Climate integrated prediction method and system based on the fusion of multiple artificial intelligence models

Abstract

The present application discloses a climate integrated prediction method and system based on the fusion of multiple artificial intelligence models, which belongs to the field of climate prediction. It includes: selecting elements that are physically correlated with the prediction target as prediction factors; selecting multiple artificial intelligence models, and pre-training the multiple artificial intelligence models based on historical observation data sets; conducting independent return tests on the multiple artificial intelligence models obtained by pre-training, and selecting multiple target artificial intelligence models based on merit; analyzing the relative contribution of each prediction factor to the prediction target, and determining the actual prediction factor corresponding to each target artificial intelligence model. Finally, a climate integrated prediction model is constructed based on multiple target artificial intelligence models and the actual prediction factors corresponding to each target artificial intelligence model. In this way, linear and nonlinear features can be fused and predicted, effectively improving the generalization and stability of the model, and improving prediction efficiency and accuracy.

Description

Technical Field

[0001] The present application relates to the field of climate prediction, and in particular to a climate integrated prediction method and system based on the fusion of multiple artificial intelligence models. Background Art

[0002] Climate is a long-term background condition for many social production and life, and is also a strategic resource with huge potential for economic benefits. Accurate climate forecasting, on the one hand, helps to warn and respond to climate events and disasters such as floods and droughts, heat waves and cold waves, and protects the safety of people's lives and property and the safe and orderly development of work in various industries, and plays an important role in supporting and safeguarding; on the other hand, it plays an important role in the deployment and guidance of the development and utilization of renewable resources and energy, such as the high-quality layout and production of new energy such as hydropower, wind power, and solar energy.

[0003] There are two main traditional climate prediction methods: statistical prediction and dynamic prediction. Statistical prediction refers to finding the statistical laws that determine the precursor signals of climate variation based on historical observation data through mathematical statistics and other means, and then predicting the future conditions of climate elements; dynamic prediction refers to numerical simulation of the earth system based on atmospheric dynamics equations through numerical calculation methods, and obtaining numerical solutions for future climate element variations through iterative calculations.

[0004] However, in the above two schemes, the statistical prediction scheme only seeks statistical laws through historical observations and cannot predict unknown events. Its limitations are becoming more and more obvious under the current situation where the climate background is constantly changing and extreme events are more frequent. The dynamic prediction scheme has errors in the initial conditions, numerical calculations, and other processes, and the mathematical description of atmospheric physical processes is not complete, which leads to huge instability in the prediction results. In addition, the numerical dynamic model consumes a lot of computing time and resources, and has considerable disadvantages in timeliness and economy. Therefore, the existing climate prediction methods cannot meet the actual needs of current high-quality development, and it is necessary to use new means and tools to develop more efficient and economical climate prediction methods. Summary of the invention

[0005] This application provides a climate integrated prediction method and system based on the fusion of multiple artificial intelligence models, which can construct a climate integrated prediction model based on the fusion of multiple artificial intelligence models. Based on the climate integrated prediction model, climate prediction can be performed to achieve nonlinear, objective and quantitative integrated prediction of the prediction target, thereby achieving more accurate and more economical climate prediction. The technical solution includes the following contents.

[0006] In a first aspect, a climate integrated prediction method based on the fusion of multiple artificial intelligence models is provided, the method comprising:

[0007] Based on the dynamic theory and climate predictability theory, elements that are physically related to the prediction target are selected from the Earth system big data as prediction factors;

[0008] Selecting multiple artificial intelligence models, and using historical observation data sets with the same interdecadal background as the prediction target as training sets to pre-train the multiple artificial intelligence models;

[0009] Based on the historical climate data of the target period, the plurality of artificial intelligence models obtained by pre-training are respectively subjected to feedback tests, and based on the feedback test results corresponding to the various artificial intelligence models, a plurality of target artificial intelligence models are selected from the plurality of artificial intelligence models;

[0010] For each of the multiple target artificial intelligence models, analyze the relative contribution of each prediction factor to the prediction target, and select a target prediction factor from the multiple prediction factors as an actual prediction factor corresponding to the target artificial intelligence model based on the analysis result;

[0011] Based on the multiple target artificial intelligence models and the actual prediction factors corresponding to each of the target artificial intelligence models, a climate integrated prediction model based on the fusion of multiple artificial intelligence models is constructed, and the climate integrated prediction model is used to predict the prediction target.

[0012] Compared with the prior art, the above technical solution has the following beneficial effects:

[0013] (1) Select multiple artificial intelligence models for predictive modeling, and select appropriate multiple target artificial intelligence models through independent return optimization schemes to merge into the final prediction model. This model can integrate the advantages of multiple artificial intelligence models and take into account the linear and nonlinear parts of the prediction system. At the same time, the model selection references the return test results of the target period that is closer to the actual prediction target, so that the model itself is more in line with the actual situation of the prediction target time, making the final prediction model more scientific and reasonable.

[0014] (2) Based on climate dynamics and climate predictability, the present invention selects in advance the elements that are predictable for the prediction target as prediction factors, and analyzes the contributions of different prediction factors for different models through an artificial intelligence interpretation scheme, and selects the prediction factor with the highest contribution to the prediction target as the actual prediction factor, which effectively improves the physical interpretability of the model and reduces the complexity of the model.

[0015] (3) The present invention selects a training set with the same interdecadal background as the prediction target, and can focus the training on the period with the same climate background as the prediction target. In this way, the impact of difficult-to-predict interdecadal signals on the prediction model can be reduced, while solving the limitation of limited time samples of Earth system big data.

[0016] (4) The present invention optimizes the final prediction model multiple times, and the final prediction model is still an integrated prediction model that integrates multiple advantageous artificial intelligence models. Compared with a single model, this model has better generalization and stability, and compared with a large artificial intelligence model, it has better computational efficiency and interpretability, and can provide more scientific, effective and objective prediction results.

[0017] Optionally, selecting elements that are physically related to the prediction target from the Earth system big data as prediction factors includes:

[0018] Based on the singular value decomposition method and the empirical orthogonal decomposition method, elements that have a telecorrelation relationship and a lead-lag relationship with the prediction target are selected from the earth system big data as the prediction factors.

[0019] Optionally, the selecting of multiple artificial intelligence models includes:

[0020] A variety of artificial intelligence models suitable for time series fitting and prediction are selected, and the various artificial intelligence models include linear regression models, generalized linear models, support vector machine models, tree models, shallow neural network models and recurrent neural network models.

[0021] Optionally, the pre-training of the plurality of artificial intelligence models using a historical observation dataset having the same decadal background as the prediction target as a training set includes:

[0022] Obtain historical climate data for 30 years before the forecast period;

[0023] The historical climate data for 30 years before the period to be predicted is used as the historical observation data set to pre-train the multiple artificial intelligence models.

[0024] Optionally, the historical climate data of the target period is used to perform feedback tests on the plurality of artificial intelligence models obtained by pre-training respectively, and a plurality of target artificial intelligence models are selected from the plurality of artificial intelligence models based on the feedback test results corresponding to the various artificial intelligence models, including:

[0025] Dividing the historical climate data of the target period into a plurality of training sets;

[0026] For any one of the plurality of artificial intelligence models, respectively inputting the plurality of training sets into the artificial intelligence model to obtain a plurality of prediction results;

[0027] Based on the multiple prediction results and the actual observation results corresponding to the multiple training sets, the artificial intelligence model is evaluated by means of deviation correlation coefficient, root mean square error and comprehensive score to obtain the return test result corresponding to the artificial intelligence model;

[0028] Based on the return test results corresponding to the various artificial intelligence models, the plurality of artificial intelligence models are sorted;

[0029] A preset number of artificial intelligence models are selected from the sorted multiple artificial intelligence models as the multiple target artificial intelligence models.

[0030] Optionally, analyzing the relative contribution of each prediction factor to the prediction target, and selecting a target prediction factor from a plurality of prediction factors as an actual prediction factor corresponding to the target artificial intelligence model based on the analysis result, includes:

[0031] Perform interpretability analysis on each of the predictors by means of relative importance parameter analysis, local interpretable model interpreter LIME, or Shapley value to determine the relative contribution value of each of the predictors to the prediction result output by the target artificial intelligence model;

[0032] A prediction factor whose relative contribution value is greater than or equal to a preset contribution threshold is selected from the multiple prediction factors as the target prediction factor to obtain an actual prediction factor corresponding to the target artificial intelligence model.

[0033] Optionally, the climate integrated prediction model based on the fusion of multiple artificial intelligence models is constructed based on the multiple target artificial intelligence models and the actual prediction factors of each of the target artificial intelligence models, including:

[0034] Calling the model parameters of each of the target artificial intelligence models after pre-training as initialization parameters for re-training the climate integrated prediction model;

[0035] The historical climate data corresponding to the prediction factors of the 30 years before the prediction period are selected as the actual training set, and the climate integrated prediction model is trained again. After the training is completed, the climate integrated prediction model is obtained.

[0036] In a second aspect, a climate integrated prediction method based on the fusion of multiple artificial intelligence models is provided, the method comprising:

[0037] Obtaining climate data corresponding to actual prediction factors corresponding to multiple target artificial intelligence models in a target period before the period to be predicted, wherein the climate integrated prediction model includes the multiple target artificial intelligence models, and the climate integrated prediction model is constructed based on the method described in the first aspect;

[0038] Inputting the climate data corresponding to the actual prediction factors corresponding to the multiple target artificial intelligence models into the multiple target artificial intelligence models to obtain multiple prediction results;

[0039] The multiple prediction results are averaged to obtain a final integrated prediction result.

[0040] In this application, the climate data corresponding to the actual prediction factors of multiple target artificial intelligence models in the target period before the period to be predicted are first obtained, and then the climate data corresponding to the actual prediction factors of multiple target artificial intelligence models are input into multiple target artificial intelligence models, and the input climate data are processed by multiple target artificial intelligence models to obtain multiple prediction results. In this way, multiple artificial intelligence models can capture linear and nonlinear signals of climate data, and climate prediction can be performed by integrating a climate integrated prediction model that integrates multiple artificial intelligence models, so that more accurate integrated prediction results can be obtained.

[0041] In a third aspect, a climate integrated prediction system based on the fusion of multiple artificial intelligence models is provided, the system comprising:

[0042] A selection module is used to select elements that are physically related to the prediction target from the Earth system big data as prediction factors based on the dynamic theory and climate predictability theory;

[0043] A pre-training module, used to select a plurality of artificial intelligence models, and use a historical observation data set having the same interdecadal background as the prediction target as a training set to pre-train the plurality of artificial intelligence models;

[0044] A model verification model, used to perform feedback verification on the plurality of artificial intelligence models obtained by pre-training based on historical climate data of a target period, and select a plurality of target artificial intelligence models from the plurality of artificial intelligence models based on the feedback verification results corresponding to the various artificial intelligence models;

[0045] An explanatory analysis module, for analyzing, for each of the multiple target artificial intelligence models, the relative contribution of each prediction factor to the prediction target, and selecting a target prediction factor from the multiple prediction factors as an actual prediction factor corresponding to the target artificial intelligence model based on the analysis result;

[0046] A construction module is used to construct a climate integrated prediction model based on the fusion of multiple artificial intelligence models based on the multiple target artificial intelligence models and the actual prediction factors corresponding to each of the target artificial intelligence models, and the climate integrated prediction model is used to predict the prediction target.

[0047] In a fourth aspect, a computer system is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is loaded into the processor, the climate integrated prediction method based on the fusion of multiple artificial intelligence models in the first aspect or the second aspect is implemented.

[0048] It can be understood that the beneficial effects of the third and fourth aspects mentioned above can be found in the relevant description of the first aspect mentioned above, and will not be repeated here. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0050] Figure 1 It is a flow chart of a climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in an embodiment of the present application;

[0051] Figure 2 This is a specific operation flow chart of an independent reward optimization solution provided in an embodiment of the present application;

[0052] Figure 3 It is a specific operation flow chart of an interpretable solution of an artificial intelligence model provided in an embodiment of the present application;

[0053] Figure 4 It is a flowchart of another climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in an embodiment of the present application;

[0054] Figure 5 This is a flowchart of another climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in an embodiment of the present application;

[0055] Figure 6 It is a comparative schematic diagram of a feedback inspection result provided in an embodiment of the present application;

[0056] Figure 7 is a comparative schematic diagram of an explanatory analysis result provided in an embodiment of the present application;

[0057] Figure 8is a comparative schematic diagram of a prediction evaluation result provided in an embodiment of the present application;

[0058] Fig. 9 It is a structural schematic diagram of a climate integrated prediction system based on the fusion of multiple artificial intelligence models provided in an embodiment of the present application;

[0059] Fig.10 It is a structural diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0060] In order to make the objectives, technical solutions and advantages of the present application clearer, the implementation methods of the present application will be further described in detail below with reference to the accompanying drawings.

[0061] It should be understood that the "multiple" mentioned in this application refers to two or more. In the description of this application, unless otherwise specified, " / " means or, for example, A / B can mean A or B; "and / or" in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships, for example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. In addition, in order to facilitate the clear description of the technical solution of this application, the words "first" and "second" are used to distinguish between the same items or similar items with basically the same functions and effects. Those skilled in the art can understand that the words "first" and "second" do not limit the quantity and execution order, and the words "first" and "second" do not limit them to be different.

[0062] With the leapfrog improvement of computer performance, especially the continuous maturity of technologies such as GPU high-performance computing, artificial intelligence models are gradually emerging in weather forecasting. Thanks to the nonlinear fitting and prediction capabilities of artificial intelligence models, weather forecasting models based on artificial intelligence technology, such as Huawei's "Pangu Meteorological Model", NVIDIA's "FourCastNet" and Google's "GraphCast", are based on artificial intelligence technologies such as computer vision. They can complete weather forecasts in a very short time that are not inferior to existing mainstream business models, demonstrating the amazing potential of artificial intelligence technology in climate forecasting.

[0063] However, the above technologies also have defects and shortcomings. First, the technology based on computer vision itself has "artifacts", which makes the error in long-term predictions exponentially amplified, thus failing to meet the prediction needs of climate scale. Second, the model training is based on limited historical data, and has limited adaptability to future climate background changes. The training cost of the large model itself is extremely high, so it cannot be quickly updated. Third, the above large model only regards the earth system as a pure data set and highly abstracts the climate system into a huge array. The resulting prediction model lacks physical interpretability. In summary, it is of great significance to use artificial intelligence technology to build an efficient prediction system based on climate dynamics and climate predictability.

[0064] The following is a detailed explanation of the climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the embodiment of the present application.

[0065] Figure 1 This is a flow chart of a climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in an embodiment of the present application. Figure 1 , the method comprises the following steps.

[0066] Step 101: Based on the dynamics theory and climate predictability theory, select elements that are physically related to the prediction target from the Earth system big data as prediction factors.

[0067] Earth system big data refers to the massive data about the Earth system (including the atmosphere, ocean, land, etc.) collected through various observation methods and sensor networks. These include meteorological data (temperature, humidity, air pressure, wind speed, wind direction, rainfall, etc.), climate data (global temperature changes, precipitation patterns, extreme weather events, etc.), etc.

[0068] The prediction target refers to the goal of climate prediction, that is, the climate to be predicted. For example, the prediction target is to predict the abnormal summer precipitation in country A.

[0069] In the embodiment of the present application, the prediction factor refers to an element that has a physical correlation with the prediction target, that is, the prediction factor can comprehensively reflect the situation of the prediction target in some aspects.

[0070] Optionally, the operation of step 101 may be: based on the singular value decomposition method and the empirical orthogonal decomposition method, selecting elements having a telecorrelation relationship and a lead-lag relationship with the prediction target from the earth system big data as prediction factors.

[0071] As an implementation method, target data is acquired, and singular value decomposition and empirical orthogonal decomposition are performed on the target data based on the Earth system big data to obtain all elements related to the prediction target. From all elements related to the prediction target, elements that have telecorrelation relationships and lead-lag relationships with the prediction target are selected as prediction factors.

[0072] Target data refers to the data of the prediction target in history. For example, if the prediction target is to predict the abnormal precipitation in summer in country A, then the target data can be the abnormal precipitation in summer in country A in history, for example, the abnormal precipitation in summer in country A within 30 years.

[0073] Teleconnection refers to a statistically significant correlation between meteorological or climate variables between two or more geographically distant regions, which usually exhibits a long-term, stable pattern.

[0074] Lead-lag relationship refers to the temporal chronological relationship between two or more climate variables. This relationship indicates that changes in one variable can predict changes in another variable, or that changes in one variable lag behind changes in another variable. Alternatively, the cross-correlation function is a common method to detect lead-lag relationships between two time series.

[0075] In this case, elements that have a physical relationship with the prediction target are selected as prediction factors instead of putting all the elements into the artificial intelligence model for training. This can effectively simplify the artificial intelligence model, focus the prediction on the predictable part, and improve the prediction efficiency and effect.

[0076] Step 102: Select multiple artificial intelligence models, and use historical observation data sets with the same interdecadal background as the prediction target as training sets to pre-train the multiple artificial intelligence models.

[0077] The types of various artificial intelligence models are different, and various artificial intelligence models include linear and nonlinear models. In an embodiment of the present application, various artificial intelligence models suitable for time series fitting and prediction can be selected. Optionally, various artificial intelligence models can include linear regression models, generalized linear models, support vector machine models, tree models, shallow neural network models, and recurrent neural network models. Among them, recurrent neural networks include long short-term memory network models and gated recurrent neural network models.

[0078] In an embodiment of the present application, by selecting multiple artificial intelligence models of different types, the multiple artificial intelligence models can capture the linear structure and nonlinear structure of the input data, thereby performing more comprehensive feature extraction on the input data and predicting more accurate results.

[0079] The historical observation data set (training set) includes multiple training samples and sample labels. Each of the multiple training samples includes a data value corresponding to a prediction factor, and the sample label is a real observation value (that is, the actual climate) corresponding to the data value corresponding to the prediction factor.

[0080] Since the climate system has climate signals of different periods such as interdecadal variability, the above predictions focus on predictable interannual and seasonal scales, and select historical climate data with the same interdecadal background as the prediction target as the historical observation data set to pre-train multiple artificial intelligence models. In this way, more accurate prediction results can be obtained later.

[0081] Among them, the operation of using the historical observation data set with the same interdecadal background as the prediction target as the training set to pre-train multiple artificial intelligence models can be: obtaining the historical climate data for 30 years before the period to be predicted; using the historical climate data for 30 years before the period to be predicted as the historical observation data set to pre-train multiple artificial intelligence models.

[0082] According to climate research, the climate changes periodically. In the field of climate research and business, 30 years is usually regarded as a period of climate state, and it is believed that the climate in this period has a similar interdecadal background. Therefore, in the embodiment of the present application, the historical climate data of 30 years before the period to be predicted can be used as a historical observation data set to pre-train multiple artificial intelligence models.

[0083] It is worth noting that, through the above step 102, a variety of artificial intelligence models after pre-training can be obtained, and the various artificial intelligence models after pre-training are also models with preliminary prediction capabilities.

[0084] Step 103: Based on the historical climate data of the target period, the multiple artificial intelligence models obtained through pre-training are respectively subjected to feedback tests, and based on the feedback test results corresponding to the various artificial intelligence models, multiple target artificial intelligence models are selected from the multiple artificial intelligence models.

[0085] The historical climate data of the target period is used to perform feedback tests on the multiple artificial intelligence models. In an embodiment of the present application, the historical climate data of the target period can be the climate data of 5 time steps before the period to be predicted. For example, the historical climate data of the target period can be the climate data within 5 years before the period to be predicted, and for another example, the historical climate data of the target period can be the climate data of 5 months before the period to be predicted.

[0086] In this case, by performing a return test on the various artificial intelligence models obtained by pre-training, the prediction capabilities of various artificial intelligence models can be known. Then, by selecting multiple target artificial intelligence models from the multiple artificial intelligence models, the target artificial intelligence model that can accurately predict the climate can be determined, thereby improving the prediction accuracy of the prediction target.

[0087] Optionally, the operation of step 103 may include the following steps (1) to (5).

[0088] (1) Divide the historical climate data of the target period into multiple training sets.

[0089] In the embodiment of the present application, the historical climate data of the target period can be divided into multiple training sets according to the time series. For example, if the historical climate data of the target period is the climate data within 5 years before the period to be predicted, then the historical climate data of the target period can be divided into 5 training sets in units of time, and each training set includes 1 year of climate data.

[0090] (2) For any one of the multiple artificial intelligence models, multiple training sets are input into the artificial intelligence model to obtain multiple prediction results.

[0091] In this case, an artificial intelligence model makes multiple independent predictions based on multiple training sets to obtain multiple prediction results. In other words, for each training set in multiple training sets, the training set is input into the artificial intelligence model, and the artificial intelligence model processes the input data and outputs the prediction result corresponding to the training set. For multiple training sets, there are multiple prediction results corresponding to the multiple training sets.

[0092] It is worth noting that by executing the above step (2) for each of the multiple artificial intelligence models, multiple prediction results corresponding to each artificial intelligence model can be obtained, that is, each artificial intelligence model corresponds to multiple prediction results.

[0093] (3) Based on multiple prediction results and actual observation results corresponding to multiple training sets, the artificial intelligence model is evaluated by means of deviation correlation coefficient, root mean square error and comprehensive score to obtain the return test results corresponding to the artificial intelligence model.

[0094] The Anomaly Correlation Coefficient (ACC) is used to evaluate the similarity between two time series, usually by comparing the anomalies of the two time series to measure their correlation. It is often used to evaluate the performance of a forecasting model.

[0095] The value range of the anomaly correlation coefficient is between -1 and 1. ACC=1 means that the anomalies between the predicted value and the actual observed value are exactly the same. ACC=0 means that there is no correlation between the anomalies between the predicted value and the actual observed value. ACC=-1 means that the anomalies between the predicted value and the actual observed value are completely opposite.

[0096] Assume that there are two time series (X) and (Y), X represents the actual observation value, Y represents the predicted value, and the length of each series is n. Then the anomaly correlation coefficient can be calculated by the following process.

[0097] 1. Calculate the long-term average

[0098] The long-term average of X is: , where t represents the tth time point in the time series, Represents the actual observation value at the tth time point.

[0099] The long-term average of Y is: , Represents the predicted value at the tth time point.

[0100] 2. Calculate the anomaly

[0101] Compute the deviation of X:

[0102] Calculate the deviation of Y:

[0103] 3. Calculate the anomaly correlation coefficient

[0104]

[0105] Among them, ACC is the anomaly correlation coefficient, for and The covariance of for The standard deviation of for The standard deviation of .

[0106] The root mean square error (RMSE) is used to measure the square root of the average of the squares of the differences between the predicted values ​​and the actual observed values, which is used to evaluate the difference between the predicted values ​​and the actual observed values. The smaller the RMSE, the smaller the difference between the predicted values ​​and the actual observed values, and thus the higher the prediction accuracy of the model.

[0107] Assume that there are two time series (X) and (Y), X represents the actual observation value, Y represents the predicted value, and the length of each series is n. Then the root mean square error between X and Y can be determined by the following formula (1).

[0108] (1)

[0109] Here, RMSE is the root mean square error between X and Y.

[0110] Performance Score (PS) is a method for evaluating the comprehensive performance of multiple indicators or factors. Specifically, the scores of multiple indicators are combined into a single score to evaluate the overall performance. Generally, the comprehensive score is determined by a weighted scoring method or a standardized weighted average method. For example, in an embodiment of the present application, the deviation correlation coefficient and the root mean square error can be combined into a single score to obtain a comprehensive score.

[0111] Optionally, the operation of step (3) may be: for any one of the multiple training sets, based on the prediction results and actual observation results corresponding to the training set, the prediction results of the training set on the artificial intelligence model are evaluated by means of deviation correlation coefficient, root mean square error and comprehensive score to obtain the evaluation results corresponding to the training set; the evaluation results corresponding to the multiple training sets are averaged to obtain the return test results corresponding to the artificial intelligence model.

[0112] It should be understood that when evaluating by means of the deviation correlation coefficient, the root mean square error and the comprehensive score, each training set will correspond to three evaluation results.

[0113] In addition, when there are three evaluation results corresponding to each training set, the same evaluation results corresponding to multiple training sets can be averaged to obtain the return test result corresponding to this artificial intelligence model. In other words, the deviation correlation coefficients corresponding to multiple training sets are first averaged to obtain the deviation correlation coefficient corresponding to this artificial intelligence model; then the root mean square errors corresponding to multiple training sets are averaged to obtain the root mean square error corresponding to this artificial intelligence model; finally, the comprehensive scores corresponding to multiple training sets are averaged to obtain the comprehensive score corresponding to this artificial intelligence model.

[0114] The evaluation results corresponding to multiple training sets can be weighted averaged to obtain the return test results corresponding to the artificial intelligence model. For example, a higher weight can be set for the evaluation results corresponding to the training set close to the time period to be predicted, so that the evaluation pays more attention to the prediction results of the artificial intelligence model for the recent period, thereby improving the accuracy of the model evaluation.

[0115] It is worth noting that by performing the above step (3) on the prediction results corresponding to each of the multiple artificial intelligence models, the return test results corresponding to each of the multiple artificial intelligence models can be obtained.

[0116] (4) Rank various AI models based on their corresponding return test results.

[0117] The return test results can include deviation correlation coefficient, root mean square error and comprehensive score.

[0118] In this case, the operation of step (4) can be: based on the deviation correlation coefficients corresponding to various artificial intelligence models, sort the multiple artificial intelligence models in order from large to small to obtain a first sorting result; based on the root mean square errors corresponding to various artificial intelligence models, sort the multiple artificial intelligence models in order from small to large to obtain a second sorting result; based on the comprehensive scores corresponding to various artificial intelligence models, sort the multiple artificial intelligence models in order from large to small to obtain a third sorting result; based on the first sorting result, the second sorting result and the third sorting result, adjust the order of the multiple artificial intelligence models to obtain the target sorting result.

[0119] The target ranking result is the ranking result of multiple artificial intelligence models obtained by combining the above three evaluation results. It should be understood that the artificial intelligence models with high rankings are models that appear to be excellent in all three evaluation indicators.

[0120] In this way, each evaluation indicator has a corresponding ranking. Subsequently, by integrating the model ranking results corresponding to multiple evaluation indicators, a more accurate target ranking result can be determined, and then a more accurate target artificial intelligence model can be selected.

[0121] (5) Selecting a preset number of artificial intelligence models from the sorted plurality of artificial intelligence models as multiple target artificial intelligence models.

[0122] The preset number can be set in advance, for example, the preset number can be set to 3.

[0123] In this case, it is equivalent to obtaining an artificial intelligence model that performs well in all three evaluation indicators from the target ranking results as the target artificial intelligence model, so that a more accurate target artificial intelligence model can be obtained.

[0124] In an embodiment of the present application, an independent return optimization scheme is designed in step 103. The scheme is based on the pre-training scheme in step 102. According to the actual prediction task, independent multi-model returns are performed for nearly 5 time steps (such as the past 5 years or the past 5 months) of the prediction period, and the independent return results are evaluated by evaluation methods such as correlation coefficient, root mean square error, and comprehensive score. The model with the highest return effect is selected as a component of the actual prediction model, thereby forming a prediction model that is more suitable for the current prediction task.

[0125] For example, Figure 2 This is a specific operation flow chart of an independent reward optimization solution provided in the embodiment of the present application, see Figure 2 ,The specific operation of the independent return optimization scheme includes the following steps (1) - (3).

[0126] (1) Identify multiple alternative artificial intelligence models and predictor data sets suitable for the prediction target.

[0127] (2) Based on the corresponding prediction factors of the 5 time steps before the period to be predicted (corresponding to Figure 2 There are 5 sets of training sets and their corresponding prediction results, which are compared with the actual observation results (corresponding Figure 2 The T-1 historical observations, T-2 historical observations, T-3 historical observations, T-4 historical observations, T-5 historical observations, etc. are used for evaluation and analysis to obtain multiple evaluation results and average them to obtain the return test results, in order to examine the actual fitting return capabilities of different models in the recent period.

[0128] For any one of the multiple artificial intelligence models, each of the five training sets is input into the artificial intelligence model, and the artificial intelligence model processes the input data to obtain five sets of prediction results. In other words, for each of the five training sets, one training set is input into an artificial intelligence model, and the artificial intelligence model processes the input data and outputs the prediction results corresponding to the training set. Five sets of prediction results can be obtained by performing predictions based on the five training sets by the artificial intelligence model.

[0129] Then, the artificial intelligence model is evaluated based on these 5 sets of prediction results and the actual observation results corresponding to the 5 sets of training sets to obtain 5 sets of evaluation results. Then, the 5 sets of evaluation results are averaged to obtain the return test results corresponding to this artificial intelligence model.

[0130] (3) Based on the return test results, several target artificial intelligence models with better return effects are selected as the final climate integrated prediction model that integrates multiple artificial intelligence models.

[0131] Step 104: For each target artificial intelligence model among the multiple target artificial intelligence models, the relative contribution of each prediction factor to the prediction target is analyzed, and based on the analysis results, a target prediction factor is selected from the multiple prediction factors as the actual prediction factor corresponding to the target artificial intelligence model.

[0132] For each prediction factor, the relative contribution of a prediction factor to the prediction target refers to the influence of this prediction factor on the prediction results of the target artificial intelligence model. It should be understood that the greater the influence of a prediction factor on the prediction results of the target artificial intelligence model, the greater the relative contribution of this prediction factor to the prediction target, and thus this prediction factor is more important for model prediction.

[0133] In this case, by analyzing the contribution of different predictors for different models and selecting the target predictor as the actual predictor corresponding to the target artificial intelligence model, the physical interpretability of the model is effectively improved and the complexity of the model is reduced.

[0134] Optionally, the operation of step 104 may include the following steps (1) and (2).

[0135] (1) Perform interpretability analysis on each predictor through relative importance parameter analysis, local interpretable model interpreter LIME, or Shapley value to determine the relative contribution of each predictor to the prediction result output by the target artificial intelligence model.

[0136] Relative importance parameter analysis is used to evaluate the relative contribution of multiple independent variables (multiple predictors) in explaining the variation of the dependent variable (prediction target). In general, the relative importance of each predictor can be calculated by calculating standardized regression coefficients and variance decomposition.

[0137] The Local Interpretable Model Explainer (LIME) is a tool for explaining the prediction results of machine learning models. Its core idea is to approximate the behavior of complex models in a local range and explain a single prediction by generating a simple, interpretable model. Its basic working principle is: first generate a series of perturbation samples, and then use a complex model to predict the perturbation samples. Then, train a simple explanation model (such as a linear model or decision tree) based on the prediction results of the perturbation samples. Finally, use the trained explanation model to explain the prediction results.

[0138] The Shapley value is used to measure the contribution of each feature to the model prediction. In the embodiment of the present application, the Shapley value is used to measure the contribution of each prediction factor to the model prediction.

[0139] It is worth noting that since different target artificial intelligence models may be different types of models, there are different interpretability methods for different types of models. Therefore, for different target artificial intelligence models, the corresponding explanatory analysis method can be selected to perform explanatory analysis on each predictor. In this way, a more accurate relative contribution value can be determined. For example, for a linear regression model, the relative importance parameter analysis method can be selected to determine the relative contribution value. For another example, for a recurrent neural network, the LIME method can be selected to determine the relative contribution value.

[0140] (2) Select a predictor whose relative contribution value is greater than or equal to a preset contribution threshold from multiple predictors as the target predictor to obtain the actual predictor corresponding to the target artificial intelligence model.

[0141] The preset contribution threshold may be set in advance, for example, the preset contribution threshold may be set larger.

[0142] In this case, when the relative contribution value of a prediction factor to the prediction target is greater than or equal to the preset contribution threshold, it means that the greater the relative contribution of this prediction factor to the prediction target, it means that this prediction factor is more important for the prediction of this target artificial intelligence model. In this case, this prediction factor can be used as the target prediction factor, that is, an actual prediction factor corresponding to this target artificial intelligence model is obtained.

[0143] When the relative contribution value of a prediction factor to the prediction target is less than the preset contribution threshold, it means that the relative contribution of this prediction factor to the prediction target is small, which means that this prediction factor is not important for the prediction of this target artificial intelligence model. In this case, this prediction factor can be ignored for this target artificial intelligence model.

[0144] In this case, by executing the above steps (1) and (2) for each target artificial intelligence model, the actual prediction factor corresponding to each target artificial intelligence model can be obtained, and each prediction factor is a feature that contributes greatly to the prediction of the corresponding target artificial intelligence model. In this way, when climate prediction is performed based on multiple target artificial intelligence models in the future, relatively accurate results can be predicted.

[0145] In an embodiment of the present application, step 104 provides an artificial intelligence model interpretation scheme, which can perform an explanatory analysis on the relative contributions of the prediction factors corresponding to each target artificial intelligence after pre-training, and select multiple prediction factors with greater relative contributions to the prediction target as the actual prediction factors of the corresponding target artificial intelligence model.

[0146] For example, Figure 3 This is a flowchart of the specific operation of an artificial intelligence model interpretable solution provided in an embodiment of the present application. Figure 3 , Figure 3 The specific operation process of the interpretable scheme includes the following steps (1) to (3).

[0147] (1) Based on several target AI models with better return effects selected by independent return optimization schemes, AI model interpretability analysis is performed on each prediction factor.

[0148] (2) Conduct interpretable analysis on the recent independent return training results of different artificial intelligence models, and rank them according to the relative contribution of different predictors in the training models to the prediction results.

[0149] (3) Based on the results of the interpretable analysis, select the prediction factors with higher relative contributions to different models as the actual prediction factor set of the final integrated prediction model that integrates multiple artificial intelligence models (that is, the actual prediction factors corresponding to each target artificial intelligence model).

[0150] Step 105: Based on multiple target artificial intelligence models and the actual prediction factors corresponding to each target artificial intelligence model, a climate integrated prediction model based on the fusion of multiple artificial intelligence models is constructed, and the climate integrated prediction model is used to predict the prediction target.

[0151] In this case, the climate integrated prediction model is equivalent to being based on the multiple target artificial intelligence models, and the multiple target artificial intelligence models are different types of models. Therefore, a climate integrated prediction model integrating multiple artificial intelligence models can be constructed. In this way, the linear and nonlinear characteristics of climate data can be captured through multiple artificial intelligence models, thereby more comprehensively capturing data characteristics, and then the climate integrated prediction model performs more accurate climate predictions.

[0152] Optionally, the operation of step 105 can be: calling the model parameters of each target artificial intelligence model after pre-training as the initialization parameters for training the climate integrated prediction model; selecting the historical climate data corresponding to the prediction factors of the 30 years before the prediction period as the actual training set, and re-training the climate integrated prediction model, and obtaining the climate integrated prediction model after the training is completed.

[0153] In this case, the data values ​​of the actual prediction factors corresponding to each target artificial intelligence model are selected from the historical observation data set as the actual training set, and the climate integrated prediction model of the integrated multiple artificial intelligence models is re-trained based on the initialization parameters to improve the prediction accuracy of the climate integrated prediction model.

[0154] In an embodiment of the present application, the climate integrated prediction method based on the fusion of multiple artificial intelligence models can effectively utilize the nonlinear quantitative prediction capabilities of artificial intelligence models while retaining the linear prediction capabilities of some models, thereby performing fusion prediction and integrated prediction of linear and nonlinear characteristics in the climate system, effectively improving the generalization and stability of the model; the embodiment of the present application is also based on climate dynamics and climate predictability research, through independent return optimization modules and artificial intelligence model interpretation modules, to screen, simplify and optimize the fusion model, effectively reduce the number of prediction model parameters, improve prediction efficiency, reduce resource consumption, focus the prediction on the predictable part, and effectively improve the objectivity and accuracy of the prediction.

[0155] Above Figure 1 In the embodiment, a climate integrated prediction model integrating multiple artificial intelligence models can be trained, and the climate integrated prediction model can be used for climate prediction later. Figure 4 A flowchart of another climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the embodiment of the present application is provided in FIG. Figure 4 , Figure 4 The following steps are included.

[0156] Step 401: Obtain climate data corresponding to actual prediction factors corresponding to multiple target artificial intelligence models in a target period before the period to be predicted, where the climate integrated prediction model includes multiple target artificial intelligence models.

[0157] The target period can be set in advance, for example, the target period can be set to 2 years. As an example, if the prediction target is the abnormal precipitation in the summer of 2023 in country A, then the period to be predicted is the summer of 2023, and the target period can be 2021 and 2022.

[0158] Continuing with the above example, we can obtain the specific data of the actual prediction factors corresponding to each target artificial intelligence model in 2021 and 2022, so as to predict the abnormal precipitation in the summer of 2023 based on this.

[0159] The climate integrated prediction model is based on the above Figure 1 The embodiment provided is constructed by a climate integrated prediction method that integrates multiple artificial intelligence models.

[0160] Above Figure 1The process of constructing a climate integrated prediction model integrating multiple artificial intelligence models in the embodiment is as follows: first, select prediction factors that are physically correlated with the prediction target; select multiple artificial intelligence models, and pre-train the multiple artificial intelligence models based on the historical observation data set of the 30 years before the prediction period; perform independent return tests on the multiple artificial intelligence models obtained by pre-training, and select multiple target artificial intelligence models with the highest return test results; for multiple target artificial intelligence models, perform explanatory analysis on each prediction factor, determine the relative contribution of each prediction factor to the prediction target, and select the prediction factor with relatively high contribution as the actual prediction factor corresponding to the corresponding target artificial intelligence model. Finally, the climate integrated prediction model is constructed based on multiple target artificial intelligence models and the actual prediction factors corresponding to each target artificial intelligence model, and then the model parameters during pre-training and the climate data corresponding to each actual prediction factor are called to perform integrated training on the climate integrated prediction model again, and the trained model is the climate integrated prediction model.

[0161] Step 402: Input the climate data corresponding to the actual prediction factors corresponding to the multiple target artificial intelligence models into the multiple target artificial intelligence models to obtain multiple prediction results.

[0162] For any target artificial intelligence model among multiple target artificial intelligence models, the climate data corresponding to the actual prediction factor corresponding to this target artificial intelligence model is input into this target artificial intelligence model, and the climate data corresponding to the input actual prediction factor is processed by this target artificial intelligence model to output the corresponding prediction result.

[0163] By performing the above operations on each of the multiple target artificial intelligence models, multiple prediction results can be output accordingly through the multiple target artificial intelligence models.

[0164] Step 403: Averaging the multiple prediction results to obtain a final integrated prediction result.

[0165] In this case, it is equivalent to integrating the prediction results output by multiple target artificial intelligence models to determine the final predicted result. In this way, integrating multiple prediction results can obtain a more accurate integrated prediction result.

[0166] Optionally, the operation of step 403 may be: performing weighted averaging on the multiple prediction results to obtain a final integrated prediction result.

[0167] In the embodiment of the present application, the prediction weight of each target artificial intelligence model can be set, that is, the degree of emphasis on the prediction results of each target artificial intelligence model can be set. For example, if a target artificial intelligence model can better capture the nonlinear characteristics of climate data, then the prediction results of this target artificial intelligence model can be given more attention, so the prediction weight of this target artificial intelligence model can be set larger.

[0168] In this way, by taking a weighted average of multiple prediction results, a more accurate integrated prediction result can be obtained.

[0169] In an embodiment of the present application, the climate data corresponding to the actual prediction factors corresponding to the multiple target artificial intelligence models in the target period before the period to be predicted are first obtained, and then the climate data corresponding to the actual prediction factors corresponding to the multiple target artificial intelligence models are input into the multiple target artificial intelligence models, and the input climate data is processed by the multiple target artificial intelligence models to obtain multiple prediction results. In this way, multiple artificial intelligence models can capture the linear and nonlinear signals of climate data, and a relatively accurate integrated prediction result can be obtained by performing climate prediction through a climate integrated prediction model that integrates multiple artificial intelligence models.

[0170] Combine the following Figure 5 The climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the embodiment of the present application is exemplified. Figure 5 A flowchart of another climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in an embodiment of the present application.

[0171] For example, taking March 2023 as the starting time, a cross-season climate integrated forecast based on the fusion of multiple artificial intelligence models is conducted for the precipitation anomaly at National Meteorological Station A in the summer of 2023 (June 2023, July 2023, and August 2023).

[0172] like Figure 5 As shown, the specific operation process of the cross-seasonal climate integrated forecast of summer precipitation in Country A in 2023 based on the fusion of multiple artificial intelligence models includes the following parts.

[0173] (1) According to the research results of dynamic theory and climate predictability theory, it can be seen that the signals affecting the summer precipitation anomaly in country A mainly come from the tropical region and the mid- and high-latitude regions of the Northern Hemisphere. The corresponding major large-scale Earth system elements are the sea surface temperature (SST) in the tropical region and the 500hPa potential height field (Z500) in the mid- and high-latitude regions of the Northern Hemisphere.

[0174] Among the above two factors, there are several obvious climate phenomena. Among them, the climate phenomena that have the most direct impact on summer precipitation in country A are El Nino, Arctic Oscillation (AO), North Atlantic Oscillation (NAO) and tropical Indian Ocean Basin Mode (IOBM). The descriptive indexes of the corresponding climate phenomena are Nino3.4 index, AO index, NAO index and IOBM index. In summary, the embodiment of the present application selects the above four indexes in winter nearly 30 years ago as the prediction factor set, that is, the prediction factors include Nino3.4 index, AO index, NAO index and IOBM index.

[0175] (2) The prediction of precipitation series can be abstracted as time series prediction. Taking into account the applicable scenarios and interpretability of the artificial intelligence model itself, the embodiment of the present application selects six artificial intelligence models, namely, multiple linear regression (LR), ridge regression (Ridge), LASSO regression (LASSO), support vector machine (SVR), long short-term memory neural network (LSTM), and gated recurrent neural network (GRU) as candidates for the final fusion model.

[0176] (3) An independent report test is conducted on the precipitation anomaly at station A in country A in the first five time steps of the summer of 2023, that is, the summer of 2018 to 2022. The five-year independent report results are compared and evaluated with the actual observation results to obtain five groups of report test results, and the five groups of report test results are averaged. The evaluation indicators include the anomaly correlation coefficient (ACC), the root mean square error (RMSE) and the comprehensive score of precipitation anomaly (PS). According to the above steps, there are a total of six groups of training results for evaluation. Figure 6 is a comparative schematic diagram of a feedback inspection result provided in an embodiment of the present application, such as Figure 6 As shown, it can be found that for this prediction task, the multiple linear regression model, ridge regression model, and long short-term memory neural network model have more outstanding return effects, which are specifically manifested as higher ACC, lower RMSE, and higher PS score. In summary, the multiple linear regression model, ridge regression model, and long short-term memory neural network model can be selected as the final climate integrated prediction model integrating multiple artificial intelligence.

[0177] (4) Conduct interpretability analysis on the above three models combined with independent return results. Figure 7 is a comparative schematic diagram of an explanatory analysis result provided in an embodiment of the present application, such as Figure 7 As shown in the figure, for the multivariate linear regression model, the relative importance parameter (R2) was used for evaluation, and it was found that the Nino3.4 index and the AO index had the highest relative contribution. For the ridge regression model, the Shapley value was used for evaluation, and it was found that the Nino3.4 index, the AO index, and the IOBM index had higher relative contributions. For the long short-term memory neural network model, LIME was used for analysis, and it was found that the Nino3.4 index and the NAO index had higher relative contributions.

[0178] (5) Based on the above steps, the final climate integrated prediction model integrating multiple artificial intelligence models is:

[0179] A: Multivariate linear model, predictors are: Nino3.4 index and AO index.

[0180] B: Ridge regression model, with the predictors being: Nino3.4 index, AO index and IOBM index.

[0181] C: Long short-term memory neural network model, prediction factors are: Nino3.4 index and NAO index.

[0182] Using historical observation data from 1993 to 2022, the three artificial intelligence models mentioned above were trained again to construct the final climate integrated prediction model.

[0183] After the climate integrated prediction model is constructed, the climate integrated prediction model can be used to make an integrated prediction of the precipitation anomaly at National Meteorological Station A in the summer of 2023 (June 2023, July 2023, and August 2023).

[0184] Specifically, the climate indices corresponding to the winters of 2021 and 2022 were used as actual prediction factors and input into the multivariate linear model, ridge regression model and long short-term memory neural network model respectively, and three groups of prediction results for the summer precipitation anomaly in country A in 2023 were obtained. The three groups of results were averaged, and finally, the cross-seasonal climate integrated prediction results of the summer precipitation anomaly in country A based on the fusion of multiple artificial intelligence models such as multivariate linear regression model, ridge regression model and long short-term memory deep learning network model were obtained.

[0185] Figure 8 is a comparative schematic diagram of a prediction evaluation result provided in an embodiment of the present application, such as Figure 8As shown, the real-time prediction and evaluation results of the precipitation anomaly in country A in the summer of 2023 by the method provided in the embodiment of the present application and the business systems of various countries are respectively shown, wherein the selected business modes are: Meteorological Bureau of country A mode, Meteorological Bureau of country B mode, Meteorological Bureau of country C mode, Meteorological Bureau of country D mode, Meteorological Bureau of country E mode, Meteorological Bureau of country F mode, Meteorological Bureau of country G mode, Meteorological Bureau of country K mode. The starting time of the mode is selected as March 2023, and the evaluation index is the PS score. It can be found that the climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the embodiment of the present application (scoring 76.1 points) has a more obvious improvement compared with the current mainstream business dynamic model of multiple countries (average score of 69.2 points).

[0186] Fig. 9 Schematic diagram of a climate integrated prediction system based on the fusion of multiple artificial intelligence models provided in the embodiment of the present application. Fig. 9 The system includes: a selection module 901, a pre-training module 902, a model verification module 903, an explanatory analysis module 904 and a construction module 905.

[0187] A selection module 901 is used to select elements that are physically related to the prediction target from the earth system big data as prediction factors based on the dynamic theory and the climate predictability theory;

[0188] A pre-training module 902 is used to select multiple artificial intelligence models and use the historical observation data set with the same interdecadal background as the prediction target as the training set to pre-train the multiple artificial intelligence models;

[0189] The model verification model 903 is used to perform feedback verification on the multiple artificial intelligence models obtained by pre-training based on the historical climate data of the target period, and select multiple target artificial intelligence models from the multiple artificial intelligence models based on the feedback verification results corresponding to the various artificial intelligence models;

[0190] An explanatory analysis module 904 is used to analyze the relative contribution of each prediction factor to the prediction target for each target artificial intelligence model in the multiple target artificial intelligence models, and select the target prediction factor from the multiple prediction factors as the actual prediction factor corresponding to the target artificial intelligence model based on the analysis result;

[0191] Construction module 905 is used to construct a climate integrated prediction model based on the fusion of multiple artificial intelligence models based on multiple target artificial intelligence models and the actual prediction factors corresponding to each target artificial intelligence model. The climate integrated prediction model is used to predict the prediction target.

[0192] Optionally, the selection module 901 is used to:

[0193] Based on the singular value decomposition method and the empirical orthogonal decomposition method, elements that have telecorrelation relationships and lead-lag relationships with the prediction targets are selected from the earth system big data as prediction factors.

[0194] Optionally, the pre-training module 902 is used to:

[0195] Select a variety of artificial intelligence models suitable for time series fitting and prediction, including linear regression models, generalized linear models, support vector machine models, tree models, shallow neural network models and recurrent neural network models.

[0196] Optionally, the pre-training module 902 is used to:

[0197] Obtain historical climate data for 30 years before the forecast period;

[0198] The historical climate data for 30 years before the prediction period is used as the historical observation data set to pre-train the various artificial intelligence models.

[0199] Optionally, the model checking module 903 is used to:

[0200] Divide the historical climate data of the target period into multiple training sets;

[0201] For any one of the multiple artificial intelligence models, multiple training sets are respectively input into the artificial intelligence model to obtain multiple prediction results;

[0202] Based on the multiple prediction results and the actual observation results corresponding to the multiple training sets, the artificial intelligence model is evaluated by means of deviation correlation coefficient, root mean square error and comprehensive score to obtain the return test result corresponding to the artificial intelligence model;

[0203] Based on the return test results corresponding to various artificial intelligence models, the various artificial intelligence models are ranked;

[0204] A preset number of artificial intelligence models are selected from the sorted plurality of artificial intelligence models as multiple target artificial intelligence models.

[0205] Optionally, the interpretive analysis module 904 is used to:

[0206] Perform interpretability analysis on each predictor by relative importance parameter analysis, local interpretable model interpreter LIME, or Shapley value to determine the relative contribution of each predictor to the prediction results output by the target AI model;

[0207] A predictor whose relative contribution value is greater than or equal to a preset contribution threshold is selected from multiple predictors as the target predictor to obtain the actual predictor corresponding to the target artificial intelligence model.

[0208] Optionally, the construction module 905 is used to:

[0209] The model parameters of each target artificial intelligence model after pre-training are used as initialization parameters for re-training the climate integrated prediction model;

[0210] The historical climate data corresponding to the prediction factors of the 30 years before the prediction period are selected as the actual training set, and the climate integrated prediction model is trained again. After the training is completed, the climate integrated prediction model is obtained.

[0211] In an embodiment of the present application, the climate integrated prediction method based on the fusion of multiple artificial intelligence models can effectively utilize the nonlinear quantitative prediction capabilities of artificial intelligence models while retaining the linear prediction capabilities of some models, thereby performing fusion prediction and integrated prediction of linear and nonlinear characteristics in the climate system, effectively improving the generalization and stability of the model; the embodiment of the present application is also based on climate dynamics and climate predictability research, through independent return optimization modules and artificial intelligence model interpretation modules, to screen, simplify and optimize the fusion model, effectively reduce the number of prediction model parameters, improve prediction efficiency, reduce resource consumption, focus the prediction on the predictable part, and effectively improve the objectivity and accuracy of the prediction.

[0212] It should be noted that the climate integrated prediction system based on the fusion of multiple artificial intelligence models provided in the above embodiment only uses the division of the above-mentioned functional modules as an example when performing climate prediction. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0213] The functional units and modules in the above embodiments may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit, and the above integrated units may be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the protection scope of the embodiments of the present application.

[0214] The climate integrated prediction system based on the fusion of multiple artificial intelligence models and the climate integrated prediction method embodiment based on the fusion of multiple artificial intelligence models provided in the above embodiments belong to the same concept. The specific working process of the units and modules in the above embodiments and the technical effects brought about can be found in the method embodiment part and will not be repeated here.

[0215] Fig.10 A schematic diagram of the structure of a computer system provided in an embodiment of the present application. Fig.10 As shown, the computer system 1000 includes: a processor 100, a memory 101, and a computer program 102 stored in the memory 101 and executable on the processor 100. When the processor 100 executes the computer program 102, the steps of the climate integrated prediction method based on the fusion of multiple artificial intelligence models in the above-mentioned embodiment are implemented.

[0216] Exemplarily, the memory 101 can be used to store relevant programs of the climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the embodiments of the present application; the processor 100 can call the relevant programs of the climate integrated prediction method based on the fusion of multiple artificial intelligence models stored in the memory 101 to execute the climate integrated prediction method based on the fusion of multiple artificial intelligence models in the embodiments of the present application.

[0217] In this embodiment, the system can be divided into functional modules according to the above method example. For example, each functional module can be corresponded to, or two or more functions can be integrated into one processing module. The above integrated module can be implemented in the form of hardware. It should be noted that the division of the system in this embodiment is schematic and is only a logical function division. There may be other division methods in actual implementation.

[0218] In the case of dividing each functional module according to each function, the system may also include a selection module, a pre-training module, a model verification module, an explanatory analysis module, a construction module, etc. It should be noted that all relevant contents of each step involved in the above method embodiment can be referred to the functional description of the corresponding functional module, which will not be repeated here.

[0219] It should be understood that the system provided in this embodiment is used to execute the above-mentioned optimal climate mode artificial intelligence identification method for climate prediction, and thus can achieve the same effect as the above-mentioned implementation method.

[0220] In the case of an integrated unit, the system may include a processing module and a storage module. The processing module may be a processor or a controller, which may implement or execute various exemplary logic blocks, modules and circuits shown in combination with the disclosure of the present application. The processor may also be a combination that implements a computing function, such as a combination of one or more microprocessors, a combination of digital signal processing (DSP) and a microprocessor, etc. The storage module may be a memory.

[0221] In addition, the system provided in the embodiments of the present application may specifically be a chip, component or module, and the chip may include a connected processor and memory; wherein the memory is used to store instructions, and when the processor calls and executes the instructions, the chip can execute a climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the above embodiments.

[0222] The present application also provides a computer-readable storage medium, which stores a computer program code. When the computer program code runs on a computer, the computer executes the above-mentioned related method steps to implement a climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the above embodiment. Among them, the computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, digital versatile disks (DVD), compact disc read-only memory (CD-ROM), microdrives and magneto-optical disks, read-only memory (ROM), random access memory (RAM), erasable programmable read-only memory (EPROM), electrically erasable programmable read only memory (EEPROM), dynamic random access memory (DRAM), video random access memory (VRAM), flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0223] The present application also provides a computer program product. When the computer program product runs on a computer, it enables the computer to execute the above-mentioned related steps to implement a climate integrated prediction method based on the fusion of multiple artificial intelligence models provided in the above embodiment.

[0224] Among them, the computer-readable storage medium, computer program product or chip provided in this application is used to execute the corresponding method provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding method provided above, and will not be repeated here.

[0225] Through the description of the above implementation methods, technical personnel in the relevant field can understand that for the convenience and simplicity of description, only the division of the above-mentioned functional modules is used as an example. In actual applications, the above-mentioned functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above.

Claims

1. A climate ensemble prediction method based on the fusion of multiple artificial intelligence models, characterized in that: The method comprises: Based on the dynamic theory and climate predictability theory, elements that are physically related to the prediction target are selected from the Earth system big data as prediction factors; Selecting multiple artificial intelligence models, and using historical observation data sets with the same interdecadal background as the prediction target as training sets to pre-train the multiple artificial intelligence models; Based on the historical climate data of the target period, the plurality of artificial intelligence models obtained by pre-training are respectively subjected to feedback tests, and based on the feedback test results corresponding to each of the artificial intelligence models, a plurality of target artificial intelligence models are selected from the plurality of artificial intelligence models; For each of the multiple target artificial intelligence models, analyze the relative contribution of each prediction factor to the prediction target, and select a target prediction factor from the multiple prediction factors as an actual prediction factor corresponding to the target artificial intelligence model based on the analysis result; Based on the multiple target artificial intelligence models and the actual prediction factors corresponding to each of the target artificial intelligence models, a climate integrated prediction model based on the fusion of multiple artificial intelligence models is constructed, and the climate integrated prediction model is used to predict the prediction target; Based on the dynamics theory and climate predictability theory, the elements that are physically related to the prediction target are selected from the Earth system big data as prediction factors, including: Based on the singular value decomposition method and the empirical orthogonal decomposition method, elements having a telecorrelation relationship and a lead-lag relationship with the prediction target are selected from the earth system big data as the prediction factors; The method of performing feedback tests on the plurality of artificial intelligence models obtained by pre-training based on the historical climate data of the target period, and selecting a plurality of target artificial intelligence models from the plurality of artificial intelligence models based on the feedback test results corresponding to each of the artificial intelligence models, includes: Dividing the historical climate data of the target period into a plurality of training sets; For any one of the plurality of artificial intelligence models, the plurality of training sets are respectively input into the artificial intelligence model to obtain a plurality of prediction results, wherein the plurality of training sets correspond to the plurality of prediction results one by one; For any one of the multiple training sets, based on the prediction results and actual observation results corresponding to the training set, the prediction results of the training set on the artificial intelligence model are evaluated by means of deviation correlation coefficient, root mean square error and comprehensive score to obtain the evaluation results corresponding to the training set; the same evaluation results corresponding to the multiple training sets are weighted averaged to obtain the return test results corresponding to the artificial intelligence model; Based on the return test results corresponding to the various artificial intelligence models, the plurality of artificial intelligence models are ranked; A preset number of artificial intelligence models are selected from the sorted plurality of artificial intelligence models as the plurality of target artificial intelligence models.

2. The method according to claim 1, characterized in that The selection of multiple artificial intelligence models includes: A variety of artificial intelligence models suitable for time series fitting and prediction are selected, and the various artificial intelligence models include linear regression models, generalized linear models, support vector machine models, tree models, shallow neural network models and recurrent neural network models.

3. The method according to claim 1, characterized in that The method of using the historical observation data set having the same decadal background as the prediction target as a training set to pre-train the multiple artificial intelligence models includes: Obtain historical climate data for 30 years before the forecast period; The historical climate data for 30 years before the period to be predicted is used as the historical observation data set to pre-train the multiple artificial intelligence models.

4. The method according to claim 1, characterized in that The analyzing the relative contribution of each prediction factor to the prediction target, and selecting a target prediction factor from a plurality of prediction factors as an actual prediction factor corresponding to the target artificial intelligence model based on the analysis result, includes: Perform interpretability analysis on each of the predictors by means of relative importance parameter analysis, local interpretable model interpreter LIME, or Shapley value to determine the relative contribution value of each of the predictors to the prediction result output by the target artificial intelligence model; A prediction factor whose relative contribution value is greater than or equal to a preset contribution threshold is selected from the multiple prediction factors as the target prediction factor to obtain an actual prediction factor corresponding to the target artificial intelligence model.

5. The method according to claim 1, characterized in that The method of constructing a climate integrated prediction model based on the fusion of multiple artificial intelligence models based on the multiple target artificial intelligence models and the actual prediction factors of each target artificial intelligence model includes: Calling the model parameters of each of the target artificial intelligence models after pre-training as initialization parameters for re-training the climate integrated prediction model; The historical climate data corresponding to the prediction factors of the 30 years before the prediction period are selected as the actual training set, and the climate integrated prediction model is trained again. After the training is completed, the climate integrated prediction model is obtained.

6. A climate ensemble prediction method based on the fusion of multiple artificial intelligence models, characterized in that: The method comprises: Obtaining climate data corresponding to actual prediction factors corresponding to multiple target artificial intelligence models in a target period before the period to be predicted, wherein the climate integrated prediction model includes the multiple target artificial intelligence models, and the climate integrated prediction model is constructed based on the method described in any one of claims 1 to 5; Inputting the climate data corresponding to the actual prediction factors corresponding to the multiple target artificial intelligence models into the multiple target artificial intelligence models to obtain multiple prediction results; The multiple prediction results are averaged to obtain a final integrated prediction result.

7. A climate integrated prediction system based on the fusion of multiple artificial intelligence models, characterized in that: The system comprises: A selection module is used to select elements that are physically related to the prediction target from the Earth system big data as prediction factors based on the dynamic theory and climate predictability theory; A pre-training module, used to select a plurality of artificial intelligence models, and use a historical observation data set having the same interdecadal background as the prediction target as a training set to pre-train the plurality of artificial intelligence models; A model verification model, used to perform feedback verification on the plurality of artificial intelligence models obtained by pre-training based on historical climate data of a target period, and select a plurality of target artificial intelligence models from the plurality of artificial intelligence models based on the feedback verification results corresponding to the various artificial intelligence models; An explanatory analysis module, for analyzing, for each of the multiple target artificial intelligence models, the relative contribution of each prediction factor to the prediction target, and selecting a target prediction factor from the multiple prediction factors as an actual prediction factor corresponding to the target artificial intelligence model based on the analysis result; A construction module, used to construct a climate integrated prediction model based on the fusion of multiple artificial intelligence models based on the multiple target artificial intelligence models and the actual prediction factors corresponding to each of the target artificial intelligence models, wherein the climate integrated prediction model is used to predict the prediction target; The selection module is specifically used to: select elements that have a telecorrelation relationship and a lead-lag relationship with the prediction target from the earth system big data as the prediction factors based on the singular value decomposition method and the empirical orthogonal decomposition method; The model verification model is specifically used to: divide the historical climate data of the target time period into multiple training sets; for any one of the multiple artificial intelligence models, input the multiple training sets into the artificial intelligence model respectively to obtain multiple prediction results, and the multiple training sets correspond to the multiple prediction results one by one; for any one of the multiple training sets, based on the prediction results and actual observation results corresponding to the training set, the prediction results of the training set on the artificial intelligence model are evaluated by means of deviation correlation coefficient, root mean square error and comprehensive score to obtain the evaluation results corresponding to the training set; the same evaluation results corresponding to the multiple training sets are weighted averaged to obtain the return verification results corresponding to the artificial intelligence model; based on the return verification results corresponding to various artificial intelligence models, the multiple artificial intelligence models are sorted; and a preset number of artificial intelligence models are selected from the sorted multiple artificial intelligence models as the multiple target artificial intelligence models.

8. A computer system, characterized in that: It includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, it implements the climate integrated prediction method based on the fusion of multiple artificial intelligence models according to any one of claims 1 to 5 or claim 6.

Images