A statistical call method for local background error covariance based on local weather types
Through the local background error covariance statistical call method based on local weather type, the problem of static and dynamic background error covariance reflecting the current weather situation and local representativeness is solved, and a more accurate numerical weather forecast is achieved. It is in the forecast of strong convection and multi-scale weather systems.
Patent Information
- Application Number
- CN202410951221.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-16
- Publication Date
- 2024-11-22
- Estimated Expiration
- 2044-07-16
AI Technical Summary
The existing static and dynamic background error covariance methods are insufficient in reflecting the current weather situation and local weather types, which affects the accuracy of numerical weather forecasts.
The local background error covariance statistical calling method based on local weather type is used, and the local background error covariance matrix reflecting the error characteristics of different scales is generated through hierarchical clustering and multi-scale decomposition, and the corresponding call is flow-dependently during the assimilation process. error covariance.
Improve the effect of assimilation analysis and the accuracy of numerical forecasts, especially in the forecast of strong convection and multi-scale weather systems, the representativeness and flow dependence of error characteristics are enhanced, and the global static background error matrix is avoided. question.
Smart Images

Figure CN118503616B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of atmospheric science, and in particular to a local background error covariance statistics calling method based on local weather types. Background Art
[0002] Numerical weather forecasting is an important technical support for current meteorological services and decision-making. It provides quantitative information on the future status of meteorological elements. Numerical weather forecasting is an initial boundary value problem, especially short-term forecasting, which requires a higher degree of accuracy of initial conditions. Therefore, how to make accurate initial values becomes the key to the success of numerical weather forecasting.
[0003] At present, the way to make accurate initial forecast values is to use data assimilation technology to fuse the model forecast with actual observations, and use the assimilated analysis field corrected by observations as the initial value. Therefore, the algorithm used in data assimilation and a large number of detailed settings become the key to determining the accuracy and rationality of the initial value.
[0004] The mainstream data assimilation methods today are variational assimilation and ensemble Kalman filter assimilation. One of the common points between the two is that they both use background error covariance. Background error covariance includes the spatial correlation, temporal correlation (four-dimensional) between errors, and the correlation between errors in different forecast variables. Since our observations of the atmosphere cannot cover all space, time, and meteorological elements, we can only propagate the correction information of the observed area to the unobserved area and non-directly observed variables through the background error covariance. Therefore, the accuracy and rationality of the background error covariance become one of the key factors for the assimilation effect.
[0005] The types of background error covariance can be divided into static background error covariance, dynamic background error covariance and mixed background error covariance according to the source of the error samples. Among them, static, dynamic and mixed represent that the error samples come from historical forecasts, current real-time forecasts and a mixture of the former two.
[0006] At present, there are two forms of using static background error covariance. One is to obtain information such as error variance, horizontal characteristic scale, vertical spatial distribution, and balance relationship between variables through statistical error samples in traditional three-dimensional variational calculus, and then reconstruct and use background error covariance in the assimilation process through recursive filtering and other techniques. For example, the NMC method. This method estimates the error through the difference in forecasts of different time periods. For example, for any moment, its error can be estimated by the difference between the 12h forecast and the 24h forecast at that moment. Then, such paired forecasts are performed multiple times over a period of time to form a statistically significant error sample set. Finally, a series of statistical methods are used to extract information such as error variance, error-related characteristic scale, and vertical error spatial distribution between different variables. The traditional static background error covariance has a feature in analyzing increments, that is, the spatial distribution characteristics of errors (Gaussian type) are simulated through recursive filtering, thereby generating an isotropic increment shape. In other words, the incremental spatial distribution generated by single-element observations is similar to concentric circles, and the increment size decreases from the observation point to the surrounding areas. When this property is applied to areas with large changes in meteorological elements, unreasonable analysis results will be generated. For example, on the front of a cold front, there is strong cyclonic shear on both sides of the front, with northerly winds behind the cold front and southwesterly winds in front of the front. If the predicted front position is accurate but weak, and observations behind the front produce an increment that strengthens the northerly wind, then the reasonable increment is to produce an increment that strengthens the southwesterly wind in front of the front. However, according to the isotropic increment characteristics of the static background error covariance, northerly wind increments will also be produced in front of the front. Then the assimilation result is that the cyclonic shear of the predicted front is further weakened, and the position of the front is moved to a slightly downstream position. In this case, the assimilation analysis does not improve the quality of the forecast field. This type of situation is not uncommon in the forecast of severe weather processes, so it is very necessary to improve the background error covariance.
[0007] The other is the simulated static background error covariance. In mixed variational assimilation and ensemble variational assimilation, the background error covariance is simulated by directly using a large number of historical samples. Compared with the standard dynamic background error covariance, the static background error covariance simulated by large samples is several times the number of samples of the dynamic background error covariance. For example, the number of samples of the dynamic background error covariance is usually 30-40, and can reach 80 in a few cases, but the static background error covariance simulated by large samples requires at least more than 100 samples, and sometimes the number of samples can reach 800. The dynamic background error covariance uses real-time ensemble forecast samples, while the samples for simulating the static background error covariance are historical forecast samples. The generation method is similar to the aforementioned NMC method, but unlike the NMC method, which does not consider the real-time weather situation at all, the static background error covariance simulated by large samples usually selects samples similar to the current weather situation to construct a sample set, so it has flow-dependent characteristics and anisotropic spatial correlation structure characteristics to a certain extent. The disadvantage of the static background error covariance simulated by large samples is that the samples are global samples. When the forecast area is large, it is difficult to take care of various local weather types. For example, at the time of analysis, the subtropical high in South China controls the weather-scale trough in North China, and there is a low vortex in the southwest. However, the samples of historical statistics can only match the subtropical high in South China and the trough in North China at the same time. In this case, the background error of large sample simulation will be greatly limited in its ability to characterize the low vortex forecast error in the southwest. In addition, both static sample sampling methods do not consider the correspondence between the error sample and the weather situation at the time of assimilation. Therefore, the main defect of using the static background error covariance is that it is difficult to reasonably characterize the error distribution characteristics of the current weather situation.
[0008] The core difference between the dynamic background error covariance and the static one is that it uses the ensemble forecasting technology to generate a certain number of samples at the time of assimilation, and directly uses these samples to realize the analysis of flow dependence. The dynamic background error covariance is usually used in ensemble Kalman filtering and ensemble variational hybrid assimilation. The advantage is that it can effectively characterize the spatial, temporal and inter-variable distribution relationship of the error when the difference between the forecast and the observation is not too large. The disadvantage of the traditional dynamic background error covariance mainly occurs in the case of large forecast deviations. For example, the convective weather system predicted by the model is very weak, while the observed convection is very strong. Then the dynamic background error covariance cannot reflect the positive correlation between the latent heat release and the upward motion that dominates the middle and upper layers of the strong convection in the convective zone, but will show a relatively random relationship. Then there may be a negative correlation, that is, through the observed radar radial wind, a wind field analysis with enhanced upward motion is obtained, but through the unreasonable dynamic background error covariance, a temperature increment with weakened latent heat release is obtained. Then the analysis field after assimilation has a dynamic incoordination, which will lead to the problem that the forecast error is difficult to decrease rapidly in subsequent forecasts. In another case, when convection is observed but the forecast is false, since no member predicts a convective system, the ensemble forecast hydrometeoric dispersion is 0 in the observed convective area, resulting in no increment in the observed convective area, which affects the effect of assimilation analysis. Since the predictability of severe convective weather systems is relatively low, the above situation often occurs in actual forecasts. Therefore, dealing with the situation when the dynamic background error covariance is not representative enough has become a problem that must be solved in data assimilation.
[0009] In summary, the advantage of the traditional static background error covariance is that it can reflect the large-scale error correlation characteristics of all weather processes over a long period of time, but the disadvantage is that it cannot reflect the error correlation structure corresponding to the current weather situation in real time, so it is difficult to obtain targeted assimilation analysis results, which affects subsequent forecasts. The advantage of the traditional dynamic background error covariance is that if there is no systematic deviation between the mean of the ensemble forecast and the observation (the difference between the two has a mean of 0 in the statistical distribution), then the dynamic background error covariance estimated by the ensemble forecast can effectively characterize the real-time characteristics of the error, but the disadvantage is that if there is a systematic deviation, such as the overall intensity of the weather system predicted by the model is weak, or the position variation is large, then the dynamic background error covariance estimated by the ensemble forecast will give an incorrect error correlation relationship. The existing method of simulating the static background error covariance through a large sample set can find forecast error samples under historically similar weather situations when the forecast area is small, but the disadvantage is that when the forecast area is large, it can often only consider the similarity of the key focus areas, and the representativeness of the remaining areas is insufficient. Summary of the invention
[0010] Purpose of the invention: The purpose of the present invention is to address the deficiencies of the prior art and provide a method for statistically calling local background error covariance based on local weather types. The background error covariance corresponding to different local weather types is statistically calculated through historical samples and called in a flow-dependent form, thereby improving the representativeness of the static background error covariance in each area of the model, providing reliable support for improving the assimilation analysis effect and ultimately improving the accuracy of numerical forecasts.
[0011] Technical solution: The local background error covariance statistics calling method based on local weather types of the present invention comprises the following steps:
[0012] S1. Determine the local grid size and resolution of large, medium and small scales according to the grid resolution and region size of the forecast model;
[0013] S2, using a period of reanalysis data and other assimilated data, and filtering them based on the meteorological scale classification to separate the large-scale and mesoscale components;
[0014] S3, using large-scale and mesoscale local grids to sample local weather types for the large-scale and mesoscale parts of S2, respectively, to form original unclassified local type samples;
[0015] S4, using hierarchical clustering to perform bottom-up cluster analysis on the local type samples obtained in S3, and obtain a hierarchical clustering tree that meets the spatial correlation coefficient as the reference standard;
[0016] S5. Based on the classification results of S4, model forecasts are carried out in the same time period as the reanalysis data of S2, using time lag ensemble and model perturbation ensemble methods to obtain large-scale and medium-scale local error samples;
[0017] S6, based on the classification results of S4 and the error samples obtained in S5, calculate the error variance and reliability scale of each model variable under each type, and calculate the fitting function of the local background error covariance matrix of each type;
[0018] S7, using the time lag ensemble, multi-physical process ensemble, physical process parameter and tendency random perturbation ensemble, simulate the period when convection occurs in the S2 reanalysis data, and use small-scale local grids to sample the simulated samples; then use the methods of S4 to S6 to perform cluster analysis on each convection scale sample, and perform error variance and reliable correlation coefficient scale statistics on the ensemble forecast samples corresponding to each type, and calculate the corresponding local background error covariance matrix fitting function;
[0019] S8. In actual forecasting, the classification in S4 and S7 is used to identify the large-scale, medium-scale and small-scale local weather types at each grid point in the forecast area and mark them at the corresponding grid points;
[0020] S9. In the local assimilation system, the background error covariance corresponding to the identified local weather type is called, and assimilation calculation is performed to obtain the corresponding assimilation analysis field.
[0021] A further preferred technical solution of the present invention is that in step S1, N x ×N y The large-scale local grid is N lx ×N ly , the mesoscale local grid is N mx ×N my , the small-scale local grid is N sx ×N sy .
[0022] Preferably, in step S2, reanalysis data and other assimilated data for a period of time are used, and the data are filtered based on the meteorological scale classification to separate the large-scale and medium-scale parts. The specific method is as follows:
[0023] S21. Determine the statistical period and download the reanalysis data set X = (x1, x2, x3, ..., x n ), where x n It means that the data set has a total of n samples x at different times, and x contains information about the east-west wind field component u, the north-south wind field component v, temperature T, specific humidity q, and air pressure p variables;
[0024] S22, use bandpass filtering or Fourier decomposition to divide X into large-scale X L and mesoscale X M Two parts, where the lower limit of the large-scale wavelength is set to 500km-1000km as needed, the upper limit of the mesoscale band is consistent with the lower limit of the large-scale wavelength, and the wavelength of the mesoscale lower limit is set to 30km-100km as needed;
[0025] S23, for the same model forecast data X as S21 period f =(x f 1,x f 2,x f 3,…,x f n ), decompose it into large-scale X in the same way as S22 f L and mesoscale X f M Two parts.
[0026] Preferably, in step S3, the specific method of using large-scale and medium-scale local grids to respectively perform local weather type sampling on the large-scale and medium-scale parts in S2 is:
[0027] S31. Get X L A data file of a time period;
[0028] S32, use N lx ×N ly The large-scale local grid of traverses each grid point, each time including the model height layer of the current grid point and the n above z Layer and below z Layer, n z The size of determines the thickness of the local grid;
[0029] S33. For the grid point (i, j, k) close to the boundary, that is, taking this point as the center, it is impossible to obtain N completely. lx ×N ly The western boundary of the local grid is defined as Max(1, iN lx / 2), the eastern boundary is min(N x , i+N lx / 2), the southern boundary is Max(1, jN ly / 2), the northern boundary is min(Ny,j+N ly / 2); the upper and lower boundaries are processed in the same way, that is, the size of the local grid is cut off according to the boundary, and the upper boundary is min(N z , k+n z ), the lower boundary is max(1, kn z ), where N z is the total number of vertical layers in the mode area determined by S1;
[0030] S34, through the method of steps S31 to S33, traverse X L All the data of the time are collected and stored;
[0031] S35, continue to process X through the method of steps S31 to S34 M , X f L and X f M data and store it.
[0032] Preferably, step S4 uses hierarchical clustering to perform bottom-up cluster analysis on local type samples, and the specific method is:
[0033] S41, standardize the local type samples obtained and remove the dimensions of each variable;
[0034] S42, use the spatial correlation coefficient, the maximum distance between clusters, the minimum distance, the average distance of the cluster or the center distance as one of the judgment criteria for all X L Conduct hierarchical cluster analysis on local type samples and record the spatial correlation coefficients at each level;
[0035] S43, cluster analysis ends when the spatial correlation coefficient of clusterable clusters is lower than a certain value. The correlation coefficient threshold is adjusted up and down according to computing resources and actual application scenarios, and the minimum is not less than 0.5;
[0036] S44. For a type that contains more than 3 original samples, the first three eigenvectors and eigenvalues of the sample set are calculated as the type representative. If it contains 2 original samples, the average value of the samples is used as the type representative for later clustering and type identification.
[0037] S45, using different judgment criteria, repeating steps S42 to S44, and performing type testing with the labeled results, and finally selecting a suitable judgment criterion to complete the final clustering;
[0038] S46. Based on the final clustering results, generate X L Cluster tree of local weather types;
[0039] S47, for X M , repeat steps S42 to S45 to form a classification and clustering tree of mesoscale local weather types;
[0040] S48. For the types at the boundary, cluster analysis is performed only on samples at the same position. The other operation steps are the same as S42 to S47.
[0041] Preferably, the specific method for obtaining large-scale and medium-scale local error samples in step S5 is:
[0042] S51. For the period selected in step S2, use the model forecast to perform ensemble forecast simulation, including time-lagged ensemble forecast, multi-physical process ensemble forecast, random physical process parameterization tendency and parameter perturbation ensemble. The number of ensemble forecast members is set according to the actual situation and computing and storage resource conditions, but the model forecast samples of the reanalysis data at each time of the period selected in step S2 should be no less than 30.
[0043] S52, using the same local sample sampling method as step S3, perform large-scale and medium-scale local sampling on all model forecast results, and mark the corresponding relationship between each local forecast sample and the reanalysis data;
[0044] S53. For each large-scale and medium-scale type classified in step S4, calculate the difference between the original type samples contained in each type and its corresponding forecast.
[0045] Preferably, in step S6, the fitting function of the local background error covariance matrix of each type is calculated, and the specific method is:
[0046] S61. For any local weather type obtained in step S4, calculate the standard deviation of each variable at each grid point on the corresponding local grid;
[0047] S62, calculating the correlation coefficients of each variable at each point of the local weather type and each variable at other points, and establishing a local background error correlation coefficient matrix;
[0048] S63, calculating the β value of each element of the background error correlation coefficient matrix according to the following formula, and multiplying it by the corresponding single element;
[0049]
[0050] Where N ρ is the number of samples, ρ is the correlation coefficient, d is the difference between the forecast and the observation, subscripts 1 and 2 represent two points with a distance r, I() is the consistency function of the correlation coefficient and the difference between the forecast and the observation, and the meaning of I() is that if the error signs of the variables at the two positions are opposite, the error correlation coefficient must be negative to update and correct the pattern variable at the second position through the observation at the first position. If the correlation coefficient is positive at this time, the error will inevitably increase, which is defined as unreasonable.
[0051] S64. Calculate the fitting coefficient of the fitting function of the obtained correlation coefficient matrix according to the following formula:
[0052]
[0053] where a pq are the coefficients of longitude (x), latitude (y), and altitude (z), n, m, and l are the orders of the polynomials, ρ is the statistical correlation coefficient, p is the number of equations, and q = (k-1)mn+(j-1)n+i;
[0054] The coefficients of each order are obtained by solving the following equation:
[0055]
[0056] S65. For all local weather types in S4, repeat steps S61 to S64 to obtain all fitting coefficients.
[0057] Preferably, step S7 is to establish and classify small-scale local weather types, and the specific method is as follows:
[0058] S71, select the time with precipitation in step S2, filter it, remove the part other than the small scale, and set the upper limit of the small scale wavelength to the lower limit of the medium scale wavelength, which is equivalent to the small scale part X f S =X f -X f L -X f M ;
[0059] S72. Use local sample collection method, using small-scale local grid N sx ×N sy Extract small-scale local weather type samples from all forecasts. The small-scale local weather samples contain cloud microphysical variables. The boundary processing method is the same as that in step S3.
[0060] S73, using a cluster analysis method to perform a hierarchical cluster analysis on the obtained small-scale weather samples to form small-scale types and cluster trees;
[0061] S74, based on the method of marking sample time, time-mark each small-scale sample, and group the samples with the same time;
[0062] S75. For each group of small-scale samples assigned, calculate the error between any two samples;
[0063] S76. Calculate the standard deviation and error correlation coefficient matrix of each grid point of each small-scale local weather type, as well as the β value, and then calculate the fitting function coefficient of the correlation coefficient matrix.
[0064] Preferably, in step S8, the large-scale, medium-scale and small-scale local weather types of each grid point in the forecast area are identified and marked at the corresponding grid point. The specific method is:
[0065] S81. Decompose the current forecast field into three parts: large scale, medium scale and small scale;
[0066] S82. Sample large-scale, medium-scale and small-scale local weather types in the forecast field;
[0067] S83. For any grid point that is not on the boundary, that is, contains the complete N lx ×N ly , N mx ×N my or N sx ×N syFor the area of the local grid, calculate its similarity with each type in the cluster tree of the corresponding scale, take the type with the highest similarity to the predicted local weather situation in the cluster tree as the type of the local weather environment where the grid point is located, and mark the type label in the type cluster tree on the grid point. For the small-scale local weather type, detect whether there is a large deviation between the convection information on the current local grid and the observed convection information. If there is a large deviation, retrieve the type with the highest consistency between the convection information and the observed information in the cluster tree, and mark the type on the grid point;
[0068] S84. For the grid points at the boundary, only the types at the same area position in the clustering tree are compared, and the results are recorded at the grid points.
[0069] Preferably, step S9 calls the background error covariance corresponding to the identified local weather type in the local assimilation system, and uses the Local DA algorithm to perform assimilation calculation. The specific method is:
[0070] S91, after the Local DA enters the update process of a certain model grid point variable, retrieve the recorded local weather type labels of the grid point at large scale, medium scale and small scale;
[0071] S92, according to the local weather type labels of each scale, reading the standard deviation and fitting function coefficient of each variable on the local weather grid corresponding to the recorded label;
[0072] S93, based on the root mean square of the estimated background error covariance Calculate the following formula to calculate ρ○C correl Each element of:
[0073]
[0074] Where S is X o The diagonal matrix of the mean square error of each model variable is included, ρ is the localization coefficient matrix, and the calculation formula is:
[0075]
[0076] where a pq are the coefficients of longitude x, latitude y, and altitude z, n, m, and l are the orders of the polynomials, ρ is the statistical correlation coefficient, p is the number of equations, q = (k-1)mn+(j-1)n+i;
[0077] Its adjustment factor Where N is X o The number of variables included, trace() represents the trace of the calculated matrix, The trace of the matrix is calculated by The sum of the squares of all elements of is obtained;
[0078] S94, adding the calculated correlation coefficient matrices of each scale;
[0079] S95. Call the standard deviation of each scale, convert it into variance and find the average of the sum of the variances of each scale, then square it to get the mean square error of each scale, and then use it to replace the corresponding elements of S, and complete Calculation.
[0080] Beneficial effects: The present invention aims to improve the situation that the traditional static background error covariance cannot adapt to the real-time weather conditions and the dynamic background error covariance has large deviations. The present invention is based on local weather types and can provide each grid point with error-related information reflecting the weather system in which it is located, so it is not affected by the size of the forecast area; it should be noted that the present invention does not expect to produce improvements when the mean of the ensemble forecast is not much different from the observation.
[0081] The present invention solves the problem that the global static background error covariance matrix is too large to be directly stored, and realizes the storage of the explicit background error covariance matrix in a "local" form, avoiding the representative problem caused by reconstructing the background error covariance through a large number of assumptions.
[0082] The present invention takes into account the multi-scale properties of errors and can effectively reflect the spatial distribution of errors of different scales and the relationship between variables compared to the traditional single-scale static background error covariance.
[0083] The present invention classifies error statistics for different weather types and geographical areas, and can better reflect the forecast error characteristics under various weather types compared to the traditional global average or latitude average static background error covariance.
[0084] In the assimilation process, the present invention can call the forecast error and covariance of the type matching or similar to the actual local weather type in the database to realize the flow-dependent application of the static background error covariance.
[0085] The present invention can continuously add different local weather types to the database, and by comparing with the local weather types in the existing clustering tree, expand new types or merge existing types, thereby realizing dynamic updating of the local weather type database. BRIEF DESCRIPTION OF THE DRAWINGS
[0086] Figure 1 Schematic diagram of hierarchical clustering scheme based on spatial correlation coefficient of local weather types, where the vertical axis is the spatial correlation coefficient, the horizontal axis is the local weather type, and A to H represent different local weather types respectively;
[0087] Figure 2Schematic diagram of the model layer of local weather type classification in the vertical direction, where k represents a certain model layer;
[0088] Figure 3 Schematic diagram of the scheme for generating weather scale error samples for a single time period;
[0089] Figure 4 Schematic diagram of the scheme for generating convective scale error samples for a single time period;
[0090] Figure 5 Schematic diagram of local weather type classification, error characteristic statistics and assimilation application process of quasi-static multi-scale background error covariance. DETAILED DESCRIPTION
[0091] The technical solution of the present invention is described in detail below with reference to the accompanying drawings, but the protection scope of the present invention is not limited to the embodiments.
[0092] Embodiment: A local background error covariance statistics calling method based on local weather type.
[0093] The background error covariance used in data assimilation methods is mainly divided into two types: static background error covariance and dynamic background error covariance. There are significant differences between the two in terms of sample composition, statistical methods, and application methods. The mixed background error covariance balances the advantages and disadvantages of dynamic and static covariances, making the assimilation analysis results statistically significantly better than the worst case of using static or dynamic alone.
[0094] It should be pointed out that for the dynamic background error covariance, the impact of mixing the static background error covariance is 1) reducing its negative impact when the dynamic background error covariance deviation is large; 2) reducing its positive effect when the dynamic background error covariance quality is high. The above situation is particularly common when severe convection occurs, mainly because the static background error covariance cannot reflect the real-time error characteristics of severe convection. Therefore, for the mixed background error covariance, there is room for improvement in its static part, that is, considering different weather conditions in the statistics and application of the static background error covariance to enhance its representativeness. For the traditional variational ensemble assimilation framework, due to its global statistical characteristics, the existing flow-dependent application of the static background error covariance can only have good error representativeness locally, because the same weather situation in a large range is difficult to repeat in large quantities. The local ensemble assimilation method represented by LETKF and Local DA provides a possible solution to the above problem, that is, developing local background error covariance based on various weather types and applying it according to the specific weather type in a local area at the time of assimilation analysis.
[0095] In addition, due to the obvious differences in the forecast error characteristics and spatial scales of weather systems of different scales, for example, the error correlation between the rising motion and horizontal airflow in severe convection is significantly different from the correlation between the rising motion and horizontal motion in the temperate cyclone of the weather scale. The former mostly corresponds to the convergence and divergence of horizontal airflow, while the latter mostly corresponds to the strength of cyclonic shear. Moreover, the horizontal error correlation scale in the severe convection area is mostly in the range of a few kilometers to more than ten kilometers, while the error correlation of the weather scale system may reach the order of hundreds of kilometers. Therefore, considering multi-scale factors in the background error covariance will more accurately and reasonably characterize the characteristics of the actual error. Existing multi-scale assimilation all considers the problem of sub-scale localization. For example, when assimilating radar data, a smaller localization scale is used, and when assimilating sounding data in clear sky areas, a larger localization distance is used. At the same time, sub-scale statistics and application of error samples have also been a research hotspot in recent years. The advantage of this method is that it no longer needs to manually distinguish the error correlation scales of different regions, but directly determines the variance and related scales of each scale through sample statistics.
[0096] The design of the present invention is applicable to the local range of weather types at different scales, classifies the local weather types at each scale, and then counts the background error covariance corresponding to each local weather type, so as to realize the flow-dependent application of the local static background error covariance in local assimilation algorithms such as Local DA, so that the error feature representativeness of the static background error covariance in assimilation is improved, thereby improving the effect of assimilation analysis, and finally achieving more accurate numerical weather forecasting.
[0097] The local background error statistics and applications involved in the present invention include two parts: type-based statistics and flow-dependent applications. Therefore, the statistics and application techniques of the above two background error covariances are involved. In addition, since the present invention also involves error statistics of different scales, it also involves some techniques of multi-scale background error covariance. Currently, the assimilation method that can be applied to the present invention is Local DA.
[0098] This embodiment provides a method for calling local background error covariance statistics based on local weather types, and the steps are as follows:
[0099] The local background error covariance statistics calling method based on the local weather type includes the following steps:
[0100] S1. According to the grid resolution and area size (N x and N y ), determine the "local" size of large scale, medium scale and small scale (large scale grid number: N lx and N ly , the number of mesoscale grids: N mx and N my , number of small-scale grids: N sxand N sy , by default, they are all set to 11×11, must be an odd number, and can be set to as needed) and resolution (for example: large-scale grid resolution 50km, medium-scale grid resolution 12km, small-scale grid resolution 3km), where the forecast mode can be selected as WRF mode.
[0101] First, we need to define local areas at different scales. For the local assimilation analysis method, "local area" is, in principle, an area defined within a certain radius centered on the current analysis point. In this embodiment, the specific content of "local area" is an area centered on the current analysis point and containing N lx ×N lx Grid points, where N lx and N lx are the number of points in the latitudinal and longitudinal directions respectively. The scope of "local" varies for different scales. Considering the characteristics of the weather scale system with a large range but slow changes, the area corresponding to "local" is larger, but the resolution is relatively low. The "local" corresponding to the convective scale is relatively small in scope but has a relatively high resolution.
[0102] Assume N lx ×N lx The weather scale is 11×11, and the weather scale range corresponds to 1000km, so the resolution of the weather scale "local" is about 100km. Taking into account the particularity of the convective system, its "local" is defined in the area with "clouds" and its surroundings, determined by information such as satellites, radars, or model hydrometeors. In addition, each scale has its corresponding upper and lower limits, so the information contained in the "local" actually corresponds to information within a scale interval. How to select this scale interval and the size of the "local" so that the weather type finally separated is representative and physically meaningful depends on the application scenario. Usually, large-scale information can be determined by reanalysis data such as FNL and ERA5. Therefore, for the weather scale, the "local" can correspond to an area with a length and width of 500km-1000km, and the grid resolution ranges from 50km to 100km, which is about 2-4 times the grid spacing of the 0.25° or 0.5° resolution of the reanalysis data. For the mesoscale "local", the corresponding resolution can be increased to about 12km, which is 4 times the resolution of the domestic CMA-MESO operational forecast and assimilation analysis, and close to the resolution of the national operational mesoscale ensemble forecast, which can reasonably characterize the local mesoscale weather system. Moreover, at the above resolution, with reanalysis data or assimilated data that can be used as a benchmark or "true value", the forecast error information can be estimated more accurately. For the small-scale "local", the resolution is about 1km, which can cover small-scale systems such as tornado-related mesocyclones and gust fronts.
[0103] Through three layers of "local" construction at different scales, the error characteristics of different scales can be captured more reasonably, improving the representativeness of the background error covariance.
[0104] S2. Use reanalysis data (FNL or ERA5) over a period of time and other assimilated data (such as the assimilated analysis field of CMA-MESO) to filter them based on the meteorological scale classification and separate the large-scale and mesoscale parts.
[0105] The specific method is:
[0106] S21. Determine the statistical period and download the reanalysis data set X = (x1, x2, x3, ..., x n ), where x n It means that the data set has a total of n samples x at different times, and x contains information about the east-west wind field component u, the north-south wind field component v, temperature T, specific humidity q, and air pressure p variables;
[0107] S22, use bandpass filtering or Fourier decomposition to divide X into large-scale X L and mesoscale X M Two parts, where the lower limit of the large-scale wavelength is set to 500km-1000km as needed, the upper limit of the mesoscale band is consistent with the lower limit of the large-scale wavelength, and the wavelength of the mesoscale lower limit is set to 30km-100km as needed;
[0108] S23, for the same model forecast data X as S21 period f =(x f 1,x f 2,x f 3,…,x f n ), decompose it into large-scale X in the same way as S22 f L and mesoscale X f M Two parts.
[0109] S3, using large-scale and medium-scale local grids to sample local weather types for large-scale and medium-scale parts in S2, respectively, to form original unclassified local type samples; the specific method of using large-scale and medium-scale local grids to sample local weather types for large-scale and medium-scale parts in S2 in step S3 is:
[0110] S31. Get X L A data file of a time period;
[0111] S32, use N lx ×N lyThe large-scale local grid of traverses each grid point, each time including the model height layer of the current grid point and the n above z Layer and below z Layer, n z The size of determines the thickness of the local grid;
[0112] S33. For the grid point (i, j, k) close to the boundary, that is, taking this point as the center, it is impossible to obtain N completely. lx ×N ly The western boundary of the local grid is defined as Max(1, iN lx / 2), the eastern boundary is min(N x , i+N lx / 2), the southern boundary is Max(1, jN ly / 2), the northern boundary is min(Ny,j+N ly / 2); the upper and lower boundaries are processed in the same way, that is, the size of the local grid is cut off according to the boundary, and the upper boundary is min(N z , k+n z ), the lower boundary is max(1, kn z ), where N z is the total number of vertical layers in the mode area determined by S1;
[0113] S34, through the method of steps S31 to S33, traverse X L All the data of the time are collected and stored;
[0114] S35, continue to process X through the method of steps S31 to S34 M , X f L and X f M data and store it.
[0115] S4, using hierarchical clustering to perform bottom-up cluster analysis on the local type samples obtained in S3, and obtain a hierarchical clustering tree that meets the spatial correlation coefficient as the reference standard;
[0116] The classification of local weather types is achieved through a hierarchical clustering method (bottom-up), the purpose of which is to match the local weather types and the background error covariance with the actual and forecast weather types as closely as possible. Based on the spatial correlation coefficients between different local weather types, a tree structure of local weather types can be established to form a type database of local weather types, such as Figure 1As shown. Based on this database, the numerical model forecast or observation sample is matched with the weather type by the preset correlation coefficient. For example, when the preset correlation degree is 80%, the type with the highest correlation coefficient with the forecast sample and exceeding 80% is searched from the level in the database where the corresponding correlation coefficient is greater than 80%. If the correlation does not exceed 80%, the correlation coefficient threshold is lowered for searching until the correlation coefficient between the sample and the database record type matches the preset threshold. The last layer of bottom-up hierarchical clustering contains all the original samples. At this time, the corresponding background error covariance is a completely static type, which is similar to the static covariance properties used in traditional three-dimensional variational methods.
[0117] Defining the elements included in the local weather type is also the key to statistics. Considering the three-dimensional structure of the atmosphere, classification statistics are performed based on multiple model layers. That is, assuming that there are Nz layers in the vertical direction of the model area, classification statistics of the Nz layer are required. Specifically, when the kth layer of the statistical model is used, the temperature, pressure, wind, and humidity of the nz layers above and below the layer (u, v, T, qv, p) are included. Figure 2 As shown. This approach is beneficial to the continuity of weather types between layers in the final classification, because only one layer of the statistical samples of the upper and lower layers is different. The reduction in computational cost can be achieved by taking into account the similarity of samples in two adjacent layers. The specific value of nz and whether a unified value is used for each layer depends on specific needs. Usually, under the condition of sufficient computing power, statistics are performed on each model layer including the nz=2 layers above and below it. In view of the characteristics of vertical development of convective systems, in the convective zone, the model variables such as temperature, pressure, wind, and humidity on the entire air column are counted, including the vertical velocity w and the mixing ratio of hydrometeors such as rain (qr), snow (qs), and sleet (qg).
[0118] Step S4 uses hierarchical clustering to perform bottom-up cluster analysis on local type samples. The specific process is as follows: Figure 3 The method is:
[0119] S41, standardize the local type samples obtained and remove the dimensions of each variable;
[0120] S42, use the spatial correlation coefficient, the maximum distance between clusters, the minimum distance, the average distance of the cluster or the center distance as one of the judgment criteria for all X L Conduct hierarchical cluster analysis on local type samples and record the spatial correlation coefficients at each level;
[0121] S43, cluster analysis ends when the spatial correlation coefficient of clusterable clusters is lower than a certain value. The correlation coefficient threshold is adjusted up and down according to computing resources and actual application scenarios, and the minimum is not less than 0.5;
[0122] S44. For a type that contains more than 3 original samples, the first three eigenvectors and eigenvalues of the sample set are calculated as the type representative. If it contains 2 original samples, the average value of the samples is used as the type representative for later clustering and type identification.
[0123] S45, using different judgment criteria, repeating steps S42 to S44, and performing type testing with the labeled results, and finally selecting a suitable judgment criterion to complete the final clustering;
[0124] S46. Based on the final clustering results, generate X L Cluster tree of local weather types;
[0125] S47, for X M , repeat steps S42 to S45 to form a classification and clustering tree of mesoscale local weather types;
[0126] S48. For the types at the boundary, cluster analysis is performed only on samples at the same position. The other operation steps are the same as S42 to S47.
[0127] The generation method of the convection scale error samples is as follows: Figure 4 Similar to the weather scale scheme, in order to increase the types of samples as much as possible, it is necessary to include cases of various weather processes and different seasons.
[0128] S5. Based on the classification results of S4, model forecasts are performed in the same time period as the reanalysis data of S2, and large-scale and medium-scale local error samples are obtained using a time lag set and a model perturbation set. The specific method for obtaining large-scale and medium-scale local error samples in step S5 is:
[0129] S51. For the period selected in step S2, use the model forecast to perform ensemble forecast simulation, including time-lagged ensemble forecast, multi-physical process ensemble forecast, random physical process parameterization tendency and parameter perturbation ensemble. The number of ensemble forecast members is set according to the actual situation and computing and storage resource conditions, but the model forecast samples of the reanalysis data at each time of the period selected in step S2 should be no less than 30.
[0130] S52, using the same local sample sampling method as step S3, perform large-scale and medium-scale local sampling on all model forecast results, and mark the corresponding relationship between each local forecast sample and the reanalysis data;
[0131] S53. For each large-scale and medium-scale type classified in step S4, calculate the difference between the original type samples contained in each type and its corresponding forecast.
[0132] S6, based on the classification result of S4 and the error samples obtained in S5, the error variance and reliability scale of each pattern variable under each type are calculated, and the fitting function of the local background error covariance matrix of each type is calculated; the fitting function of the local background error covariance matrix of each type is calculated in step S6, and the specific method is:
[0133] S61. For any local weather type obtained in step S4, calculate the standard deviation of each variable at each grid point on the corresponding local grid;
[0134] S62, calculating the correlation coefficients of each variable at each point of the local weather type and each variable at other points, and establishing a local background error correlation coefficient matrix;
[0135] S63, calculating the β value of each element of the background error correlation coefficient matrix according to the following formula, and multiplying it by the corresponding single element;
[0136]
[0137] Where N ρ is the number of samples, ρ is the correlation coefficient, d is the difference between the forecast and the observation, subscripts 1 and 2 represent two points with a distance r, I() is the consistency function of the correlation coefficient and the difference between the forecast and the observation, and the meaning of I() is that if the error signs of the variables at the two positions are opposite, the error correlation coefficient must be negative to update and correct the pattern variable at the second position through the observation at the first position. If the correlation coefficient is positive at this time, the error will inevitably increase, which is defined as unreasonable.
[0138] S64. Calculate the fitting coefficient of the fitting function of the obtained correlation coefficient matrix according to the following formula:
[0139]
[0140] where a pq are the coefficients of longitude (x), latitude (y), and altitude (z), n, m, and l are the orders of the polynomials, ρ is the statistical correlation coefficient, p is the number of equations, and q = (k-1)mn+(j-1)n+i;
[0141] The coefficients of each order are obtained by solving the following equation:
[0142]
[0143] S65. For all local weather types in S4, repeat steps S61 to S64 to obtain all fitting coefficients.
[0144] S7, using the time lag set, multi-physical process set, physical process parameters and tendency random disturbance set method, simulate the period when convection occurs in the S2 reanalysis data, and use small-scale local grids to sample the simulated samples; then use the methods of S4 to S6 to perform cluster analysis on each convection scale sample, and perform error variance and reliable correlation coefficient scale statistics on the ensemble forecast samples corresponding to each type, and calculate the corresponding local background error covariance matrix fitting function; step S7 is the establishment and classification of small-scale local weather types, and the specific method is:
[0145] S71, select the time with precipitation in step S2, filter it, remove the part other than the small scale, and set the upper limit of the small scale wavelength to the lower limit of the medium scale wavelength, which is equivalent to the small scale part X f S =X f -X f L -X f M ;
[0146] S72. Use local sample collection method, using small-scale local grid N sx ×N sy Extract small-scale local weather type samples from all forecasts. The small-scale local weather samples contain cloud microphysical variables. The boundary processing method is the same as that in step S3.
[0147] S73, using a cluster analysis method to perform a hierarchical cluster analysis on the obtained small-scale weather samples to form small-scale types and cluster trees;
[0148] S74, based on the method of marking sample time, time-mark each small-scale sample, and group the samples with the same time;
[0149] S75. For each group of small-scale samples assigned, calculate the error between any two samples;
[0150] S76. Calculate the standard deviation and error correlation coefficient matrix of each grid point of each small-scale local weather type, as well as the β value, and then calculate the fitting function coefficient of the correlation coefficient matrix.
[0151] S8. In actual forecasting, the classification in S4 and S7 is used to identify the large-scale, medium-scale and small-scale local weather types at each grid point in the forecast area and mark them at the corresponding grid points. In step S8, the large-scale, medium-scale and small-scale local weather types at each grid point in the forecast area are identified and marked at the corresponding grid points. The specific method is:
[0152] S81. Decompose the current forecast field into three parts: large scale, medium scale and small scale;
[0153] S82. Sample large-scale, medium-scale and small-scale local weather types in the forecast field;
[0154] S83. For any grid point that is not on the boundary, that is, contains the complete N lx ×N ly , N mx ×N my or N sx ×N sy For the area of the local grid, calculate its similarity with each type in the cluster tree of the corresponding scale, take the type with the highest similarity to the predicted local weather situation in the cluster tree as the type of the local weather environment where the grid point is located, and mark the type label in the type cluster tree on the grid point. For the small-scale local weather type, detect whether there is a large deviation between the convection information on the current local grid and the observed convection information. If there is a large deviation, retrieve the type with the highest consistency between the convection information and the observed information in the cluster tree, and mark the type on the grid point;
[0155] S84. For the grid points at the boundary, only the types at the same area position in the clustering tree are compared, and the results are recorded at the grid points.
[0156] S9. In the local assimilation system, the background error covariance corresponding to the identified local weather type is called, and assimilation calculation is performed to obtain the corresponding assimilation analysis field.
[0157] The implementation of local multi-scale static error covariance in the local assimilation algorithm, taking the Local DA algorithm as an example, Figure 5 As shown, the weather type classification, error statistics and assimilation application process of the quasi-static multi-scale local error covariance with flow-dependent characteristics are given. Figure 5 The left column and the middle column in the lower middle part correspond to the previous local weather type cluster analysis and forecast error statistics scheme. The assimilation application process is located in Figure 5 The rightmost column.
[0158] Step S9 calls the background error covariance corresponding to the identified local weather type in the local assimilation system and uses the Local DA algorithm to perform assimilation calculation. The specific method is:
[0159] S91, after the Local DA enters the update process of a certain model grid point variable, retrieve the recorded local weather type labels of the grid point at large scale, medium scale and small scale;
[0160] S92, according to the local weather type labels of each scale, reading the standard deviation and fitting function coefficient of each variable on the local weather grid corresponding to the recorded label;
[0161] S93, according to the calculation formula of the root mean square CT of the estimated background error covariance, calculate ρ○C in the following formula correl Each element of:
[0162]
[0163] Where S is X o The diagonal matrix of the mean square error of each model variable is included, ρ is the localization coefficient matrix, and the calculation formula is:
[0164]
[0165] where a pq are the coefficients of longitude x, latitude y, and altitude z, n, m, and l are the orders of the polynomials, ρ is the statistical correlation coefficient, p is the number of equations, q = (k-1)mn+(j-1)n+i;
[0166] Its adjustment factor Where N is X o The number of variables included, trace() represents the trace of the calculated matrix, The trace of the matrix is calculated by The sum of the squares of all elements of is obtained;
[0167] S94, adding the calculated correlation coefficient matrices of each scale;
[0168] S95. Call the standard deviation of each scale, convert it into variance and find the average of the sum of the variances of each scale, then square it to get the mean square error of each scale, and then use it to replace the corresponding elements of S, and complete Calculation.
[0169] The effects that can be achieved through this embodiment are:
[0170] (1) Compared with the traditional static background error covariance, the local static background error of the present invention has a more reasonable flow dependence. Here is an example reflecting the flow dependence characteristics of the local background error covariance, a local grid of 9×9. In the local weather type near the center of a typhoon, the wind field presents a cyclonic circulation, the temperature decreases from the center to the surrounding, and the humidity center is located in the northeast quadrant of the typhoon center, that is, the humidity decreases from the northeast to the southwest. At this time, if a positive temperature increment is added to the center of the region, the local background error covariance will produce a cyclonic circulation increment in the wind field, corresponding to the enhancement of the typhoon warm center. At the same time, the local background error covariance produces a positive increment in humidity in the northeast quadrant and a negative increment in the southwest quadrant, increasing the original humidity gradient near the typhoon center. The above increments are consistent with the distribution characteristics of the elements near the original typhoon center, which is a manifestation of the flow dependence characteristics of the local background error covariance. If it is a traditional single variable static background error covariance, it will only produce an increment in temperature. If the equilibrium relationship between potential temperature and stream function is considered, a weak cyclonic wind field increment will be generated, but there will still be no water vapor increment, or the water vapor increment and temperature increment will be in phase (the constraint condition that the saturation specific humidity is not changed by adjusting the temperature is used), which cannot reflect the spatial distribution characteristics of the local weather system in terms of humidity. Therefore, the use of the local background error covariance described in the present invention will help improve the quality of assimilation analysis, thereby improving the forecast effect.
[0171] (2) The present invention has high scalability and portability. The present invention is not bound to a specific numerical weather forecast model and can be applied to different types of regional numerical forecast models, such as WRF, ARPS, etc., and has high portability. The present invention is designed with the possibility of adding local weather types in the future in mind. If it is found that the local weather type encountered in the current forecast does not have a high similarity in the original clustering tree, such as a corresponding type exceeding 0.9, then the new local weather type can be added to the existing clustering tree without recalculating the entire clustering tree. Therefore, the present invention has high scalability.
[0172] As described above, although the present invention has been shown and described with reference to specific preferred embodiments, it should not be construed as limiting the present invention itself. Various changes in form and details may be made without departing from the spirit and scope of the present invention as defined in the appended claims.
Claims
1. A local background error covariance statistics calling method based on local weather types, characterized in that: The steps include: S1. Determine the local grid size and resolution of large, medium and small scales according to the grid resolution and region size of the forecast model; N x ×N y The large-scale local grid is N lx ×N ly , the mesoscale local grid is N mx ×N my , the small-scale local grid is N sx ×N sy ; S2. Use reanalysis data from a period of time and filter them based on meteorological scale classification to separate large-scale and medium-scale parts. The specific method is: S21. Determine the statistical period and download the reanalysis data set X = (x1, x2, x3, ..., x n ), where x n It means that the data set has a total of n samples x at different times, and x contains information about the east-west wind field component u, the north-south wind field component v, temperature T, specific humidity q, and air pressure p variables; S22, use bandpass filtering or Fourier decomposition to divide X into large-scale X L and mesoscale X M Two parts, where the lower limit of the large-scale wavelength is set to 500km-1000km, the upper limit of the mesoscale band is consistent with the lower limit of the large-scale wavelength, and the wavelength of the mesoscale lower limit is set to 30km-100km; S23, for the same model forecast data X as S21 period f =(x f 1,x f 2,x f 3,…,x f n ), decompose it into large-scale X in the same way as S22 f L and mesoscale X f M Two parts; S3, using large-scale and medium-scale local grids to sample local weather types for the large-scale and medium-scale parts of S2, respectively, to form original unclassified local type samples; the specific method is: S31. Get X L A data file of a time period; S32, use N lx ×N ly The large-scale local grid of traverses each grid point, each time including the model height layer of the current grid point and the n above z Layer and below z Layer, n z The size of determines the thickness of the local grid; S33. For the grid point (i, j, k) close to the boundary, that is, taking this point as the center, it is impossible to obtain N completely. lx ×N ly The western boundary of the local grid is defined as Max(1, iN lx / 2), the eastern boundary is min(N x , i+N lx / 2), the southern boundary is Max(1, jN ly / 2), the northern boundary is min(Ny, j+N ly / 2); the upper and lower boundaries are processed in the same way, that is, the size of the local grid is cut off according to the boundary, and the upper boundary is min(N z , k+n z ), the lower boundary is max(1, kn z ), where N z is the total number of vertical layers in the mode area determined by S1; S34, through the method of steps S31 to S33, traverse X L All the data of the time are collected and stored; S35, continue to process X through the method of steps S31 to S34 M , X f L and X f M and store the data; S4. Use hierarchical clustering to perform bottom-up cluster analysis on the local type samples obtained in S3 to obtain a hierarchical clustering tree that meets the spatial correlation coefficient as the reference standard; the specific method is: S41, standardize the local type samples obtained and remove the dimensions of each variable; S42, using the spatial correlation coefficient as the judgment standard for all X L Conduct hierarchical cluster analysis on local type samples and record the spatial correlation coefficients at each level; S43, cluster analysis ends when the spatial correlation coefficient of clusterable clusters is lower than a certain value. The correlation coefficient threshold is adjusted up and down according to computing resources and actual application scenarios, and the minimum is not less than 0.5; S44. For a type that contains more than 3 original samples, the first three eigenvectors and eigenvalues of the sample set are calculated as the type representative. If it contains 2 original samples, the average value of the samples is used as the type representative for later clustering and type identification. S45. Generate X based on the final clustering result L Cluster tree of local weather types; S46, for X M , repeat steps S42 to S44 to form a classification and clustering tree of mesoscale local weather types; S47, for the types at the boundary, cluster analysis is performed only on samples at the same position, and the operation steps are the same as S42 to S46; S5. Based on the classification results of S4, model forecasts are carried out in the same time period as the reanalysis data of S2, using time-lagged ensemble forecasts, multi-physical process ensemble forecasts, random physical process parameterization tendencies and parameter perturbation ensemble methods to obtain large-scale and medium-scale local error samples; the specific methods are as follows: S51. For the period selected in step S2, use the model forecast to perform ensemble forecast simulation, including time-lagged ensemble forecast, multi-physical process ensemble forecast, random physical process parameterization tendency and parameter perturbation ensemble. The number of ensemble forecast members is set according to the actual situation and computing and storage resource conditions, but the model forecast samples of the reanalysis data at each time of the period selected in step S2 should be no less than 30. S52, using the same local sample sampling method as step S3, perform large-scale and medium-scale local sampling on all model forecast results, and mark the corresponding relationship between each local forecast sample and the reanalysis data; S53, for each large-scale and medium-scale type classified in step S4, calculating the difference between the original type sample contained in each type and its corresponding forecast; S6. Based on the classification results of S4 and the error samples obtained in S5, the error variance and reliability scale of each model variable under each type are calculated, and the fitting function of the local background error covariance matrix of each type is calculated; the specific method is: S61. For any local weather type obtained in step S4, calculate the standard deviation of each variable at each grid point on the corresponding local grid; S62, calculating the correlation coefficients of each variable at each point of the local weather type and each variable at other points, and establishing a local background error correlation coefficient matrix; S63, calculating the β value of each element of the background error correlation coefficient matrix according to the following formula, and multiplying it by the corresponding single element; Where N ρ is the number of samples, ρ is the correlation coefficient, d is the difference between the forecast and the observation, subscripts 1 and 2 represent two points with a distance r, I() is the sign consistency function of the correlation coefficient and the difference between the forecast and the observation, and the meaning of I() is that if the error signs of the variables at the two positions are opposite, the error correlation coefficient must be negative to update and correct the pattern variable at the second position through the observation at the first position. If the correlation coefficient is positive at this time, the error will inevitably increase, which is defined as unreasonable; S64. Calculate the fitting coefficient of the fitting function of the obtained correlation coefficient matrix according to the following formula: where a pq are the coefficients of longitude x, latitude y, and altitude z, n, m, and l are the orders of the polynomials, ρ is the statistical correlation coefficient, p is the number of equations, q = (k-1)mn+(j-1)n+i; The coefficients of each order are obtained by solving the following equation: S65, for all local weather types in S4, repeat steps S61 to S64 to obtain all fitting coefficients; S7, using the time lag set, multi-physical process set, physical process parameters and tendency random disturbance set method, simulate the period when convection occurs in the S2 reanalysis data, and use small-scale local grids to sample the simulated samples; then use the methods of S4 to S6 to perform cluster analysis on each convection scale sample, and perform error variance and reliable correlation coefficient scale statistics on the ensemble forecast samples corresponding to each type, and calculate the corresponding local background error covariance matrix fitting function; step S7 is the establishment and classification of small-scale local weather types, and the specific method is: S71, select the time with precipitation in step S2, filter it, remove the part other than the small scale, and set the upper limit of the small scale wavelength to the lower limit of the medium scale wavelength, which is equivalent to the small scale part X f S =X f -X f L -X f M ; S72. Use local sample collection method, using small-scale local grid N sx ×N sy Extract small-scale local weather type samples from all forecasts. The small-scale local weather samples contain cloud microphysical variables. The boundary processing method is the same as that in step S3. S73, using a cluster analysis method to perform a hierarchical cluster analysis on the obtained small-scale weather samples to form small-scale types and cluster trees; S74, based on the method of marking sample time, time-mark each small-scale sample, and group the samples with the same time; S75. For each group of small-scale samples assigned, calculate the error between any two samples; S76, calculating the standard deviation and error correlation coefficient matrix of each grid point of each small-scale local weather type, as well as the β value, and then calculating the fitting function coefficient of the correlation coefficient matrix; S8. In actual forecasting, the classification in S4 and S7 is used to identify the large-scale, medium-scale and small-scale local weather types at each grid point in the forecast area and mark them at the corresponding grid points; the specific method is as follows: S81. Decompose the current forecast field into three parts: large scale, medium scale and small scale; S82. Sample large-scale, medium-scale and small-scale local weather types in the forecast field; S83. For any grid point that is not on the boundary, that is, contains the complete N lx ×N ly , N mx ×N my or N sx ×N sy For the area of the local grid, calculate its similarity with each type in the cluster tree of the corresponding scale, take the type with the highest similarity to the predicted local weather situation in the cluster tree as the type of the local weather environment where the grid point is located, and mark the type label in the type cluster tree on the grid point. For the small-scale local weather type, detect whether there is a large deviation between the convection information on the current local grid and the observed convection information. If there is a large deviation, retrieve the type with the highest consistency between the convection information and the observed information in the cluster tree, and mark the type on the grid point; S84, for a grid point at the boundary, only types at the same area in the cluster tree are compared, and the result is recorded at the grid point; S9. In the local assimilation system, the background error covariance corresponding to the identified local weather type is called and assimilated to obtain the corresponding assimilation analysis field. The specific method is as follows: S91, after the Local DA enters the update process of a certain model grid point variable, retrieve the recorded local weather type labels of the grid point at large scale, medium scale and small scale; S92, according to the local weather type labels of each scale, reading the standard deviation and fitting function coefficient of each variable on the local weather grid corresponding to the recorded label; S93, based on the root mean square of the estimated background error covariance Calculate the following formula to calculate ρ ○ C correl Each element of: Where S is X o The diagonal matrix of the mean square error of each model variable is included, ρ is the localization coefficient matrix, and the calculation formula is: where a pq are the coefficients of longitude x, latitude y, and altitude z, n, m, and l are the orders of the polynomials, ρ is the statistical correlation coefficient, p is the number of equations, q = (k-1)mn+(j-1)n+i; Its adjustment factor Where N is X o The number of variables included, trace() represents the trace of the calculated matrix, The trace of the matrix is calculated by The sum of the squares of all elements of is obtained; S94, adding the calculated correlation coefficient matrices of each scale; S95. Call the standard deviation of each scale, convert it into variance and find the average of the sum of the variances of each scale, then square it to get the mean square error of each scale, and then use it to replace the corresponding elements of S, and complete Calculation.
Citation Information
Patent Citations
Rapid updating mixing assimilation method based on time lag set
CN105447593A
Global medium-term numerical forecast GRAPES_GFS
CN110703357A
Method for inhibiting false convection
CN113406590A
Local set data assimilation method
CN113792262A
Hybrid assimilation system and method based on ensemble Kalman filtering framework
CN114048433A
