A multi-mode forecast correction method and system based on ensemble learning

Abstract

The present application relates to the technical field of forecast correction, in particular to a multi-mode forecast correction method and system based on ensemble learning, comprising the following steps: S1. mapping the multi-mode original forecast field of a target area to a phase space to generate a phase space state vector including key elements; S2. calculating the state mutual coupling matrix between modes by a transfer entropy algorithm based on the phase space state vector of S1; S3. when the maximum entropy value in the state mutual coupling matrix of S2 exceeds a bit threshold value, triggering a phase space trajectory corrector to reconstruct a forecast trajectory and output a correction result, otherwise directly outputting an optimal ensemble of original forecasts. The present application effectively identifies systematic risks such as typhoon peripheral vortex pollution, frontal resonance error and other difficult-to-detect systematic risks by a single mode, and realizes scientific quantification of correction triggering.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of forecast correction, in particular to a multi-mode forecast correction method and system based on ensemble learning. BACKGROUND

[0002] With the development of numerical weather prediction technology, global and regional numerical models (ECMWF, GFS, CMA, etc.) are widely used in weather prediction. However, due to differences in initial conditions, physical process parameterization and resolution between different models, their output results often show inconsistencies, especially in severe weather systems such as strong convection, typhoon and frontal surface, the forecast bias is more significant. In order to improve the reliability of the forecast, the multi-model integrated correction method has become a research hotspot.

[0003] The existing multi-model fusion correction method mainly includes simple average, weighted average, Bayesian model average method, linear regression correction, etc. Most of the traditional methods are based on static weights or linear correlation, which is difficult to identify and quantify the dynamic causal relationship between models, especially in the case of system mutation or model resonance, it is difficult to effectively capture the evolution of coupling strength. The current commonly used methods are mostly based on threshold or empirical rule to determine whether to correct, lack of interpretable dynamic trigger mechanism, it is difficult to ensure that the correction scheme achieves a balance between physical consistency and numerical stability.

[0004] In weather systems with obvious spatial organization structure such as typhoon and frontal surface, the conventional ensemble prediction error not only reflects the numerical deviation, but also reflects the trajectory drift and structure mismatch, and the existing correction method usually cannot realize the targeted correction of these structure errors. SUMMARY

[0005] The present application provides a multi-mode forecast correction method and system based on ensemble learning, aiming to build a multi-mode forecast correction framework based on ensemble learning, and realize the organic integration of system-level coupling identification and physical consistency correction.

[0006] A multi-mode forecast correction method based on ensemble learning, comprising the following steps:

[0007] S1. Map the multi-mode original prediction field of the target area to the phase space to generate a phase space state vector including key elements; the key elements include temperature elements, humidity elements and wind field elements;

[0008] S2. Based on the phase space state vector of S1, calculate the state mutual coupling matrix between models by transfer entropy algorithm;

[0009] S3. When the maximum entropy value in the state mutual coupling matrix of S2 exceeds the bit threshold, trigger the phase space trajectory corrector to reconstruct the prediction trajectory and output the correction result, otherwise directly output the optimal ensemble of the original prediction.

[0010] Optionally, the key elements include temperature elements, humidity elements, and wind field elements, and the wind field elements include zonal wind and meridional wind.

[0011] Optionally, the S1 includes element field tensor construction and time lag embedding reconstruction.

[0012] The element field tensor construction specifically includes extracting temperature elements, humidity elements, zonal wind elements, and meridional wind elements in a target region from a plurality of numerical weather prediction models to form a plurality of element fields, uniformly organizing the extracted element fields according to spatial positions, element types, and model sources, and constructing a four-dimensional tensor data structure to represent original prediction information of the plurality of models.

[0013] The time lag embedding reconstruction specifically includes, for each grid point in the four-dimensional tensor data structure, using Takens embedding theory to extract a historical sequence over a period of time for the same grid position, and generating a time lag phase space vector group including a current state and a plurality of lag states.

[0014] Optionally, the S1 further includes attractor projection standardization, specifically including projecting the time lag phase space vector obtained by each time lag embedding reconstruction into a standard attractor manifold constructed in advance from historical observation data to generate a standardized phase space state vector, and the projection process is realized based on a nonlinear mapping method of Hilbert-Schmidt independence criterion.

[0015] Optionally, the S2 includes model state discretization and joint probability reconstruction.

[0016] The model state discretization includes discretizing the phase space state vector of each prediction model in the target region, and the discretization process calculates the number of states to be divided according to the historical observation entropy value, so that the state division can match the complexity of the weather system.

[0017] The joint probability reconstruction includes, after completing the state discretization, constructing a joint state transition statistical model between each pair of prediction models, selecting a combination of historical states of a group of target models and current state sequences of source models, and statistically calculating the frequency of occurrence of future states of the target models in each interval corresponding to the current state sequence combination, and normalizing the statistical frequency into a joint state transition probability.

[0018] Optionally, the transfer entropy algorithm includes calculating a transfer entropy value of the source model to the target model based on the joint state transition probability result, and the transfer entropy value is used to measure the degree of reduction of uncertainty of the future state of the target model by the source model, i.e., the strength of information coupling between the two.

[0019] Optionally, the S2 further comprises filling the transition entropy calculation results between all the prediction modes into a two-dimensional matrix according to the combination of the source mode and the target mode, to construct a complete state mutual coupling matrix, each element of the state mutual coupling matrix representing the information influence strength of one prediction mode on another prediction mode, and the diagonal elements being forced to be zero to eliminate self-coupling effects.

[0020] Optionally, the S3 specifically comprises:

[0021] S31, when the maximum entropy value of the state mutual coupling matrix satisfies the condition bit, positioning the corresponding spatial grid area, and determining the weather system type of the target area, representing the maximum entropy value of the state mutual coupling matrix;

[0022] If the spatial grid area satisfies: , it is determined as a typhoon system;

[0023] If the spatial grid area satisfies: , it is determined as a frontal system;

[0024] wherein, represents the 850 hPa relative vorticity, represents the temperature gradient, represents the 500 hPa frontal function;

[0025] S32, physical constraint reconstruction: according to the identified system type, different physical constraint correction strategies are executed:

[0026] S321, typhoon system correction: spiral conservation trajectory correction is executed, and the correction control equation is:

[0027] ;

[0028] wherein, represents the wind speed vector, represents the vortex vector, represents the external forcing term, represents the wind stress tensor, represents the entropy value threshold amount, represents the control path ring, represents the total derivative varying with the air mass;

[0029] S322, frontal system correction: baroclinic energy redistribution is executed, and the correction control equation is:

[0030] ; wherein, represents the effective potential energy, represents the thermal diffusion coefficient, represents the potential temperature, represents the entropy-energy coupling coefficient, and represents the vertical potential temperature gradient;

[0031] S33, the simulation result after trajectory correction is taken as a virtual new mode, and the prediction mode without triggering correction is integrated and output, and the variational integrated optimization is performed.

[0032] A multi-mode prediction correction system based on ensemble learning is used to realize the multi-mode prediction correction method based on ensemble learning, comprising the following modules:

[0033] A phase space construction module: the multi-mode original prediction field data of the target area is mapped to the phase space to generate a phase space state vector including temperature elements, humidity elements and wind field elements;

[0034] A coupling calculation module: based on the phase space state vector, a state mutual coupling matrix between modes is calculated by a transfer entropy algorithm;

[0035] A trajectory correction module: when the maximum entropy value in the state mutual coupling matrix exceeds a preset bit threshold, a trajectory correction operation is triggered to dynamically reconstruct the original prediction trajectory;

[0036] An integrated output module: after the trajectory correction is completed, the correction result is output, and if the correction is not triggered, the optimal integrated result of the original prediction is directly output.

[0037] The beneficial effects of the present application are:

[0038] The present application introduces a time delay embedding and attractor projection mechanism in the field of multi-mode ensemble prediction, reconstructs the state trajectory based on Takens embedding theory, and constructs a differential homeomorphism mapping function through an independence criterion, so that the states of each prediction mode are projected into the standard attractor manifold constructed by historical observations, and the structural distortion problem caused by the deviation of the original prediction between the mode systems is eliminated. The mechanism can dynamically adjust the embedding dimension and the structure of the mapping function, and maintain the physical coherence and topological consistency of the prediction field under different meteorological conditions.

[0039] The present application further introduces a transfer entropy theory on the basis of the traditional ensemble correction method, constructs a mutual coupling matrix based on multi-order state joint probability, and can capture the nonlinear dependence and causal information transmission relationship between modes. Through state discrete adaptive division, three-order joint indicator function statistics and diagonal suppression mechanism, the systematic risks such as typhoon peripheral vortex pollution and frontal resonance error which are difficult to be detected by a single mode are effectively identified, and the scientific quantization of correction triggering is realized.

[0040] ​The present application dynamically identifies the potential unstable area by comparing the maximum transfer entropy value of the state mutual coupling matrix with a set threshold value, and matches different physical correction equations based on the system type, introduces the helicity conservation trajectory adjustment for the typhoon system, and adopts the baroclinic energy redistribution mechanism for the frontal system, further introduces Delta TE as a modulation factor, realizes the continuous response of the correction strength and the inconsistency between the modes, and breaks through the traditional binary trigger mode. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description are only a part of the present application, and other drawings can also be obtained by those skilled in the art without any creative effort on the basis of these drawings.

[0042] Figure 1 The correction method flowchart of the embodiment of the present application is shown in the figure.

[0043] Figure 2 The correction system function module schematic diagram of the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0044] The present application will be described in detail below in combination with the drawings and specific embodiments. For some known technologies, other alternative ways can also be adopted by those skilled in the art to implement them; and the drawings are only used to more specifically describe the embodiments, and are not intended to specifically limit the present application.

[0045] As shown in the figure, a multi-mode forecast correction method based on ensemble learning includes the following steps: Figure 1

[0046] S1. Map the multi-mode original prediction field of the target area to the phase space to generate a phase space state vector including key elements; the key elements include temperature elements, humidity elements, wind field elements.

[0047] S1 specifically includes the following steps:

[0048] S11. The temperature field , humidity field , zonal wind field and meridional wind field of ECMWF model, GFS model and CMA model in the target area are composed into a four-dimensional tensor: ; wherein, represents the total number of spatial grids, 4 represents four elements , is the number of models.

[0049] ​In the forecast area, each numerical weather prediction model (ECMWF model, GFS model, CMA model) outputs a two-dimensional spatial field of multiple meteorological elements (temperature, humidity, wind field, etc.), that is, a grid data covering the entire geographical area. Each element field is understood as a matrix of size , where and represent the number of grids in the longitude and latitude directions, respectively. The specific construction process of the four-dimensional tensor is as follows:

[0050] S111. Classification by element: For each model, extract four key meteorological elements: temperature , humidity , zonal wind , and meridional wind . Each element is a two-dimensional matrix of size .

[0051] S112. Element dimension concatenation: Concatenate the four elements under the same model along the "element dimension" to form a three-dimensional tensor: ; this indicates that at each spatial grid point , it contains the values of 4 meteorological elements.

[0052] S113. Model dimension concatenation: Further concatenate the three-dimensional tensors constructed by the above different models along the "model dimension" to form the final four-dimensional tensor: ; where represents the number of models participating in integration. The final tensor records the numerical value of each model and each meteorological element at each position , serving as the basic data structure for subsequent phase space analysis and multi-model fusion.

[0053] This four-dimensional tensor structure realizes the systematic organization of the "spatial position-meteorological element-numerical model" three-dimensional dimension, which not only preserves the independence of different models, but also supports subsequent unified phase space embedding and correction processing by grid point.

[0054] S12, time lag embedding reconstruction: For each grid point , generate a time lag phase space vector according to the Takens embedding theorem:

[0055] ;

[0056] where represents the time lag embedding phase space state vector of grid point at time , represents the synthetic element vector of grid point at time , Optimal time lag, denoted as, is the embedding dimension, denoted as, is the embedding delay, denoted as, is the wind field embedding element (horizontal wind speed module), is calculated based on mutual information criterion, Attractor volume, denoted as, is the geometric volume occupied by the attractor in the state space, estimated based on box counting method, is the maximum Lyapunov exponent, denoted as, is obtained by Wolf algorithm, characterizing the divergence rate of the system to initial perturbation,

[0057] 1. Optimal time lag is calculated by mutual information method to calculate the degree of information redundancy between independent variables, for example, in the temperature element time series, the mutual information function :

[0058] ; where, ; in the mutual information calculation, denotes the joint probability distribution function of variables and , that is, the probability of variables taking values and at the same time in the time series. When , , denotes the joint probability of temperature being at the current time and being after hours, which is used to describe the time dependence between variables.

[0059] 2. Attractor volume is calculated: the first local minimum of is selected as the optimal time lag . The volume covered by the phase space trajectory is estimated by box counting method, the embedding vector is projected into high-dimensional space, divided into fixed size hypercubes (boxes), and the number of occupied boxes is counted to estimate the attractor dimension :

[0060] ; denotes the edge length in box counting, that is, when the state trajectory is divided into a cubic grid in the phase space, the size of each grid element.

[0061] 3. Lyapunov exponent is calculated: Wolf et al. algorithm is used to track any two initial similar states in the embedding space, and observe their divergence degree in the time evolution process, which is estimated by the following formula:​​​

[0062] ; the maximum of the multiple iterations is taken as the largest Lyapunov exponent.

[0063] The "time-delayed embedding reconstruction" is performed to transform the original time series of meteorological elements into the form of phase space trajectories, which can be used for dynamic analysis. The atmospheric system is essentially a high-dimensional nonlinear dynamic system, whose current state is not only affected by the current input, but also closely related to the past state. Through time-delayed embedding, an embedding space can be reconstructed, so that the nonlinear dynamic structure hidden in the original time series is explicit, thereby providing a basis for subsequent inter-model transition analysis and critical point identification.

[0064] The traditional multi-model correction method only considers the prediction error at a certain time section, ignoring the time series evolution characteristics. By introducing the state at multiple times (such as , ), the evolution trajectory of the system can be analyzed as a whole, so as to capture the trend, periodicity or mutation characteristics in the weather evolution process.

[0065] Embedding dimension and time delay are selected based on attractor theory and information theory criteria, avoiding subjective errors caused by artificial parameter setting, making the embedded trajectory more consistent with the topological structure of the original dynamic system, and helping to accurately restore the intrinsic dynamics of the weather system.

[0066] To establish a unified representation for attractor matching and correction, the multi-dimensional vector obtained by time-delayed embedding can be used as the input for subsequent attractor manifold projection and state matching, which is conducive to establishing a state alignment reference system between different numerical models, thereby improving the physical consistency and accuracy of multi-model fusion correction.

[0067] This step converts the original meteorological field from "point value sequence" to "dynamic trajectory", which is the key bridge between statistical prediction methods and nonlinear system theory.

[0068] S13, Attractor Projection: Project onto the reference attractor manifold constructed by historical observation data to generate a standardized phase space state vector: ; where represents the differential homeomorphism mapping function based on the Hilbert-Schmidt independence criterion (HSIC), which is used to eliminate systematic bias between multi-model data.

[0069] " Attractor projection" is a very critical step in the entire multi-model prediction correction method, and the essential purpose is to project the phase space trajectory , mapped (projected) into a reliable, standard reference space built from historical observations, to remove systematic biases between different models.

[0070] An "attractor manifold" is a collection of states that describe the long-term evolution of a dynamical system, revealing the typical evolution paths in the state space. The reference attractor manifold is built from a large number of historical observations; reflects the statistical regularities and dynamical characteristics of the evolution of various meteorological elements (temperature, humidity, wind speed, etc.) in the real atmosphere; and can be regarded as an "ideal state trajectory library" for correcting trajectory deviations produced by prediction models. Different prediction models (such as ECMWF, CMA, etc.) have certain systematic biases in their output of meteorological elements due to the use of different physical parameterization schemes.

[0071] For example: under the same initial conditions, a certain model may consistently overestimate humidity or underestimate wind speed. If these biased values are used directly for phase space analysis, it will affect the subsequent transition probability and critical point determination. Therefore, the phase space state vector output by the model needs to be projected into a "reliable reference system", i.e., the reference attractor manifold mentioned above, to remove the inherent systematic bias of the model.

[0072] Through a differential homeomorphism mapping , specifically using the mathematical method of Hilbert-Schmidt Independence Criterion (HSIC), to construct the projection function , HSIC is used to measure the nonlinear dependence between two variables, by optimizing a mapping so that the statistical distribution of the projected state vector and the historical attractor is most consistent; this process is to find a nonlinear coordinate transformation to "deform" the model prediction value into the distribution of historical observations. After projection, the obtained state vector is: , this vector:

[0073] retains the information of the original trajectory;

[0074] is corrected to a scale consistent with historical observations;

[0075] can be uniformly compared with other model outputs and historical evolution trajectories;

[0076] is the core input for subsequent calculation of transition probability matrix.

[0077] The differential homeomorphism mapping function of Hilbert-Schmidt Independence Criterion (HSIC) maps the phase space state vector of multiple models to the reference attractor manifold built from historical observations, obtaining the standardized state vector . This scheme specifically includes the following steps:

[0078] a) Establish the target mapping structure: This involves mapping the phase space state vector under the current mode. attractor trajectory samples corresponding to historical observation data As input-output pairs, a nonlinear mapping function to be learned is constructed. Its goal is to maintain the dynamic consistency of the state trajectory after projection, while aligning the statistical structures of different data sources.

[0079] b) Introduce HSIC as a dependency maximization criterion: to ensure the mapped state vector with target attractor sample Statistically consistent, the HSIC index is used as a measure of nonparametric dependence. HSIC measures whether two variables are independent in the kernel space, and is defined as:

[0080] ;in, For joint nuclear covariance operators, It represents the Hilbert-Schmidt norm; the larger the HSIC value, the stronger the Hilbert-Schmidt norm. and The greater the dependence (i.e., the more similar the structures), in this invention, .

[0081] c) Construct the kernel function and optimize the mapping: Use the Gaussian kernel function. As the basis for kernel mapping, the Gram matrix is ​​calculated for both the input sample and the target sample:

[0082] Input sample: ;

[0083] Target sample: ;

[0084] Based on this, the HSIC loss function is defined as follows: ,in: For a centralized matrix tr The objective of trace operation optimization is to minimize the loss function, and the parameterized solution is then obtained. This allows the mapped state to depend to the greatest extent possible on the historical attractor trajectory in the kernel space.

[0085] d) Implementation of the mapping structure: This mapping function A neural network based on a multilayer perceptron (MLP) structure is used; the activation function is ReLU or Tanh.

[0086] After initializing the weights, the HSIC loss function is iteratively optimized using gradient descent.

[0087] The final output is a reversible, continuous, and differentiable transformation function with the property of differential homeomorphism.

[0088] e) Through the above projection process, the phase space state vectors output by all modes are uniformly mapped into an observation-dominated phase space reference structure, with the following effects:

[0089] Eliminate systematic bias of different modes in temperature, humidity, wind speed, and other elements;

[0090] Restore the true physical structure of each mode prediction sequence.

[0091] S2. Based on the phase space state vector of S1, calculate the state mutual coupling matrix between modes by the transfer entropy algorithm.

[0092] S2 specifically includes the following steps:

[0093] S21, mode state discretization, for each prediction mode , its phase space state vector is discretized into state intervals according to the equal probability quantile, where: ; wherein, represents the observation entropy value calculated based on the historical observation sample, represents the number of state division intervals, dynamically matches the system complexity, represents the upward rounding operation.

[0094] Mode state discretization is used to convert continuous phase space state vectors into a finite number of discrete state intervals, thereby providing a basic semantic unit for subsequent probability calculation and information theory metrics.

[0095] Since the state vector output by the weather mode is continuous, in order to calculate the probability of state transition between modes, it is necessary to first divide these continuous values into countable discrete state categories. By equal probability quantile method, it is ensured that the number of samples in each state interval is close, avoiding some intervals being too sparse or too dense, and improving the stability of probability estimation.

[0096] The number of intervals Q in this step is dynamically calculated according to the historical observation entropy value, ensuring that the division accuracy adapts to the "information richness" of the system itself under different weather conditions. The calculation accuracy of the transfer entropy is matched with the complexity, without distortion or overfitting. Each mode state is divided into Q discrete intervals, and the transfer relationship between multiple modes is finally mapped to the information flow intensity (i.e. transfer entropy) between mode pairs, and these entropy values form the final mutual coupling matrix. Mode state discretization provides the counting basis for joint probability reconstruction of discrete state space in step b.

[0097] S22, joint probability reconstruction: calculate the joint probability of source mode to target mode Joint state transition probability:

[0098] ;

[0099] where, denotes the target mode at time , denotes the target mode history state sequence length , denotes the source mode current and lag state sequence length , denotes the interval of discretized state, is an indicator function, taking value 1 if condition is satisfied, otherwise 0, denotes the number of valid samples. In specific implementation, set , denotes the 3rd order history state of target mode and 2nd order state of source mode.

[0100] Traditional state transition modeling method is based on the first-order Markov assumption, i.e. future state only relates to current state. However, atmospheric system has obvious memory and cross-scale lag effect, future state of a mode is often influenced by its own history state and current or lag state of other modes. Therefore, multi-order state (i.e. , ) must be introduced to reconstruct the joint dependence structure closer to the real coupling mechanism.

[0101] Since the phase space state vector is continuous value, directly constructing joint probability density function has dimension disaster and sample sparsity problem in high-dimensional space, therefore, discretizing state vector into finite state set first, then estimating joint probability through counting method (i.e. indicator function statistics) is a practical and stable way.

[0102] Subsequent need to use transfer entropy to measure the causal coupling strength between different modes, the essence of transfer entropy is the difference of conditional information entropy, its calculation depends on the accurate estimation of joint and conditional probability. Only the complete joint probability containing is obtained, the conditional probability can be further derived and the transfer entropy can be calculated.

[0103] S23, transfer entropy calculation: according to the above joint state transition probability, calculate the transfer entropy from mode to :

[0104] ; used to measure the degree of uncertainty reduction of future state of source mode on the basis of known history state of target mode, i.e. the causal coupling strength between them.

[0105] S24, inter-coupling matrix generation: all the inter-mode transition entropy results are constructed into a state inter-coupling matrix:

[0106] ;

[0107] and the main diagonal elements are forced to be zero, that is, ; to exclude the pseudo-coupling effect caused by the memory of the mode itself.

[0108] The inter-coupling matrix is a two-dimensional structure that describes the state dependence strength and information transmission direction between multiple numerical prediction modes. Each element of it represents the influence degree of the source mode on the state evolution of the target mode, which is quantified by the transition entropy calculated in the previous step. The value in the pth row and qth column of the matrix represents the state coupling influence of mode on mode .

[0109] The matrix is asymmetric because the influence has directionality (for example, ECMWF's influence on CMA ≠ CMA's influence on ECMWF);

[0110] The diagonal line is forced to be zero, which aims to eliminate the historical memory of the mode itself to prevent mistaking the continuation within the mode as external coupling influence.

[0111] It can quantify the strength and direction of the interdependence between modes; provide criteria for subsequent judgment of whether to trigger critical point strengthening correction; support dynamic weighting or mode selection strategy in the process of model integration. This matrix is a "mode influence map" that reveals which mode is dominating, which mode is being affected, and the degree of information coupling between them.

[0112] S3. When the maximum entropy value in the state inter-coupling matrix of S2 exceeds the bit threshold, trigger the phase space trajectory corrector to reconstruct the prediction trajectory and output the correction result, otherwise directly output the original prediction optimal integration.

[0113] Typhoon is a typical low-level vortex system, and its most obvious structural feature is in the middle and lower troposphere, especially at 850 hPa, which is often considered as the most active layer of tropical cyclone activity. At 850 hPa, the wind field structure at the center of the typhoon is concentrated and the shear is small, which is convenient for accurately capturing the rotation intensity. Relative vorticity can effectively reflect the rotation of the low-level wind field and is an important indicator for measuring the development and intensity of typhoon.

[0114] Frontal system is usually formed in mid-latitude area, which is a typical baroclinic structure. The vertical structure is most obvious in the middle layer, and the 500-hPa layer is the most significant level of the intersection of cold and warm air masses and the baroclinicity of the frontal slope. Frontogenesis function can comprehensively reflect the temperature gradient, wind field convergence and inclination, and is the key dynamic index to judge the generation and maintenance of the frontal surface. This layer can effectively strip the influence of surface interference, making the identification process more stable and generalizable.

[0115] Therefore, S3 specifically comprises the following steps:

[0116] S31, system type identification: when the maximum transition entropy value of the state mutual coupling matrix meets the condition:

[0117] bit, the corresponding space grid area is located, and the weather system type of the area is judged;

[0118] If the area meets: , it is determined as a typhoon system;

[0119] If it meets: , it is determined as a frontal system.

[0120] Wherein, represents the 850-hPa relative vorticity, represents the temperature gradient, represents the 500-hPa frontogenesis function, represents the maximum transition entropy value in the state mutual coupling matrix.

[0121] Based on the coupling analysis of a large number of historical weather processes, it is found that when the inter-model transition entropy exceeds 2.5 bits, it corresponds to a significantly unstable area, such as the generation of typhoon, the outbreak of frontal surface and other severe weather evolution stages. The background transition entropy between models usually fluctuates between 1.0-2.0 bits, and 2.5 bits as a boundary can effectively distinguish between "normal inconsistency" and "physical structural anomaly", preventing excessive correction.

[0122] S32, physical constraint reconstruction: according to the identified system type, different physical constraint correction strategies are executed:

[0123] ① Typhoon system correction: helicity conservation trajectory correction: the correction control equation is:

[0124] ;

[0125] Wherein, represents the wind velocity vector, represents the vortex vector, represents the external forcing term, represents the wind stress tensor,​ represents the entropy value exceeding the threshold amount, represents the control path loop, represents the total derivative with respect to the air mass.

[0126] The modified control equation physically represents the rate of change of the dot product of the wind velocity vector and the vortex vector (i.e., the "helicity integral") along a control path over time, which is equal to the sum of two parts:

[0127] One part comes from the helicity change of the internal forcing term of the atmospheric system;

[0128] The other part is caused by the helicity term of the wind stress modulated by the coupling enhancement signal (i.e., ΔTE).

[0129] That is, "the evolution of the helical structure in the core area of the typhoon is not only determined by the traditional physical forcing term, but also affected by the uncertainty enhancement from multi-model coupling, which is quantified by the entropy value exceeding the threshold amount ΔTE and applied to the helicity structure of the wind stress field, thereby driving the path to make limited adjustments." The introduction of the helicity integral is based on the modeling of the essential characteristics of typhoons, which are strong rotation-dominated atmospheric structures whose movement paths are subject to the symmetry and conservation of the overall vortex structure. Therefore, compared with other systems, the prediction bias is more likely to accumulate in the destruction of the helical structure or the misjudgment of the forcing. Using the helicity integral as the core variable can reflect the state changes within the entire vortex ring, with global, continuous, and conservation constraints. The addition of the ΔTE term realizes the dynamic quantification of the correction strength, and ΔTE is the mode coupling strength exceeding the threshold value calculated by the inter-coupling matrix in the previous stage, representing the degree of prediction inconsistency or mutual interference between the current modes. Introducing ΔTE as a modulation coefficient into the wind stress term helps to make the correction strategy not only based on the physical state but also based on the risk identification results of multi-model collaborative diagnosis, enhancing the adaptability and credibility of the prediction.

[0130] The external forcing term and the wind stress term respectively bear two sources of power for path adjustment:

[0131] 1. The external forcing term reflects the helicity change caused by traditional heat, moisture, or terrain;

[0132] 2. The wind stress helicity term is introduced as a correction driving term weighted by ΔTE, which is used to correct the "trajectory drift" problem caused by system inconsistency.

[0133] Compared with the traditional method of relying only on physical numerical output for path correction, this equation integrates system identification, strength quantification, physical regulation, and path feedback in a complete chain.

[0134] ② Frontal system correction: baroclinic energy redistribution: the second modified control equation is:

[0135] ; wherein, represents the effective potential energy, represents the thermal diffusion coefficient, represents the potential temperature, represents the entropy-energy coupling coefficient, represents the vertical potential temperature gradient.

[0136] S33, result set integration output: the simulation results after trajectory correction are taken as a virtual new mode, and the original mode is integrated and output. The following variational integration optimization is performed:

[0137] ;

[0138] wherein, represents the forecast output of the th mode, represents the corrected trajectory (a new virtual mode), represents the observation value vector, represents the mode weight vector, represents the smoothing constraint coefficient, represents the L1 regularization of the weight space gradient, which is used to suppress resonance error and improve integration stability.

[0139] The modified control equation two physically represents the evolution process of the effective potential energy (E) over time, which is jointly affected by three mechanisms:

[0140] The advection effect of wind speed on energy distribution (the first term): represents the spatial movement of effective potential energy that occurs with wind flow, which dominates the horizontal propagation and development of the frontal system.

[0141] Energy redistribution caused by thermal diffusion (the second term): describes the local energy dispersion caused by the diffusion of heat in the atmosphere, which helps to simulate the evolution of the energy tail of the rear of the front.

[0142] Vertical energy injection caused by system instability enhancement (the third term): this term is the introduced term of the present application, which consists of two parts:

[0143] 1. Vertical potential temperature gradient, which represents the instability degree of the atmospheric stratification and is the physical basis of the frontal structure;

[0144] 2. Transferred entropy threshold value (ΔTE), which reflects the coupling enhancement of the current numerical modes in this region, i.e., the enhancement of system uncertainty.

[0145] The product of the two is modulated by the coefficient β, thereby realizing an energy correction mechanism based on the inconsistency strength of the modes.

[0146] Frontal system is essentially a typical baroclinic unstable structure, and the energy accumulation and release in frontal zone depends on the effective potential energy formed by the vertical temperature gradient, so the regulation of effective potential energy plays a key role in the prediction of frontal path and intensity. The equation is exactly the core variable of effective potential energy, which accurately cuts into the physical core of the evolution of frontal system. The introduction of Delta TE reflects the feedback of system coupling information on energy structure, and after the strong coupling area between multiple modes is identified by the mutual coupling matrix, Delta TE is introduced as a regulating term, so that the physical correction is not only limited to the meteorological field itself, but also can respond to the inconsistency of numerical mode. In the frontal system, this feedback mechanism can effectively avoid the false energy enhancement or weakening caused by mode resonance.

[0147] The traditional scheme often ignores the vertical disturbance error caused by multi-mode inconsistency, resulting in a wide frontal surface and a blurred path. The joint action of Delta TE and the vertical potential temperature gradient in the present application can more realistically reflect the driving force of the current stratification instability on the evolution of the front, improving the accuracy of frontal front positioning and intensity prediction. Beta, as the entropy energy coupling coefficient, can be optimized and set according to historical return data to ensure that Delta TE is neither too aggressive nor ineffective in different weather types.

[0148] As shown in Figure 2 , a multi-mode prediction correction system based on ensemble learning is used to implement the above method, comprising the following modules:

[0149] Phase space construction module: mapping the multi-mode original prediction field data of the target area to the phase space to generate a phase space state vector including temperature elements, humidity elements and wind field elements;

[0150] Coupling calculation module: calculating the state mutual coupling matrix between modes based on the phase space state vector through the transfer entropy algorithm;

[0151] Trajectory correction module: when the maximum entropy value in the state mutual coupling matrix exceeds the preset bit threshold, triggering trajectory correction operation to dynamically reconstruct the original prediction trajectory;

[0152] Integrated output module: after the trajectory correction is completed, output the correction result, if the correction is not triggered, directly output the optimal integrated result of the original prediction.

[0153] The present application covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present application. In order to make the public have a thorough understanding of the present application, specific details are described in the following preferred embodiments of the present application, and the present application can also be completely understood without the description of these details for those skilled in the art. In addition, in order to avoid unnecessary confusion to the essence of the present application, well-known methods, processes, procedures, elements and circuits are not described in detail.

[0154] The above merely describes the preferred embodiments of the present application, and it should be pointed out that, for those skilled in the art, some improvements and refinements can be made without departing from the principles of the present application, and these improvements and refinements should also be considered as falling within the protection scope of the present application.

Claims

1. A multi-model forecast correction method based on ensemble learning, characterized in that, The method comprises the following steps: S1. mapping a multi-mode original prediction field of a target area to a phase space to generate a phase space state vector comprising key elements; the key elements comprise temperature elements, humidity elements, and wind field elements; S2. calculating a state mutual coupling matrix between modes by a transfer entropy algorithm based on the phase space state vector of S1; S3. when a maximum entropy value in the state mutual coupling matrix of S2 exceeds a bit threshold value, triggering a phase space trajectory corrector to reconstruct a prediction trajectory and output a correction result, otherwise directly outputting an original prediction optimal integration, specifically comprising: S31, when the maximum entropy value of the state mutual coupling matrix satisfies the condition bit, positioning The corresponding space grid area, and judge the target area weather system type, Indicates the maximum entropy value of the state mutual coupling matrix; If the space grid area satisfies: ; then it is determined as a typhoon system; If the spatial grid region satisfies: ; then the front system is determined. wherein, represents 850 hPa relative vorticity, represents temperature gradient, represents 500 hPa frontogenesis function; S32. physical constraint reconstruction: according to the identified weather system type, different physical constraint correction strategies are executed: S321. typhoon system correction: spiral conservation trajectory correction is executed, and the correction control equation is: ； wherein, represents a wind velocity vector, represents a vorticity vector, represents an external forcing term, represents a wind stress tensor, represents an entropy value threshold quantity, represents a control path loop, represents a total derivative with respect to the air mass; S322. frontal system correction: baroclinic energy redistribution is executed, and the correction control equation is: ; where, represents the effective bit energy, represents the thermal diffusivity, represents the bit temperature, represents the entropy-energy coupling coefficient, and , represents the vertical bit temperature gradient; S33. taking the simulation result after trajectory correction as a virtual new mode, integrating the prediction mode without triggering correction, and executing variational integration optimization. 2.The multi-model forecast adjustment method based on ensemble learning according to claim 1, wherein, The key elements comprise temperature elements, humidity elements, and wind field elements, and the wind field elements comprise zonal wind and meridional wind. 3.The multi-model forecast correction method based on ensemble learning according to claim 2, wherein, The S1 comprises element field tensor construction and time lag embedding reconstruction. The element field tensor construction specifically comprises extracting temperature elements, humidity elements, zonal wind elements and meridional wind elements in a target area from a plurality of numerical weather prediction modes to form a plurality of element fields, uniformly organizing the extracted element fields according to spatial positions, element types and mode sources to construct a four-dimensional tensor data structure representing multi-mode original prediction information. The time lag embedding reconstruction specifically comprises, for each grid point in the four-dimensional tensor data structure, using Takens embedding theory to extract a historical sequence over a period of time for the same grid position to generate a time lag phase space vector group comprising a current state and a plurality of lag states.

4. The multi-model forecast adjustment method based on ensemble learning according to claim 3, wherein, The S1 further comprises attractor projection standardization, specifically comprising projecting each time lag embedding reconstruction obtained time lag phase space vector into a standard attractor manifold constructed from historical observation data to generate a standardized phase space state vector, and the projection process is realized based on a nonlinear mapping method of Hilbert-Schmidt independence criterion.

5. The multi-model forecast adjustment method based on ensemble learning according to claim 1, wherein, The S2 comprises mode state discretization and joint probability reconstruction. The mode state discretization comprises discretizing the phase space state vector of each prediction mode in the target area, and the discretization process calculates the number of states to be divided according to historical observation entropy values, so that the state division can match the complexity of the weather system; The joint probability reconstruction comprises constructing a joint state transition statistical model between each pair of prediction modes after completing state discretization, selecting a group of historical states of target modes and current state sequences of source modes, statistically calculating the frequency of future states of the target modes corresponding to the current state sequence combination occurring in each interval, and normalizing the statistical frequency to a joint state transition probability.

6. The multi-model forecast adjustment method based on ensemble learning according to claim 5, wherein, The transfer entropy algorithm includes calculating a transfer entropy value of the source mode to the target mode based on a joint state transition probability result, and the transfer entropy value is used to measure a degree of uncertainty reduction of a future state of the source mode to the target mode, that is, a strength of information coupling between the source mode and the target mode.

7. The ensemble learning based multi-model forecast adjustment method of claim 6, wherein, The S2 further includes filling the calculation results of the transfer entropy between all the prediction modes into a two-dimensional matrix in a combination of the source mode and the target mode, and constructing a complete state mutual coupling matrix, each element of the state mutual coupling matrix representing an information influence strength of one prediction mode on another prediction mode, and diagonal elements being forced to be zero to eliminate self-coupling effects.

8. An ensemble learning based multi-model forecast adjustment system for implementing an ensemble learning based multi-model forecast adjustment method according to any one of claims 1-7, characterized in that, The method comprises the following modules: A phase space construction module: mapping multi-mode original prediction field data of a target area to a phase space to generate a phase space state vector including temperature elements, humidity elements and wind field elements; A coupling calculation module: calculating a state mutual coupling matrix between modes by a transfer entropy algorithm based on the phase space state vector; A trajectory correction module: triggering a trajectory correction operation to dynamically reconstruct an original prediction trajectory when a maximum entropy value in the state mutual coupling matrix exceeds a preset bit threshold; An integrated output module: outputting a correction result after the trajectory correction is completed, and directly outputting an optimal integrated result of the original prediction if the correction is not triggered.

Images