WRF model dynamic tuning method based on multi-dimensional parameter sensitivity test and dual-objective optimization

Abstract

The invention relates to the technical field of meteorological numerical forecasting, and provides a WRF model dynamic tuning method based on a multi-dimensional parameter sensitivity test and dual-objective optimization. The objective of the invention is to solve the problems of insufficient prediction accuracy and limited service application caused by parameter fragmentation, low test efficiency, one-sided evaluation mechanism and computing resource waste in existing WRF model parameter tuning. An automatic sensitivity test is realized by constructing a multi-dimensional parameter system and developing a Python interface program, RMSE, MAE and CC indexes are fused to generate a sensitivity index SI with a weight, and key parameter combinations are screened. And a calculation efficiency factor is introduced to establish a prediction precision-calculation efficiency double-objective optimization model, a particle swarm optimization algorithm is adopted to solve a Pareto optimal solution, and dynamic balance of prediction precision and calculation efficiency in a complex scene is realized. The method covers the number of grids, spin-up time and various parameterization schemes, breaks through the limitation of traditional single parameter adjustment and optimization, and improves the prediction accuracy of extreme weather and complex terrain regions.

Description

Technical Field

[0001] This invention relates to the field of meteorological numerical forecasting technology, specifically to a dynamic optimization method for WRF models based on multi-dimensional parameter sensitivity experiments and bi-objective optimization, applicable to diverse meteorological scenarios such as complex terrain and extreme weather. Background Technology

[0002] Numerical weather prediction uses large computers to perform numerical calculations on atmospheric motion and evolution based on actual atmospheric conditions, solving dynamic, thermodynamic, and microphysical equations to predict atmospheric conditions and weather phenomena over a future period. The WRF model, as an advanced mesoscale numerical weather prediction model, has wide applications in both operational meteorological forecasting and scientific research.

[0003] In numerical weather prediction, the accuracy of models is affected by a variety of factors. Among these, the factors that significantly impact forecast results, particularly in numerical model settings, include microphysical parameterization schemes, longwave radiation parameterization schemes, shortwave radiation parameterization schemes, boundary layer schemes, land surface process schemes, and cumulus parameterization schemes. Although numerical weather prediction models are constantly evolving and improving, discrepancies between forecasts and actual weather conditions still exist. For example, in complex terrain areas, existing numerical weather prediction models struggle to accurately capture local meteorological changes; and the accuracy of forecasts for seasonal transitions and extreme weather events needs further improvement.

[0004] The current WRF model faces the following technical bottlenecks in numerical weather prediction: 1) Parameter fragmentation: Existing technologies mostly use single parameter or small-scale parameter combination optimization (such as adjusting only the microphysics scheme), without forming a systematic optimization system covering the number of grids, spin-up time, and multiple parameterization schemes, resulting in the neglect of parameter interaction effects in complex scenarios; 2) Low experimental efficiency: The experimental method of manually adjusting parameters one by one is time-consuming and labor-intensive, and it is difficult to ensure the uniqueness of variables. For example, a single grid resolution experiment requires manual modification of the configuration file and repeated running of the model, and the cycle of a single parameter full combination experiment can be as long as several weeks; 3) One-sided evaluation mechanism: Relying on a single index (such as using only RMSE) to evaluate the impact of parameters cannot fully reflect the forecast accuracy, bias, and correlation, which may lead to misjudgment of key parameters; 4) When pursuing high accuracy by adopting high-resolution grids or complex parameterization schemes, computational efficiency is often ignored. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of insufficient forecast accuracy and limited business applications caused by parameter fragmentation, low experimental efficiency, one-sided evaluation mechanism, and waste of computational resources in the parameter tuning of existing WRF models.

[0006] To achieve the above objectives, the present invention employs the following technical means:

[0007] This invention provides a dynamic tuning method for a WRF model based on multi-dimensional parameter sensitivity experiments and dual-objective optimization, comprising the following steps:

[0008] Step 1: Select multiple parameter factors that are relevant to the application scenario and construct a systematic multi-dimensional parameter system.

[0009] Step 2: Automated implementation of standardized univariate sensitivity tests. A Python interface program was developed based on the WRF model. The script controls the combination of multi-dimensional parameter variables in batches, passes different parameter values ​​in sequence, and submits all test combinations at the same time to generate corresponding prediction results.

[0010] Step 3: Multi-indicator quantitative evaluation and sensitivity index screening. By integrating three types of indicators, namely root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (CC), a weighted sensitivity index SI is constructed. Key parameters are determined by ranking the SI, and the optimal combination of key sensitive parameters and forecast results is obtained.

[0011] Step 4: Dual-objective optimization mechanism for forecast accuracy and computational efficiency. Based on the determined optimal parameter combination, a computational efficiency factor (running time) is introduced to establish a dual-objective optimization model for forecast accuracy and computational efficiency. The Pareto optimal solution is obtained through the Particle Swarm Optimization (PSO) algorithm, achieving a dynamic balance between forecast accuracy and computational efficiency in operational applications.

[0012] In the above scheme, step 1 specifically includes the following steps:

[0013] Step 1.1: Parameter Dimension Expansion

[0014] Establish a multi-dimensional parameter set that includes spatial resolution parameters, number of grids, time initialization parameters (spin-up time, 6h / 12h / 24h), boundary condition parameters (time resolution, 1h / 3h / 6h), and at least two physical process parameterization schemes (such as the microphysics scheme WSM6 / Thompson and the cumulus scheme Kain-Fritsch / BMJ).

[0015] This breaks through the traditional "single parameter trial and error" model and improves forecast accuracy through multi-dimensional parameter collaborative optimization.

[0016] Step 1.2: Scenario-based filtering logic

[0017] Parameters are dynamically selected based on the characteristics of the forecast scene. For example, for complex terrain areas, the grid resolution + land surface process scheme (Noah / RUC) is selected as the focus; for heavy precipitation forecasts, the combination of microphysical scheme + cumulus scheme is focused.

[0018] This reduces the amount of experimental calculations and improves optimization efficiency.

[0019] In the above scheme, step 2 specifically includes the following steps:

[0020] Step 2.1: Develop a Python interface program based on the WRF model to design a standardized univariate sensitivity test procedure. In each set of experiments, a single parameter variable is precisely controlled through the program interface (e.g., changing only the number of grids or adjusting only a certain parameterization scheme), while other parameters remain at their default values ​​to ensure the uniqueness of the experimental variables.

[0021] By standardizing the experimental procedures, the reproducibility and scalability of multi-parameter sensitivity analysis can be achieved, making it applicable to any region and application scenario.

[0022] Step 2.1: Automated Batch Execution

[0023] This system uses Python to automate batch processing, simultaneously submitting calculations for multiple parameter combinations to obtain corresponding forecast results. Compared to manual experiments, automated batch processing reduces the cycle of multi-dimensional parameter variable full combination experiments from 72 hours to 24 hours, while avoiding rework and time consumption caused by human configuration errors.

[0024] In the above scheme, step 3 specifically includes the following steps:

[0025] Step 3.1: Calculate the three types of indicators—Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Correlation Coefficient (CC)—for each combination of multi-dimensional parameter variables. Then, construct a weighted sensitivity index SI by fusing these three indicators: RMSE, MAE, and CC.

[0026] SI = α × RMSE + β × MAE + γ × CC

[0027] Where α, β, and γ are scene weights (e.g., in precipitation forecasting, α = 0.4, β = 0.3, and γ = 0.3).

[0028] Step 3.2: Determine the key parameters by sorting them by SI values, and select the parameter combination with the highest ranking.

[0029] This approach breaks through the limitations of a single indicator, quantifying the impact of parameters from three dimensions: accuracy (RMSE), bias (MAE), and correlation (CC), thus avoiding misjudgments of parameters with "high RMSE but high correlation" in traditional methods. Furthermore, the selected key parameters can be further reused in establishing forecasting systems for different regions in other similar scenarios.

[0030] In the above scheme, step 4 specifically includes the following steps:

[0031] Step 4.1: Introduce a computational efficiency factor (running time T), and compare the running time T under the optimal parameter combination obtained in Step 3.2 with T0 under the original default parameters to calculate the efficiency loss rate:

[0032] Loss = (T - T0) × 100% / T0

[0033] Step 4.2: If the efficiency loss exceeds a preset threshold of 30%, then activate the dual-objective optimization model.

[0034] By using the particle swarm optimization algorithm, while ensuring that the increase in SI value does not exceed 5%, a suboptimal parameter combination with better computational efficiency is searched, thus achieving a dynamic balance between accuracy and efficiency.

[0035] Because the present invention employs the above-mentioned technical means, it has the following beneficial effects:

[0036] 1. Improve forecast accuracy: By conducting comprehensive and systematic sensitivity tests on multiple key parameters, we can gain a deeper understanding of the impact of each parameter on forecast results. This allows for precise adjustment of model parameters, significantly improving the accuracy of operational numerical weather prediction models in forecasting meteorological elements. This provides more reliable forecast data for meteorological services. For example, in tests in complex terrain areas, compared with traditional empirical optimization methods, this invention reduces the RMSE of precipitation forecasts by 25%-32% and the RMSE of temperature forecasts by 18%-24%. In typhoon track forecasts, the 24-hour forecast error is reduced by 15-20 kilometers, providing more reliable forecast data for meteorological services. Especially in the forecasting of severe weather such as rainstorms and typhoons, it can more accurately predict their occurrence time, location, and intensity, providing strong support for disaster prevention and mitigation.

[0037] 2. Enhanced Model Adaptability: The method of this invention can find the most suitable parameter combination through sensitivity experiments based on different geographical regions, seasonal characteristics, and meteorological conditions, enabling operational numerical weather prediction models to better adapt to complex and ever-changing realities. For example, in mountainous scenarios, by optimizing the combination of grid number and land surface process schemes, the accuracy of local temperature forecasts can be improved by 20%. The method also enhances the model's forecasting capabilities in different scenarios, such as in mountainous and coastal areas with complex terrain, and in weather forecasts for different seasons, achieving better forecast results.

[0038] 3. Establish a standardized optimization system: Breaking away from traditional reliance on experience, establish a standardized system covering the entire process from parameter selection, experimental design, quantitative evaluation to dynamic optimization, which is highly replicable. It can also generate parameter optimization templates for different climate zones (such as the South China monsoon region and the Northwest arid region), promoting the transformation of numerical weather prediction operations from "experience-driven" to "data-driven."

[0039] 4. Efficient utilization of computing resources: Through a dual-objective optimization mechanism, computing efficiency is significantly improved while ensuring forecast accuracy. For example, in high-resolution forecast scenarios, parameter combination optimization reduces single-machine running time from 11 hours to 5 hours while maintaining forecast accuracy, thereby significantly reducing hardware costs and energy consumption for business operations. Attached Figure Description

[0040] Figure 1 This is a simplified flowchart of the present invention. Detailed Implementation

[0041] The embodiments of the present invention will be described in detail below. Although the present invention will be described and illustrated in conjunction with some specific embodiments, it should be noted that the present invention is not limited to these embodiments. On the contrary, any modifications or equivalent substitutions made to the present invention should be covered within the scope of the claims of the present invention.

[0042] Furthermore, to better illustrate the present invention, numerous specific details are set forth in the following detailed embodiments. Those skilled in the art will understand that the present invention can be practiced without these specific details.

[0043] This invention provides a dynamic tuning method for a WRF model based on multi-dimensional parameter sensitivity experiments and dual-objective optimization, comprising the following steps:

[0044] Step 1: Select multiple parameter factors that are relevant to the application scenario and construct a systematic multi-dimensional parameter system.

[0045] Step 2: Automated implementation of standardized univariate sensitivity tests. A Python interface program was developed based on the WRF model. The script controls the combination of multi-dimensional parameter variables in batches, passes different parameter values ​​in sequence, and submits all test combinations at the same time to generate corresponding prediction results.

[0046] Step 3: Multi-indicator quantitative evaluation and sensitivity index screening. By integrating three types of indicators, namely root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (CC), a weighted sensitivity index SI is constructed. Key parameters are determined by ranking the SI, and the optimal combination of key sensitive parameters and forecast results is obtained.

[0047] Step 4: Dual-objective optimization mechanism for forecast accuracy and computational efficiency. Based on the determined optimal parameter combination, a computational efficiency factor (running time) is introduced to establish a dual-objective optimization model for forecast accuracy and computational efficiency. The Pareto optimal solution is obtained through the Particle Swarm Optimization (PSO) algorithm, achieving a dynamic balance between forecast accuracy and computational efficiency in operational applications.

[0048] In the above scheme, step 1 specifically includes the following steps:

[0049] Step 1.1: Parameter Dimension Expansion

[0050] Establish a multi-dimensional parameter set that includes spatial resolution parameters, number of grids, time initialization parameters (spin-up time, 6h / 12h / 24h), boundary condition parameters (time resolution, 1h / 3h / 6h), and at least two physical process parameterization schemes (such as the microphysics scheme WSM6 / Thompson and the cumulus scheme Kain-Fritsch / BMJ).

[0051] This breaks through the traditional "single parameter trial and error" model and improves forecast accuracy through multi-dimensional parameter collaborative optimization.

[0052] Step 1.2: Scenario-based filtering logic

[0053] Parameters are dynamically selected based on the characteristics of the forecast scene. For example, for complex terrain areas, the grid resolution + land surface process scheme (Noah / RUC) is selected as the focus; for heavy precipitation forecasts, the combination of microphysical scheme + cumulus scheme is focused.

[0054] This reduces the amount of experimental calculations and improves optimization efficiency.

[0055] In the above scheme, step 2 specifically includes the following steps:

[0056] Step 2.1: Develop a Python interface program based on the WRF model to design a standardized univariate sensitivity test procedure. In each set of experiments, a single parameter variable is precisely controlled through the program interface (e.g., changing only the number of grids or adjusting only a certain parameterization scheme), while other parameters remain at their default values ​​to ensure the uniqueness of the experimental variables.

[0057] By standardizing the experimental procedures, the reproducibility and scalability of multi-parameter sensitivity analysis can be achieved, making it applicable to any region and application scenario.

[0058] Step 2.1: Automated Batch Execution

[0059] This system uses Python to automate batch processing, simultaneously submitting calculations for multiple parameter combinations to obtain corresponding forecast results. Compared to manual experiments, automated batch processing reduces the cycle of multi-dimensional parameter variable full combination experiments from 72 hours to 24 hours, while avoiding rework and time consumption caused by human configuration errors.

[0060] In the above scheme, step 3 specifically includes the following steps:

[0061] Step 3.1: Calculate the three types of indicators—Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Correlation Coefficient (CC)—for each combination of multi-dimensional parameter variables. Then, construct a weighted sensitivity index SI by fusing these three indicators: RMSE, MAE, and CC.

[0062] SI = α × RMSE + β × MAE + γ × CC

[0063] Where α, β, and γ are scene weights (e.g., in precipitation forecasting, α = 0.4, β = 0.3, and γ = 0.3).

[0064] Step 3.2: Determine the key parameters by sorting them by SI values, and select the parameter combination with the highest ranking.

[0065] This approach breaks through the limitations of a single indicator, quantifying the impact of parameters from three dimensions: accuracy (RMSE), bias (MAE), and correlation (CC), thus avoiding misjudgments of parameters with "high RMSE but high correlation" in traditional methods. Furthermore, the selected key parameters can be further reused in establishing forecasting systems for different regions in other similar scenarios.

[0066] In the above scheme, step 4 specifically includes the following steps:

[0067] Step 4.1: Introduce a computational efficiency factor (running time T), and compare the running time T under the optimal parameter combination obtained in Step 3.2 with T0 under the original default parameters to calculate the efficiency loss rate:

[0068] Loss = (T - T0) × 100% / T0

[0069] Step 4.2: If the efficiency loss exceeds a preset threshold of 30%, then activate the dual-objective optimization model.

[0070] By using the particle swarm optimization algorithm, while ensuring that the increase in SI value does not exceed 5%, a suboptimal parameter combination with better computational efficiency is searched, thus achieving a dynamic balance between accuracy and efficiency.

[0071] Example 1:

[0072] The following steps are included in the WRF model tuning process for complex terrain areas:

[0073] Step 1: Constructing a multi-dimensional parameter set

[0074] The core parameters were determined as follows: grid resolution (3km / 9km), land surface process scheme (Noah / RUC), and boundary layer scheme (YSU / MYJ). Other parameters were set to default values.

[0075] Scene adaptation: Add terrain filtering parameters (filter coefficient 0.1 / 0.3 / 0.5) for mountainous terrain.

[0076] Step 2: Execution of systematic univariate sensitivity test

[0077] Based on the numerical weather prediction model WRF, an automated experimental interface was developed using Python scripts. This interface precisely transmits parameter values ​​to the numerical weather prediction model, enabling multiple sensitivity experiments. In each experiment, the univariate principle was strictly followed, changing only one parameter while keeping others constant. Batch scripting was used to automate model execution and result collection.

[0078]

[0079]

[0080] Step 3: Multi-indicator evaluation and SI calculation

[0081] Forecast results from each sensitivity experiment were collected and compared with actual observation data from China Meteorological Administration's surface meteorological observation stations and radiosonde stations after spatiotemporal matching. A comprehensive sensitivity assessment system was constructed using three evaluation indicators: root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (CC).

[0082] The evaluation was conducted under the following conditions: 3km grid resolution + Noah land surface process scheme + YSU boundary layer scheme + filter coefficient 0.1. The precipitation RMSE = 12.5m, MAE = 8.3mm, CC = 0.82. The calculated SI = 0.4×(12.5 / 15) + 0.3×(8.3 / 10) + 0.3×0.82 = 0.71.

[0083] The lower the SI value, the higher the accuracy. By comparing the SI values ​​of 18 experiments, the lowest SI value (i.e. the highest prediction accuracy) was identified as the key sensitive combination.

[0084] Step 4: Dual-objective optimization, balancing dynamic tuning and computational efficiency

[0085] Based on the parameter sensitivity analysis results obtained in step 3, the numerical prediction model is optimized as follows:

[0086] Parameters that significantly affect the forecast results are adjusted to their optimal values. For example, the land surface process scheme is fixed as the Noah scheme, the boundary layer scheme is MYJ, the filter coefficient is 0.1, the number of grids is set to 3km×3km, and the numerical forecast parameter settings are reconstructed.

[0087] Introduce a computational efficiency evaluation metric, such as the computational efficiency loss rate when the model's runtime is T1 under this parameter combination, compared to T0 under the original default parameters.

[0088] If the efficiency loss exceeds a preset threshold (e.g., 30%), a dual-objective optimization model is initiated. Using the particle swarm optimization algorithm, a suboptimal parameter combination with better computational efficiency is searched while ensuring that the SI value increase does not exceed 5%. For example, the number of grids is adjusted to 9km×9km, and the boundary layer scheme is optimized to the YSU scheme. Ultimately, the model running time is shortened by 25% while the forecast accuracy decreases by only 3%, achieving a dynamic balance between accuracy and efficiency.

[0089] In summary, the present invention has the following characteristics:

[0090] This invention addresses the problems of systematic deficiencies, low experimental efficiency, biased evaluation, and waste of computational resources in existing WRF model parameter tuning by integrating multi-dimensional parameter sensitivity experiments with a dual-objective optimization mechanism.

[0091] 1. This invention solves the problem of parameter fragmentation in existing technologies by constructing a systematic multi-dimensional parameter system (step 1), thereby improving forecast accuracy. Specifically, step 1 includes parameter dimension expansion (such as spatial resolution, number of grids, time initialization parameters, boundary condition parameters, and physical process parameterization schemes) and scenario-based filtering logic (such as dynamically filtering parameter combinations for complex terrain areas or heavy precipitation forecasts). This technique avoids neglecting complex scenarios by covering the interactive effects of multi-dimensional parameters, thus achieving multi-parameter collaborative optimization. Logical analysis process: Existing technologies only optimize single parameters, leading to the neglect of parameter interaction effects; this invention, by systematically constructing a multi-dimensional parameter system, ensures that all relevant parameters are included in the optimization scope, solving the fragmentation problem, and thus improving forecast accuracy through synergistic effects.

[0092] 2. This invention solves the problem of low experimental efficiency by automating standardized univariate sensitivity tests (step 2), achieving the effects of shortening the experimental cycle, avoiding human error, and enabling reproducible applications. Specifically, step 2 includes developing a Python interface program to control the combination of parameter variables (step 2.1) and automating batch execution (step 2.2), ensuring that each set of experiments only changes a single parameter variable. This technique uses scripts to submit experimental combinations in batches, avoiding the time-consuming and error-prone manual adjustments. Logical analysis: Existing manual experimental methods are inefficient and prone to errors; this invention, through standardized processes and automated interfaces, achieves the uniqueness and parallel processing of experimental variables, solving the efficiency bottleneck problem, thereby significantly improving experimental speed and supporting large-scale applications in multiple regions.

[0093] 3. This invention solves the problem of one-sided evaluation by using multi-indicator quantitative assessment and sensitivity index screening (step 3), achieving the effects of comprehensively quantifying the impact of parameters, avoiding misjudgment, and enhancing parameter reusability. Specifically, step 3 integrates root mean square error (RMSE), mean absolute error (MAE), and correlation coefficient (CC) to construct a weighted sensitivity index SO (step 3.1), and determines key parameters by ranking SO (step 3.2). This technique breaks through the limitations of single indicators through comprehensive evaluation of multi-dimensional indicators. Logical analysis process: Existing reliance on a single indicator (such as only RMSE) easily leads to misjudgment of highly correlated parameters; this invention solves the problem of one-sided evaluation by weighting and integrating three types of indicators to quantify the impact from three dimensions: accuracy, bias, and correlation, thereby accurately screening key parameters and supporting their reuse in similar scenarios.

[0094] 3. This invention solves the problem of wasted computational resources by employing a dual-objective optimization mechanism of forecast accuracy and computational efficiency (step 4), achieving a dynamic balance between forecast accuracy and computational efficiency. Specifically, step 4 introduces a computational efficiency factor (such as running time), establishes a dual-objective optimization model (step 4.1), and uses the particle swarm optimization algorithm to search for the Pareto optimal solution (step 4.2). This technique optimizes efficiency while ensuring accuracy, avoiding resource waste caused by high-precision solutions. Logical analysis: Existing technologies often neglect efficiency in pursuit of accuracy; this invention, through a dual-objective model and optimization algorithm, adjusts parameter combinations within a preset threshold, solving the problem of resource waste, thereby achieving synergistic optimization of accuracy and efficiency in operational applications.

Claims

1. A WRF model dynamic tuning method based on multi-dimensional parameter sensitivity test and double-target optimization, characterized in that, The method comprises the following steps: Step 1: selecting a plurality of parameter factors related to an application scenario, and constructing a multi-dimensional parameter system comprising a spatial resolution parameter, a grid number, a time initialization parameter, a boundary condition parameter and at least two types of physical process parameterization schemes; Step 2: standardized single-variable sensitivity test automation, developing a Python interface program based on the WRF model, controlling a plurality of multi-dimensional parameter variable combinations in batches through a script, sequentially passing different parameter values, and simultaneously submitting all test combinations to generate corresponding prediction results; Step 3: multi-index quantitative evaluation and sensitivity index screening, constructing a weighted sensitivity index SI by fusing root mean square error RMSE, mean absolute error MAE and correlation coefficient CC, and determining key parameters by SI sorting to obtain a key sensitive parameter and an optimal parameter combination of a prediction result; Step 4: a prediction accuracy-computing efficiency double-target optimization mechanism, introducing a computing efficiency factor based on the determined optimal parameter combination, establishing a prediction accuracy-computing efficiency double-target optimization model, and solving a Pareto optimal solution by a particle swarm optimization algorithm PSO to realize dynamic balance between prediction accuracy and computing efficiency in business application.

2. The method of claim 1, wherein, Step 1 specifically comprises the following steps: Step 1.1: parameter dimension expansion A multi-dimensional parameter set comprising a spatial resolution parameter, a grid number, a time initialization parameter, a boundary condition parameter and at least two types of physical process parameterization schemes is established; Step 1.2: scenario-based screening logic Parameters are dynamically screened according to prediction scenario characteristics, including: complex terrain area key selection of grid resolution and land surface process scheme combination; strong precipitation prediction focusing on the combination of microphysical scheme and cumulus scheme.

3. The method of claim 1, wherein, Step 2 specifically comprises the following steps: Step 2.1: developing a Python interface program based on the WRF model, designing a standardized single-variable sensitivity test process, accurately controlling a single parameter variable, and keeping other parameters at default values to ensure the uniqueness of the test variable; Step 2.2: automated batch running Automated batch running is realized through a python program, and the corresponding prediction results are obtained by simultaneously submitting the calculation of a plurality of parameter combinations.

4. The method of claim 1, wherein, Step 3 specifically comprises the following steps: Step 3.1: calculating the fused root mean square error RMSE, mean absolute error MAE and correlation coefficient CC of each multi-dimensional parameter variable combination, and constructing a weighted sensitivity index SI by fusing the root mean square error RMSE, mean absolute error MAE and correlation coefficient CC: SI = α × RMSE + β × MAE + γ × CC Wherein, α, β, γ are scene weights; Step 3.2: determining key parameters by SI value sorting, and selecting the parameter combination with the highest sorting.

5. The method of claim 1, wherein, Step 4 specifically comprises the following steps: Step 4.1: introducing a computing efficiency factor, running time T, comparing the running time T of the optimal parameter combination obtained in step 3.2 with T0 under the original default parameters, and calculating the computing efficiency loss rate: Loss = (T - T0) × 100% / T0 Step 4.2: If the efficiency loss exceeds the preset threshold 30%, start the dual-objective optimization model, search for a more optimal parameter combination through the particle swarm algorithm under the premise that the SI value increase does not exceed 5%.

Images