Ensemble forecast multi-source error cooperative disturbance method based on error characteristic evolution and dispersion resonance

Abstract

The invention provides an ensemble forecast multi-source error collaborative disturbance method based on error characteristic evolution and dispersion resonance, which comprises the following steps of: selecting three single error sources of an initial value, a side boundary and a physical process, and generating a disturbance set of a single error source and a multi-error source combination by adopting a disturbance method; calculating ensemble forecast error energy of the single error source and the multiple error sources, decomposing the ensemble forecast error energy into different spatial scales, and analyzing error evolution characteristics of the single error source and the multiple error sources; in combination with an ESA technology, sensitive areas and sensitive elements of single-error-source and multi-error-source disturbance errors are explored, error evolution characteristics of multi-error-source disturbance and single-error-source disturbance are compared, an interaction mechanism between disturbance of different error sources is analyzed, a dispersion resonance model is constructed, and the disturbance of the multi-error-source disturbance and the disturbance of the single error source are calculated. And evaluating the synergistic effect of the multi-error-source combined disturbance in the sensitive area or the sensitive element, and constructing the collaborative disturbance of the multi-source error. According to the invention, a key mechanism of single-source error growth and an interaction mechanism of multi-source errors are disclosed, and the prediction precision of ensemble prediction is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of numerical weather prediction, and particularly relates to a multi-source error collaborative perturbation method for ensemble forecasting based on error characteristic evolution and dispersion resonance. Background Art

[0002] Ensemble forecasting is a method that reflects the uncertainty of weather prediction by generating multiple different forecast results. The core lies in quantifying the probability distribution through multi-member simulations. Ensemble forecasting has the following characteristics: Multi-perturbation generation: introducing small perturbations to the initial conditions, boundary conditions, physical process parameterization schemes, etc., or using different numerical models to generate multiple forecast members; Probabilistic output: the differences between members can reflect the probability distribution of the forecast quantity (such as precipitation probability, typhoon path possibility), rather than a single deterministic conclusion; Coping with uncertainty: covering both initial field errors (such as observation errors, insufficient model initialization) and the randomness of the model physical process (such as physical process parameterization schemes).

[0003] However, in convective-scale forecasting, the error sources are complex (such as initial value, lateral boundary, physical process error sources). Existing technologies are difficult to effectively separate the contributions of different error sources, and lack quantitative analysis of the interaction mechanism of multi-source perturbations. Randomly combined multi-source error perturbation combinations may lead to unreasonable distributions of meteorological elements that do not conform to physical laws or unreasonable time tendencies, affecting the error estimation of ensemble forecasting. Therefore, it is difficult to determine which perturbation combination can make the ensemble dispersion resonance grow, resulting in limitations in the design of ensemble dispersion and poor collaborative effects of multi-source perturbations. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a multi-source error collaborative perturbation method for ensemble forecasting based on error characteristic evolution and dispersion resonance.

[0005] Technical Solution: A multi-source error collaborative perturbation method for ensemble forecasting based on error characteristic evolution and dispersion resonance, comprising the following steps: S1. Set the initial value, lateral boundary, and physical process as single error sources, and respectively generate a perturbation ensemble of single error sources and a perturbation ensemble of multi-error sources by using the perturbation methods of the initial value, lateral boundary, and physical process; S2. Calculate the error energy RMDTE of the single-error-source ensemble forecasting and the multi-error-source ensemble forecasting respectively, decompose the RMDTE on different spatial scales by using a scale separation method, and respectively analyze the error evolution characteristics of the single-error-source perturbation and the multi-error-source perturbation; S3. Use the RMDTE combined with the ensemble sensitivity analysis method to respectively obtain the sensitive regions where the errors of each single-error-source perturbation and the multi-error-source perturbation grow rapidly and the corresponding sensitive elements; S4. Compare the error evolution characteristics, sensitive regions, and differences in sensitive elements between multi-error-source perturbations and single-error-source perturbations, and analyze the interaction mechanism between different single-error-source perturbations in multi-error-source perturbations; S5. Construct a dispersion resonance model to dynamically evaluate the synergistic effect of multi-source perturbation combinations in sensitive regions or sensitive elements. Adjust the perturbation phase according to the evaluation results and the interaction mechanism between different single-error-source perturbations in multi-error-source perturbations, and construct a synergistic perturbation set for multi-error sources.

[0006] Specifically, in step S1, the initial value perturbation methods include the local growth mode breeding method, the singular vector method, rescaled ensemble transformation, and ensemble Kalman filtering; the lateral boundary perturbation method includes dynamic downscaling; the physical process perturbation methods include stochastic physical process parameterization tendency perturbation, stochastic kinetic energy backscatter scheme, and stochastic parameter perturbation.

[0007] Preferably, in step S1, the initial value perturbation method is the local growth mode breeding method; the lateral boundary perturbation method is dynamic downscaling; the physical process perturbation method is stochastic parameter perturbation.

[0008] Specifically, in step S2, the calculation formula for the error energy RMDTE is: , where: is the error energy, is the total error energy, , and respectively represent the differences in zonal wind U, meridional wind V, and temperature T between the perturbed forecast and the unperturbed control forecast; is the specific heat capacity in isobaric dry air; is the reference temperature; i , j represent the horizontal grid points of the model area, t represents the integration time; p is the air pressure corresponding to the model layer, n is the number of ensemble members, l represents the number of model layers used for vertical integration.

[0009] Specifically, in step S2, analyzing the error evolution characteristics of single-error-source perturbations and multi-error-source perturbations includes: based on the perturbation ensemble, analyzing the error evolution characteristics of single-error-source perturbations and multi-error-source perturbations from two aspects of RMDTE and spatial scale, and summarizing the evolution laws of the zonal and meridional propagation, amplitude peaks, and amplitude valleys of the errors of each single-error-source perturbation and multi-error-source perturbation.

[0010] Specifically, step S3 includes: using the RMDTE obtained in step S2, in the peak region of the RMDTE, calculating the results of the ensemble sensitivity analysis of the RMDTE with the 500 hPa geopotential height, 850 hPa wind field, convective available potential energy, and vertical wind shear between high and low altitudes, and identifying the sensitive regions where the perturbation errors of each single error source and multi-error source grow rapidly and the corresponding sensitive elements according to the results of the ensemble sensitivity analysis.

[0011] Specifically, in step S4, analyzing the interaction mechanism between different single error source perturbations in the multi-error source perturbation includes: analyzing the error evolution characteristics, sensitive regions, and differences in sensitive variables between single-source errors and multi-source errors, and establishing the interaction mechanism between different single error source perturbations in the multi-error source perturbation according to the gain or suppression effect brought by the multi-error source perturbation relative to the single error source perturbation.

[0012] Specifically, step S5 includes: Using the perturbation ensemble of single error source and multi-error source combinations, evaluating the synergy effect of the multi-error source combination in the sensitive region or sensitive element by calculating the perturbation ensemble dispersion of the interaction between each single error source perturbation in the multi-error source combination perturbation. If the perturbation ensemble dispersion is less than or equal to 0, adjust the phase of the combined perturbation according to the interaction mechanism between different single error source perturbations in the multi-error source perturbation, and recalculate the perturbation ensemble dispersion until the perturbation ensemble dispersion is greater than 0 to obtain the synergy perturbation ensemble of the multi-error source.

[0013] Specifically, in step S5, the calculation formula for the perturbation ensemble dispersion is: , In the formula: is the perturbation ensemble dispersion, is the multi-error source combination perturbation, and are two different single error source perturbations, is the number of samples.

[0014] Specifically, in step S2, the spatial scales include the medium-γ, medium-β, and medium-α scales.

[0015] Beneficial effects: Compared with the prior art, the remarkable effects of the present invention are as follows: The present invention respectively adopts corresponding perturbation methods for three single error sources, namely the initial value, the lateral boundary, and the physical process, to generate perturbation sets of single error sources and combinations of multiple error sources. Then, by calculating the error energy of the forecasts of single error sources and multiple error source sets, and combining scale separation to decompose the error energy into different spatial scales, the evolution characteristics of the perturbation errors of single error sources and multiple error sources are analyzed from two aspects: the amplitude and structure of the perturbations. Then, combined with the Ensemble Sensitivity Analysis (ESA) technique, the sensitive regions where each single error source grows rapidly and the matching sensitive elements are explored. By comparing the error evolution characteristics of the combined perturbations of multiple error sources and the perturbations of single error sources, the interaction mechanism between different single error source perturbations in the combined perturbations of multiple error sources is analyzed. Finally, a dispersion resonance model is constructed to achieve the purpose of dynamically evaluating the synergistic effect of the combined perturbations of multiple error sources in sensitive regions or sensitive elements. According to the evaluation results, the perturbation phase is adjusted, and finally, a coordinated perturbation of multi-source errors is constructed. Through the multi-scale separation of multi-source errors, ensemble sensitivity analysis, and the dispersion resonance model, the present invention reveals the key mechanism for the growth of single-source errors and the interaction mechanism of multi-source errors, which is beneficial to constructing a reasonable and coordinated coordinated perturbation of multi-source errors, thereby improving the error prediction accuracy and early warning timeliness of the convective-scale ensemble forecasting system for severe convective weather. Brief Description of the Drawings

[0016] Figure 1 is the flowchart of the method of the present invention. Detailed Embodiments

[0017] The following further describes a preferred embodiment of the present invention with reference to the drawings.

[0018] Please refer to Figure 1 As shown, the present invention provides an ensemble forecasting multi-source error coordinated perturbation method based on error characteristic evolution and dispersion resonance, including the following steps: S1. Set the initial value, the lateral boundary, and the physical process as single error sources, and respectively use the perturbation methods of the initial value, the lateral boundary, and the physical process to generate perturbation sets of single error sources and perturbation sets of multiple error sources combined by multiple single error sources. There are 7 groups in total, namely: single initial value perturbation, single lateral boundary perturbation, single physical process perturbation, initial value + lateral boundary perturbation, initial value + physical process perturbation, lateral boundary + physical process perturbation, and initial value + lateral boundary + physical process perturbation.

[0019] The specific perturbation methods adopted are as follows: Initial value perturbation method: Localized Growth Model Breeding Method (LBGM), Singular Vector Method, Re-scaled Ensemble Transform, and Ensemble Kalman Filter (EnKF).

[0020] Lateral boundary perturbation method: Dynamic downscaling.

[0021] Physical process perturbation methods: Stochastic Physical Process Parameterization Tendency Perturbation (SPPT), Stochastic Kinetic Energy Backscatter Scheme (SKEB), and Stochastic Parameter Perturbation (SPP).

[0022] In the present invention, one method is selected for each of the three single error sources respectively.

[0023] For the initial value perturbation method, the local growth mode breeding method is selected, and the specific description is as follows.

[0024] The Breeding Growth Mode (BGM) method first superimposes an arbitrary perturbation on the initial field of the model, integrates the initial field of the control experiment and the initial field after the superimposed perturbation simultaneously. After integrating for a period of time, the control forecast and the perturbation forecast are obtained. The scale of the perturbation is adjusted to the same order of magnitude as the initial perturbation, and then the analyzed perturbation is superimposed on the new atmospheric initial field, and the breeding is repeated continuously until the fastest growing mode required is obtained. The BGM method scales the perturbation at the end of the breeding cycle: , In the formula: c is the adjustment coefficient, related to the model layer k ; represents the forecast perturbation, represents the perturbation after scale adjustment; c(k) is the preset root mean square of the perturbation e 0 (k) is the ratio of e n ( k ) to the root mean square of the forecast perturbation

[0025] For the Local Breeding Growth Mode (LBGM) method, the statistical range of e 0 (k) and e n ( k ) is adjusted from the complete forecast area of the traditional BGM method to the neighborhood area of each grid point, and the neighborhood radius is r , and the scaling factor c(k) is adjusted to: , In the formula: The subscript t represents the perturbation coefficient at t time.

[0026] For the lateral boundary perturbation method, the dynamic downscaling is selected, and the specific description is as follows.

[0027] Dynamic downscaling directly uses the analysis perturbations of the global ensemble forecasting system, interpolates and superimposes them on the initial field of the model, thereby providing initial values and lateral boundary perturbations for regional ensemble forecasting. This method is simple and easy to implement, has a low computational cost, and has good performance.

[0028] For the physical process perturbation method, random parameter perturbation is selected, and the specific description is as follows.

[0029] The random parameter perturbation method characterizes the uncertainty of the model by perturbing the empirical and adjustable parameters in the sub-grid parameterization scheme, and can better reflect the forecasting uncertainty of small-scale systems. The random parameter perturbation method believes that the magnitude of the model error is proportional to the magnitude of the parameter. A random number (usually between -0.5 and 0.5) is multiplied by the uncertain parameter in the parameterization scheme through the multiplier method to construct the perturbation. The general expression is: , In the formula: is the perturbed parameter value, is the unperturbed parameter value, is a two-dimensional random perturbation field with time and space correlation generated by the random mode generator.

[0030] The expansion in the spectral space is: , In the formula: k，l respectively represent the zonal x direction and the meridional y direction wave number components (there are a total of K+1 x-direction wave numbers, y a total of L+1 y-direction wave numbers), and the Fourier mode forms a set of orthogonal basis functions in the rectangular region ( 0 < x < X, 0 < y < Y ).

[0031] Each spectral coefficient evolves according to the following first-order autoregressive equation: , In the formula: is the linear autoregressive coefficient, is the decorrelation time scale, is the model time step; is the noise amplitude dependent on the wave number; is a Gaussian white noise process with a mean of 0 and a standard deviation of 1. The resulting perturbation mode is uniform in space, and its horizontal length scale is L , and the grid perturbation variance is ; is the effective radial wave number, is the spectral variance, generating Gaussian perturbations with zero mean and variance of at each grid point. That is, the perturbation mode depends entirely on three parameters: the grid standard deviation , the length scale , and the decorrelation time . At the same time, since the Gaussian distribution can lead to extremely large values, the range of the random numbers needs to be restricted.

[0032] S2. Calculate the error energy RMDTE of the single error source ensemble prediction and the multi - error source ensemble prediction respectively. RMDTE is the weighted average of the total error energy (DTE) in the vertical direction, and the calculation formula is: , where: is the error energy, is the total error energy, , and represent the differences in the zonal wind U, meridional wind V, and temperature T between the perturbed prediction and the unperturbed control prediction respectively; is the specific heat capacity in the constant - pressure dry air, ; is the reference temperature, ; i , j represent the horizontal grid points of the model area, t represents the integration time; p is the air pressure corresponding to the model layer, n is the number of ensemble members, l represents the number of model layers used for vertical integration.

[0033] Since the occurrence and development of severe convective weather are often affected by the combined influence of environmental factors at different altitudes and different scales, a weighting function is introduced in the vertical direction of calculating DTE for RMDTE, so as to be able to describe the error distribution characteristics of the entire atmosphere layer.

[0034] Then, a scale - separation method, such as the discrete cosine transform, is used to decompose RMDTE at different spatial scales (medium - γ, medium - β, and medium - α) to quantify the perturbations at different scales. Finally, based on the obtained ensemble sample set above, the error evolution characteristics of the single - error - source perturbation and the multi - error - source perturbation are analyzed from two aspects: the perturbation amplitude (RMDTE) and the structure (spatial scale). That is, by analyzing the propagation of RMDTE at different scales in the zonal and meridional directions and its variation with time, the laws of the zonal - meridional propagation and the evolution of the amplitude peaks (valleys) of the single - error - source perturbation and the multi - error - source perturbation are summarized.

[0035] S3. Use the RMDTE combined with the Ensemble Sensitivity Analysis method (ESA) to obtain the sensitive regions and corresponding sensitive elements where the perturbation errors of each single error source and the perturbation errors of multiple error sources grow rapidly. Specifically, in the peak region of the RMDTE, calculate the ensemble sensitivity analysis results of the RMDTE with forecast variables such as the 500 hPa geopotential height, 850 hPa wind field, convective available potential energy, and vertical wind shear between high and low altitudes, and identify the sensitive regions and corresponding sensitive elements where the perturbation errors of each single error source and the perturbation errors of multiple error sources grow rapidly according to the ensemble sensitivity analysis results.

[0036] S4. By comparing the error evolution characteristics of multi-source perturbations and single-source perturbations, explore the relative contribution laws of different single error source perturbations to the temporal and spatial variations of errors. According to the gain or suppression effect brought by multi-error source perturbations compared with single-error source perturbations, establish the interaction mechanism between different single error source perturbations in the multi-error source combination, that is, establish a physical conceptual model of multi-source perturbation interaction to deepen the understanding of the growth of different perturbations in convective-scale weather.

[0037] S5. Construct a dispersion resonance model, and use the perturbation ensembles of single error sources and multi-error source combinations to dynamically evaluate the synergy effect of multi-source perturbation combinations in sensitive regions or sensitive elements, that is, by adjusting the perturbation ensemble dispersion of the interaction between each single error source perturbation in the multi-error source combination perturbation S , so that the perturbation effect of multi-error sources develops in a synergistic direction. The following is the calculation of the perturbation ensemble dispersion: , In the formula: is the perturbation ensemble dispersion, is the multi-error source combination perturbation, and are two different single error source perturbations, is the number of samples.

[0038] If , it means that the multi-error source combination perturbation can make the ensemble dispersion resonate and grow. If , then it is necessary to analyze the perturbation phase differences of different single error sources in the sensitive region according to the error evolution characteristics of single error sources and the interaction mechanism between different single error source perturbations in multi-error source perturbations, and adjust the phase of the combined perturbation by modifying the sign of the random number in the perturbation method so that smoothly changes in the direction of resonant perturbation until the perturbation ensemble dispersion is greater than 0, and a synergistic perturbation ensemble of multi-error sources is obtained.

[0039] The following specifically illustrates the above method in a specific implementation scenario: Application scenario: Ensemble forecasting of isolated convective systems.

[0040] Step 1: Download the ECMWF global ensemble forecast products provided by TIGGE (with a resolution of ) to provide the ensemble analysis perturbations required for the dynamic downscaling method.

[0041] Based on the WRF model, design 7 groups of perturbation ensembles containing single error sources and combinations of multiple error sources. Among them, the initial value perturbation adopts the LBGM method, the lateral boundary perturbation adopts the dynamic downscaling method, and the physical process perturbation adopts the SPP method. Each group of ensembles contains 16 ensemble members.

[0042] Step 2: Calculate the RMDTE of each single error source ensemble forecast and multi-error source ensemble forecast. When calculating, integrate over 30 model levels. Use the discrete cosine transform to decompose the RMDTE on medium γ (2 - 20 km range), medium β (20 - 200 km range), and medium α (200 - 2000 km range) scales. Based on the obtained ensemble sample set, discuss the zonal and meridional propagation characteristics and the characteristics of temporal evolution of various single error source perturbations and multi-error source perturbations from the two perspectives of the obtained RMDTE and different spatial scales.

[0043] Step 3: Use the RMDTE obtained in Step 2. In the peak region of the RMDTE, calculate the ESA results of the RMDTE and forecast variables such as the 500 hPa geopotential height, 850 hPa wind field (U / V), convective available potential energy (CAPE), and vertical shear between high and low altitudes. According to the positive and negative regions and the magnitude values of the sensitivity values of each variable, identify the sensitive regions (i.e., positive and negative large value regions) where the perturbations of each single error source and multi-error source grow rapidly and their matching sensitive elements or diagnostic quantities.

[0044] Step 4: Compare the spatio-temporal evolution laws of the RMDTE of different scales of multi-error sources and single error sources and their sensitive regions for different forecast variables, analyze the differences in the perturbation errors of the two in terms of amplitude, zonal and meridional propagation, as well as sensitive regions and sensitive variables, explore the gain or suppression effects generated by the interaction of multiple single error sources, and explore the interaction mechanism of perturbations from different sources.

[0045] Step 5: To reduce the antagonistic (inhibitory) effect between multi-error source perturbations, adjust the perturbation ensemble dispersion of the above multi-error source combined perturbations according to the evolution laws of each source error . That is, if , then it is necessary to explore the phase difference of the corresponding single error source perturbation according to the error evolution characteristics to adjust the phase of the multi-error source combined perturbation (such as changing the inverse phase to the same phase), and recalculate until the multi-error source combined perturbation causes dispersion resonance ( ). Based on SMaximization principle, optimize the weights and phases of the combined perturbations of multiple error sources to generate a more reasonable set of collaborative perturbations of multiple error sources.

Claims

1. A multi-source error collaborative perturbation method for ensemble forecasting based on error feature evolution and dispersion resonance, characterized in that It includes the following steps: S1. Set the initial value, lateral boundary, and physical process as single error sources, and use the perturbation methods of the initial value, lateral boundary, and physical process respectively to generate the perturbation sets of single error sources and multi-error sources; S2. Calculate the error energy RMDTE of the single error source set prediction and the multi-error source set prediction respectively, decompose the RMDTE at different spatial scales by using the scale separation method, and analyze the error evolution characteristics of the single error source perturbation and the multi-error source perturbation respectively; S3. Use the RMDTE combined with the ensemble sensitivity analysis method to obtain the sensitive regions and corresponding sensitive elements where the errors of each single error source perturbation and multi-error source perturbation grow rapidly respectively; S4. Compare the differences in the error evolution characteristics, sensitive regions, and sensitive elements between the multi-error source perturbation and the single error source perturbation, and analyze the interaction mechanism between different single error source perturbations in the multi-error source perturbation; S5. Construct a dispersion resonance model to dynamically evaluate the synergistic effect of the multi-source perturbation combination in the sensitive region or sensitive element, adjust the perturbation phase according to the evaluation result and the interaction mechanism between different single error source perturbations in the multi-error source perturbation, and construct a synergistic perturbation set of multi-error sources.

2. The multi-source error collaborative perturbation method for ensemble prediction according to claim 1, wherein: In the step S1, the initial value perturbation methods include the local growth mode breeding method, the singular vector method, the re-scaled ensemble transformation, and the ensemble Kalman filter; the lateral boundary perturbation method includes dynamic downscaling; the physical process perturbation methods include the stochastic physical process parameterization tendency perturbation, the stochastic kinetic energy backscatter scheme, and the stochastic parameter perturbation.

3. The multi-source error collaborative perturbation method for ensemble forecasting according to claim 2, characterized in that: In the step S1, the initial value perturbation method is the local growth mode breeding method; the lateral boundary perturbation method is dynamic downscaling; the physical process perturbation method is the stochastic parameter perturbation.

4. The multi-source error collaborative perturbation method for ensemble prediction according to claim 1, wherein: In the step S2, the calculation formula of the error energy RMDTE is: ， wherein: is the error energy, is the total error energy, , and respectively represent the differences in the zonal wind U, meridional wind V, and temperature T between the perturbed forecast and the unperturbed control forecast; is the specific heat capacity in constant-pressure dry air; is the reference temperature; i , j represent the horizontal grid points of the pattern area, t represent the integration time; p is the air pressure corresponding to the pattern layer, n is the number of ensemble members, l represent the number of pattern layers used for vertical integration.

5. The multi-source error collaborative perturbation method for ensemble forecasting according to claim 1, characterized in that: In the step S2, analyzing the error evolution characteristics of the single error source perturbation and the multi-error source perturbation includes: based on the perturbation set, analyzing the error evolution characteristics of the single error source perturbation and the multi-error source perturbation from two aspects of RMDTE and spatial scale, and summarizing the evolution laws of the zonal and meridional propagation, amplitude peaks, and amplitude valleys of the errors of each single error source perturbation and multi-error source perturbation.

6. The multi-source error collaborative perturbation method for ensemble forecasting according to claim 1, wherein: The step S3 includes: using the RMDTE obtained in the step S2, in the peak region of the RMDTE, calculate the ensemble sensitivity analysis results of the RMDTE with the 500 hPa geopotential height, 850 hPa wind field, convective available potential energy, and vertical wind shear between high and low altitudes, and identify the sensitive regions where each single error source and multi-error source grow rapidly and the corresponding sensitive elements according to the ensemble sensitivity analysis results.

7. The multi-source error collaborative perturbation method for ensemble prediction according to claim 1, characterized in that: In the step S4, analyzing the interaction mechanism between different single error source perturbations in the multi-error source perturbation includes: analyzing the differences in the evolution characteristics, sensitive regions, and sensitive variables between the single-source error and the multi-source error, and establishing the interaction mechanism between different single error source perturbations in the multi-error source perturbation according to the gain or suppression effect brought by the multi-error source perturbation relative to the single error source perturbation.

8. The multi-source error collaborative perturbation method for ensemble prediction according to claim 1, characterized in that: The step S5 includes: Using a set of perturbations composed of a single error source and multiple error sources, the synergy effect of the multiple error source combination in the sensitive region or sensitive element is evaluated by calculating the dispersion of the set of perturbations of the interaction between the single error source perturbations in the multiple error source combination perturbation. If the dispersion of the set of perturbations is less than or equal to 0, the phase of the combined perturbation is adjusted according to the interaction mechanism of the different single error source perturbations in the multiple error source perturbation, and the dispersion of the set of perturbations is recalculated until the dispersion of the set of perturbations is greater than 0, and a set of cooperative perturbations of the multiple error sources is obtained.

9. The multi-source error collaborative perturbation method for ensemble prediction according to claim 1, characterized in that: In the step S5, the calculation formula for the dispersion of the set of perturbations is: ， Wherein: is the disturbance set dispersion, is the combined disturbance of multiple error sources, and are two different single error source disturbances, is the number of samples.

10. The multi-source error collaborative perturbation method for ensemble forecasting according to claim 1, wherein: In the step S2, the spatial scales include the medium γ, medium β, and medium α scales.

Images