Multi-Mode Ensemble Meteorological Prediction Method and System Based on Predictable Signal Probability Mapping

Abstract

The present invention relates to the field of meteorological services, and provides a multi-model ensemble meteorological prediction method and system based on predictable signal probability mapping. The method includes: S1, identifying predictable modes of climate variables based on historical observation data; S2, extracting predictable components and residual components from observation results and model prediction results; S3, establishing probability mapping relationships between model predictable components and observation predictable components, and between model residual components and observation residual components; S4, performing bias correction on the predictable components and residual components in the model prediction results; S5, determining the weight coefficients of multi-model ensemble averaging to obtain meteorological prediction results. The ensemble averaging method provided by the present invention performs ensemble based on the bias correction of different prediction components of the model, takes into account the performance differences of dynamic climate models for prediction biases from different sources, can more reasonably and accurately exploit the advantages of specific climate models, and improve prediction accuracy.

Description

Technical Field

[0001] The present invention relates to the field of meteorological services, and particularly to a multi-model ensemble meteorological prediction method and system based on predictable signal probability mapping. Background Art

[0002] Extreme climate events such as high temperature, low temperature, rainstorm and flood often cause major meteorological disasters and have an important impact on social economy and people's production and life. If accurate predictions of extreme climate events can be made two weeks to one month (or one season) in advance, it is crucial for preventing and reducing the possible impacts of related meteorological disasters.

[0003] Currently, dynamic climate models are the main tools for predicting extreme climate events on the sub-seasonal to seasonal (i.e., two weeks to month to season) time scale. However, due to the existence of model initial errors and model itself errors, there is still a large uncertainty in sub-seasonal to seasonal predictions based on dynamic climate models. To solve the problem of large prediction uncertainty of a single dynamic climate model, researchers have proposed a multi-model ensemble prediction strategy. Multi-model ensemble prediction has the advantage of considering both initial value uncertainty and model uncertainty, and is thus considered an effective way to improve the accuracy of climate prediction and has also become the mainstream trend of current meteorological operational prediction on the sub-seasonal to seasonal time scale. However, multi-model ensemble averaging often has a smoothing effect on the prediction results, resulting in a significant reduction in the amplitude of the prediction results after ensemble averaging, especially the prediction ability for extreme climate events is significantly insufficient and it is difficult to meet the actual business service requirements.

[0004] Regarding the problem of multi - model ensemble prediction bias, previous studies have proposed the variance correction method, that is, using the variance ratio between historical observation data and the prediction results of dynamic climate models as the weight coefficient, and then multiplying or dividing the prediction results of dynamic climate models by a certain multiple. Due to the complexity of the internal interaction process of the climate system, the evolution of the prediction bias of dynamic climate models often has non - linear characteristics. However, the variance correction method can only perform simple linear correction on the prediction bias of dynamic climate models and is powerless for its non - linear evolution. In addition, some researchers have used the non - parametric percentile mapping method to correct the prediction bias of dynamic climate models. This method corrects the bias between the predicted probability density function distribution of the model and the observed probability density function distribution. However, climate variables are often controlled by multiple external forcing factors and physical processes, showing obvious multi - temporal and multi - spatial scale variability characteristics. For different temporal and spatial scale variabilities, the sources of their model prediction biases are often different, so there may be different probability density function distribution characteristics. However, the non - parametric percentile mapping method can only correct the bias of the overall probability density distribution of climate variables and cannot distinguish the influence of different prediction bias sources. Therefore, the above - mentioned methods have limited correction effects on multi - model ensemble prediction bias and there is great room for improvement. In addition, the current prediction bias correction methods all correct the bias of single - point data and do not consider the spatio - temporal continuity of model prediction bias, resulting in fragmented correction effects. Summary of the Invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and provide a multi - model ensemble meteorological prediction method and system based on predictable signal probability mapping. Through the separation of predictable signals and the probability mapping relationship between prediction results and observation results, effective correction of prediction biases from different sources and improvement of the accuracy of multi - model ensemble prediction are achieved.

[0006] The present invention adopts the following technical solutions:

[0007] On the one hand, the present invention provides a multi - model ensemble meteorological prediction method based on predictable signal probability mapping, including:

[0008] S1. Based on historical observation data, identify the predictable modes of specific climate variables;

[0009] S2. Through spatial projection of the predictable modes, extract the predictable components and residual components in the observation results and the prediction results of dynamic climate models;

[0010] S3. Respectively obtain the cumulative probability density function distributions of the predictable components and residual components obtained in step S2, and establish the probability mapping relationships between the predictable components of the dynamic climate model and the observed predictable components, and between the residual components of the dynamic climate model and the observed residual components;

[0011] S4. According to the probability mapping relationship obtained in step S3, perform bias correction on the predictable component and the residual component in the prediction result of the dynamic climate model respectively;

[0012] S5. According to the regression relationship between the prediction bias correction result of the dynamic climate model of different components and the observation result, determine the weight coefficients of the multi-model ensemble average, and obtain the multi-model ensemble meteorological prediction result accordingly.

[0013] In any of the possible implementation manners as described above, a further implementation manner is provided. In step S1, the historical observation data is selected as the long-sequence meteorological variable observation data, and the empirical orthogonal function decomposition is performed on the long-sequence meteorological variable observation data to identify the predictable mode of the specific climate variable.

[0014] In any of the possible implementation manners as described above, a further implementation manner is provided, where the meteorological variables include air temperature and precipitation.

[0015] In any of the possible implementation manners as described above, a further implementation manner is provided. Step S2 is specifically as follows:

[0016] S21. Based on the meteorological variable observation data and the empirical orthogonal function decomposition (abbreviated as EOF) method, obtain the predictable mode of the meteorological variable:

[0017] X o = V o T o ;

[0018] Among them, the subscript o represents observation; X o represents the meteorological variable observation value; V o represents the spatial eigenvector of the EOF decomposition, that is, the EOF mode; T o represents the time coefficient corresponding to the EOF mode; Based on the Kaiser criterion, the first h EOF modes when the cumulative variance contribution rate reaches 80% are considered as predictable modes, then the predictable component of the observation result is:

[0019]

[0020] Among them, the superscript p represents the predictable component, the subscript i represents the i-th EOF mode, h is the total number of EOF modes used to construct the predictable component, V o，i and T o，i respectively represent the spatial eigenvector and the time coefficient corresponding to the i-th EOF mode in the observation result;

[0021] The formula of the Kaiser criterion is:

[0022]

[0023] Among them, λ m is the eigenvalue of the retained EOF mode, and S k is the k-th variance of the decomposed meteorological variable;

[0024] The remaining component is the residual component, denoted as The superscript n represents the residual component;

[0025]

[0026] S22. Project the prediction result X of the dynamic climate model m onto the predictable mode V o to obtain the predictable component in the prediction result and its residual component

[0027]

[0028] The subscript m represents the dynamic climate model;

[0029] Traverse all dynamic climate models to obtain the predictable component and residual component of the meteorological variable prediction result under each dynamic climate model.

[0030] For any possible implementation manner as described above, a further implementation manner is provided. Specifically, step S3 is as follows:

[0031] S31. For the predictable components and residual components in the observation results and the prediction results of the dynamic climate model obtained in step S2, calculate their respective probability density functions f; the calculation formula of the probability density function f is:

[0032]

[0033] x represents the meteorological variable value, Δx represents the value interval of the meteorological variable value, n represents the number of samples where the meteorological variable value appears in the interval [x, x + Δx], N is the total number of samples, and f(x, x + Δx) represents the occurrence probability of the meteorological variable value in the value range [x, x + Δx];

[0034] S32. According to the probability density function obtained in step S31, obtain the cumulative probability density function F(x) to make the mapping relationship between the observation results and the prediction results of the dynamic climate model unique; the calculation formula of the cumulative probability density function F(x) is:

[0035]

[0036] The cumulative probability density function F(x) is the integral of the probability density from the infinitesimal value of the meteorological variable value to the value x;

[0037] S33. Establish the mapping relationships between the predictable components of the dynamic climate model prediction results and the predictable components of the observation results, and between the residual components of the dynamic climate model prediction results and the residual components of the observation results respectively:

[0038]

[0039] Among them, for the meteorological variable value x m (t) of the dynamic climate model at time t, its predictable component has a cumulative probability density of The cumulative probability density of the corresponding observation result is According to the mapping relationship between the two, the following equation holds:

[0040]

[0041] The residual component has a cumulative probability density of The cumulative probability density of the corresponding observation result is According to the mapping relationship between the two, the following equation holds:

[0042]

[0043] For any of the possible implementation methods described above, a further implementation method is provided. Specifically, step S4 is as follows:

[0044] S41. Project the meteorological variable prediction value x m (t) of the dynamic climate model onto the first h predictable component modes of the corresponding model to obtain the predictable component of the dynamic climate model, and calculate the residual component m through the difference between x and Among them, the cumulative probability corresponding to the predictable component is denoted as The cumulative probability corresponding to the residual component is denoted as

[0045] S42. According to the mapping relationship between the observation results and the dynamic climate model prediction results, in the cumulative probability distribution of the predictable components of the observation results, obtain the value of the meteorological variable x corresponding to the cumulative probability of , denoted as which is the prediction deviation correction result of the predictable component corresponding to x m (t);

[0046] S43. In the cumulative probability density function distribution of the residual components of the observation results, obtain the value of the meteorological variable x corresponding to the cumulative probability of , denoted as That is x m (t) corresponds to the prediction deviation correction result of the residual component;

[0047] S44. Traverse all dynamic climate models in steps S41 - S43, and then obtain the deviation correction results of the predictable components and residual components of all dynamic climate models.

[0048] For any of the possible implementation manners described above, a further implementation manner is provided. Step S5 is specifically as follows:

[0049] Based on the deviation correction in step S4, perform an ensemble average on the deviation correction results of the predictable components and residual components of all dynamic climate models at time t to obtain the multi - model ensemble average prediction result at time t:

[0050]

[0051] where M is the total number of dynamic climate models; a j is the weight coefficient of the predictable component of the j - th dynamic climate model, and its magnitude is determined by establishing a linear regression relationship between the predictable component of the prediction result of this model and the predictable component of the observation result; b j is the weight coefficient of the residual component of the j - th dynamic climate model, and its magnitude is determined by establishing a linear regression relationship between the residual component of the prediction result of this model and the residual component of the observation result.

[0052] For any of the possible implementation manners described above, a further implementation manner is provided. For the dynamic climate model, several of the China Meteorological Administration CMA - CPSv3, European Centre for Medium - Range Weather Forecasts ECMWF5, UK GloSea6, US CFSv2, and Japan JMA_CPS3 are selected.

[0053] On the other hand, the present invention also provides a multi - model ensemble meteorological prediction system based on predictable signal probability mapping. The system adopts the above - mentioned method. The system includes:

[0054] A historical observation data collection module for collecting historical meteorological variable observation data;

[0055] A climate variable predictable mode extraction module for extracting the predictable modes of specific climate variables according to the historical meteorological variable observation data;

[0056] A predictable component and residual component extraction module for extracting the predictable components and residual components in the observation result and the prediction result of the dynamic climate model by performing a spatial projection on the predictable modes;

[0057] A probability mapping relationship establishment module, which is used to establish the probability mapping relationships between the predictable components of the dynamic climate model and the observable predictable components, and between the residual components of the dynamic climate model and the observable residual components;

[0058] A predictable component and residual component deviation correction module, which is used to respectively correct the deviations of the predictable components and residual components in the prediction results of the dynamic climate model according to the probability mapping relationships;

[0059] A weight coefficient determination module, which determines the weight coefficients of the multi-model ensemble average according to the regression relationships between the prediction deviation correction results of the dynamic climate model of different components and the observation results;

[0060] A prediction module, which predicts future meteorology by using the multi-model ensemble average method according to the weight coefficients.

[0061] For any possible implementation manner as described above, a further implementation manner is provided. The system further includes a prediction result evaluation module, which evaluates the accuracy of the prediction results by using the time correlation coefficient method and the space correlation coefficient method.

[0062] The beneficial effects of the present invention are as follows:

[0063] 1. The present invention provides a method for correcting the prediction deviation of a dynamic climate model based on the probability mapping of predictable signals, so as to improve the prediction accuracy of the dynamic climate model.

[0064] 2. A multi-model ensemble prediction method based on deviation correction is provided to improve the prediction ability of the multi-model ensemble for extreme climate events.

[0065] 3. The predictable modes of specific climate variables are identified through long-term historical observation data. On the one hand, the authenticity and stability of the predictable modes are ensured. On the other hand, the same predictable modes are used as the basis for both observation and model prediction, ensuring the consistency of physical processes and deviation sources, making the established probability mapping relationships more reasonable and accurate, thereby improving the accuracy of prediction deviation correction. Through empirical orthogonal function decomposition and predictable mode identification, not only the predictable components and residual components with different prediction deviation sources are distinguished, but also the dimension of the original data is reduced, ensuring the spatio-temporal continuity of the prediction deviation, and at the same time improving the efficiency of deviation correction. Description of the Drawings

[0066] Figure 1 The figure shows a schematic flowchart of a multi-model ensemble meteorological prediction method based on the probability mapping of predictable signals according to an embodiment of the present invention.

[0067] Figure 2 The figure shows a flowchart for obtaining the predictable components, residual components and their probability density functions in the observation and model prediction results in the embodiment.

[0068] Figure 3 The figure shows the flowchart of bias correction for the predictable component and the residual component of the prediction results of the dynamic climate model in the embodiment.

[0069] Figure 4 The figure shows the prediction skills of the multi-model ensemble for the temperature anomaly in East Asia from 1991 to 2020: (a) before bias correction; (b) after bias correction; (c) the distribution of the difference between after bias correction and before bias correction; in (a) and (b), a positive value indicates the existence of prediction skills, and a negative value indicates no prediction skills; in (c), a positive value indicates that bias correction has a promoting effect on prediction skills, and a negative value indicates no promoting effect.

[0070] Figure 5 The figure shows the comparison between the multi-model ensemble prediction results and the observed results of extreme high temperatures (unit: °C) in July - August 2022 in the embodiment: (a) the prediction results before bias correction; (b) the prediction results after bias correction; (c) the observed results; the box indicates the Yangtze River Basin (25° - 40°N, 90° - 120°E), and R represents the spatial correlation coefficient between the prediction results and the observed results. Detailed implementation manners

[0071] The specific embodiments of the present invention will be described in detail below with reference to specific drawings. It should be noted that the technical features described in the following embodiments or the combinations of technical features should not be considered as isolated, and they can be combined with each other to achieve better technical effects.

[0072] As Figure 1 shown, an embodiment of the present invention provides a multi-model ensemble meteorological prediction method based on predictable signal probability mapping, including:

[0073] S1. Based on historical observation data, identify the predictable modes of specific climate variables;

[0074] S2. Through spatial projection of the predictable modes, extract the predictable components and residual components in the observed results and the prediction results of the dynamic climate model;

[0075] S3. Respectively obtain the cumulative probability density function distributions of the predictable components and residual components obtained in step S2, and establish the probability mapping relationships between the predictable components of the dynamic climate model and the observed predictable components, and between the residual components of the dynamic climate model and the observed residual components;

[0076] S4. According to the probability mapping relationships obtained in step S3, respectively perform bias correction on the predictable components and residual components in the prediction results of the dynamic climate model;

[0077] S5. Determine the weight coefficients of the multi-model ensemble average according to the regression relationship between the prediction deviation correction results of the dynamic climate models of different components and the observed results, and obtain the meteorological prediction results of the multi-model ensemble based on this.

[0078] The following takes air temperature as an example to illustrate the specific implementation manner of the present invention in detail:

[0079] 1. Obtain the observed data and the prediction data of the dynamic climate model

[0080] Obtain the historical observed data of air temperature since 1961; obtain the historical retrospective and real-time prediction air temperature data of multiple dynamic climate models such as CMA-CPSv3 of the China Meteorological Administration, FGOALS-f2 of the Institute of Atmospheric Physics, Chinese Academy of Sciences, ECMWF5 of the European Centre, GloSea6 of the United Kingdom, CFSv2 of the United States, and JMA_CPS3 of Japan since 1991.

[0081] 2. Obtain the predictable components and residual components in the observed results and the prediction results of the dynamic climate model

[0082] Figure 2 Shows the specific process of obtaining the predictable components and residual components of air temperature and their probability density functions.

[0083] First, based on the air temperature observed data and the empirical orthogonal function decomposition (referred to as EOF) method, obtain the predictable modes of air temperature:

[0084] X o =V o T o , (1)

[0085] Among them, the subscript o represents observation; X o represents the observed value of air temperature; V o represents the spatial eigenvector of the EOF decomposition, that is, the EOF mode; T o represents the time coefficient corresponding to the EOF mode. The first h EOF modes when the cumulative variance contribution rate reaches 80% are considered predictable modes, then the predictable component of the observed result is:

[0086]

[0087] Among them, the superscript p represents the predictable component, the subscript i represents the i-th EOF mode, h is the total number of EOF modes used to construct the predictable component, V o，i and T o，i respectively represent the spatial eigenvector and time coefficient corresponding to the i-th EOF mode in the observed result.

[0088] The remaining part is the residual component, denoted as

[0089]

[0090] The superscript n represents the residual component.

[0091] Secondly, project the prediction result (X m ) of the dynamical climate model onto the predictable mode V o , and then obtain the predictable component and its residual component in the prediction result

[0092]

[0093] Here, the subscript m represents the dynamical climate model. Based on Equation 4-5, all dynamical climate models are traversed to obtain the predictable component and the residual component of the air temperature prediction result for each model.

[0094] 3. Obtain the probability function distributions of the predictable components and the residual components in the observation results and the prediction results of the dynamical climate model

[0095] For the predictable components and the residual components in the observation results and the prediction results of the dynamical climate model obtained in Step 2, calculate their respective probability density functions (f). The calculation formula for the probability density function is:

[0096]

[0097] x represents the air temperature value, Δx represents the value interval of the air temperature value, n represents the number of samples where the air temperature value appears in the range of [x, x + Δx], N is the total number of samples, and f(x, x + Δx) represents the occurrence probability of the air temperature value in the range of [x, x + Δx].

[0098] Subsequently, further convert the probability density function into a cumulative probability density function (F) so that the mapping relationship between the observation results and the prediction results of the dynamical climate model is unique. The calculation formula for the cumulative probability density function is:

[0099]

[0100] That is to say, the cumulative probability density function F(x) is the integral of the probability density from the infinitesimal value of the air temperature to the value x.

[0101] 4. Obtain the probability mapping relationship between the observation results and the prediction results of the dynamical climate model

[0102] Based on Step 3, establish the mapping relationships between the predictable component of the prediction result of the dynamical climate model and the predictable component of the observation result, and between the residual component of the prediction result of the dynamical climate model and the residual component of the observation result respectively: ​

[0103]

[0104] Among them, for the air temperature value x of the dynamic climate model at time t m (t), its predictable component The cumulative probability density is The cumulative probability density corresponding to the observation result is According to the mapping relationship between the two, the following formula holds:

[0105]

[0106] Residual component The cumulative probability density is The cumulative probability density corresponding to the observation result is According to the mapping relationship between the two, the following formula holds:

[0107]

[0108] 5. Obtain the bias correction result predicted by the dynamic climate model

[0109] As Figure 3 shown, taking the air temperature prediction value x of a certain dynamic climate model at time t m (t) as an example, the specific prediction bias correction process is introduced.

[0110] First, project the air temperature prediction value x of the dynamic climate model m (t) onto the first h predictable component modes of the corresponding model to obtain the predictable component of the dynamic climate model And calculate the residual component through the difference between x m (t) and Denoted as Among them, the cumulative probability corresponding to the predictable component is denoted as The residual component The corresponding cumulative probability is denoted as Then, according to the mapping relationship (formula 8) between the observation result and the prediction result of the dynamic climate model, in the cumulative probability distribution of the predictable component of the observation result, obtain the value of the air temperature x corresponding to the cumulative probability of , denoted as This is the prediction bias correction result of the predictable component corresponding to x m (t).

[0111] Similarly, in the cumulative probability density function distribution of the residual component of the observation result, find the value of the air temperature x corresponding to the cumulative probability of , denoted as This is x m(t) corresponds to the prediction deviation correction result of the residual component.

[0112] Apply the above operations to all dynamic climate models, and then obtain the deviation correction results of the predictable components and residual components of all dynamic climate models.

[0113] 6. Obtain the multi-model ensemble prediction result

[0114] On the basis of deviation correction, perform an ensemble average on the deviation correction results of the predictable components and residual components of all dynamic climate models at time t to obtain the multi-model ensemble average prediction result at time t:

[0115]

[0116] where M is the total number of dynamic climate models; a j is the weight coefficient of the predictable component of the j-th dynamic climate model, and its value is determined by establishing a linear regression relationship between the predictable component of the prediction result of this model and the predictable component of the observation result; b j is the weight coefficient of the residual component of the j-th dynamic climate model, and its value is determined by establishing a linear regression relationship between the residual component of the prediction result of this model and the residual component of the observation result.

[0117] Verification:

[0118] To verify the improvement effect of the present invention on the prediction skill of climate variables, Figure 4 The prediction skills of summer temperature in the East Asian region before and after deviation correction and their differences are given. The dynamic climate models used include CMA-CPSv3 of the China Meteorological Administration, FGOALS-f2 of the Institute of Atmospheric Physics, Chinese Academy of Sciences, ECMWF5 of the European Centre, GloSea6 of the United Kingdom, CFSv2 of the United States, and JMA_CPS3 of Japan. The training stage is from 1991 to 2005, and the verification period is from 2006 to 2020. Here, the Temporal Correlation Coefficient (TCC) is used to evaluate the prediction skill of the model. TCC reflects the degree of consistency between the prediction result and the observation result, and the calculation formula is as follows:

[0119]

[0120] where F represents the predicted value, O represents the observed value, and the overline represents the multi-year time average; t represents time (year), t = 1, 2,......, N, and N is the total number of samples. The value range of TCC is between ±1. TCC = 1 means that the prediction result is completely consistent with the observation result. TCC > 0 means that the dynamic climate model has prediction skill, and TCC < 0 means that the dynamic climate model has no prediction skill.

[0121] Calculation formula for spatial correlation coefficient:

[0122]

[0123] Among them, F represents the predicted value, O represents the observed value, and the overline represents the spatial area average; i represents the number of meridional grid points, i = 1, 2,......, K, where K is the total number of meridional grid points; j represents the number of zonal grid points, j = 1, 2,......, L, where L is the total number of zonal grid points. The range of R is between ±1. R = 1 indicates that the spatial distribution patterns of the prediction result and the observed result are exactly the same. The closer R is to 1, the more similar the spatial distribution patterns of the prediction result and the observed result are.

[0124] From Figure 4 It can be seen that after bias correction of the East Asian summer temperature using the present invention, the prediction skill is significantly improved compared with that before bias correction. Especially in the regions where the TCC skill is positive (>0) before bias correction, the improvement effect is the most obvious. For the regions where the TCC skill is negative (<0) before bias correction, it indicates that the dynamic climate model has limited prediction ability for the temperature in these regions, which is caused by defects such as the initialization and physical processes of the dynamic climate model itself. Therefore, it does not belong to the problems to be solved by the present invention.

[0125] In July - August 2022, extreme high - temperature events beyond records occurred in the Yangtze River Basin of China, with abnormally high temperatures (see Figure 5 c)). Using the present invention application, bias correction was performed on the ensemble - mean prediction results of the dynamic climate model ( Figure 5 b)). Comparing with before bias correction ( Figure 5 a)), the prediction results of the multi - model ensemble mean for the extreme high - temperature in the Yangtze River Basin after bias correction are more consistent with the observed results. The spatial correlation coefficient increased from the original 0.41 to 0.77, that is, it increased by 88%.

[0126] In December 2022, serious extreme low - temperature events occurred in the central and eastern regions of China, with abnormally low temperatures. Using the present invention application, bias correction was performed on the ensemble - mean prediction results of the dynamic climate model. Comparing with before bias correction, the prediction results of the multi - model ensemble mean for the extreme low - temperature in the central and eastern regions of China after bias correction are more consistent with the observed results. The spatial correlation coefficient increased from the original 0.37 to 0.87, that is, it increased by 1.35 times.

[0127] A multi - model ensemble prediction method based on predictable signal probability mapping proposed by the present invention realizes effective correction of prediction biases from different sources and improvement of the accuracy of multi - model ensemble prediction through the separation of predictable signals and the probability mapping relationship between prediction results and observed results.

[0128] On the one hand, the present invention provides a method for correcting the prediction bias of a dynamic climate model based on predictable signal probability mapping. First, empirical orthogonal function decomposition is performed on historical long-term observational data to identify the predictable modes of specific climate variables; second, through spatial projection of the predictable modes, the predictable components and their residual components in the observational data and the dynamic climate model prediction results are obtained respectively; then, probability mapping relationships are established between the model predictable components and the observational predictable components, and between the model residual components and the observational residual components, and the prediction bias of the model is effectively corrected through this mapping relationship. Since the model hindcast data often has a time length of only 20 - 30 years, identifying the predictable modes of specific climate variables through historical long-term observational data, on the one hand, ensures the authenticity and stability of the predictable modes, and on the other hand, both observation and model prediction use the same predictable modes as the basis, ensuring the consistency of physical processes and bias sources, making the established probability mapping relationship more reasonable and accurate, thereby improving the accuracy of prediction bias correction. Through empirical orthogonal function decomposition and predictable mode identification, not only are the predictable components and residual components with different prediction bias sources distinguished, but also the dimension of the original data is reduced, ensuring the spatio-temporal continuity of the prediction bias and improving the efficiency of bias correction at the same time.

[0129] On the other hand, the present invention also provides a multi-model ensemble prediction method based on bias correction. On the basis of correcting the prediction bias of a single model, weight coefficients are established respectively according to the regression relationships between the predictable components of each model and the observational predictable components, and between the model residual components and the observational residual components, and then the multi-model ensemble average prediction result is obtained. By allocating weights according to the performance capabilities of each model for the predictable components and residual components, especially giving a larger weight to a model with excellent prediction performance in a certain aspect (predictable components or residual components), and a smaller weight to a model with poor prediction performance in a certain aspect. Due to the differences in the dynamic frameworks and physical processes of the models, the prediction biases from different sources have model dependencies. The ensemble average method provided by the present invention performs an ensemble on the basis of correcting the biases of different prediction components of the models, taking into account the performance differences of the dynamic climate models for prediction biases from different sources, and thus can more reasonably and accurately exploit the advantages of specific climate models, thereby improving prediction accuracy.

[0130] Although several embodiments of the present invention have been given in this text, those skilled in the art should understand that the embodiments in this text can be changed without departing from the spirit of the present invention. The above embodiments are only exemplary and should not be used as a limitation of the scope of the rights of the present invention.

Claims

1. A multi-mode ensemble meteorological forecasting method based on predictable signal probability mapping, characterized in that: The method comprises: S1. Identify predictable modes of specific meteorological variables based on historical observations; S2. extracting predictable components and residual components from observation results and dynamic climate model prediction results by spatially projecting the predictable modes; S3, respectively obtaining the cumulative probability density function distribution of the predictable component and the residual component obtained in step S2, and establishing a probability mapping relationship between the predictable component of the dynamic climate model and the observed predictable component, and between the residual component of the dynamic climate model and the observed residual component; S4, according to the probability mapping relationship obtained in step S3, respectively correcting the deviation of the predictable component and the residual component in the prediction result of the dynamic climate model; S5. Determine the weight coefficient of the multi-model ensemble average based on the regression relationship between the prediction bias correction results of the dynamic climate model of different components and the observation results, and obtain the multi-model ensemble meteorological forecast results based on it; Step S3 is specifically as follows: S31, for the predictable component and the residual component in the observation result and the dynamic climate model prediction result obtained in step S2, respectively calculate the respective probability density functions f; the calculation formula of the probability density function f is: x represents the value of the meteorological variable, Δx represents the interval of the meteorological variable value, n represents the number of samples in which the meteorological variable value appears in the interval [x, x+Δx], N is the total number of samples, and f(x, x+Δx) represents the probability of occurrence of the meteorological variable value in the range [x, x+Δx]; S32, according to the probability density function obtained in step S31, a cumulative probability density function F(x) is obtained, so that the mapping relationship between the observation results and the dynamic climate model prediction results is unique; the calculation formula of the cumulative probability density function F(x) is: The cumulative probability density function F(x) is the integral of the probability density when the meteorological variable value is infinitesimal to the value x; S33, respectively establish mapping relationships between the predictable components of the dynamic climate model prediction results and the predictable components of the observation results, and between the residual components of the dynamic climate model prediction results and the residual components of the observation results: Among them, for the meteorological variable value x of the dynamic climate model at time t m (t), its predictable component The cumulative probability density of The cumulative probability density of the corresponding observation is According to the mapping relationship between the two, the following formula holds: Residual Component The cumulative probability density of The cumulative probability density of the corresponding observation is According to the mapping relationship between the two, the following formula holds:

2. The multi-mode ensemble meteorological forecasting method based on predictable signal probability mapping according to claim 1, characterized in that: In step S1, the historical observation data uses long-sequence meteorological observation data, and the long-sequence meteorological observation data is decomposed by empirical orthogonal functions to identify predictable modes of specific climate variables.

3. The multi-mode ensemble meteorological forecasting method based on predictable signal probability mapping according to claim 1, characterized in that: The meteorological variables include temperature and precipitation.

4. The multi-mode ensemble meteorological forecasting method based on predictable signal probability mapping according to claim 1, characterized in that: Step S2 is specifically as follows: S21. Based on the meteorological variable observation data and the empirical orthogonal function decomposition method EOF, the predictable modes of meteorological variables are obtained: X o =V o T o ; Wherein, subscript o represents observation; X o Represents the observed value of meteorological variables; V o represents the spatial eigenvector of EOF decomposition, i.e., EOF mode; T o Represents the time coefficient corresponding to the EOF mode; based on the Kaiser standard, the first h EOF modes when the cumulative variance contribution rate reaches 80% are considered to be predictable modes, then the predictable component of the observation result for: Where, the superscript p represents the predictable component, the subscript i represents the i-th EOF mode, h is the total number of EOF modes used to construct the predictable component, V o，i and T o，i They represent the spatial eigenvector and time coefficient corresponding to the i-th EOF mode in the observation results respectively; The remaining component is the residual component, denoted as The superscript n represents the residual component; S22. The prediction results of the dynamic climate model are X m Projection to the predictable mode V o Get the predictable component in the prediction result and its residual Subscript m represents the dynamical climate model; All dynamical climate models are traversed to obtain the predictable component and residual component of the meteorological variable prediction results in each dynamical climate model.

5. The multi-mode ensemble meteorological forecasting method based on predictable signal probability mapping according to claim 1, characterized in that: Step S4 is specifically as follows: S41. The predicted value x of the meteorological variable in the dynamic climate model m (t) is projected onto the first h predictable component modes of the corresponding model to obtain the predictable component of the dynamical climate model And through x m (t) and The residual component is calculated by the difference between Among them, the predictable component The corresponding cumulative probability is recorded as Residual Component The corresponding cumulative probability is recorded as S42. According to the mapping relationship between the observation results and the prediction results of the dynamic climate model, the cumulative probability distribution of the predictable components of the observation results is obtained as The value of the meteorological variable x corresponding to the time is recorded as That is x m (t) The forecast deviation correction result corresponding to the predictable component; S43. In the cumulative probability density function distribution of the residual component of the observation result, the cumulative probability is obtained as The value of the meteorological variable x corresponding to the time is recorded as That is x m (t) The prediction deviation correction result of the corresponding residual component; S44, traverse steps S41-S43 through all dynamic climate models, and then obtain the deviation correction results of the predictable components and residual components of all dynamic climate models.

6. The multi-mode ensemble meteorological forecasting method based on predictable signal probability mapping according to claim 5, characterized in that: Step S5 is specifically as follows: On the basis of the bias correction in step S4, the bias correction results of the predictable components and residual components of all dynamic climate models at time t are averaged to obtain the multi-model ensemble average prediction result at time t: Where M is the total number of dynamical climate models; a j is the weight coefficient of the predictable component of the jth dynamic climate model, and its size is determined by establishing a linear regression relationship between the predictable component of the model prediction result and the predictable component of the observation result; b j is the weight coefficient of the residual component of the jth dynamic climate model, and its size is determined by establishing a linear regression relationship between the residual component of the model prediction result and the residual component of the observation result.

7. The multi-mode ensemble meteorological forecasting method based on predictable signal probability mapping according to claim 1, characterized in that: The dynamic climate model is selected from among CMA-CPSv3 of China Meteorological Administration, FGOALS-f2 of Institute of Atmospheric Physics of Chinese Academy of Sciences, ECMWF5 of Europe, GloSea6 of UK, CFSv2 of USA and JMA_CPS3 of Japan.

8. A multi-mode ensemble weather forecast system based on predictable signal probability mapping, characterized in that: The system adopts the method according to any one of claims 1 to 7, and the system comprises: Historical observation data collection module, used to collect historical meteorological variable observation data; A climate variable predictable mode extraction module, used to extract the predictable mode of a specific climate variable based on the historical meteorological variable observation data; A predictable component and residual component extraction module extracts predictable components and residual components from observation results and dynamic climate model prediction results by spatially projecting the predictable mode; A probability mapping relationship establishment module is used to establish a probability mapping relationship between a predictable component of a dynamic climate model and an observed predictable component, and between a residual component of a dynamic climate model and an observed residual component; A predictable component, that is, a residual component deviation correction module, is used to perform deviation correction on the predictable component and the residual component in the prediction result of the dynamic climate model according to the probability mapping relationship; The weight coefficient determination module determines the weight coefficient of the multi-model ensemble average according to the regression relationship between the prediction bias correction results of the dynamic climate model of different components and the observation results; The prediction module predicts future weather conditions using a multi-model ensemble average method based on the weight coefficients.

9. The multi-mode ensemble weather forecasting system based on predictable signal probability mapping according to claim 8, characterized in that: The system also includes a prediction result evaluation module, which uses a time correlation coefficient method and a space correlation coefficient method to evaluate the accuracy of the prediction result.

Images