Subseasonal climate objective prediction method and system based on machine learning and multi-scale circulation

Abstract

The invention discloses a sub-seasonal climate objective prediction method and system based on machine learning and multi-scale circulation, and the method comprises the steps: carrying out the conversion of meteorological element data of a meteorological station in a target region, and obtaining the main modes and time coefficient sequences of climate anomalies in different prediction periods through EOF decomposition; the method comprises the following steps: intelligently screening atmosphere reanalysis circulation data and sub-seasonal climate mode circulation prediction data to establish a dynamic circulation prediction factor set; modeling and evaluating the time coefficient corresponding to the main mode of the climate anomaly by using a random forest, and constructing random forest climate objective prediction models of different prediction time periods; based on the dynamic circulation prediction factor set, taking a random forest climate objective prediction model as an initial prediction result, and utilizing a gradient boosting tree to construct a gradient boosting tree climate objective prediction model; performing EOF reconstruction on an objective model prediction result; forming an integrated prediction result; and correction is carried out to realize climate anomaly space field prediction in different prediction time periods.

Description

Technical Field

[0001] The present invention belongs to the technical field of climate prediction, and in particular relates to a method and system for objectively predicting sub-seasonal climate based on machine learning and multi-scale circulation. Background Art

[0002] With the rapid development of the economy and society, the demand for climate forecasting services in various fields is growing, especially in natural disaster prevention, agricultural production planning, energy management and other aspects. Sub-seasonal climate forecast, that is, the forecast on a time scale of 11-60 days, as the hub of medium- and short-term forecasts and long-term climate forecasts, plays a vital role in helping governments and relevant departments shift from passive response to disasters to active disaster prevention and reduction. The extended effectiveness of sub-seasonal climate forecasts can effectively improve the early warning capabilities and emergency response levels of meteorological disaster risks, and buy time for local governments' disaster prevention and reduction work. It can be seen that sub-seasonal climate forecasts have significant social application value.

[0003] As the key interface of the weather and climate forecasting system, the weather and climate change laws are particularly complex because the effect of the initial atmospheric conditions on the sub-seasonal time scale is constantly attenuating, while the influence of the external forcing signal of the underlying surface has not yet fully emerged. Therefore, the prediction results of the sub-seasonal climate model often have shortcomings such as large prediction bias and unstable prediction effect. In addition, the sub-seasonal climate model is still insufficient in spatial precision, which limits the accurate grasp of the spatial anomaly distribution characteristics of climate in my country, especially in provinces, autonomous regions and municipalities with significant regional climate differences, thus affecting the practicality of the prediction results.

[0004] With the increasing demand for personalized and refined climate forecasting services from all walks of life, the traditional qualitative and subjective climate forecasting methods can no longer meet the current business and service needs. Therefore, the development of intelligent and objective sub-seasonal climate forecasting technology is of great significance to improving the level of short-term climate forecasting in my country, and can also better serve all aspects of social economy and provide strong support for responding to the challenges brought by climate change. Summary of the invention

[0005] In view of the climate prediction problems existing in the prior art, such as the limited reference of sub-seasonal climate models and the insufficient intelligence level of climate prediction technology, the purpose of the present invention is to overcome the defects of the prior art and propose an objective prediction method and system for sub-seasonal climate based on machine learning and multi-scale circulation.

[0006] In view of this, the present invention proposes an objective prediction method for sub-seasonal climate based on machine learning and multi-scale circulation, including:

[0007] Step 1) converting the meteorological element data of the meteorological observation station in the target area to be predicted into the meteorological element anomaly field, performing EOF decomposition, and obtaining the main modes of climate anomalies and their time coefficient sequences in different prediction periods;

[0008] Step 2) Using atmospheric reanalysis circulation data and sub-seasonal climate model circulation prediction data, correlation analysis and intelligent screening of key information of multi-time scale circulation are carried out to establish a set of dynamic circulation prediction factors for the previous period and the same period;

[0009] Step 3) Based on the dynamic circulation prediction factor set, random forest is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies, and a random forest climate objective prediction model for different prediction periods is constructed;

[0010] Step 4) Based on the dynamic circulation prediction factor set and taking the random forest objective climate prediction model as the initial prediction result, the gradient boosting tree is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies in step 1), and a gradient boosting tree climate objective prediction model is constructed;

[0011] Step 5) EOF reconstruction is performed on the prediction results of the objective prediction model to obtain the spatial prediction field of climate element anomalies in different prediction periods and different prediction combination methods;

[0012] Step 6) According to the historical return test results of different forecast combination methods and different sub-seasonal climate model elements, the forecast results of multiple methods are ranked and screened, and then dynamically weighted integrated to form an integrated forecast result;

[0013] Step 7) Based on the sliding high correlation between the anomalies of climate elements at different time scales in the target area in the early stage and the anomalies of climate elements in the predicted period, the actual amplitude of the anomalies of climate elements in various regions in the target area in the early stage is used to correct the predicted amplitude of the prediction results of each region, so as to realize the prediction of the spatial field of climate anomalies in different prediction periods.

[0014] Preferably, the step 1) performs EOF decomposition to obtain the main modes of climate anomalies in different forecast periods and their time coefficient sequences; including:

[0015] EOF decomposition is performed, and the spatial distribution patterns that pass the North significance test in different forecast periods are taken to determine the main modes of climate anomalies and the corresponding time coefficient series.

[0016] Preferably, the key information of multi-time-scale circulation in step 2) includes: the high and low-level atmospheric reanalysis circulation fields of the first 10 days, the first 20 days and / or the first 30 days of the forecast period and the high and low-level circulation forecast fields of multiple domestic and foreign sub-seasonal climate numerical models during the same period.

[0017] Preferably, the step 2) comprises:

[0018] Combined with the previous and concurrent main circulation influencing systems of the target area to be predicted, the circulation analysis area is divided into one or several large sub-areas, and the sub-area blocks are not connected to each other;

[0019] Based on the analysis of the correlation between the climate anomaly characteristics of the target area in the forecast period and the multi-time scale circulation in the previous and concurrent periods, the technology of target objective detection based on the connected domain extraction algorithm is used to dynamically identify irregular circulation areas whose correlation with the time coefficient corresponding to the main mode of climate anomaly in the target area reaches the threshold in sub-regional blocks, and the average value of the above-mentioned area is used as the prediction factor to establish a dynamic circulation prediction factor dataset for multiple time scales in different forecast periods within the sub-seasonal time scale.

[0020] Preferably, in the modeling process of step 3) and step 4), a dynamic parameter adjustment strategy based on Bayesian optimization is adopted, and in the evaluation of the prediction model, the minimum average root mean square error of leave-one-out cross-validation is used as the selection criterion for the optimal model.

[0021] Preferably, the step 6) comprises:

[0022] Step 6-1) Obtain prediction methods of various combinations of machine learning algorithms and multi-time scale circulations and prediction products of multiple sub-seasonal climate model elements;

[0023] Step 6-2) Conduct independent sample tests on the prediction results of different combination prediction methods and different sub-seasonal climate model elements in the past X years, and select the top three prediction methods with the highest average anomaly sign consistency rate in the independent sample test to enter the integrated solution candidate;

[0024] Step 6-3) Combine the test scores to obtain the weights k of different prediction methods in different prediction periods i :

[0025]

[0026] Among them, AS i is the sign agreement rate of anomalies entering the integrated alternatives method, i = 1, 2, 3, representing the top three respectively;

[0027] Step 6-4) Combine the dynamic weights to obtain the integrated prediction solutions for different prediction periods and obtain the integrated prediction result y:

[0028]

[0029] Among them, y i is the prediction result of the final integrated alternative solution method, and k1 is the weight of the prediction method with the highest test score.

[0030] Preferably, the deviation sign consistency rate AS in step 6-2) is:

[0031]

[0032] Where N is the total number of meteorological stations or grid points participating in the scoring, N t The AS value is between [0%, 100%] and is the number of meteorological stations or grid points whose predicted anomalies are the same as the actual anomalies or one of them is 0.

[0033] On the other hand, the present invention proposes an objective sub-seasonal climate prediction system based on machine learning and multi-scale circulation, comprising:

[0034] EOF decomposition module, used to convert the meteorological element data of the meteorological observation station in the target area to be predicted into the meteorological element anomaly field, and perform EOF decomposition to obtain the main modes of climate anomalies and their time coefficient sequences in different prediction periods;

[0035] The dynamic circulation prediction factor set establishment module is used to use the atmospheric reanalysis circulation data and the sub-seasonal climate model circulation prediction data to conduct correlation analysis and intelligent screening of key information of multi-time scale circulation, and establish the previous and concurrent dynamic circulation prediction factor sets;

[0036] The random forest climate objective prediction model building module is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies based on the dynamic circulation prediction factor set and use random forests to build random forest climate objective prediction models for different prediction periods;

[0037] The gradient boosting tree objective climate prediction model construction module is used to build a gradient boosting tree objective climate prediction model based on the dynamic circulation prediction factor set and the random forest objective climate prediction model as the initial prediction result. The gradient boosting tree is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies in the EOF decomposition module, and the gradient boosting tree is used to build the gradient boosting tree objective climate prediction model.

[0038] EOF reconstruction module, used to perform EOF reconstruction on the prediction results of the objective prediction model, and obtain the spatial prediction field of climate element anomalies with different prediction combination methods in different prediction periods;

[0039] The prediction result integration module is used to dynamically weight the prediction results of multiple methods after ranking and screening according to the historical return test results of different prediction combination methods and different sub-seasonal climate model elements to form an integrated prediction result;

[0040] The correction module is used to correct the predicted amplitude of the prediction results of each region based on the sliding high correlation between the anomalies of climate elements at different time scales in the target area in the early stage and the anomalies of climate elements in the predicted period, and use the actual amplitude of the anomalies of climate elements in various regions in the target area in the early stage to realize the prediction of the spatial field of climate anomalies in different prediction periods.

[0041] Compared with the prior art, the advantages of the present invention are:

[0042] 1. The present invention integrates and applies multi-time scale and multi-source data, and adds prediction factors of "multi-time scale, multi-climate model, and multi-circulation elements" to the correlation method, which fully considers and dynamically captures the more detailed circulation anomaly characteristics that affect the climate of the target area.

[0043] 2. The present invention combines the advantages of machine learning algorithms in processing large amounts of high-dimensional data with complex features and processing nonlinear relationships, and uses random forest and gradient boosting tree ensemble algorithms to build models, giving full play to the role of multi-time scale circulation factors and enhancing the interpretability, prediction stability and accuracy of the prediction model.

[0044] 3. The present invention is based on the spatial distribution optimization of the persistence of climate anomaly characteristics. The temperature trend forecast for the next 1-30 days in Hainan Province is verified by the rolling season. The result shows that the correlation method of the present invention improves the spatial correlation coefficient score of the integrated prediction results of temperature anomalies from April to September 2024 by 0.132, effectively grasping the spatial distribution characteristics of temperature anomalies in different prediction periods.

[0045] 4. The correlation method of the present invention is conducive to improving the accuracy and refinement of sub-seasonal climate forecasts, can well grasp the spatial distribution characteristics of regional climate anomalies, greatly improve the intelligence and objectivity of sub-seasonal climate forecasts, optimize the work effect and efficiency of sub-seasonal climate forecasts and climate services, and has high application value and promotion potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0046] Figure 1 It is a flow chart of the objective prediction method of sub-seasonal climate based on machine learning and multi-scale circulation of the present invention;

[0047] Figure 2 It is a comparison of the prediction effects before and after the optimization of the spatial distribution of temperature anomalies from April to September 2024;

[0048] Figure 3 It is a comparison of the independent sample test results of the temperature between the present invention and different sub-seasonal climate modes from April to September, which is the average of the three years from 2021 to 2023;

[0049] FIG4 is a comparison of the real-time temperature prediction test results of the present invention and different sub-seasonal climate models from April to September 2024, wherein FIG4(a) is the anomaly sign consistency rate, and FIG4(b) is the anomaly spatial correlation coefficient;

[0050] Figure 5 is a comparison of the independent sample test results of the temperature from April to September in the three years from 2021 to 2023 between the present invention and a single machine learning objective method, wherein Figure 5(a) is the sign consistency rate of the anomaly, and Figure 5(b) is the spatial correlation coefficient of the anomaly. DETAILED DESCRIPTION

[0051] The correlation method of the present invention utilizes atmospheric reanalysis data and multiple sub-seasonal climate model data from home and abroad. By analyzing the correlation between the main modes of climate anomalies in the target area during the forecast period and the different circulation fields of the upper and lower layers in the previous and concurrent periods, the circulation in the key areas of the previous period and the circulation in the key areas of the climate model in the concurrent period are intelligently identified and screened, and a set of dynamic circulation prediction factors at multiple time scales is established. On this basis, two machine learning integrated algorithms, random forest and gradient boosting tree, are used to construct an objective climate prediction model, and an objective sub-seasonal climate prediction method based on machine learning and multi-scale circulation is designed. The method flow is as follows: Figure 1 shown.

[0052] The technical solution of the present invention is described in detail below with reference to the accompanying drawings and embodiments.

[0053] Example 1

[0054] Embodiment 1 of the present invention proposes an objective prediction method for sub-seasonal climate based on machine learning and multi-scale circulation. Each step is described in detail as follows.

[0055] (1) EOF decomposition of meteorological element anomaly fields in different forecast periods

[0056] Taking the average value of the last 30 years as the climate benchmark, the meteorological element data of the target area meteorological observation station to be predicted are converted into the meteorological element anomaly / anomaly percentage field, and EOF decomposition is performed to obtain the main modes of climate anomalies and corresponding time coefficient series that pass the North significance test in different forecast periods. For example, the temperature anomaly field in a certain area, after EOF decomposition, may obtain three spatial distribution types, i.e. the main modes of climate anomalies through the North significance test, such as the consistent type in the whole area, the high in the east and the low in the west, and the high in the south and the low in the north.

[0057] (2) Intelligent screening of key information on multi-time scale circulation

[0058] Based on the high and low atmospheric reanalysis circulation fields of different previous time scales (such as the first 10 days, 20 days, and 30 days of the predicted period), and the high and low circulation prediction fields of domestic and foreign sub-seasonal climate numerical models, first of all, in order to ensure that the selected circulation factors have certain physical significance, combined with the main circulation influencing systems of the target area under analysis, the circulation analysis area is roughly divided into one or several large sub-areas, such as the western Pacific subtropical high pressure area, the mid-latitude westerly belt, the India-Myanmar trough area, etc. The area range of the divided sub-regional blocks is significantly larger than the traditional definition, and the sub-regional blocks are not connected to each other; then, the target objective detection technology based on the connected domain extraction algorithm is used to dynamically identify irregular circulation areas with significantly high correlation (reaching the significance level of 0.05) with the main modal time coefficients of the climate anomaly in the target area according to the sub-regional blocks, and the regional average value of the circulation in the significantly high correlation area is used as the prediction factor to establish a dynamic circulation prediction factor data set for different prediction periods within the sub-seasonal time scale.

[0059] The use of this intelligent and dynamic method to obtain the set of circulation prediction factors can effectively overcome the limitations of the traditional method that relies on the fixed position and regional rules of the weather system circulation to screen the key circulation prediction signals, which causes the key prediction signals to be weakened or the key factors to be reduced, and improve the reliability of the prediction factors. For example, the monitoring area specified by the National Climate Center for the Western Pacific Subtropical High Operation is (10°N-50°N, 90°E-180°). However, the impact of the Western Pacific Subtropical High on the climate anomaly characteristics of the target area in different forecast periods is not evenly distributed in the entire monitoring area. The traditional monitoring area screening of prediction factors is likely to cause the selected key prediction signals to be interfered, thereby reducing the reliability and stability of the prediction factors.

[0060] (3) Dynamic modeling of random forest algorithm based on Bayesian optimization

[0061] Based on the dynamic circulation prediction factor dataset, the random forest ensemble algorithm is used to construct and optimize the time coefficients corresponding to the main modes of climate anomalies in different forecast periods. In this process, a dynamic parameter adjustment strategy based on Bayesian optimization is adopted to build an objective prediction model that adapts to the time coefficients of the main modes of climate anomalies in different forecast periods. In the evaluation of the prediction model, the minimum average root mean square error of the leave-one-out cross-validation is used as the selection criterion for the optimal model. There are 5 parameters used for hyperparameter tuning: n_estimators, max_depth, max_features, min_samples_leaf, and min_impurity_decrease.

[0062] (4) Dynamic modeling of gradient boosting tree algorithm based on Bayesian optimization

[0063] Based on the dynamic circulation prediction factor dataset, the random forest objective prediction model after hyperparameter tuning is used as the initial prediction parameter init of the gradient boosting tree. The gradient boosting tree ensemble algorithm is used to build and optimize the time coefficients corresponding to the main modes of climate anomalies in different prediction periods. The dynamic parameter adjustment strategy based on Bayesian optimization is also adopted to build an objective prediction model that adapts to the time coefficients of the main modes of climate anomalies in different prediction periods. In the evaluation of the prediction model, the minimum average root mean square error of the leave-one-out cross-validation is also used as the selection criterion for the optimal model. There are 8 parameters used for hyperparameter tuning, including n_estimators, learning_rate, criterion, loss, max_depth, max_features, subsample and min_impurity_decrease.

[0064] By constructing an objective climate prediction model based on a machine learning algorithm with a multi-time-scale dynamic prediction factor set and an optimal combination of dynamic hyperparameters, the objectivity and intelligence of the climate prediction model are greatly improved, and the applicability and flexibility of the prediction model in different prediction periods are enhanced.

[0065] (5) EOF reconstruction of the spatial prediction field of climate anomalies

[0066] Combined with the main spatial modal distribution characteristics of climate anomalies, the EOF reconstruction of the prediction results of the corresponding time coefficient objective prediction model is carried out to obtain the spatial prediction field of climate element anomalies with different prediction combination methods in different prediction periods.

[0067] (6) Dynamic weight integration of multi-method prediction results

[0068] According to the prediction results of different prediction combination methods and the historical return test results of multiple sub-seasonal climate model element prediction products, the weight factor combination of different prediction periods is obtained, and the dynamic weight integration of different prediction results is performed to make full use of the prediction results with strong prediction ability and reduce the impact of bad prediction results. Main methods:

[0069] ① Obtain prediction methods using a variety of combinations of machine learning algorithms and multi-time scale circulation, as well as prediction products for multiple sub-seasonal climate model elements.

[0070] ② Conduct independent sample test and evaluation on different forecasting methods in recent years. The forecasting methods with the top three average anomaly sign consistency rates (AS) in independent sample tests are selected as candidates for integrated solutions.

[0071]

[0072] In formula (1), N is the total number of meteorological stations or grid points participating in the scoring, N tIt is the number of meteorological stations or grid points whose predicted and actual anomalies (anomaly percentage) have the same sign or one of the two anomalies is 0. The AS value is between [0%, 100%]. The larger the AS value, the better the prediction method can predict the trend of climate anomalies in a certain area.

[0073] ③ Combine the test scores to obtain the weights k of different prediction methods in different prediction periods i .

[0074]

[0075] In formula (2), AS i is the anomaly sign consistency rate of the alternatives entering the integration, i = 1, 2, 3, representing the top three respectively.

[0076] ④ Combine the dynamic weight combination to obtain the integrated prediction scheme for different prediction periods and get the integrated prediction result y.

[0077]

[0078] In formula (3), y i is the prediction result of the ensemble alternatives, and k1 is the weight of the prediction method with the highest test score. It is mainly based on two considerations: First, when the deviation signs of the prediction results of the three alternative methods are consistent, it means that the prediction results are relatively certain, multiplying The magnitude of climate anomaly prediction can be appropriately increased; second, when the anomaly signs of the prediction results of the three methods are inconsistent, multiply by The weight of the prediction method with the highest AS score can be appropriately increased.

[0079] (7) Spatial distribution optimization based on the persistence of climate anomaly characteristics

[0080] Fully examine the persistence signals of climate anomaly characteristics of each region in the target area to be predicted and the climate anomaly characteristics of different time scales in the previous period (such as the 10 days, 20 days, 30 days, etc. in the previous period to be predicted). In view of the uneven spatial distribution of the temporal persistence of climate anomalies in each region in the target area, combined with the high sliding correlation between climate anomalies of different time scales in the previous period in each region and climate anomalies in the predicted period, the actual amplitude of climate anomalies / anomaly percentages in each region in the target area in the previous period of the prediction is used to intelligently and objectively correct the prediction amplitude of the integrated prediction results of each region, so as to achieve effective optimization of the spatial prediction field of climate anomalies in different prediction periods.

[0081] Example 2

[0082] Embodiment 2 of the present invention provides an objective prediction system for sub-seasonal climate based on machine learning and multi-scale circulation, which is implemented based on the method steps of Embodiment 1. The system includes:

[0083] EOF decomposition module, used to convert the meteorological element data of the meteorological observation station in the target area to be predicted into the meteorological element anomaly field, perform EOF decomposition, and obtain the main modes of climate anomalies and their time coefficient sequences in different prediction periods;

[0084] The dynamic circulation prediction factor set establishment module is used to use the atmospheric reanalysis circulation data and the sub-seasonal climate model circulation prediction data to conduct correlation analysis and intelligent screening of key information of multi-time scale circulation, and establish the previous and concurrent dynamic circulation prediction factor sets;

[0085] The random forest climate objective prediction model building module is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies based on the dynamic circulation prediction factor set and use random forests to build random forest climate objective prediction models for different prediction periods;

[0086] The gradient boosting tree objective climate prediction model construction module is used to build a gradient boosting tree objective climate prediction model based on the dynamic circulation prediction factor set and the random forest objective climate prediction model as the initial prediction result. The gradient boosting tree is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies in the EOF decomposition module, and the gradient boosting tree is used to build the gradient boosting tree objective climate prediction model.

[0087] EOF reconstruction module, used to perform EOF reconstruction on the prediction results of the objective prediction model, and obtain the spatial prediction field of climate element anomalies with different prediction combination methods in different prediction periods;

[0088] The prediction result integration module is used to dynamically weight the prediction results of multiple methods after ranking and screening according to the historical return test results of different prediction combination methods and different sub-seasonal climate model elements to form an integrated prediction result;

[0089] The correction module is used to correct the predicted amplitude of the prediction results of each region based on the sliding high correlation between the anomalies of climate elements at different time scales in the target area in the early stage and the anomalies of climate elements in the predicted period, and use the actual amplitude of the anomalies of climate elements in various regions in the target area in the early stage to realize the prediction of the spatial field of climate anomalies in different prediction periods.

[0090] Sub-seasonal temperature forecast case application

[0091] The objective prediction method of subseasonal climate based on machine learning and multi-scale circulation is applied to the prediction of the average temperature trend of Hainan Island in the next 1-30 days from April to September. The average of 1991-2020 represents the annual average value, and the temperature anomaly / abnormal value is relative to the average of 1991-2020; the two quantitative evaluation indicators of anomaly sign consistency rate (AS) and anomaly spatial correlation coefficient (ACC) are used to test the prediction application effect from the two aspects of trend prediction and spatial distribution prediction.

[0092]

[0093] In formula (4), ΔR f and are the predicted and average values ​​of the mean temperature anomaly, ΔR0 and is the observed value and average value of the average temperature anomaly, and N is the total number of stations or grid points actually participating in the assessment. The ACC value is between [-1,1]. The larger the ACC value is, the better the prediction method can grasp the spatial distribution characteristics of climate anomalies in a certain area.

[0094] The area selected for atmospheric circulation analysis in this prediction case is 60°S-60°N, 60°E-210°E. The early circulation field of the prediction period selects 200hPa zonal wind field, 500hPa zonal wind field, 850hPa meridional wind field, 925hPa relative humidity field, and sea level pressure field, a total of six high and low atmospheric reanalysis circulation fields. The circulation prediction field of the prediction period selects four high and low circulation prediction fields of 500hPa potential height field, 850hPa zonal wind field, 850hPa meridional wind field, and sea level pressure field from three sub-seasonal climate models of CMA-CPSv3, ECMWF S2S, and NCEP-CFSv2. The climate model temperature prediction product also selects three sub-seasonal climate models of CMA-CPSv3, ECMWF S2S, and NCEP-CFSv2. The prediction performance advantages of the association method of the present invention are introduced from three aspects below.

[0095] (1) Comparison before and after spatial distribution optimization

[0096] like Figure 2 As shown in the figure, the spatial distribution optimization scheme based on the persistence of climate anomaly characteristics has improved the average spatial correlation coefficient ACC score of the real-time prediction of the average temperature anomaly in the next 1-30 days for each pentad from April to September 2024 by 0.132 on average compared with the correction before, and the correction skill is basically positive for each pentad. It can be seen that the optimization of the spatial distribution of climate anomalies in Hainan Island by the association method of the present invention has greatly improved the prediction skill of the spatial distribution characteristics of temperature anomalies in the next 1-30 days for each pentad.

[0097] (2) Comparison with temperature prediction results from multiple climate models

[0098] like Figure 3 As shown, in the independent sample test of the temperature anomaly prediction for the next 1-30 days for each hou from April to September in 2021-2023, the average AS score of the average temperature anomaly for the next 1-30 days for each hou predicted by the present invention is 81.2%, which is higher than that of the three sub-seasonal climate models of CMA-CPSv3, ECMWF S2S, and NCEP-CFSv2; its average ACC score is 0.262, which is significantly higher than that of the three sub-seasonal climate models of CMA-CPSv3, ECMWF S2S, and NCEP-CFSv2.

[0099] As shown in FIG4 , a comparison is made between the real-time temperature prediction test results of the present invention and different sub-seasonal climate models from April to September 2024, wherein FIG4 (a) is the anomaly sign consistency rate, and FIG4 (b) is the anomaly spatial correlation coefficient.

[0100] In the real-time prediction test of temperature anomalies in the next 1-30 days for each pentad from April to September 2024, the average AS score of the average temperature anomaly predicted by the present invention for the next 1-30 days for each pentad is 93.2%, which is higher than the three sub-seasonal climate models of CMA-CPSv3, ECMWF S2S, and NCEP-CFSv2; its average ACC score is 0.494, which is significantly improved compared with the three sub-seasonal climate models of CMA-CPSv3, ECMWF S2S, and NCEP-CFSv2, and except for the temperatures reported for the next 1-30 days in individual pentads in May and June, the ACC scores of the temperature anomaly prediction of the present invention are basically stably maintained as positive values, that is, the prediction skills are all positive, with the highest reaching 0.923.

[0101] (3) Comparison with the results of a single objective method

[0102] As shown in Figure 5, the independent sample test effect comparison of the temperature of the present invention and a single machine learning objective method for the three-year average of April to September from 2021 to 2023 is shown, where Figure 5 (a) is the anomaly sign consistency rate, and Figure 5 (b) is the anomaly spatial correlation coefficient. In the independent sample test of the temperature anomaly prediction for the next 1-30 days from April to September from 2021 to 2023, the average AS score of the present invention for predicting the average temperature anomaly for the next 1-30 days is 81.2%, which is higher than the objective prediction model constructed by a single gradient boosting tree or random forest algorithm; its average ACC score is 0.262, which is also higher than the objective prediction model based on a single machine learning algorithm.

[0103] In summary, the correlation method of the present invention gives full play to the advantages of different time scale circulation data and different climate model data in the prediction of climate anomalies in different prediction periods, effectively reduces the prediction instability that may be caused by relying on a single climate model or a single machine learning algorithm, improves the prediction accuracy of climate anomaly trends, and especially significantly improves the prediction skills of spatial distribution of climate anomalies. It has the characteristics of high prediction accuracy, good grasp of spatial distribution, high prediction stability and high intelligence.

[0104] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention is described in detail with reference to the embodiments, it should be understood by those skilled in the art that any modification or equivalent replacement of the technical solutions of the present invention does not depart from the spirit and scope of the technical solutions of the present invention and should be included in the scope of the claims of the present invention.

Claims

1. An objective prediction method for sub-seasonal climate based on machine learning and multi-scale circulation, including: Step 1) converting the meteorological element data of the meteorological observation station in the target area to be predicted into the meteorological element anomaly field, performing EOF decomposition, and obtaining the main modes of climate anomalies and their time coefficient sequences in different prediction periods; Step 2) Using atmospheric reanalysis circulation data and sub-seasonal climate model circulation prediction data, correlation analysis and intelligent screening of key information of multi-time scale circulation are carried out to establish a set of dynamic circulation prediction factors for the previous period and the same period; Step 3) Based on the dynamic circulation prediction factor set, random forest is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies, and a random forest climate objective prediction model for different prediction periods is constructed; Step 4) Based on the dynamic circulation prediction factor set, the random forest climate objective prediction model is used as the initial prediction result, and the gradient boosting tree is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies in step 1), and a gradient boosting tree climate objective prediction model is constructed; Step 5) EOF reconstruction is performed on the prediction results of the objective prediction model to obtain the spatial prediction field of climate element anomalies in different prediction periods and different prediction combination methods; Step 6) According to the historical return test results of different forecast combination methods and different sub-seasonal climate model elements, the forecast results of multiple methods are ranked and screened, and then dynamically weighted integrated to form an integrated forecast result; Step 7) Based on the sliding high correlation between the anomalies of climate elements at different time scales in the target area in the early stage and the anomalies of climate elements in the predicted period, the actual amplitude of the anomalies of climate elements in various regions in the target area in the early stage is used to correct the predicted amplitude of the prediction results of each region, so as to realize the prediction of the spatial field of climate anomalies in different prediction periods.

2. The objective prediction method for subseasonal climate based on machine learning and multi-scale circulation according to claim 1 is characterized in that: The step 1) performs EOF decomposition to obtain the main modes of climate anomalies and their time coefficient sequences in different forecast periods; including: EOF decomposition is performed, and the spatial distribution patterns that pass the North significance test in different forecast periods are taken to determine the main modes of climate anomalies and the corresponding time coefficient series.

3. The objective prediction method for subseasonal climate based on machine learning and multi-scale circulation according to claim 1 is characterized in that: The key information of multi-time scale circulation in step 2) includes: the high and low atmospheric reanalysis circulation fields in the first 10 days, the first 20 days and / or the first 30 days of the forecast period and the high and low circulation forecast fields of multiple domestic and foreign sub-seasonal climate numerical models in the same period.

4. The objective prediction method for subseasonal climate based on machine learning and multi-scale circulation according to claim 3 is characterized in that: The step 2) comprises: Combined with the previous and concurrent main circulation influencing systems of the target area to be predicted, the circulation analysis area is divided into one or several large sub-areas, and the sub-area blocks are not connected to each other; Based on the analysis of the correlation between the climate anomaly characteristics of the target area in the forecast period and the multi-time scale circulation in the previous and concurrent periods, the technology of target objective detection based on the connected domain extraction algorithm is used to dynamically identify irregular circulation areas whose correlation with the time coefficient corresponding to the main mode of climate anomaly in the target area reaches the threshold in sub-regional blocks, and the average value of the above-mentioned area is used as the prediction factor to establish a dynamic circulation prediction factor dataset for multiple time scales in different forecast periods within the sub-seasonal time scale.

5. The objective prediction method for subseasonal climate based on machine learning and multi-scale circulation according to claim 1, characterized in that: In the modeling process of step 3) and step 4), a dynamic parameter adjustment strategy based on Bayesian optimization is adopted. In the evaluation of the prediction model, the minimum average root mean square error of leave-one-out cross-validation is used as the selection criterion for the optimal model.

6. The objective prediction method for subseasonal climate based on machine learning and multi-scale circulation according to claim 1, characterized in that: The step 6) comprises: Step 6-1) Obtain prediction methods of various combinations of machine learning algorithms and multi-time scale circulations and prediction products of multiple sub-seasonal climate model elements; Step 6-2) Conduct independent sample tests on the prediction results of different combination prediction methods and different sub-seasonal climate model elements in the past X years, and select the top three prediction methods with the highest average anomaly sign consistency rate in the independent sample test to enter the integrated solution candidate; Step 6-3) Combine the test scores to obtain the weights k of different prediction methods in different prediction periods i : Among them, AS i is the sign agreement rate of anomalies entering the integrated alternatives method, i = 1, 2, 3, representing the top three respectively; Step 6-4) Combine the dynamic weights to obtain the integrated prediction solutions for different prediction periods and obtain the integrated prediction result y: Among them, y i is the prediction result of the final integrated alternative solution method, and k1 is the weight of the prediction method with the highest test score.

7. The objective prediction method for sub-seasonal climate based on machine learning and multi-scale circulation according to claim 5, characterized in that: The deviation sign consistency rate AS of step 6-2) is: Where N is the total number of meteorological stations or grid points participating in the scoring, N t The AS value is between [0%, 100%] and is the number of meteorological stations or grid points whose predicted anomalies are the same as the actual anomalies or one of them is 0.

8. An objective sub-seasonal climate prediction system based on machine learning and multi-scale circulation, characterized in that: include: EOF decomposition module, used to convert the meteorological element data of the meteorological observation station in the target area to be predicted into the meteorological element anomaly field, perform EOF decomposition, and obtain the main modes of climate anomalies and their time coefficient sequences in different prediction periods; The dynamic circulation prediction factor set establishment module is used to use the atmospheric reanalysis circulation data and the sub-seasonal climate model circulation prediction data to conduct correlation analysis and intelligent screening of key information of multi-time scale circulation, and establish the previous and concurrent dynamic circulation prediction factor sets; The random forest climate objective prediction model building module is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies based on the dynamic circulation prediction factor set and use random forests to build random forest climate objective prediction models for different prediction periods; The gradient boosting tree objective climate prediction model construction module is used to build a gradient boosting tree objective climate prediction model based on the dynamic circulation prediction factor set and the random forest objective climate prediction model as the initial prediction result. The gradient boosting tree is used to model and evaluate the time coefficients corresponding to the main modes of climate anomalies in the EOF decomposition module, and the gradient boosting tree is used to build the gradient boosting tree objective climate prediction model. EOF reconstruction module, used to perform EOF reconstruction on the prediction results of the objective prediction model, and obtain the spatial prediction field of climate element anomalies with different prediction combination methods in different prediction periods; The prediction result integration module is used to dynamically weight the prediction results of multiple methods after ranking and screening according to the historical return test results of different prediction combination methods and different sub-seasonal climate model elements to form an integrated prediction result; and The correction module is used to correct the predicted amplitude of the prediction results of each region based on the sliding high correlation between the anomalies of climate elements at different time scales in the target area in the early stage and the anomalies of climate elements in the predicted period, and use the actual amplitude of the anomalies of climate elements in various regions in the target area in the early stage to realize the prediction of the spatial field of climate anomalies in different prediction periods.

Images