Explainable and machine learning based climate model preference and ensemble prediction method

Abstract

The present application relates to an interpretable and machine learning-based climate model preferably combined with a collection estimation method, belonging to the technical field of climate models, the method comprising: obtaining area-weighted average ozone column total of reanalysis data and CMIP6 model data, and constructing a data set; based on the data set, respectively adopting a traditional method and a decision tree SHAP method to screen the CMIP6 model; based on the CMIP6 model data screened by the traditional method and the decision tree SHAP method, a machine learning model is constructed; based on the constructed machine learning model, the optimal screening method and the corresponding optimal machine learning model are determined; based on the optimal screening method and the corresponding optimal machine learning model, the Antarctic ozone under different greenhouse gas emission scenarios is estimated. Through the present application, more accurate Antarctic ozone recovery estimation results can be obtained.

Description

Technical Field

[0001] This invention relates to the field of climate model technology, and in particular to a climate model optimization and ensemble prediction method based on interpretability and machine learning. Background Technology

[0002] The Antarctic ozone hole was a globally significant environmental crisis in the 1980s, and signs of recovery began to appear at the beginning of this century. However, global warming, by influencing atmospheric circulation in the Southern Hemisphere stratosphere, can affect Antarctic ozone, thus increasing the uncertainty of ozone changes. Under the backdrop of climate change, the state of Antarctic ozone recovery and when it will return to pre-depletion levels have become crucial scientific questions that urgently need to be addressed.

[0003] The Coupled Model Intercomparison Project Phase 6 (CMIP6) provided ozone simulations for the next century from multiple climate models under different warming scenarios. However, discrepancies exist among the models, necessitating effective evaluation to select the best-performing models. Traditional model selection methods, however, rely heavily on evaluation metrics (such as mean squared error or correlation coefficients) and lack unified standards. Furthermore, they focus only on the performance of individual models, failing to capture the combined relationships between different models. Therefore, traditional selection methods have limitations. Additionally, multi-model ensemble results (MMEs) often serve as strong references for future projections, but existing MME methods frequently employ equal weighting or tend to assign higher weights to better models, neglecting nonlinear combinations between different models. Therefore, effective improvements to multi-model ensemble methods are needed.

[0004] In recent years, artificial intelligence (AI) technology has developed rapidly and has been widely applied in many fields, especially the Shapley Additive Explanations (SHAP), an interpretable AI method proposed in recent years. Based on the SHAP index, this method can quantitatively characterize the weighted contribution of each input variable to the model's prediction results, directly identifying the main variables affecting the prediction results and significantly improving the model's transparency and reliability. Introducing decision tree algorithms and SHAP values ​​into the model selection process may yield better selection results than traditional methods. Furthermore, machine learning (ML) based on nonlinear mapping relationships can assign appropriate weights to different climate models and correct systematic biases between models and reanalysis data, contributing to more accurate prediction results. Summary of the Invention

[0005] To select the CMIP6 model family with superior simulation capabilities, this invention constructs a novel method for climate model optimization based on the SHAP value index of interpretable AI methods. Furthermore, it utilizes machine learning methods to improve the traditional multi-model ensemble, achieving bias correction and obtaining the optimal machine learning prediction model. Finally, based on the prediction model, more accurate Antarctic ozone recovery forecasts are obtained.

[0006] To achieve the above objectives, the present invention provides the following solution:

[0007] Climate model optimization and ensemble prediction methods based on interpretability and machine learning include:

[0008] Obtain the area-weighted average total ozone column concentration from reanalysis data and CMIP6 model data, and construct a dataset;

[0009] Based on the dataset, the CMIP6 patterns were filtered using both traditional methods and the decision tree SHAP method.

[0010] A machine learning model is constructed based on CMIP6 pattern data selected using traditional methods and the SHAP decision tree method.

[0011] Based on the constructed machine learning model, the optimal screening method and the corresponding optimal machine learning model are determined;

[0012] Based on the optimal screening method and the corresponding optimal machine learning model, the recovery of Antarctic ozone under different greenhouse gas emission scenarios is predicted.

[0013] Optionally, obtaining the total area-weighted average ozone column includes:

[0014] Calculate the total ozone column volume using CMIP6 ozone mixing ratio data;

[0015] Based on the total ozone column TCO3 and CMIP6 calculated from the reanalysis data, an area-weighted average is performed to obtain the area-weighted average total ozone column.

[0016] Optionally, screening CMIP6 patterns using traditional methods includes:

[0017] Based on the dataset, the Taylor Skill Score (TSS) and the Comprehensive Rating Index (M) for each CMIP6 mode were calculated. R ;

[0018] The CMIP6 models are ranked according to their Taylor skill scores and comprehensive rating indicators, and the average ranking of the two indicators is taken to select the top-ranked models.

[0019] The expression for the comprehensive rating index is:

[0020] ;

[0021] Where n is the number of patterns, r IVS and r RMSE It is a ranking of the interannual variability index IVS and root mean square error RMSE of the model.

[0022] Optionally, the SHAP decision tree method can be used to filter CMIP6 patterns, including:

[0023] Based on the dataset, using the European Mid-Term Fifth Generation Reanalysis (ERA5) data as a reference and the CMIP6 model data as input, a decision tree model was trained.

[0024] Based on the SHAP value output by the decision tree model, the CMIP6 pattern is filtered using a preset filtering method.

[0025] Optionally, the preset filtering method includes:

[0026] Single-time filtering: Sort all patterns from largest to smallest based on the absolute value of SHAP corresponding to each pattern, and filter out the top-ranked patterns in one go;

[0027] Filter from best to worst: Remove the pattern with the largest absolute SHAP value each time;

[0028] Poor-to-good filtering: Each time, the pattern with the smallest absolute SHAP value is removed;

[0029] For both the best-to-worst and worst-to-best screening, the CMIP6 patterns need to be screened multiple times. In each screening process, the absolute SHAP value of each pattern is calculated first. After multiple screenings are completed, the average absolute SHAP value of each pattern in all screening processes is calculated. Then, the average absolute SHAP values ​​are sorted from largest to smallest to obtain the patterns with the highest ranking.

[0030] Optionally, a machine learning model is constructed based on CMIP6 pattern data selected using traditional methods and the SHAP decision tree method; including:

[0031] The machine learning model was trained using the selected CMIP6 pattern data as input and the reanalysis data as a reference.

[0032] The primary metric S is used to tune the parameters of the trained machine learning model.

[0033] Optionally, the expression for the main indicator S is:

[0034] ;

[0035] in, As an indicator of interannual variability, Mean square error, The correlation coefficient.

[0036] Optionally, based on the constructed machine learning model, determining the optimal screening method and the corresponding optimal machine learning model includes:

[0037] For each constructed machine learning model, calculate the MSE, R, and IVS of each model on the test dataset, and calculate the corresponding main metrics.

[0038] Obtain the average value of the main indicators of the model under each screening method, and obtain the main indicators corresponding to each screening method. The screening method with the smallest main indicator is the optimal screening method.

[0039] In machine learning models built based on the optimal selection method, the model with the smallest main index is further selected as the optimal model based on the main index of each model.

[0040] The beneficial effects of this invention are as follows:

[0041] Based on the SHAP value index in interpretable AI methods, this invention develops three new artificial intelligence screening methods. Compared with traditional screening methods, the new methods have the smallest deviation in the CMIP6 patterns selected, and the optimal screening method can obtain the optimal pattern family for set prediction.

[0042] This invention constructs multiple machine learning models based on the results of four screening methods, and through evaluation, finally determines the optimal screening method and its corresponding optimal machine learning model. The machine learning improved multi-mode set (MME-ML) can achieve bias correction, making the new multi-mode set results more reliable.

[0043] Based on the constructed optimal model, this invention provides more accurate recovery prediction results under different greenhouse gas emission scenarios. Both the improved MME-ML and the traditional MME method show that higher emission scenarios correspond to earlier recovery times, and compared with the traditional method, the results obtained by the new method have a significant delay of 7-11 years. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a schematic diagram of the process for the optimal selection and ensemble prediction of climate models based on interpretable and machine learning, according to an embodiment of the present invention.

[0046] Figure 2 The following are the multi-mode ensemble prediction results of Antarctic ozone in the SON months (September-November) according to an embodiment of the present invention; wherein, (a) is the recovery curve predicted by the general multi-mode ensemble (MME) method, and the filled interval is the recovery curve corresponding to the ±σ region; (b) is the recovery curve predicted by the multi-mode ensemble (MME-ML-ANN) method based on the optimal machine learning model - artificial neural network (ANN); (c) is the recovery year predicted by the general multi-mode ensemble (MME) method, and the error bar is the recovery year corresponding to the ±σ region; (d) is the recovery year predicted by the multi-mode ensemble (MME-ML-ANN) method based on the optimal machine learning model - artificial neural network (ANN). Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0049] like Figure 1 As shown, this embodiment proposes a climate model optimization and ensemble prediction method based on interpretability and machine learning, including:

[0050] S1. Obtain the weighted average total ozone column from the reanalysis data and the CMIP6 model data, and construct the dataset;

[0051] S2. Based on the dataset, the CMIP6 pattern is filtered using traditional methods;

[0052] S3. Based on the dataset, the CMIP6 pattern is filtered using the decision tree SHAP method;

[0053] S4. Construct a machine learning model based on CMIP6 pattern data selected using traditional methods and the SHAP decision tree method;

[0054] S5. Based on the constructed machine learning model, determine the optimal screening method and the corresponding optimal machine learning model;

[0055] S6. Based on the optimal screening method and the corresponding optimal machine learning model, the recovery prediction of Antarctic ozone under different greenhouse gas emission scenarios is carried out.

[0056] To obtain superior CMIP6 models and reduce interference from models with significant biases, this embodiment employs a traditional screening method based on Taylor skill scores, mean squared error, and interannual variability, as well as a screening method based on interpretable AI and decision tree SHAP values. To integrate the advantages of different CMIP6 models and obtain the most suitable model combination for machine learning models, this embodiment constructs multiple machine learning models based on different screening results, improves the multi-model ensemble method, and uses comprehensive indicators to evaluate the screening method and machine learning models, selecting the optimal screening method and its corresponding optimal model. Finally, based on the optimal screening method and optimal model, CMIP6 multi-model ensemble prediction is conducted to obtain more accurate Antarctic ozone recovery prediction results.

[0057] Further, obtaining the weighted average total ozone column includes:

[0058] Calculate the total ozone column volume using CMIP6 ozone mixing ratio data;

[0059] Based on the total ozone column volume in the reanalysis data and the TCO3 calculated from CMIP6, an area-weighted average is performed to obtain the weighted average total ozone column volume.

[0060] Specifically, in step S1 of this embodiment, a dataset is constructed based on the acquired reanalysis data and CMIP6 ozone mixing ratio data, and appropriate preprocessing and dataset partitioning are performed. The specific operations are as follows:

[0061] S101. Total column ozone (TCO3) is a crucial indicator for assessing the recovery of the Antarctic ozone layer. This embodiment uses the area-weighted mean (AWM) of the total column ozone (TCO3) in the Antarctic region (60°S-90°S) as the research indicator. First, the total column ozone (TCO3) was calculated using 13 available CMIP6 ozone mixing ratio data points. Then, the TCO3 data from the ERA5 reanalysis data and the TCO3 calculated from CMIP6 were respectively averaged area-weighted within the Antarctic region.

[0062]

[0063] The formula for calculating the total amount of ozone column in the CMIP6 mode ozone mixing ratio stratified data is as above.

[0064] In the formula: O3 is the ozone volume mixing ratio (mol / mol); P is the atmospheric pressure (hPa); T0: 273.15 (K); R: 8.314 (J / (mol・K)); M0, P0, and g0 are the molar mass of air, standard atmospheric pressure, and gravitational acceleration, respectively.

[0065] ;

[0066] The formula for calculating the area-weighted average (AWM) of TCO3 is as follows: where n lat It is the number of latitudinal grid points, n lon It is the number of longitude grid points, A ij It is the area of ​​each grid point, dot ij That is the total amount of ozone column at each grid point.

[0067] S102. Divide the weighted average total ozone column (AWM-TCO3) of the ERA5 reanalysis data and the CMIP6 model data into training set, validation set and test set, respectively, with the proportions being 70%, 15% and 15% of the original sequence.

[0068] Furthermore, the traditional methods for screening CMIP6 patterns include:

[0069] Based on the dataset, the Taylor skill score and comprehensive rating index for each CMIP6 mode were calculated.

[0070] The CMIP6 models are ranked according to their Taylor skill scores and comprehensive rating indicators, and the average ranking of the two indicators is taken to select the top-ranked models.

[0071] Specifically, in step S2 of this embodiment, based on the segmented ERA5 reanalysis data and CMIP6 pattern data, and using the ERA5 reanalysis data as a reference, the CMIP6 patterns are screened using a traditional screening method. The specific steps are as follows:

[0072] S201. Traditional screening of CMIP6 models based on various indicators.

[0073] The Taylor-diagram provides a concise overview of the correspondence between the CMIP6 model and observed data. The Taylor Skill Score (TSS) is calculated based on σ and R from the Taylor-diagram.

[0074] ;

[0075] The interannual variability skill score (IVS) is used to quantify the consistency between the CMIP6 model and observed interannual variations.

[0076] ;

[0077] In the formula, m represents the model, o represents the observation. Since ERA5 reanalysis data has good applicability, ERA5 data is used instead of observation here. R0 is the maximum possible value of the correlation coefficient, equal to 0.999. TSS varies between 0 and 1, with models closer to 1 performing better. Similarly, the closer IVS is to 0, the smaller the interannual variation between the model and observations, and the better the model's performance.

[0078] Comprehensive rating indicator M R The simulation capability of a model is quantified by combining multiple evaluation indicators, and the calculation formula is as follows. Where n is the number of models, r... IVS and r RMSE This refers to the ranking of the pattern's IVS and RMSE. This embodiment uses these indicators for traditional screening of CMIP6 patterns. R The closer it is to 1, the better the pattern performs.

[0079] ;

[0080] Using the weighted total ozone column (AWM-TCO3) of the split (training set + validation set) ERA5 reanalysis data and CMIP6 model data, the TSS and M including IVS and RMSE of each CMIP6 model were calculated. R Indicators. Subsequently, based on the TSS and M of each model. R The results were sorted separately, and the average ranking of the two indicators was taken to finally select the top-ranked CMIP6 modes.

[0081] Furthermore, the SHAP decision tree method is used to filter CMIP6 patterns, including:

[0082] Based on the dataset, using the ERA5 reanalysis data as a reference and the CMIP6 pattern data as input, a decision tree model is trained.

[0083] Based on the SHAP value output by the decision tree model, the CMIP6 pattern is filtered using a preset filtering method.

[0084] Specifically, in step S3 of this embodiment, since the SHAP value can characterize the weighted contribution of the climate model, the larger its absolute value, the greater the model's influence on the prediction (i.e., the better the model), and vice versa. Therefore, the SHAP value method based on the decision tree model is used to screen the CMIP6 model, and the specific steps are as follows:

[0085] S301. Based on the pre-divided data (training set + validation set), using the ERA5 reanalysis data as a reference and the CMIP6 pattern data as input, first train the model using the training set, then fine-tune the parameters using the validation set to construct a decision tree model.

[0086] S302. Based on the SHAP value output by the decision tree model, develop three methods: Single Screening, Best-to-Worst Screening, and Worst-to-Best Screening to screen CMIP6 patterns, in order to solve the problem that SHAP values ​​are easily affected by extreme patterns (including extremely good or extremely bad patterns).

[0087] (1) Single Screening: Sort all patterns from largest to smallest based on their SHAP absolute value, and select the top-ranked patterns at once. (2) Best-to-Worst Screening: Remove the pattern with the largest SHAP absolute value each time to avoid over-reliance on a few high-impact patterns and improve its generalization ability. (3) Worst-to-Best Screening: Remove the pattern with the smallest SHAP absolute value each time, that is, remove the pattern with the least contribution and the most noise.

[0088] For (2) and (3), multiple screenings of the CMIP6 patterns are required (i.e., multiple elimination operations as described above). In each screening process, the absolute SHAP value of each pattern is calculated first. After multiple screenings are completed, the average absolute SHAP value of each pattern in all screening processes is calculated. Then, the average absolute SHAP values ​​are sorted from largest to smallest to obtain the patterns with the highest ranking, thereby ensuring statistical significance and eliminating the influence of extreme patterns.

[0089] Furthermore, a machine learning model is constructed based on the CMIP6 pattern data selected using traditional methods and the SHAP decision tree method; including:

[0090] The machine learning model was trained using the selected CMIP6 pattern data as input and ERA5 reanalysis data as a reference.

[0091] Use key metrics to tune the parameters of the trained machine learning model.

[0092] Specifically, in step S4 of this embodiment, based on the CMIP6 patterns obtained through four screening methods—a traditional method in step S2 and three decision tree SHAP value methods in step S3—different machine learning (ML) models are constructed to improve the multi-pattern ensemble method. The specific operations are as follows:

[0093] S401. Using CMIP6 pattern data obtained through four screening methods as input and ERA5 reanalysis data as reference, machine learning (ML) models such as generalized linear regression (GLM), random forest (RF), extreme gradient boosting (XGBoost), long short-term memory network (LSTM), and artificial neural network (ANN) are trained using the training set.

[0094] S402. On the validation set, the primary metric S (Score) is used to tune the machine learning (ML) model, ensuring that the tuning process considers the errors, correlations, and consistency of interannual variations among the data. The expression for the S metric is as follows:

[0095] ;

[0096] When IVS and MSE are smaller, and R is closer to 1, the predicted magnitude and trend of the reanalyzed data are more consistent. Therefore, the smaller S is, the better the model performance.

[0097] Table 1 shows the evaluation results of the four screening methods using different metrics. The evaluation metrics for each screening method in the table are the average of the evaluation metrics of the five machine learning models built based on that method on the test set. In the table, S = IVS × MSE × (1-R); as shown below:

[0098] Table 1. Evaluation results of the four screening methods using different indicators.

[0099] index Traditional Screening Single Screening Best-to-Worst Screening Worst-to-Best Screening MSE 246.836 241.057 243.641 231.435 R 0.789 0.768 0.756 0.786 IVS×10 0.725 1.207 0.604 0.687 S 3.771 6.750 3.593 3.435

[0100] Furthermore, based on the constructed machine learning model, the optimal screening method and the corresponding optimal machine learning model are determined, including:

[0101] For each constructed machine learning model, calculate the MSE, R, and IVS of each model on the dataset, and calculate the corresponding main metrics.

[0102] Obtain the average value of the main indicators of the model under each screening method, and obtain the main indicators corresponding to each screening method. The screening method with the smallest main indicator is the optimal screening method.

[0103] In machine learning models built based on the optimal selection method, the model with the smallest main index is further selected as the optimal model based on the main index of each model.

[0104] Specifically, in step S5 of this embodiment, based on the main indicator S described in step S4, the optimal screening method and the corresponding optimal machine learning model are determined. The specific operation is as follows:

[0105] S501. For the five machine learning models (4×5=20 models in total) constructed for the four screening methods, calculate the MSE, R, and IVS of each model on the test set, and calculate the corresponding main index S. Then, take the average S of the five models under each screening method to obtain the S corresponding to each screening method, and select the screening method with the smallest S as the optimal screening method.

[0106] S502. Among the five machine learning models constructed based on the optimal selection method, the model with the smallest S index is further selected as the optimal model based on the S index of each model.

[0107] Specifically, in step S6 of this embodiment, the optimal screening method and optimal model obtained in S5 are used to predict the recovery of Antarctic ozone under different greenhouse gas emission scenarios (SSPs). The specific steps are as follows:

[0108] S601. Under each greenhouse gas emission scenario (SSP), the CMIP6 model obtained by the optimal screening method is input into the optimal ML model to obtain the Antarctic ozone recovery prediction curve using machine learning improved multi-model ensemble (MME-ML).

[0109] S602. Using the area-weighted total ozone column (AWM-TCO3) data from the actual ERA5 reanalysis data as a benchmark, determine the Antarctic ozone recovery year obtained by the optimal machine learning model corresponding to the optimal screening method under different SSPs.

[0110] The final estimated recovery curve and recovery year results for Antarctic ozone are as follows: Figure 2 As shown. Figure 2 (a) and Figure 2 In (b), the purple and black solid lines represent the European Reanalysis 5 (ERA5) data and the 9-year boxcar filter values, respectively. The black dashed line is the 1980 reference value. The dark blue, light blue, orange, and red solid lines represent the 9-year boxcar filter predictions for different greenhouse gas emission scenarios (SSP126, SSP245, SSP370, and SSP585), respectively. Comparing these two figures, it is clear that compared to the general multi-model ensemble method (MME), the machine learning-improved ensemble prediction method (MME-ML-ANN) significantly corrects the simulation bias of the CMIP6 model, resulting in more reliable predictions.

[0111] Figure 2 (c) and Figure 2 In (d), the Antarctic ozone recovery years for SSP126, SSP245, SSP370, and SSP585 are represented by dark blue circles, light blue squares, orange triangles, and red rhombuses, respectively. Figure 2As shown in (c), under the low-to-medium emission scenarios (SSP126, SSP245) predicted by the MME method, the Antarctic ozone recovery years are 2056 and 2051, respectively. Using σ as the standard deviation of all models from the multi-model average, the ±1σ recovery years are 2037-2078 and 2039-2081, respectively. Under the higher emission scenarios (SSP370, SSP585), the Antarctic ozone recovery years are 2050 and 2049, respectively, with ±1σ recovery years of 2034-2070 and 2035-2068, respectively. Figure 2 The MME-ML-ANN projections in (d) indicate that the recovery year for Antarctic ozone is 2065 under scenario SSP126, 2062 under SSP245, and 2057 and 2056 under SSP370 and SSP585, respectively. The improved MME-ML-ANN results show a significant delay of 7-11 years compared to the traditional MME.

[0112] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A climate model optimization and ensemble prediction method based on interpretability and machine learning, characterized in that, include: Obtain the area-weighted average total ozone column concentration from reanalysis data and CMIP6 model data, and construct a dataset; Based on the dataset, the CMIP6 patterns were filtered using both traditional methods and the decision tree SHAP method. A machine learning model is constructed based on CMIP6 pattern data selected using traditional methods and the SHAP decision tree method. Based on the constructed machine learning model, the optimal screening method and the corresponding optimal machine learning model are determined; Based on the optimal screening method and the corresponding optimal machine learning model, the recovery prediction of Antarctic ozone under different greenhouse gas emission scenarios is carried out. Obtaining the total area-weighted average ozone column includes: Calculate the total ozone column volume using CMIP6 ozone mixing ratio data; Based on the total ozone column TCO3 and CMIP6 calculated from the reanalysis data, an area-weighted average was performed to obtain the area-weighted average total ozone column. Traditional methods for screening CMIP6 patterns include: Based on the dataset, the Taylor Skill Score (TSS) and the Comprehensive Rating Index (M) for each CMIP6 mode were calculated. R ; Based on the Taylor Skill Score (TSS) and the overall rating index (M) for each model R The two indicators are sorted separately, and the average ranking of the two indicators is taken to finally select the top-ranked CMIP6 modes. The comprehensive rating indicator M R The expression is: ； Where n is the number of patterns, r IVS and r RMSE It is a ranking of the interannual variability indices IVS and root mean square error RMSE of the model. The SHAP decision tree method was used to filter CMIP6 patterns, including: Based on the dataset, using the reanalysis data therein as a reference and CMIP6 pattern data as input, a decision tree model is trained. Based on the SHAP value output by the decision tree model, the CMIP6 pattern is filtered using a preset filtering method; A machine learning model is constructed based on CMIP6 pattern data selected using traditional methods and the SHAP decision tree method; including: The machine learning model was trained using the selected CMIP6 pattern data as input and the reanalysis data as a reference. The primary metric S is used to tune the parameters of the trained machine learning model.

2. The climate model optimization and ensemble prediction method based on interpretability and machine learning according to claim 1, characterized in that, The preset filtering method includes: Single-time filtering: Sort all patterns from largest to smallest based on the absolute value of SHAP corresponding to each pattern, and filter out the top-ranked patterns in one go; Filter from best to worst: Remove the pattern with the largest absolute SHAP value each time; Poor-to-good filtering: Each time, the pattern with the smallest absolute SHAP value is removed; For both the best-to-worst and worst-to-best screening, the CMIP6 patterns need to be screened multiple times. In each screening process, the absolute SHAP value of each pattern is calculated first. After multiple screenings are completed, the average absolute SHAP value of each pattern in all screening processes is calculated. Then, the average absolute SHAP values ​​are sorted from largest to smallest to obtain the patterns with the highest ranking.

3. The climate model optimization and ensemble prediction method based on interpretability and machine learning according to claim 1, characterized in that, The expression for the main indicator S is: ； in, As an indicator of interannual variability, Mean square error, The correlation coefficient.

4. The climate model optimization and ensemble prediction method based on interpretability and machine learning according to claim 1, characterized in that, Based on the constructed machine learning model, the optimal screening method and the corresponding optimal machine learning model are determined, including: For each constructed machine learning model, calculate the MSE, R, and IVS of each model on the test dataset, and calculate the corresponding main metrics. Obtain the average value of the main indicators of the model under each screening method, and obtain the main indicators corresponding to each screening method. The screening method with the smallest main indicator is the optimal screening method. In machine learning models built based on the optimal selection method, the model with the smallest main index is further selected as the optimal model based on the main index of each model.

Images