Domain adaptation weather forecasting method based on multi-intelligent large model fusion

Abstract

The application provides a field adaptation weather forecast method based on multi-intelligent large model fusion, generates a forecast result by using multiple AI weather models based on the same GFS initial field, solves an optimal fusion weight by constructing a fusion model and an optimization objective function containing a data fitting term and a physical constraint term, adopts a differentiated fusion strategy according to different forecast time periods, dynamically adjusts the weight in combination with historical forecast performance and real-time observation data, and outputs a forecast product after quality control. The application has the beneficial effects that the advantages of each model are fully utilized, the forecast precision is improved, the field adaptability and physical consistency are enhanced, the forecast stability is improved, full-time coverage is realized, the calculation efficiency is high, and the business demand is met.

Description

Technical Field

[0001] This invention belongs to the field of meteorological forecasting technology, and in particular relates to a domain-adaptive meteorological forecasting method based on the fusion of multiple intelligent large models. Background Technology

[0002] With the rapid development of artificial intelligence technology, deep learning-based AI weather forecasting models such as FengWu, FourCastNet, FuXi, and Pangu have made breakthrough progress in the field of weather forecasting. These AI models, by deeply mining massive amounts of historical meteorological data, can quickly generate global weather forecast products, demonstrating potential to surpass traditional numerical models in terms of computational efficiency and forecast accuracy. However, existing technologies have the following problems: Limitations of Single Model Performance: Due to differences in training data sources, network architecture design, and optimization strategies, different AI meteorological models exhibit significant heterogeneity in forecast performance across different forecast timeframes (short-term, medium-term, and long-term), meteorological elements (temperature, precipitation, wind speed), and geographical regions. A single model struggles to maintain optimal forecast performance across all scenarios. Insufficient Domain Adaptability: While existing general-purpose AI meteorological models can provide global forecast products, they exhibit significant shortcomings in adaptability to specific business domains, such as refined short-term forecasts, stable medium-term forecasts, and long-term climate trend predictions. This makes it difficult to meet the differentiated needs of scenarios like refined forecasts and severe weather warnings. Lack of Physical Consistency: Purely data-driven AI models often generate non-physical interpretations due to a lack of atmospheric dynamic constraints. Insufficient characterization of atmospheric physical processes leads to physical inconsistencies in forecast results, severely restricting the physical reliability and business application value of forecasts. Insufficient forecast stability: The forecast results of a single model are easily affected by factors such as training data distribution shifts and fluctuations in network parameter initialization, resulting in insufficient time series stability of the forecast results, making it difficult to meet the stringent requirements of meteorological operations for forecast continuity and reliability. Limitations in the effectiveness of fusion methods: Current multi-model fusion methods mostly rely on simple weighted averaging or voting mechanisms, lacking dynamic adaptive weight learning mechanisms and physical process constraints. This makes it difficult to achieve dynamic complementarity of the advantages of each model, and also fails to ensure the physical rationality of the fusion results and the improvement of forecast accuracy. Summary of the Invention

[0003] In view of this, the present invention aims to propose a domain-adaptive weather forecasting method based on the fusion of multiple intelligent large models, so as to solve at least one of the problems existing in the above-mentioned prior art.

[0004] To achieve the above objectives, the technical solution of the present invention is implemented as follows: A domain-adaptive weather forecasting method based on multi-intelligent large model fusion includes the following steps: S1. Based on the initial field data of the unified global forecast system, multiple AI meteorological models are used to make weather forecasts and generate a set of forecast results. S2. Based on the set of forecast results, construct a fusion forecast model: Based on the fusion weight, sum the forecast values ​​of each model by weight, and then add the bias correction term to obtain the fusion forecast result; S3. Based on data fitting and physical constraints, construct an objective function to optimize the fusion weights and bias correction terms; S4. Using a constrained optimization algorithm, minimize the objective function and solve for the fusion weights and bias correction terms; S5. Adopt differentiated fusion strategies according to different forecast lead periods; S6. Dynamically optimize the fusion weights based on historical forecast performance and real-time observation data; S7. Perform quality control and physical consistency checks on the fused forecast results.

[0005] Furthermore, in step S1, the meteorological elements predicted by the model include temperature, humidity, zonal wind, meridional wind, and geopotential. The model's prediction results are in the GRIB2 standard format, with a spatial resolution of 0.25°×0.25° and 13 vertical layers.

[0006] Furthermore, in step S2, the expression for the fused forecast result is as follows: ; In the formula, The optimal forecast result after fusion. The total number of models participating in the fusion. Let p be the fusion weight coefficient of the i-th model element. Let be the predicted value of element p for the i-th model at time t. This is the deviation correction term for element p.

[0007] Furthermore, in step S3, the expression for the objective function is as follows: ; In the formula, For data fitting terms, For physical constraint terms, , These are the weighting coefficients. This is the result of GFS numerical weather prediction. This is the physical consistency constraint function.

[0008] Furthermore, in step S4, the fusion weights and bias correction terms are solved, with the following expressions: ; In the formula, Let be the optimal weight of the i-th model for element p. This is the deviation correction term for the optimal element p.

[0009] Furthermore, in step S5, the differentiated fusion strategy includes: For short-term forecasts with high temporal resolution, the weights are biased towards models with excellent short-term performance, with less weight given to physical constraints, prioritizing accuracy. For medium-term forecasts with medium temporal resolution, the weights are evenly distributed to balance accuracy and stability. For long-term forecasts with low temporal resolution, the weights are biased towards models with long-term stability, with more weight given to physical constraints, prioritizing the reasonableness of the trend.

[0010] Furthermore, in step S6, the fusion weights are dynamically optimized, as shown in the following expression: ; In the formula, Let be the final dynamic fusion weight of the i-th model for element p at time t. Based on the weighting coefficient, For weighting adjustment factors based on historical forecasting techniques, This is a weighting adjustment factor based on real-time observation and comparison; ; In the formula, For the current moment, The moment when the i-th model performs optimally. Standard deviation; ; In the formula, Let i be the current forecast value of the i-th model. The current actual observation value, This represents the standard deviation of the observation error.

[0011] Furthermore, in step S7, quality control and physical consistency checks are performed on the fused forecast results, including: Outlier detection: Identify and correct forecast values ​​that exceed reasonable ranges; Physical consistency verification: Check whether the fusion results meet physical constraints; Spatial continuity check: Ensure that the forecast results are spatially continuous; Temporal consistency check: Ensure that the forecast results are temporally continuous.

[0012] Furthermore, in step S3, the physical constraints include geostrophic wind balance constraints and potential vortex conservation constraints. The expression is as follows: ; In the formula, For Coriolis parameters, It is a unit vector in the vertical direction. This is the actual wind field vector. This is the geostrophic wind field vector; Potential vortex conservation constraint The expression is as follows: ; In the formula, For potential vortex, It is the total derivative.

[0013] Compared with existing technologies, the domain-adaptive weather forecasting method based on multi-intelligent large model fusion described in this invention has the following beneficial effects: (1) The forecast accuracy has been significantly improved. By using a multi-model complementary fusion strategy, the technical strengths of each AI meteorological model are fully utilized. Compared with a single model, the ACC (Anomaly Correlation Coefficient) of the fusion forecast is increased by more than 5%, and the RMSE (Root Mean Square Error) is reduced by more than 8%, achieving a substantial breakthrough in forecast accuracy.

[0014] (2) It has outstanding domain adaptability. It adopts a dual strategy of time-segmented fusion and spatial multi-scale fusion, which can adopt differentiated fusion schemes for different forecast time segments and geographical features to achieve accurate cross-scenario adaptation and meet the specific needs of diverse business scenarios such as refined forecasting and disaster early warning.

[0015] (3) Ensure consistency of physical processes. The geostrophic wind balance constraint and potential vortex conservation constraint mechanism are introduced. All constraint terms are constructed based on core meteorological elements and their spatiotemporal derivatives to ensure that the fusion results strictly follow the laws of atmospheric physics, effectively avoid non-physical generation, and significantly improve the physical reliability and operational application value of the forecast results.

[0016] (4) The stability of the forecast results is significantly enhanced. The systematic bias of the single model is effectively diluted by the synergistic fusion of multiple models, which greatly improves the time series stability and robustness of the forecast results. The system’s annual normal operation rate is ≥99.8%, and the forecast product integrity rate is ≥99.95%, meeting the business continuity requirements.

[0017] (5) Dynamic adaptive optimization capability: Equipped with an adaptive weight optimization engine and a real-time feedback adjustment mechanism, it can dynamically adjust the fusion weight allocation based on historical forecast performance and real-time observation data, so as to achieve continuous iterative optimization of forecast performance and ensure the long-term applicability of the model.

[0018] (6) Full-time coverage capability: Through the spatiotemporal splicing and fusion of multi-model forecast results, it achieves seamless full-time coverage from 000 to 360 hours, meeting the timeliness requirements of different business scenarios.

[0019] (7) High computational efficiency. Compared with parallel computing of multiple models running independently, the fusion method only needs to perform post-processing fusion of the output results of each model, reducing the consumption of computing resources by more than 60%. It can complete the full-domain fusion computing within 30 minutes to meet the real-time response requirements of the business side. Attached Figure Description

[0020] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall process described in an embodiment of the present invention; Figure 2 This is a schematic diagram of the multi-model fusion architecture described in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the construction of the physical constraint objective function according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the time-segmented fusion strategy described in an embodiment of the present invention; Figure 5 This is a schematic diagram of the adaptive weight optimization process described in an embodiment of the present invention; Figure 6 This is a schematic diagram of the quality control and product output process according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the business operation cycle process described in an embodiment of the present invention; Figure 8 This is a schematic diagram of the RMSE evaluation of the 2-meter temperature fusion forecast according to an embodiment of the present invention. Detailed Implementation

[0021] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0022] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0023] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art will understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0024] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0025] like Figures 1 to 7 As shown, the domain-adaptive weather forecasting method based on multi-intelligent large model fusion has the following specific steps: S1. Based on the initial field data of the unified global forecast system, multiple AI meteorological models are used to make weather forecasts and generate a set of forecast results. S2. Based on the set of forecast results, construct a fusion forecast model: Based on the fusion weight, sum the forecast values ​​of each model by weight, and then add the bias correction term to obtain the fusion forecast result; S3. Based on data fitting and physical constraints, construct an objective function to optimize the fusion weights and bias correction terms; S4. Using a constrained optimization algorithm, minimize the objective function and solve for the fusion weights and bias correction terms; S5. Adopt differentiated fusion strategies according to different forecast lead periods; S6. Dynamically optimize the fusion weights based on historical forecast performance and real-time observation data; S7. Perform quality control and physical consistency checks on the fused forecast results.

[0026] In a preferred embodiment of the present invention, in step S1, based on the same GFS initial field, N AI meteorological models are used to make forecasts respectively, resulting in a set of forecast results from N models. The expression is as follows: ; In the formula, Let represent the forecast result of the i-th AI model at time t, where i∈[1,N], N≥2. The forecast result of each AI model includes core meteorological elements: temperature T, humidity q or relative humidity RH, zonal wind U, meridional wind V, geopotential height Z or geopotential Φ.

[0027] In this embodiment, the N AI meteorological models include, but are not limited to: FengWu model: 000-240 hour forecast data, 6-hour interval; FengWu v2 model: 000-240 hour forecast data, 6-hour interval; FourCastNet v2 model: 000-240 hour forecast data, 6-hour interval; FuXi model: 000-360 hour forecast data, 6-hour interval; Pangu model: 000-192 hour forecast data, 6-hour interval.

[0028] All AI model forecasts were performed in the GRIB2 standard format with a spatial resolution of 0.25°×0.25°, a global grid of 1440×721 points, and 13 vertical layers: 1000, 925, 850, 700, 600, 500, 400, 300, 250, 200, 150, 100, and 50 hPa.

[0029] In a preferred embodiment of the present invention, the specific operation flow of step S1 is as follows: (1) Obtain initial field data from the Global Forecasting System (GFS), wherein the initial field data includes, but is not limited to, three-dimensional spatial distribution data of meteorological elements such as temperature field, humidity field, wind field, geopotential height field, etc.; (2) The initial field data is format-converted and preprocessed to adapt it to the input requirements of each AI meteorological model. Specifically, the GFS initial field data is converted into the grid format, variable naming conventions, and unit system required by each model; (3) The initial field data after format conversion is used as input to drive multiple independently trained meteorological models to predict short, medium and long-term meteorological elements respectively; (4) Output the forecast results of each model for key variables such as temperature, humidity, wind field and precipitation, and form a forecast result set MAIM^{AI}MAI.

[0030] It should be noted that the above list of models is only an example. Those skilled in the art can select other AI meteorological models to participate in the fusion according to actual needs, as long as the model can generate meteorological forecast results based on the GFS initial field.

[0031] In a preferred embodiment of the present invention, in step S2, a fusion forecast model is constructed. The expression for the fusion forecast result is as follows: ; In the formula: This indicates the best forecast result after fusion, and p represents the forecast elements, including core meteorological elements such as temperature T, humidity q or relative humidity RH, zonal wind U, meridional wind V, and geopotential height Z. This represents the predicted value of the i-th AI model for element p at time t; This represents the fusion weight coefficient of the i-th AI model for feature p, satisfying the following constraints: , ; This represents the deviation correction term for element p.

[0032] In this embodiment, the fusion weighting coefficient The initial values ​​can be allocated using an equal-weighted method, i.e. Then, the optimal weights are solved through subsequent optimization steps. (Bias correction term) Used to correct systematic biases, its initial value can be set to zero or obtained through historical forecast error statistics. It should be noted that the above fusion expression uses a linear weighted fusion method. Those skilled in the art can also use nonlinear fusion methods, such as neural network fusion or support vector regression fusion, as long as effective fusion of multi-model forecast results can be achieved.

[0033] In a preferred embodiment of the present invention, in step S3, an optimization objective function including physical constraints is constructed. The expression is as follows: ; In the formula: the first term is the data fitting term: This indicates the difference between the fused forecast results and the GFS forecast field. The first term is the GFS numerical prediction result, serving as the physical constraint baseline; the second term is the physical constraint term: This represents physical consistency constraints, constructed for core meteorological elements, including: Geostrophic wind balance constraints: Constraining the actual wind field based on the relationship between geopotential height gradient and horizontal wind speed. The deviation from the geostrophic wind field is expressed as follows: ; In the formula, For Coriolis parameters, It is a unit vector in the vertical direction. This is the actual wind field vector. Let be the geostrophic wind field vector. The physical significance of the geostrophic wind balance constraint lies in the fact that, in large-scale motion at mid-to-high latitudes, the actual wind field should approximately satisfy the geostrophic balance relationship, that is, the Coriolis force and the pressure gradient force are basically in balance. By introducing this constraint, unreasonable wind field structures appearing in the fusion results can be effectively suppressed.

[0034] Potential vortex conservation constraint: Based on the conservation of potential vortex, the physical rationality of atmospheric dynamic processes is ensured, as expressed below: ; In the formula, For potential vortex, It is the total derivative; , The weighting coefficients are used to balance the importance of different constraint terms. The physical significance of the potential vortex conservation constraint lies in the fact that, under adiabatic and frictionless conditions, potential vortices are conserved along the airflow trajectory. By introducing this constraint, we can ensure that the fusion results have physical rationality in terms of atmospheric dynamic processes and avoid the occurrence of non-physical potential vortex anomalies.

[0035] In a preferred embodiment of the present invention, the weighting coefficient and The value can be adjusted according to the forecast lead time. Specifically: For short-term forecasts of 0-120 hours, Relatively large The size is relatively small, so priority should be given to ensuring forecast accuracy; For medium-term forecasts of 120-240 hours, and Relatively balanced, with balance accuracy consistent with physical properties; For long-term forecasts of 240-360 hours, The size is relatively large, so priority should be given to ensuring physical rationality and trend stability.

[0036] In a preferred embodiment of the present invention, in step S4, the objective function is minimized. Solve for the optimal fusion weight coefficients. The expression is as follows: ; In the formula, Let be the optimal weight of the i-th model for element p. This is the deviation correction term for the optimal element p; The optimization process employs constrained optimization algorithms, including but not limited to the Lagrange multiplier method, projected gradient method, interior point method, and sequential quadratic programming (SQP), while also satisfying weight normalization constraints. , .

[0037] like Figure 5 As shown, the adaptive weight optimization process includes the following steps: (1) Initialize weights ; (2) Calculate the objective function ; (3) Calculate the gradient ; (4) Solve the quadratic programming subproblem; (5) Update weights ; (6) Convergence judgment: If the condition is met, the optimal weight is output; otherwise, return to step (2) to continue the iteration.

[0038] In a preferred embodiment of the present invention, the optimization process employs a constrained optimization algorithm, including but not limited to: (1) Lagrange multiplier method: By introducing Lagrange multipliers, the constrained optimization problem is transformed into an unconstrained optimization problem; (2) Projection gradient method: During gradient descent, the updated weights are projected onto the feasible region to ensure that the constraints are met; (3) Interior point method: By introducing obstacle functions, the constraints are incorporated into the objective function, and the unconstrained optimization problem is solved iteratively; (4) Sequential Quadratic Programming (SQP): The original problem is transformed into a series of quadratic programming subproblems for solution.

[0039] The above optimization algorithm also satisfies the weight normalization constraint: .

[0040] In one specific embodiment of the present invention, sequential quadratic programming (SQP) is used for optimization, and the specific steps are as follows: (1) Set initial weights Initial deviation correction term ; (2) Calculate the current objective function value and gradient ; (3) Constructing a quadratic programming subproblem: ; in This is an approximation of the objective function using the Hessian matrix. (4) Solve the quadratic programming subproblem to obtain the weight update. ; (5) Update weights: , ; (6) Determine the convergence condition: If ,in To preset the convergence threshold, it is usually set to 10. -6 If the result is positive, then output the optimal weight; otherwise, let... Return to step (2).

[0041] It should be noted that the above optimization algorithm is only an example. Those skilled in the art can choose other optimization algorithms according to actual needs, as long as they can solve the constrained optimization problem.

[0042] In a preferred embodiment of the present invention, in step S5, a differentiated fusion strategy is adopted according to different forecast lead times.

[0043] like Figure 4 As shown, the time-leading segmented fusion strategy divides the fusion process into three stages based on the forecast lead time: a short-term forecast strategy of 0-120 hours, a 1-hour interval, and... ; 120-240 hour medium-range forecast strategy, 3-hour interval. ; Long-term forecast strategy of 240-360 hours, 6-hour interval, ; Output short-term, medium-term, and long-term products respectively.

[0044] Specifically, differentiated integration strategies include: Short-term forecasts (0-120 hours): High temporal resolution fusion with 1-hour intervals is used to optimize short-term forecast accuracy. Within this lead time range: The weights are biased towards models that perform well in short-term forecasts. Models with good short-term forecast performance, such as FengWu and FourCastNet, are given higher fusion weights. Physical constraint weighting coefficient The value is relatively small, with 0.05 being the preferred value to prioritize forecast accuracy. It has a high temporal resolution, enabling it to capture rapidly changing weather processes.

[0045] 120-240 hour medium-range forecasts: Employing fusion with a 3-hour interval at medium temporal resolution to balance forecast accuracy and stability within this lead timeframe: The weights are relatively balanced, taking into account the prediction performance of each model. Physical constraint weighting coefficient A moderate value is selected, with 0.1 being the preferred value, ensuring balance accuracy and physical consistency. It has a moderate time resolution, balancing computational efficiency and forecast accuracy.

[0046] 240-360 hour long-term forecast: Employing low temporal resolution fusion at 6-hour intervals, the focus is on ensuring the accuracy of the forecast trend within this timeframe: The weights are biased towards models with stable long-term forecast performance. Specifically, models with good long-term forecast performance, such as FuXi, are assigned higher fusion weights. Physical constraint weighting coefficient The value is relatively large, with a preferred value of 0.2, prioritizing the rationality of the trend and physical consistency. With lower temporal resolution, it focuses on capturing large-scale climate trends.

[0047] It should be noted that the above-mentioned time period division and parameter values ​​are only preferred embodiments, and those skilled in the art can make adjustments according to actual business needs and model performance characteristics.

[0048] In a preferred embodiment of the present invention, in step S6, the fusion weights are dynamically adjusted based on historical forecast performance and real-time observation data. The expression is as follows: ; In the formula: These are the basic weight coefficients, obtained through training with historical data; The weighting adjustment factor based on historical forecasting techniques is expressed as follows: ; In the formula, in the formula, For the current moment, The moment when the i-th model performs optimally. The standard deviation is denoted as σ. The physical meaning of this adjustment factor is that the prediction weight is higher near the optimal time of the model's performance, and the weight gradually decreases as the model moves away from the optimal time, exhibiting a Gaussian distribution characteristic.

[0049] The weight adjustment factor based on real-time observation comparison is expressed as follows: ; In the formula, Let i be the current forecast value of the i-th model. The current actual observation value, This represents the standard deviation of the observation error. The physical meaning of this adjustment factor is: the closer the model's predicted value is to the actual observed value, the higher its weight; the greater the deviation, the lower its weight.

[0050] In a preferred embodiment of the present invention, the specific operation process of dynamic weight adjustment is as follows: (1) Obtain historical forecast data, statistically analyze the forecast errors of each model in different time periods, and determine the optimal performance time of each model. and standard deviation ; (2) Obtain real-time observation data and calculate the deviation between the current forecast value and the observed value of each model; (3) Calculate according to the above formula and ; (4) Combining the basic weight coefficient Calculate the final dynamic fusion weights ; (5) Normalize the dynamic fusion weights to ensure .

[0051] It should be noted that the above dynamic weight adjustment mechanism can be extended according to actual business needs, such as introducing more adjustment factors, such as model uncertainty factors and regional adaptability factors, to further improve the accuracy and stability of fusion forecasts.

[0052] In a preferred embodiment of the present invention, in step S7, the fusion forecast results are subjected to quality control and physical consistency checks, such as... Figure 6 As shown, the quality control process includes: after outlier detection, the fused forecast results are determined to be within a reasonable range; if not, the outliers are corrected. If so, physical consistency verification is performed, including geostrophic wind balance and potential vortex conservation checks, to determine if they pass; if not, corrections are made. If successful, spatial continuity and temporal consistency checks are performed, and finally, the final product in GRIB2 format is output. Specifically, the quality control and physical consistency checks include the following: Outlier detection: Identifying and correcting forecast values ​​that exceed reasonable ranges; specifically: (1) Based on the physical characteristics of meteorological elements, set reasonable ranges for each element. For example, the reasonable range for temperature at 2 meters is usually -90°C to 60°C, and the reasonable range for relative humidity is 0% to 100%. (2) Check the fused forecast results grid by grid to identify outliers that exceed the reasonable range; (3) Correct outliers. Correction methods include, but are not limited to: replacing with interpolation results of adjacent grid points, replacing with climatological average values, replacing with forecast results from the previous time period, etc.

[0053] Physical consistency verification: Check whether the fusion result meets the physical constraints, specifically including: (1) Geostrophic wind balance check: Calculate the deviation between the actual wind field and the geostrophic wind field in the fusion result. If the deviation exceeds the preset threshold, it will be corrected. (2) Potential vortex conservation check: Calculate the rate of change of potential vortex over time in the fusion result. If the rate of change exceeds the preset threshold, then make corrections.

[0054] Spatial continuity check: Ensures that the forecast results are spatially continuous and avoids non-physical abrupt changes, specifically including: (1) Calculate the gradient of meteorological elements between adjacent grid points; (2) If the gradient exceeds the preset threshold, it is determined to be spatially discontinuous and smoothing is performed; (3) Smoothing methods include, but are not limited to: Gaussian smoothing, median filtering, variational analysis, etc.

[0055] Temporal consistency check: Ensures that forecast results are continuous in time and conform to atmospheric evolution patterns. Specifically, this includes: (1) Calculate the rate of change of meteorological elements between adjacent time periods; (2) If the rate of change exceeds the preset threshold, it is determined that the time is discontinuous and interpolation correction is performed; (3) Interpolation methods include, but are not limited to: linear interpolation, cubic spline interpolation, time smoothing, etc.

[0056] The fusion forecast results obtained through quality control are used as the final product output and saved as a GRIB2 format file.

[0057] Example 1: The following uses the 2-meter temperature fusion prediction as an example to explain in detail the specific implementation of the present invention. 1. Data Acquisition and Preprocessing: Acquire initial field data provided by the Global Forecast System (GFS) and convert the initial field data into a data format suitable for the input requirements of multiple independently trained large meteorological models. Specifically: (1) Download initial field data from the GFS data server, including but not limited to temperature field, humidity field, wind field, geopotential height field, etc.; (2) Perform quality control on the data, removing missing and outlier values; (3) Convert the data into the format required by each AI model, including grid resampling, variable renaming, unit conversion, etc.

[0058] 2. Multi-model forecasting: Using the format-converted initial field data as input, multiple independently trained large-scale meteorological models are driven to predict short-, medium-, and long-term meteorological elements, outputting the forecast results of each model for key variables such as temperature, humidity, wind field, and precipitation. Specifically: (1) Drive the FengWu model to generate 2-meter temperature forecasts for 000-240 hours; (2) Drive the FengWu v2 model to generate 2-meter temperature forecasts for 000-240 hours; (3) Drive the FourCastNet v2 model to generate 2-meter temperature forecasts for 000-240 hours; (4) Drive the FuXi model to generate 2-meter temperature forecasts for 000-360 hours; (5) Drive the Pangu model to generate 000-192 hour 2-meter temperature forecasts.

[0059] 3. Fusion Model Construction: Forecast values ​​are extracted from the outputs of various large meteorological models to construct the input feature space of the fusion model. Simultaneously, data with the same elements in the GFS forecast field are used as supervision label data.

[0060] A fully connected neural network model is constructed, with the output values ​​of each large meteorological model as input features. A loss function is constructed by combining the GFS forecast values ​​with data fitting terms and physical constraint terms. The model weights are optimized through iterative training to minimize forecast errors and improve the physical consistency of the fusion results.

[0061] 4. Weight Optimization and Fusion: Obtain the globally optimal model weight configuration after training. Based on this weight configuration, the 2-meter temperature forecast results of each meteorological model are weighted and fused to generate a 2-meter temperature forecast product with high accuracy, strong physical rationality and high timeliness.

[0062] 5. Quality Control and Product Output: The fused forecast results undergo quality control and physical consistency checks, including outlier detection, physical consistency verification, spatial continuity checks, and temporal consistency checks. The fused forecast results that pass quality control are used as the final product output and saved as GRIB2 format files.

[0063] 6. Accuracy assessment: such as Figure 8 As shown, the RMSE evaluation results of the fusion model AIM and GFS for 2-meter temperature forecasts indicate that, throughout the entire forecast lead time of 0-360 hours, the RMSE of the fusion model is consistently lower than that of GFS, with the advantage becoming more pronounced after 120 hours. Specifically: (1) In the 0-120 hour short-term forecast phase, the fusion model's RMSE is reduced by approximately 10-15% compared to the GFS model; (2) In the 120-240 hour medium-term forecast phase, the RMSE of the fusion model is reduced by about 15-20% compared with GFS; (3) In the 240-360 hour long-term forecast phase, the RMSE of the fusion model is reduced by about 20-25% compared with GFS.

[0064] The above results demonstrate that the multi-intelligent large-scale model fusion method described in this invention can significantly improve the accuracy of weather forecasts and meet the operational needs of different timeframes.

[0065] Example 2: Multi-element fusion forecast In this embodiment, the specific implementation of the present invention is illustrated by taking the fusion forecast of multiple elements such as temperature, humidity, and wind field as an example.

[0066] 1. Acquisition of multi-element forecast results: Following the method in Example 1, the forecast results of each AI model for elements such as temperature T, humidity q, zonal wind U, meridional wind V, and geopotential height Z are obtained.

[0067] 2. Integration of sub-elements The elements are integrated separately, specifically: (1) For the temperature element, the method described in Example 1 is used for fusion; (2) For humidity, a similar method is used for fusion, but the physical characteristics of humidity, such as relative humidity not exceeding 100%, must be considered; (3) For wind field elements, a similar method is used for fusion, while introducing geostrophic wind balance constraints; (4) For the potential height element, a similar method is used to fuse them, while the potential vortex conservation constraint is introduced.

[0068] 3. Multi-factor consistency check After each element is individually fused, a multi-element consistency check is performed to ensure that the fusion result is physically self-consistent. Specifically: (1) Check the static equilibrium relationship between the temperature field and the potential height field; (2) Check the geostrophic balance relationship between the wind field and the geopotential height field; (3) Check the thermodynamic relationship between the humidity field and the temperature field.

[0069] Example 3: Regional refined forecast In this embodiment, a detailed forecast of a specific region is used as an example to illustrate the specific implementation of the present invention.

[0070] 1. Regional Data Extraction: Extract data for the target region from global forecast results. The target region can be set according to operational needs, such as 70°E-140°E and 15°N-55°N for China.

[0071] 2. Regional Weight Optimization: The fusion weights are re-optimized based on the characteristics of the target region. Specifically: (1) Collect historical observation data of the target area; (2) Evaluate the forecasting performance of each model in the target area; (3) Adjust the fusion weights according to the regional forecast performance so that the weights are biased towards the model that performs well in the target region.

[0072] 3. Regional product output: Convert the fused forecast results into regional product formats, such as regional grid data and site forecast data, to meet regional business needs.

[0073] To verify the technical effects of the present invention, the following comparative experiments were conducted: 1. Improved forecast accuracy: By leveraging the complementary advantages of multiple models and integrating them, the technical strengths of each AI meteorological model are fully utilized. Compared with a single model, the fusion forecast has an ACC (Anomaly Correlation Coefficient) that is improved by more than 5% and an RMSE (Root Mean Square Error) that is reduced by more than 8%, achieving a substantial breakthrough in forecast accuracy.

[0074] 2. Domain adaptability: It adopts a dual strategy of time-segmented fusion and spatial multi-scale fusion, which can use differentiated fusion schemes for different forecast time segments and geographical features to achieve accurate cross-scenario adaptation and meet the specific needs of diverse business scenarios such as refined forecasting and disaster early warning.

[0075] 3. Consistency of physical processes: The introduction of geostrophic wind balance constraints and potential vortex conservation constraints ensures that the fusion results strictly follow the laws of atmospheric physics, effectively avoid non-physical generation, and significantly improve the physical reliability and operational application value of the forecast results.

[0076] 4. Forecast stability: By synergistically fusing multiple models, the systematic bias of a single model is effectively diluted, significantly improving the time series stability and robustness of the forecast results. The system's annual uptime rate is ≥99.8%, and the forecast product integrity rate is ≥99.95%, meeting the requirements of business continuity.

[0077] 5. Dynamic adaptive optimization: Equipped with an adaptive weight optimization engine and a real-time feedback adjustment mechanism, it can dynamically adjust the fusion weight allocation based on historical forecast performance and real-time observation data, so as to achieve continuous iterative optimization of forecast performance and ensure the long-term applicability of the model.

[0078] 6. Full-time coverage: By spatiotemporal splicing and fusion of multi-model forecast results, seamless full-time coverage from 000 to 360 hours is achieved to meet the timeliness requirements of different business scenarios.

[0079] 7. Computational efficiency: Compared to running multiple models in parallel computing independently, the fusion method only requires post-processing and fusion of the output results of each model, reducing computing resource consumption by more than 60%. It can complete full-domain fusion computing within 30 minutes, meeting the real-time response requirements of the business side.

[0080] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A domain-adaptive weather forecasting method based on multi-intelligent large model fusion, characterized by: Includes the following steps: S1. Based on the initial field data of the unified global forecast system, multiple AI meteorological models are used to make weather forecasts and generate a set of forecast results. S2. Based on the set of forecast results, construct a fusion forecast model: Based on the fusion weight, sum the forecast values ​​of each model by weight, and then add the bias correction term to obtain the fusion forecast result; S3. Based on data fitting and physical constraints, construct an objective function to optimize the fusion weights and bias correction terms; S4. Using a constrained optimization algorithm, minimize the objective function and solve for the fusion weights and bias correction terms; S5. Adopt differentiated fusion strategies according to different forecast lead periods; S6. Dynamically optimize the fusion weights based on historical forecast performance and real-time observation data; S7. Perform quality control and physical consistency checks on the fused forecast results.

2. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion as described in claim 1, characterized in that: In step S1, the meteorological elements predicted by the model include temperature, humidity, zonal wind, meridional wind, and geopotential. The model's prediction results are in the GRIB2 standard format, with a spatial resolution of 0.25°×0.25° and 13 vertical layers.

3. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion as described in claim 1, characterized in that: In step S2, the expression for the fused forecast result is as follows: ； In the formula, The optimal forecast result after fusion. The total number of models participating in the fusion. Let p be the fusion weight coefficient of the i-th model element. Let be the predicted value of element p for the i-th model at time t. This is the deviation correction term for element p.

4. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion as described in claim 1, characterized in that: In step S3, the expression for the objective function is as follows: ； In the formula, For data fitting terms, For physical constraint terms, , These are the weighting coefficients. This is the result of GFS numerical weather prediction. This is the physical consistency constraint function.

5. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion according to claim 1, characterized in that: In step S4, the fusion weights and bias correction terms are solved, with the following expressions: ； In the formula, Let be the optimal weight of the i-th model for element p. This is the deviation correction term for the optimal element p.

6. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion according to claim 1, characterized in that: In step S5, the differentiated fusion strategy includes: For short-term forecasts with high temporal resolution, the weights are biased towards models with excellent short-term performance, with less weight given to physical constraints, prioritizing accuracy. For medium-term forecasts with medium temporal resolution, the weights are evenly distributed to balance accuracy and stability. For long-term forecasts with low temporal resolution, the weights are biased towards models with long-term stability, with more weight given to physical constraints, prioritizing the reasonableness of the trend.

7. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion according to claim 1, characterized in that: In step S6, the fusion weights are dynamically optimized, as shown in the following expression: ； In the formula, Let be the final dynamic fusion weight of the i-th model for element p at time t. Based on the weighting coefficient, For weighting adjustment factors based on historical forecasting techniques, This is a weighting adjustment factor based on real-time observation and comparison; ； In the formula, For the current moment, The moment when the i-th model performs optimally. Standard deviation; ； In the formula, Let i be the current forecast value of the i-th model. The current actual observation value, This represents the standard deviation of the observation error.

8. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion according to claim 1, characterized in that: In step S7, quality control and physical consistency checks are performed on the fused forecast results, including: Outlier detection: Identify and correct forecast values ​​that exceed reasonable ranges; Physical consistency verification: Check whether the fusion results meet physical constraints; Spatial continuity check: Ensure that the forecast results are spatially continuous; Temporal consistency check: Ensure that the forecast results are temporally continuous.

9. The domain-adaptive weather forecasting method based on multi-intelligent large model fusion according to claim 1, characterized in that: In step S3, the physical constraints include geostrophic wind balance constraints and potential vortex conservation constraints. The expression is as follows: ； In the formula, For Coriolis parameters, It is a unit vector in the vertical direction. This is the actual wind field vector. This is the geostrophic wind field vector; Potential vortex conservation constraint The expression is as follows: ； In the formula, For potential vortex, It is the total derivative.

Images