and optimized.

[0080] The gridded data of each member and each meteorological element of the large-scale deterministic grid forecast dataset F100 of the meteorological model, and the spatial scale separation coefficient vector C corresponding to the gridded data filter , and take each element in the subvector from large to small as the cutoff spatial scale to perform two-dimensional local weighted regression on the gridded data, retain the low-pass field and recursively perform two-dimensional local weighted regression on the high-pass field in turn, and obtain a set of Cfilter The corresponding spatial scales are separated into gridded data sets. A mapping relationship between the historical prediction data set in the data set and the measured power corresponding to the same period in history is established. The future prediction data set in the data set is substituted into the mapping relationship to obtain a set of future prediction powers, and the error between the future prediction power and the measured power in the same period is further calculated. Based on the error traversal, the data set is optimized, and the ones with large errors are eliminated to obtain the traversal optimized meteorological large model deterministic grid forecast data set D100.

[0081] S104: Generate optimized future forecast meteorological elements and convert them into power.

[0082] Using the historical gridded data of the meteorological large model deterministic grid forecast dataset D100, the meteorological element forecast time series or power forecast time series of each member of D100 in a certain wind farm or photovoltaic power station is calculated through mathematical statistical downscaling methods such as bilinear interpolation or other statistical downscaling methods, or dynamic downscaling methods such as regional meteorological models, and then the meteorological element forecast time series or power forecast time series of the wind farm or photovoltaic power station is weighted averaged with weights, and the prediction deviation of the time series after weighted average is calculated. Then, the weight that minimizes the prediction deviation is calculated through mathematical statistical methods such as the least squares method. Using the historical gridded data of the meteorological large model deterministic grid forecast dataset D100, the optimal weights of each member of D100 are cross-validated to enhance the generalization ability of the optimal weights.

[0083] Using the future forecast gridded data of the meteorological large model deterministic grid forecast data set D100, the meteorological element forecast time series or power forecast time series of each member of D100 in a certain wind farm or photovoltaic power station is calculated, and then the meteorological element forecast time series or power forecast time series of the wind farm or photovoltaic power station is weighted averaged with the above optimal weights to obtain the fused meteorological element forecast time series or power forecast time series.

[0084] In addition, the mapping relationship between the historical gridded data of each member of the meteorological large model deterministic grid forecast dataset D100 and the historical measured power can be directly established. The above mapping relationship is cross-validated using the historical gridded data and historical measured power of the meteorological large model deterministic grid forecast dataset D100 to enhance the generalization ability of the mapping relationship. Then, the future forecast gridded data of the meteorological large model deterministic grid forecast dataset D100 is substituted into the above mapping relationship to obtain the optimized power prediction time series.

[0085] This embodiment provides a more detailed implementation process for steps S101 to S103.

[0086] The detailed description of step S101 is as follows: Figure 2 As shown, including:

[0087] S201: Prepare the meteorological large model ensemble forecast dataset ENS0.

[0088] This step takes the pure machine learning model GraphCast and the physical-machine learning hybrid model NeuralGCM as examples, combined with the global assimilation system JEDI, to illustrate the process of preparing the meteorological large model ensemble forecast dataset ENS0.

[0089] First, the ensemble initial field required for the ensemble forecast of the large meteorological model is produced. To generate large-scale ensemble forecasts with more than 100 ensemble members, initial perturbations can be added by using methods such as Perlin noise and RNG noise based on the deterministic forecasts of major world meteorological centers, or by using the gen_ens_pert_B module of the global assimilation system JEDI to add initial perturbations based on the pre-trained background covariance matrix.

[0090] Then, the initial field of each ensemble member is used as input, and the pure machine learning model GraphCast or the physical-machine learning hybrid model NeuralGCM is used for inference and forecasting to obtain a set of ensemble forecast datasets ENS0.

[0091] If the physical-machine learning hybrid model NeuralGCM is used, methods such as Perlin noise method and RNG noise method can be used to add random disturbances to the time prediction tendencies of variables such as temperature anomalies, atmospheric wind field vorticity, and divergence, thereby disturbing the physical processes of the large meteorological model.

[0092] Then, the meteorological model of each ensemble member is inferred and forecasted, and random disturbances are added by using methods such as Perlin noise method and RNG noise method. The above steps are repeated many times to achieve the effect of expanding the ensemble discretization.

[0093] S202: Perform collective normalization and forecasting to obtain a deterministic forecast data set F0.

[0094] This step takes the pure machine learning model GraphCast and the physical-machine learning hybrid model NeuralGCM as examples, combined with the global assimilation system JEDI, to illustrate the process of ensemble assimilation based on the meteorological large model ensemble forecast dataset ENS0 and obtaining the deterministic forecast dataset F0.

[0095] First, the spatial location information, time information, and variable information of the observation data are encoded into the IODA format dedicated to the JEDI system and saved. The saved IODA format observation data file is upgraded to the IODA version recognizable by the JEDI system using the ioda-upgrade.x data update program and its corresponding information conversion configuration file ObsSpaceV2-to-V3.yaml, and the available meteorological observation data set OBS0 of the global assimilation system JEDI is obtained.

[0096] Then, use JEDI's BUMP module to calculate the localization matrix L0 required to calculate the ensemble background covariance based on the ensemble forecast ENS0. You can also set the localization matrix L0 to a constant matrix based on empirical values. Use JEDI's 3DEnVar or 4DEnVar function to assimilate the meteorological observation dataset OBS0, update the forecast field of the meteorological large model, and obtain the analysis field. Use the analysis field as the input of the numerical weather forecast or the meteorological large model to forecast and obtain the forecast field F0.

[0097] In this embodiment, the detailed description of step S102 is as follows: Figure 3 As shown, including:

[0098] S301: Construct fine-tuning coefficient matrix C tuning and fine-tune the localization matrix L tuning .

[0099] This step explains how to construct the differential coefficient matrix C tuning and fine-tune the localization matrix L tuning process.

[0100] For a certain meteorological element variable Var0 in the localized matrix L0, the submatrix L0 corresponds to Var0 , the structural shape and L0 Var0 The same matrix C0 Var0 , whose value is the Gaussian weight attenuated toward the surrounding area with the new energy station as the center.

[0101] ;

[0102] Among them, A0 represents the amplitude of the weight, R is the range of influence, and r is the distance between a grid point and the new energy station; C0 is the above-mentioned Gaussian attenuation distribution within the range of R and is 0 outside the range of R.

[0103] Traverse and adjust A0, you can go from min(L0 Var0 ) / L0 center Change to max(L0 Var0 ) / L0 center ,max(L0 Var0) is the possible maximum value of the localization matrix L0, which can be a planetary scale, min(L0 Var0 ) is the minimum possible value of L0, which can be the grid spacing, L0 center is the matrix L0 Var0 The value of the site location in the new energy field.

[0104] By traversing and adjusting R, it can vary from the grid points adjacent to the new energy stations to the global scope.

[0105] Combination traversal adjusts A0 and R, all combinations of A0 and R correspond to C0 Var0 , forming a matrix L0 Var0 The fine-tuning coefficient matrix of .

[0106] Traverse the sub-matrices of all meteorological element variables in the localization matrix L0, repeat the above operation, and obtain the fine-tuning coefficient matrix C tuning .

[0107] In the above traversal adjustment process, L0 Var0 The value of each grid point is related to C0 Var0 By multiplying the values ​​of each grid point, we can get the fine-tuned localization matrix.

[0108] Traverse the sub-matrices of all meteorological element variables in the localization matrix L0, repeat the above operation, and obtain the fine-tuning localization matrix L tuning .

[0109] S302: Traverse the optimal large meteorological model deterministic grid forecast data set F100.

[0110] This step describes the process of traversing the optimal large meteorological model deterministic grid forecast data set F100.

[0111] First, fine-tune the localization matrix L tuning The submatrix of all meteorological elements corresponding to a certain combination of A0 and R is recorded as the localized matrix L1.

[0112] Then, based on the ensemble forecast ENS0 and the localization matrix L1, step S101 is repeated to obtain the corresponding large meteorological model deterministic forecast data set F1.

[0113] Then, traverse and fine-tune the localization matrix L tuning Each meteorological element variable Var0, each submatrix L0 corresponding to each combination of A0 and R Var0 , repeat the above steps to obtain the fine-tuning localization matrix L tuning The corresponding set of meteorological large model deterministic grid forecast data sets F100 all .

[0114] Then, mathematical statistical methods such as linear regression or other machine learning methods are used to establish a large-scale meteorological model deterministic grid forecast dataset F100. all The mapping relationship between the historical prediction data set of a set of sub-data sets and the corresponding measured power in the same period of history; repeat the above steps to traverse the deterministic forecast F100 of the meteorological large model all For each sub-dataset in the dataset, a mapping relationship is established between each sub-dataset and the measured power corresponding to the same period in history; the above-mentioned large meteorological model deterministic forecast F100 all The mapping relationship between all sub-datasets in the dataset constitutes the meteorological large model deterministic grid forecast dataset F100. all A mapping relationship is established between the historical forecast data set and the corresponding measured power set in the same period of history R100 all .

[0115] Then, the meteorological large model deterministic grid forecast dataset F100 all Substitute the future prediction data set of a set of sub-data sets into the mapping relationship R100 all The corresponding sub-mapping relationship is obtained to obtain the corresponding future predicted power, and the error between the future predicted power and the measured power in the same period is calculated; the above steps are repeated to traverse the meteorological large model deterministic grid forecast data set F100 all For each group of sub-datasets in the dataset, the future predicted power and predicted power error corresponding to each group of sub-datasets are obtained respectively; the above-mentioned meteorological large model deterministic grid forecast dataset F100 all The future prediction powers corresponding to all sub-data sets in the , together form a set of future prediction powers P100 all ; The above-mentioned large meteorological model deterministic grid forecast dataset F100 all The predicted power errors corresponding to all sub-data sets in the , together constitute a set of future predicted power deviations ERROR100 all .

[0116] Then, based on the error ERROR100 all Traverse the optimal F100 all , the F100 with larger error all After elimination, the optimal large-scale meteorological model deterministic grid forecast data set F100 is obtained.

[0117] In this embodiment, the detailed description of step S103 is as follows: Figure 4 As shown, including:

[0118] S401: Constructing the spatial scale separation coefficient vector C filter .

[0119] This step describes the construction of the spatial scale separation coefficient vector C filterprocess.

[0120] First, the gridded data of a certain member or a certain meteorological element of the large-scale meteorological model deterministic grid forecast data set F100 in S102 is obtained.

[0121] Then, the correlation coefficients of the historical one-dimensional time series data corresponding to a certain grid point of the gridded data and the historical one-dimensional time series data corresponding to other grid points are calculated respectively, and the relationship between the above correlation coefficients and the grid point spacing is statistically analyzed.

[0122] Then, the above steps are repeated to obtain the relationship between the average correlation coefficient of all grid points of the grid data and the grid point spacing, and the grid point spacing corresponding to the significant local extreme value is obtained, which is the correlation spatial scale separation coefficient corresponding to the grid data. The significance here can be taken as a correlation coefficient greater than a certain threshold value, such as a correlation coefficient greater than 0.5.

[0123] Then, traversing the above steps, the correlation spatial scale separation coefficients corresponding to the gridded data of all members and all meteorological elements of the meteorological large model deterministic grid forecast dataset F100 are obtained, which constitute the correlation spatial scale separation coefficient C of the meteorological large model deterministic grid forecast dataset F100. corr .

[0124] Then, power spectrum analysis is performed on the gridded data in the longitude and latitude directions respectively to obtain the power spectrum density of the gridded data in the longitude and latitude directions. The spatial scale corresponding to the local extreme value greater than the test value is the power spectrum spatial scale separation coefficient corresponding to the gridded data.

[0125] Then, traversing the above steps, the power spectrum spatial scale separation coefficients corresponding to the gridded data of all members and all meteorological elements of the meteorological large model deterministic grid forecast dataset F100 are obtained, which constitute the power spectrum spatial scale separation coefficient C of the meteorological large model deterministic grid forecast dataset F100. spectrum .

[0126] Correlation spatial scale separation coefficient C of the large-scale deterministic grid forecast dataset F100 of the meteorological model corr and power spectrum spatial scale separation coefficient C spectrum The spatial scale separation coefficient vector C that together constitutes the meteorological large model deterministic grid forecast dataset F100 filter .

[0127] S402: Calculate and optimize the scale-separated large-scale meteorological model deterministic grid forecast data set D100.

[0128] This step describes the process of calculating and optimizing the scale-separated large-scale meteorological model deterministic grid forecast data set D100.

[0129] First, the gridded data of a certain member and a certain meteorological element of the meteorological large model deterministic grid forecast dataset F100 are obtained.

[0130] Then, take the spatial scale separation coefficient vector C filter Subvector C1 corresponding to the gridded data filter .

[0131] Then, take C1 filter The element with the largest value is taken as the first cutoff spatial scale, and a two-dimensional local weighted regression (LOESS) is performed on the gridded data to obtain the low-pass field with a scale greater than the first cutoff spatial scale and the high-pass field with a scale less than the first cutoff spatial scale.

[0132] Then, take C1 filter The element with the second largest value is used as the second truncation space scale, and a two-dimensional local weighted regression (LOESS) is performed on the high-pass field whose scale is smaller than the first truncation space scale to obtain a low-pass field whose scale is larger than the second truncation space scale and a high-pass field whose scale is smaller than the second truncation space scale.

[0133] Then, take C1 from largest to smallest. filter Repeat the above steps to get a set of C1 filter Corresponding spatial scale separated grid data F1 filter .

[0134] Then, repeat the above steps to separate the gridded data of all members and meteorological elements of the meteorological large model dataset F100 in S102 into spatial scales, and obtain C filter Corresponding spatial scale-separated meteorological large model deterministic grid forecast dataset F100 filter .

[0135] Then, mathematical statistical methods such as linear regression or other machine learning methods are used to establish a large-scale meteorological model deterministic grid forecast dataset F100. filter The mapping relationship between the historical prediction data set of a group of sub-data sets and the corresponding measured power in the same period of history; repeat the above steps to traverse the meteorological large model deterministic grid forecast data set F100 filter For each sub-dataset in the dataset, a mapping relationship is established between each sub-dataset and the measured power corresponding to the same period in history; the above-mentioned meteorological large model deterministic grid forecast dataset F100 filter The mapping relationship between all sub-datasets in the dataset constitutes the meteorological large model deterministic grid forecast dataset F100. filter A mapping relationship is established between the historical forecast data set and the corresponding measured power set in the same period of history R100 filter .

[0136] Then, the meteorological large model deterministic grid forecast dataset F100 filter Substitute the future prediction data set of a set of sub-data sets into the mapping relationship R100 filter The corresponding sub-mapping relationship is obtained to obtain the corresponding future predicted power, and the error between the future predicted power and the measured power in the same period is calculated; the above steps are repeated to traverse the meteorological large model deterministic grid forecast data set F100 filter For each group of sub-datasets in the dataset, the future predicted power and predicted power error corresponding to each group of sub-datasets are obtained respectively; the above-mentioned meteorological large model deterministic grid forecast dataset F100 filter The future prediction powers corresponding to all sub-datasets in the , together form a set of future prediction powers P100 filter ; The above-mentioned large meteorological model deterministic grid forecast dataset F100 filter The predicted power errors corresponding to all sub-data sets in the , together constitute a set of future predicted power deviations ERROR100 filter .

[0137] Then, based on the error ERROR100 filter Traverse the optimal F100 filter , the F100 with larger error filter After elimination, the optimal large-scale meteorological model deterministic grid forecast data set D100 is obtained.

[0138] In this embodiment, the model used for forecasting or simulation, in addition to the meteorological large model, such as the Pangu meteorological large model, the Graphcast meteorological large model, the NeuralGCM meteorological large model, the Fuxi meteorological large model, etc., can also use the self-developed meteorological large model, or can use the traditional meteorological model, such as the global atmospheric simulation system Model for Prediction AcrossScales (MPAS) and FiniteVolume Cubed-Sphere dynamical core (FV3), etc., and also include other atmospheric models with simulation capabilities, such as various climate models, ocean models, sea-air coupling models, etc. The use of any mode or model to obtain gridded meteorological data to calculate the localized matrix or perform scale separation is within the protection scope of the present invention.

[0139] The fine-tuning weights used in S102 of this embodiment may use not only Gaussian distribution decreasing weights, but also linear decreasing weights, inverse proportional decreasing weights, exponential decreasing weights, logarithmic decreasing weights, cosine decreasing weights, polynomial decreasing weights, hyperbolic decreasing weights, and sigmoid decreasing weights, and any weight function that can characterize the decrease with distance is within the protection scope of the present invention.

[0140] In the scale separation method S103 in this embodiment, in addition to filtering using the local weighted regression method, smoothing or filtering methods such as sliding average, sliding median, sliding Gaussian weighted average, local weighted quadratic regression, robust local weighted regression, robust local weighted quadratic regression, wavelet analysis method and Savitsky-Gorey smoothing method can also be used. Any method that can be smoothed or filtered for scale separation is within the protection scope of the present invention.

[0141] This embodiment proposes a method for fine-tuning the localized matrix in data assimilation based on a large meteorological model: first, the ensemble forecast and deterministic forecast gridded data of the large meteorological model forecast are produced, then the localized matrix of data assimilation is calculated and fine-tuned based on the ensemble forecast, data assimilation is performed and the large meteorological model is inferred to obtain a set of gridded forecast data, a mapping relationship between the gridded data and the measured power is established to obtain the predicted power, the predicted power error is calculated and the gridded forecast data of the large meteorological model is optimized. This method optimizes the key technical links of the ensemble assimilation calculation of the localized matrix in the traditional numerical weather forecast and data assimilation technology, and takes the new energy power prediction error as the optimization target, organically integrating data assimilation with new energy prediction, which can effectively improve the effects of numerical weather forecast and data assimilation technology, and thus improve the accuracy of new energy meteorological element prediction and power prediction.

[0142] This embodiment proposes a method for scale separation and optimization of a large meteorological model: first, statistical analysis is performed on the power spectrum and correlation between grid points of the grid data of the preferred large meteorological model to obtain a spatial scale separation coefficient, and then, a two-dimensional local weighted regression is recursively performed on the grid data of the preferred large meteorological model according to the spatial scale separation coefficient to obtain a set of scale-separated grid data of the large meteorological model, a mapping relationship between the grid data and the measured power is established to obtain the predicted power, the predicted power error is calculated, and the grid forecast data of the large meteorological model with scale separation is optimized. This method selects the spatial scales and corresponding grid data with smaller errors and better forecast quality in each spatial scale of the large meteorological model, which can effectively remove the noise signal of the prediction of new energy meteorological elements of the large meteorological model, and further improve the accuracy of photovoltaic or wind power prediction with meteorological sources optimized by spatial scale separation as input.

[0143] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A new energy wind and solar power prediction method based on a large meteorological model, characterized in that: include: S1: Create a large meteorological model ensemble forecast dataset ENS0, perform ensemble normalization and forecast to obtain a deterministic forecast dataset F0; S2: Construct a fine-tuning localization matrix, and based on the fine-tuning localization matrix and the meteorological large model ensemble forecast dataset ENS0, traverse and optimize to obtain the meteorological large model deterministic grid forecast dataset F100; S3, performing spatial scale separation on the meteorological large model deterministic grid forecast dataset F100, and traversing and optimizing to obtain the meteorological large model deterministic grid forecast dataset D100; S4. Calculate weights using historical gridded data of the meteorological large model deterministic grid forecast data set D100, calculate power forecast time series using future forecast gridded data of the meteorological large model deterministic grid forecast data set D100, and perform weighted averaging based on the weights to obtain a fused power forecast time series.

2. The method for predicting wind and solar power of new energy based on a large meteorological model according to claim 1 is characterized in that: Step S1 specifically includes: S11, making the ensemble initial field required for the meteorological large model ensemble forecast, adding the initial disturbance on the basis of the deterministic forecast of the meteorological center, using the meteorological large model to make forecasts with the initial field of each ensemble member as input, and obtaining a set of meteorological large model ensemble forecast data sets ENS0; S12. Obtain the available meteorological observation dataset OBS0 of the global assimilation system JEDI; calculate the localization matrix L0 based on the meteorological large model ensemble forecast dataset ENS0, assimilate the meteorological observation dataset OBS0; update the forecast field of the meteorological large model to obtain the analysis field; use the analysis field as the input of the numerical weather forecast or the meteorological large model to obtain the deterministic forecast dataset F0.

3. The method for predicting wind and solar power of new energy based on a large meteorological model according to claim 2 is characterized in that: Step S2 specifically includes: S21. For a submatrix corresponding to a certain meteorological element variable in the localized matrix L0, a fine-tuning coefficient matrix is ​​constructed, whose value is a Gaussian weight centered on the new energy station and decaying toward the periphery; the amplitude and influence range of the traversal adjustment weight are combined to form a set of matrices corresponding to the Gaussian weights, forming a fine-tuning coefficient matrix C tuning ; When adjusting the amplitude and influence range of Gaussian weights by combination traversal, the fine-tuning coefficient matrix composed of each adjustment of Gaussian weights is multiplied element by element with the corresponding localization matrix, and the combination traversal obtains the fine-tuning localization matrix L tuning ; S22. Fine-tuning the localization matrix L tuning The submatrix of all meteorological elements corresponding to a certain combination of amplitude and influence range in the ensemble forecast dataset ENS0 of the meteorological model is collectively normalized and forecasted to obtain the corresponding meteorological model deterministic forecast dataset F1; Traverse all fine-tuning localization matrices L tuning The amplitude and influence range are combined to obtain the fine-tuning localization matrix L tuning The corresponding set of meteorological large model deterministic grid forecast data sets F100 all ; Establish a set of meteorological large model deterministic grid forecast data sets F100 all The first mapping relationship between the historical prediction data set and the measured power corresponding to the same period in history is used to convert the set of meteorological large model deterministic grid forecast data set F100 all Substitute the future prediction data set of the first sub-data set into the first mapping relationship to obtain a set of future prediction power, and calculate the error between the future prediction power and the actual measurement during the same period; traverse the set of meteorological large model deterministic grid forecast data sets F100 all , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set F100 is obtained.

4. The method for predicting wind and solar power of new energy based on a large meteorological model according to claim 3 is characterized in that: Step S3 specifically includes: S31. For each member and each meteorological element of the large-scale deterministic grid forecast data set F100 of the meteorological model, the correlation coefficient between each grid point and other grid points is calculated, and the relationship between the correlation coefficient and the grid point spacing is statistically analyzed to obtain the grid point spacing corresponding to the significant local extreme value of the correlation coefficient greater than the threshold value, which is used as the correlation spatial scale separation coefficient corresponding to the grid data; For the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast dataset F100 of the meteorological model, power spectrum analysis is performed in the longitude and latitude directions respectively, and the spatial scale corresponding to the local extreme value greater than the test value is obtained as the power spectrum spatial scale separation coefficient corresponding to the gridded data; The correlation spatial scale separation coefficient and the power spectrum spatial scale separation coefficient together constitute the spatial scale separation coefficient vector C of the meteorological large model deterministic grid forecast data set F100 filter ; S32, gridded data of each member and each meteorological element of the large-scale deterministic grid forecast data set F100 of the meteorological model, and the spatial scale separation coefficient vector C corresponding to the gridded data filter , and take each element in the subvector from large to small as the cutoff spatial scale to perform two-dimensional local weighted regression on the gridded data, retain the low-pass field and recursively perform two-dimensional local weighted regression on the high-pass field in turn, and obtain a set of C filter Corresponding spatial scale-separated deterministic grid forecast dataset F100 filter ; Establishing the deterministic grid forecast dataset F100 filter The second mapping relationship between the historical prediction data set and the measured power corresponding to the same period in history; the deterministic grid forecast data set F100 filter Substitute the future prediction data set of a sub-data set into the second mapping relationship to obtain a set of future prediction powers, and calculate the error between the obtained future prediction power and the actual measurement during the same period; traverse the deterministic grid forecast data set F100 filter , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set D100 is obtained.

5. A new energy wind and solar power prediction device based on a large meteorological model, characterized in that: include: Dataset module: Create a large meteorological model ensemble forecast dataset ENS0, perform ensemble normalization and forecast to obtain a deterministic forecast dataset F0; Fine-tuning localization module: construct a fine-tuning localization matrix, and based on the fine-tuning localization matrix and the meteorological large model ensemble forecast dataset ENS0, traverse and optimize to obtain the meteorological large model deterministic grid forecast dataset F100; Scale separation module: performing spatial scale separation on the meteorological large model deterministic grid forecast data set F100, and traversing and optimizing to obtain the meteorological large model deterministic grid forecast data set D100; Power prediction module: The weights are calculated using the historical gridded data of the meteorological large model deterministic grid forecast data set D100, and the power prediction time series is calculated using the future forecast gridded data of the meteorological large model deterministic grid forecast data set D100, and weighted averaging is performed according to the weights to obtain a fused power prediction time series.

6. The new energy wind and solar power prediction device based on the meteorological large model according to claim 5 is characterized in that: The dataset modules include: Ensemble forecast data set unit: produces the ensemble initial field required for the meteorological large model ensemble forecast, adds initial disturbances based on the deterministic forecast of the meteorological center, and uses the meteorological large model to make forecasts with the initial field of each ensemble member as input, thus obtaining a set of meteorological large model ensemble forecast data sets ENS0; Deterministic forecast dataset unit: obtain the available meteorological observation dataset OBS0 of the global assimilation system JEDI; calculate the localization matrix L0 based on the meteorological large model ensemble forecast dataset ENS0, assimilate the meteorological observation dataset OBS0; update the forecast field of the meteorological large model to obtain the analysis field; use the analysis field as the input of the numerical weather forecast or the meteorological large model to obtain the deterministic forecast dataset F0.

7. The new energy wind and solar power prediction device based on the meteorological large model according to claim 6 is characterized in that: The fine-tuning localization module includes: Fine-tuning localized matrix unit: for a submatrix corresponding to a meteorological element variable in the localized matrix L0, a fine-tuning coefficient matrix is ​​constructed, whose value is a Gaussian weight centered on the new energy station and decaying toward the periphery; the amplitude and influence range of the weight are combined and traversed to form a set of matrices corresponding to the Gaussian weights, forming a fine-tuning coefficient matrix C tuning ; When adjusting the amplitude and influence range of Gaussian weights by combination traversal, the fine-tuning coefficient matrix composed of each adjustment of Gaussian weights is multiplied element by element with the corresponding localization matrix, and the combination traversal obtains the fine-tuning localization matrix L tuning ; The first traversal optimization unit: for the sub-matrix of all meteorological elements corresponding to a certain combination of amplitude and influence range in the fine-tuning localization matrix, based on the sub-matrix and the meteorological large model ensemble forecast data set ENS0, ensemble assimilation and forecast are performed to obtain the corresponding meteorological large model deterministic forecast data set F1; Traverse all fine-tuning localization matrices L tuning The amplitude and influence range are combined to obtain the fine-tuning localization matrix L tuning The corresponding set of meteorological large model deterministic grid forecast data sets F100 all ; Establish a set of meteorological large model deterministic grid forecast data sets F100 all The first mapping relationship between the historical prediction data set and the measured power corresponding to the same period in history is used to convert the set of meteorological large model deterministic grid forecast data set F100 all Substitute the future prediction data set of the first sub-data set into the first mapping relationship to obtain a set of future prediction power, and calculate the error between the future prediction power and the actual measurement during the same period; traverse the set of meteorological large model deterministic grid forecast data sets F100 all , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set F100 is obtained.

8. The new energy wind and solar power prediction device based on the meteorological large model according to claim 7 is characterized in that: The scale separation module includes: Separation coefficient vector unit: for the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast data set F100 of the meteorological model, the correlation coefficient between each grid point and other grid points is calculated, and the relationship between the correlation coefficient and the grid point spacing is statistically analyzed to obtain the grid point spacing corresponding to the significant local extreme value of the correlation coefficient greater than the threshold value, which is used as the correlation spatial scale separation coefficient corresponding to the gridded data; For the gridded data of each member and each meteorological element of the large-scale deterministic grid forecast dataset F100 of the meteorological model, power spectrum analysis is performed in the longitude and latitude directions respectively, and the spatial scale corresponding to the local extreme value greater than the test value is obtained as the power spectrum spatial scale separation coefficient corresponding to the gridded data; The correlation spatial scale separation coefficient and the power spectrum spatial scale separation coefficient together constitute the spatial scale separation coefficient vector C of the meteorological large model deterministic grid forecast data set F100 filter ; Second traversal optimization unit: gridded data of each member and each meteorological element of the large meteorological model deterministic grid forecast data set F100, and the spatial scale separation coefficient vector C corresponding to the gridded data filter , and take each element in the subvector from large to small as the cutoff spatial scale to perform two-dimensional local weighted regression on the gridded data, retain the low-pass field and recursively perform two-dimensional local weighted regression on the high-pass field in turn, and obtain a set of C filter Corresponding spatial scale-separated deterministic grid forecast dataset F100 filter ; Establishing the deterministic grid forecast dataset F100 filter The second mapping relationship between the historical prediction data set and the measured power corresponding to the same period in history; the deterministic grid forecast data set F100 filter Substitute the future prediction data set of a sub-data set into the second mapping relationship to obtain a set of future prediction powers, and calculate the error between the obtained future prediction power and the actual measurement during the same period; traverse the deterministic grid forecast data set F100 filter , based on the error elimination, the traversal and optimal meteorological large model deterministic grid forecast data set D100 is obtained.

Images