CN117950087B - Artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode - Google Patents

Abstract

The present application provides an artificial intelligence downscaling climate prediction method based on large-scale optimal climate modes, using the spatiotemporal coupled modal decomposition method, based on the large-scale climate elements corresponding to the downscaled climate prediction target elements, extracting the same period large-scale optimal climate modes and time series that determine the relative tendency of large-scale climate element anomalies; using artificial intelligence model training and constructing a downscaling prediction model of the nonlinear relationship between the large-scale optimal climate mode and the relative tendency of regional refined prediction target climate element anomalies; bringing the time coefficient of the same period large-scale optimal climate mode predicted by the global climate dynamic model into the prediction model to predict the relative tendency of regional refined climate element anomalies; combined with recent background anomalies, realizing artificial intelligence downscaling climate prediction of regional refined prediction target climate element anomalies. This method can improve the regional refined climate prediction capability by establishing an efficient and accurate downscaling climate prediction model.

Description

Artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode

Technical Field

This application involves the design and application of an artificial intelligence downscaling climate prediction method based on large-scale optimal climate modes. In the refined prediction business for specific regions, this application can realize the refined, quantitative and intelligent prediction of key climate elements such as regional precipitation and temperature based on large-scale optimal climate modes and artificial intelligence models, and provide scientific and technological support for regional climate disaster warning and other services.

Background technique

Due to the complex and ever-changing topography and different climate conditions in different regions, these factors bring many challenges to regional refined climate forecasting. With the continuous development of economy and technology, the demand for refined climate forecasting services in many industries is increasing, especially in areas such as agriculture, transportation, and new energy that have high requirements for regional climate conditions. The climate forecasting service capabilities of related technologies are far from meeting actual needs.

In response to the above needs, the main technical means in the relevant technology is the downscaling climate prediction scheme, which aims to convert the output information of the global or large-scale climate model into a detailed climate prediction for a specific small-scale region, mainly including two schemes: dynamic downscaling and statistical downscaling. Dynamic downscaling is mainly achieved through regional climate models (RCMs). This type of model is built based on the global climate model and runs under the same boundary conditions. The output of a specific region has a higher spatial resolution through nested grids and other schemes. The regional climate model is built based on the atmospheric dynamics equation, which can consider the complex internal dynamic processes of the climate system. At the same time, the model has a higher resolution, so it can better capture small-scale climate characteristics and is suitable for areas with complex terrain or special climate characteristics. However, the regional climate model requires a huge amount of computing resources and computing time. At the same time, it is limited by the physical framework and boundary conditions of the global climate model, and the regional climate model has very obvious uncertainty. Statistical downscaling mainly uses statistical and mathematical methods to establish the relationship between large-scale climate elements and small-scale climate elements, and then downscales climate prediction. This program mainly includes building a regression prediction model based on linear regression relationships, classifying and predicting through historical weather types, and using historical data to correct model outputs. Compared with dynamic downscaling methods, statistical downscaling has a simpler model architecture and higher computational efficiency, and can be customized according to available data or specific needs. However, the effectiveness of statistical downscaling methods is highly dependent on high-quality and long-term historical observation data. In the context of global climate change, the limitation of being unable to predict unknown situations is becoming increasingly obvious.

Summary of the invention

The present application provides an artificial intelligence downscaling climate prediction method based on large-scale optimal climate modes. The method performs efficient and accurate nonlinear quantitative intelligent prediction through a nonlinear prediction model, thereby improving the accuracy of downscaling climate prediction.

In a first aspect, an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode is provided, the method comprising:

Based on the large-scale prediction target elements corresponding to the downscaled climate prediction target elements, the large-scale circulation field for constructing the physical statistical relationship is selected, and the corresponding climate anomaly field, anomaly relative tendency field and recent background anomaly field are determined respectively;

Performing spatiotemporal coupling decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale prediction target element and the time series corresponding to the large-scale optimal climate mode;

The time series is used as a prediction factor, the abnormal relative tendency field of the downscaled climate prediction target element is used as a prediction target, and the nonlinear prediction model to be trained is trained to obtain a nonlinear prediction model;

According to the large-scale optimal climate mode and the large-scale circulation field of the same period, the time coefficient corresponding to the large-scale circulation field of the same period is determined, and the time coefficient is introduced into the nonlinear prediction model to obtain the prediction result of the abnormal relative tendency field of the downscaled climate prediction target element;

Based on the prediction results of the recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element, a nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element is obtained.

In the above technical scheme, the artificial intelligence downscaling climate prediction method based on the large-scale optimal climate mode has less demand for computing resources than the regional climate dynamic model, and at the same time increases the prediction ability for nonlinear systems compared to the downscaling schemes in related technologies. The selection of prediction factors is based on the spatiotemporal coupling relationship between large-scale climate elements. There are strong physical constraints between the prediction factors and the prediction targets. At the same time, through the artificial intelligence model, the prediction factors of climate elements affecting different regions are nonlinearly combined and weighted, so as to build a more complete nonlinear prediction model. Based on the large-scale optimal mode, the nonlinear downscaling prediction of the downscaled climate prediction target elements can effectively grasp the overall characteristics of the large-scale background of the regional climate, and perform efficient and accurate nonlinear quantitative intelligent prediction through the nonlinear prediction model, thereby improving the accuracy of downscaled climate prediction.

In combination with the first aspect, in some possible implementations, the large-scale optimal climate mode is extracted through a spatiotemporal coupling decomposition method based on the large-scale climate elements corresponding to the downscaled climate prediction target elements and the large-scale circulation elements that determine the anomalies of the large-scale climate elements.

In conjunction with the first aspect, in some possible implementations, the large-scale prediction target element corresponding to the downscaled climate prediction target element is selected to construct a large-scale circulation field for physical statistical relationship, including:

The large-scale circulation field for constructing the physical statistical relationship is selected by using the climate dynamics theory and based on the large-scale prediction target elements corresponding to the downscaled climate prediction target elements.

In the above scheme, the large-scale circulation field can be used as the basic climate element for extracting the large-scale optimal climate mode and can be used as the source of extracting prediction factors in the subsequent process.

In combination with the first aspect, in some possible implementations, the performing spatiotemporal coupling decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale prediction target element and the time series corresponding to the large-scale optimal climate mode includes:

Performing singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode;

The abnormal relative tendency field of the large-scale circulation field is projected onto the large-scale optimal climate mode to obtain a time series corresponding to the large-scale optimal climate mode.

In the above scheme, the optimal climate mode that can best predict the climate conditions of the target climate elements is obtained through the time-space coupling decomposition method, and the time series of the abnormal relative tendency field of the large-scale circulation field is obtained through the projection method, so as to effectively extract the overall characteristics of the large-scale background of the regional climate and improve the accuracy of the refined climate prediction results of the regional climate.

In combination with the first aspect, in some possible implementations, performing singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode includes:

Performing singular value decomposition on the abnormal relative inclination field of the large-scale circulation field and the abnormal relative inclination field of the large-scale prediction target element;

The large-scale optimal climate mode is obtained by sorting according to the covariance corresponding to the singular value decomposition results.

In the above scheme, by analyzing the size of the covariance contribution, the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale circulation field can be determined, and then the regional refined target climate elements can be effectively predicted through the large-scale optimal climate mode.

In combination with the first aspect, in some possible implementations, determining the time coefficient corresponding to the large-scale circulation field in the same period according to the large-scale optimal climate mode and the large-scale circulation field in the same period includes:

Using a global climate dynamic model to predict the circulation field corresponding to the downscaled climate prediction target element, to obtain the large-scale circulation field in the same period;

The large-scale circulation field of the same period is projected onto the large-scale optimal climate mode to obtain the time coefficient corresponding to the large-scale circulation field of the same period.

In the above scheme, the global climate dynamic model is used to predict the large-scale circulation field in the same period, and the large-scale circulation in the same period is projected onto the large-scale optimal climate mode, so that the corresponding time coefficient is used as the input of the nonlinear prediction model. Since the nonlinear prediction model is trained by target climate elements of different resolutions, the obtained prediction results can accurately represent the downscaled climate prediction.

In combination with the first aspect, in some possible implementations, the prediction result of the recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element is used to obtain the nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element, including:

The recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element are added to obtain a nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element.

In the above scheme, the artificial intelligence downscaling climate prediction method based on large-scale optimal climate modes can target climate elements at different scales for prediction. It can further complete efficient and accurate artificial intelligence downscaling nonlinear predictions based on the coarse-resolution large-scale dynamic model prediction results, which can effectively meet the needs of actual business applications and provide an efficient and reliable solution for improving the level of downscaling climate prediction business.

In combination with the first aspect, in some possible implementations, the time series is used as a prediction factor, the abnormal relative tendency field of the downscaled climate prediction target element is used as a prediction target, and the nonlinear prediction model to be trained is trained to obtain a nonlinear prediction model, including:

Get AI models;

Based on the artificial intelligence model, a nonlinear prediction model to be trained including multiple decision trees is constructed; wherein the artificial intelligence model includes: a random forest regression model.

In the above scheme, the prediction factors of climate elements affecting different regions are nonlinearly combined and weighted by presetting a random forest regression model, so as to construct a nonlinear prediction model with more accurate prediction results.

In combination with the first aspect, in some possible implementations, the method further includes:

The anomaly field of any climate element is decomposed into the anomaly relative tendency field and the recent background anomaly field.

In the above scheme, the prediction target and prediction factors are decomposed on the time scale through the above process, and the prediction is concentrated on the required time scale to reduce the interference of signals of other time scales on the prediction.

In a second aspect, an artificial intelligence downscaling climate prediction device based on a large-scale optimal climate mode is provided, the device comprising:

The first determination module is used to select the large-scale circulation field for constructing the physical statistical relationship based on the large-scale prediction target elements corresponding to the downscaled climate prediction target elements, and respectively determine the corresponding climate anomaly field, anomaly relative tendency field and recent background anomaly field;

The first decomposition module is used to perform spatiotemporal coupling decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale prediction target element and the time series corresponding to the large-scale optimal climate mode;

A training module is used to train the nonlinear prediction model to be trained by taking the time series as a prediction factor and the abnormal relative tendency field of the downscaled climate prediction target element as a prediction target, so as to obtain a nonlinear prediction model;

The second determination module is used to determine the time coefficient corresponding to the large-scale circulation field at the same period according to the large-scale optimal climate mode and the large-scale circulation field at the same period, and import the time coefficient into the nonlinear prediction model to obtain the prediction result of the abnormal relative tendency field of the downscaled climate prediction target element;

The third determination module is used to obtain the nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element based on the prediction results of the recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element.

In a third aspect, an electronic device is provided, comprising a memory and a processor. The memory is used to store executable program code, and the processor is used to call and run the executable program code from the memory, so that the device executes the method in the first aspect or any possible implementation of the first aspect.

In a fourth aspect, a computer program product is provided, comprising: a computer program code, which, when executed on a computer, enables the computer to execute the method in the first aspect or any possible implementation of the first aspect.

In a fifth aspect, a computer-readable storage medium is provided, which stores a computer program code. When the computer program code runs on a computer, the computer executes the method in the above-mentioned first aspect or any possible implementation of the first aspect.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG1 is a schematic flow chart of an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode provided in an embodiment of the present application;

FIG2 is another schematic flow chart of an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode provided in an embodiment of the present application;

FIG3 is another schematic flow chart of an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode provided in an embodiment of the present application;

FIG4 is a flow chart of an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode provided in an embodiment of the present application;

FIG5 is an operation flow chart of an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode provided in an embodiment of the present application;

FIG6 is a comparison diagram of the spatial distribution of precipitation anomalies in the middle and lower reaches of the Yangtze River in summer 2019 at 2374 stations and 160 stations provided in an embodiment of the present application;

FIG7 is a structural diagram of an artificial intelligence prediction model based on random forest regression constructed in an embodiment of the present application;

FIG8 is a spatial distribution comparison diagram of the precipitation anomaly field business model prediction results of the middle and lower reaches of the Yangtze River in the summer of 2019 in the embodiment of the present application and the climate prediction results provided by the embodiment of the present application;

FIG9 is a schematic diagram of the structure of an artificial intelligence downscaling climate prediction device based on a large-scale optimal climate mode provided in an embodiment of the present application;

FIG. 10 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

Detailed ways

The technical solution in the present application will be described clearly and in detail below in conjunction with the accompanying drawings. In the description of the embodiments of the present application, unless otherwise specified, "/" means or, for example, A/B can mean A or B: "and/or" in the text 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 the description of the embodiments of the present application, "multiple" means two or more than two.

In the following, the terms "first" and "second" are used for descriptive purposes only and are not to be understood as suggesting or implying relative importance or implicitly indicating the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the features.

The technical solution provided by the embodiment of the present application is introduced below. The embodiment of the present application provides an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode. See Figure 1, which is a schematic flow chart of an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode provided by the embodiment of the present application. The method includes the following steps:

101. Based on the large-scale prediction target elements corresponding to the downscaled climate prediction target elements, the large-scale circulation field for constructing the physical statistical relationship is selected, and the corresponding climate anomaly field, anomaly relative tendency field and recent background anomaly field are determined respectively.

Among them, the downscaled climate prediction target element can be a regional refined climate prediction element, for example, it can include: regional refined rainfall prediction or temperature prediction, etc. The large-scale prediction target element corresponding to the downscaled climate prediction target element is a coarse-grained target climate element with a resolution greater than the downscaled climate prediction target element. For example, the downscaled climate prediction target element is the precipitation at 2374 stations, and the large-scale prediction target element can be the precipitation at 160 stations. Using the climate dynamics theory, based on the large-scale prediction target element corresponding to the downscaled climate prediction target element, the large-scale circulation field for the construction of physical statistical relationships is selected. That is, through the dynamics theory, in the historical return data set of the climate dynamic model, a large-scale circulation field for the construction of physical statistical relationships is selected. In this way, the large-scale circulation field can be used as the basic climate element for extracting the large-scale optimal climate mode, and as the source of extracting prediction factors in subsequent steps. For example, the tropical outgoing longwave radiation (OLR) and geopotential height @ 500hPa (Pa) at mid- and high-latitudes in the Northern Hemisphere (GeopotentialHeight @ 500hPa, Z500) data reported from March 1 each year in the historical report data set of the climate dynamic model are selected, and the abnormal relative tendency field is calculated. The abnormal relative tendency field of tropical OLR in summer and the abnormal relative tendency field of Z500 at mid- and high-latitudes in the Northern Hemisphere are taken as large-scale circulation fields. In this way, by selecting the large-scale circulation field through dynamic theory, the obtained large-scale circulation field can more accurately determine the large-scale prediction target elements corresponding to the downscaled climate prediction target elements.

In some possible implementations, the anomaly relative tendency method refers to dividing the predicted target anomaly into two parts in the prediction: the anomaly relative tendency and the corresponding recent observed background anomaly. The anomaly is predicted by predicting the anomaly relative tendency of the target seasonal average, and the prediction is focused on the anomaly relative tendency part determined by the predictable interannual variability. In step 101, the climate anomaly field, the anomaly relative tendency field, and the recent background anomaly field corresponding to the downscaled climate prediction target element, as well as the climate anomaly field, the anomaly relative tendency field, and the recent background anomaly field corresponding to the large-scale prediction target element are calculated. Among them, the climate anomaly field is the sum of the anomaly relative tendency field and the recent background anomaly field. If the anomaly relative tendency field is defined as the difference between the two adjacent anomalies of the prediction target year (for example, t+1 year) and the previous year (i.e., t year), the corresponding recent background anomaly field is the observed anomaly value of the previous year.

In some possible implementations, based on the large-scale prediction target element corresponding to the downscaled climate prediction target element, the corresponding abnormal relative tendency field and the recent background abnormal field are respectively determined; and the abnormal relative tendency field and the recent background abnormal field are added to obtain the climate abnormal field. For example, the abnormal relative tendency field and the recent background abnormal field are added to obtain the climate abnormal field.

Among them, the climate anomaly field, the anomaly relative tendency field, and the recent background anomaly field can be calculated through the following process:

First, obtain the original data of climate variables; then, subtract the climate state to obtain the anomaly (i.e., climate anomaly field); finally, combine the formula (climate anomaly field = abnormal relative tendency + abnormal relative background) to obtain the abnormal relative tendency field and the recent background anomaly field. In this way, by calculating the climate anomaly field, abnormal relative tendency field, and recent background anomaly field, it is possible to facilitate subsequent regional refined climate prediction.

102. Performing spatiotemporal coupling decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale prediction target element and the time series corresponding to the large-scale optimal climate mode.

Among them, the spatiotemporal coupling decomposition method is used to perform spatiotemporal coupling decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target elements, and obtain the large-scale optimal climate mode as spatial information and the time series as temporal information.

The summer tropical OLR anomaly relative tendency field and the anomaly relative tendency field of the mid-high latitude Z500 in the Northern Hemisphere are used as large-scale circulation fields. The summer tropical OLR anomaly relative tendency field and the anomaly relative tendency field of the mid-high latitude Z500 in the Northern Hemisphere are used to perform spatiotemporal coupling decomposition on the anomaly relative tendency field of the large-scale prediction target elements in the same period, respectively, to obtain the large-scale optimal climate mode and the time series corresponding to the large-scale optimal climate mode.

Among them, the large-scale optimal climate mode is extracted by the spatiotemporal coupling decomposition method based on the large-scale climate elements corresponding to the downscaled climate prediction target elements and the large-scale circulation elements that determine the large-scale climate element anomalies. The time series corresponding to the large-scale optimal climate mode is actually used as a prediction factor and used for prediction model training. The above-mentioned large-scale optimal climate mode is extracted by the spatiotemporal coupling decomposition method based on the large-scale climate elements corresponding to the regional refined prediction target elements and the large-scale circulation elements that determine the large-scale climate element anomalies. The prediction factor represents the cause of the large-scale climate element anomaly background corresponding to the regional refined climate element anomaly. Among them, the large-scale circulation elements that determine the large-scale climate element anomaly can be a large-scale circulation field.

In some possible implementations, in order to more accurately obtain the large-scale optimal climate mode and the corresponding time series, the above step 102 can be implemented by the steps shown in FIG. 2:

201, performing singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode.

Among them, the large-scale optimal climate mode is obtained by performing singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element. Singular Value Decomposition Analysis (SVD), also known as Maximum Covariance Analysis (MCA), is a matrix decomposition method used to reduce a matrix into its components. This method is often used in the field of meteorology for the diagnostic analysis of the spatiotemporal distribution coupling signals of two meteorological fields. The optimal climate mode of the summer large-scale atmospheric circulation anomaly relative tendency that determines the relative tendency of precipitation anomalies in the same period is extracted from historical observation data through the SVD method.

In some possible implementations, the large-scale optimal climate mode is obtained by performing singular value decomposition on the abnormal relative inclination field of the large-scale circulation field and the abnormal relative inclination field of the large-scale prediction target element, and sorting them according to the covariance corresponding to the singular value decomposition results. Here, if the singular value decomposition method is used to decompose the abnormal relative inclination field of the large-scale circulation field and the abnormal relative inclination field of the large-scale prediction target element, three matrices are obtained, namely, the S matrix (a matrix composed of eigenvectors), the U matrix (orthogonal matrix) and the V matrix (orthogonal matrix). In this way, the three matrices after singular value decomposition can be used as the results of singular value decomposition. Among them, the S matrix, the U matrix and the V matrix carry eigenvalues, and the larger the eigenvalue, the greater the contribution of the matrix to the covariance, which further indicates that the climate mode represented by the matrix can best represent the climate corresponding to the target climate element; so the matrix can be used as the optimal climate mode. In this way, since the covariance contribution of the U matrix is the largest, the U matrix can be used as the large-scale optimal climate mode, indicating that the U matrix can best reflect the climate conditions of the target climate element. For example, if the target climate element is precipitation, the El Nino mode can account for more than 30%, so this mode is a more important mode. In this way, by analyzing the size of the covariance contribution, the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale circulation field can be determined, and then the regional refined target climate elements can be effectively predicted through the large-scale optimal climate mode.

202, projecting the abnormal relative inclination field of the large-scale circulation field onto the large-scale optimal climate mode to obtain a time series corresponding to the large-scale optimal climate mode.

Among them, the projection method is used to calculate the large-scale optimal climate mode to obtain the time series. The abnormal relative tendency field of the large-scale circulation field is projected onto the large-scale optimal climate mode to obtain a time series. The value of the time series can characterize the degree of similarity; the higher the similarity, the larger the corresponding value, and the lower the similarity, the smaller the corresponding value. For example, the abnormal relative tendency field of the large-scale circulation field is the abnormal relative tendency field of OLR and Z500, and the large-scale optimal climate mode is the U matrix. Then, the abnormal relative tendency field of OLR and Z500 is projected onto the decomposed U matrix to obtain the time series. In this way, the optimal climate mode that can best predict the climate conditions of the target climate element is obtained by the spatiotemporal coupling decomposition method, and the time series of the abnormal relative tendency field of the large-scale circulation field is obtained by the projection method, so that the overall characteristics of the large-scale background of the regional climate can be effectively extracted to improve the accuracy of the refined climate prediction results of the regional climate.

103. Taking the time series as a prediction factor and the abnormal relative tendency field of the downscaled climate prediction target element as a prediction target, the nonlinear prediction model to be trained is trained to obtain a nonlinear prediction model.

The training is based on the time series and the abnormal relative tendency field of the downscaled climate prediction target element. The nonlinear prediction model is obtained by training the nonlinear prediction model to be trained by taking the time series as the prediction factor and the abnormal relative tendency field of the downscaled climate prediction target element as the training target. In this way, by constructing an artificial intelligence nonlinear prediction model, taking the time series of the large-scale optimal climate mode as the prediction factor, taking the abnormal relative tendency field of the downscaled climate prediction target element as the prediction target, and using the artificial intelligence model to train the nonlinear prediction model to be trained, the obtained nonlinear prediction model is made more perfect.

In some possible implementations, first, an artificial intelligence model is obtained; and based on the artificial intelligence model, a nonlinear prediction model to be trained including multiple decision trees is constructed; wherein the artificial intelligence model includes: a random forest regression model.

In some possible implementations, first, based on a preset random forest regression model, a nonlinear prediction model to be trained including multiple decision trees is constructed; here, a nonlinear prediction model to be trained including multiple decision trees can also be constructed based on models such as classification tree schemes, support vector machines, and recurrent neural networks. Taking the preset random forest regression model as an example, a nonlinear prediction model to be trained including multiple decision trees is constructed through multiple decision trees in the preset random forest regression model. Among them, the multiple decision trees can be the first layer regression tree, the second layer regression tree, and the third layer regression tree as shown in Figure 7. Afterwards, the time series and the abnormal relative tendency field of the downscaled climate prediction target element are used as training data sets, and the parameters of the nonlinear prediction model to be trained are adjusted to obtain the nonlinear prediction model. Here, the time series is used as a prediction factor, and the abnormal relative tendency field of the downscaled climate prediction target element is used as a prediction target. By dividing the training data set into multiple sub-data sets, the network parameters such as the weights of multiple decision trees in the nonlinear prediction model to be trained are adjusted respectively to obtain a nonlinear prediction model. In this way, the prediction factors of climate elements affecting different regions are nonlinearly combined and weighted through artificial intelligence models, thereby constructing a nonlinear prediction model with more accurate prediction results.

After taking the time series and the abnormal relative tendency field of the downscaled climate prediction target element as the training data set, first, the training data set is sampled to obtain multiple sub-data sets. Here, the training data set is randomly sampled by self-help aggregation to obtain multiple sub-data sets. For example, the same number of sub-data sets can be extracted according to the number of decision trees. Then, based on each sub-data set in the multiple sub-data sets, each decision tree in the nonlinear prediction model to be trained is trained to obtain multiple trained decision trees. Here, a decision tree is trained by a sub-data set. Since the number of sub-data sets is the same as the number of decision trees, each decision tree in the multiple decision trees is trained to obtain multiple trained decision trees. Finally, based on the multiple trained decision trees, the nonlinear prediction model is obtained. Here, the multiple trained decision trees are weighted and combined to obtain the nonlinear prediction model. In this way, by extracting each sub-data set to train a decision tree, the mean of the predictions of all decision trees is used as the generated prediction result. Compared with a single decision tree model, the nonlinear prediction model can effectively reduce the overfitting problem and provide an estimate of the importance of each prediction factor, which plays an important role in understanding the causal relationship between the prediction factor and the prediction target. In the application of downscaling climate prediction, this nonlinear prediction model can analyze the contribution of large-scale optimal climate modes affecting different refined sites, which helps to understand the causes of climate prediction results and improves the accuracy and robustness of nonlinear prediction model predictions.

104. According to the large-scale optimal climate mode and the large-scale circulation field of the same period, determine the time coefficient corresponding to the large-scale circulation field of the same period, and import the time coefficient into the nonlinear prediction model to obtain the prediction result of the abnormal relative tendency field of the downscaled climate prediction target element.

Among them, the large-scale circulation field during the same period was predicted by the global climate dynamics model.

The global climate dynamic model is used to predict the circulation field corresponding to the downscaled climate prediction target element to obtain the large-scale circulation field of the same period; then, the large-scale circulation field of the same period is projected to the large-scale optimal climate mode to obtain the time coefficient corresponding to the large-scale circulation field of the same period.

In some possible implementations, a historical return data set of a climate dynamic model is obtained, and a global climate dynamic model is used to predict the circulation field corresponding to the downscaled climate prediction target element according to the historical return data set of the climate dynamic model, so as to obtain the large-scale circulation field of the same period. The large-scale circulation field of the same period is projected to the large-scale optimal climate mode by the projection method, so as to obtain the similarity between the large-scale circulation field of the same period and the large-scale optimal climate mode, and the similarity is used as the time coefficient corresponding to the large-scale circulation field of the same period. In a specific example, if the target climate element is precipitation in 2019, that is, it is necessary to predict the precipitation in 2019, the tropical OLR in summer of that year and the Z500 data of the mid- and high-latitudes in the northern hemisphere are selected from the historical return data set of the climate dynamic model; then, according to the tropical OLR in summer of that year and the Z500 data of the mid- and high-latitudes in the northern hemisphere, the tropical OLR in summer of 2019 and the anomaly relative inclination field of the Z500 in the mid- and high-latitudes in the northern hemisphere are predicted, so as to obtain the large-scale circulation field of the same period. The time coefficient corresponding to the large-scale optimal climate mode in summer 2019 is calculated by the projection method, and the time coefficient is used as the actual prediction factor for calculating the relative tendency field of precipitation anomalies in the middle and lower reaches of the Yangtze River in summer 2019. The time coefficient as a prediction factor is input into the nonlinear prediction model to obtain the artificial intelligence downscaling prediction results of the relative tendency field of precipitation anomalies at stations in the middle and lower reaches of the Yangtze River in summer 2019, that is, the prediction results of the abnormal relative tendency field of the downscaled climate prediction target element. In this way, by using the global climate dynamic model to predict the large-scale circulation field in the same period, the large-scale circulation in the same period is projected onto the large-scale optimal climate mode, so that the corresponding time coefficient is used as the input of the nonlinear prediction model. Since the nonlinear prediction model is trained by target climate elements of different resolutions, the obtained prediction results can accurately represent the downscaled climate prediction.

105. Based on the prediction results of the recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element, a nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element is obtained.

Among them, the anomaly field of the historical observation data is calculated through the regional refined climate element historical observation data set. The anomaly field of the historical observation data = the abnormal relative tendency field of the historical downscaled climate prediction target element + the recent background anomaly field. For example, the anomaly field of the historical observation data is the sum of the relative tendency field of precipitation anomaly at station 2374 and the recent background anomaly field of precipitation at station 2374. The recent background anomaly field of the downscaled climate prediction target element is the recent background anomaly field of precipitation at station 2374. The climate prediction result is obtained by combining the recent background anomaly field of the downscaled climate prediction target element with the prediction results output by the nonlinear prediction model.

In some possible implementations, the above step 105 may be implemented by the steps shown in FIG3 :

301, adding the recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element to obtain a nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element.

Here, since the prediction results represent the abnormal relative tendency field of the downscaled climate prediction target elements output by the nonlinear prediction model, the abnormal relative tendency field of all predicted downscaled climate prediction target elements is added to the recent background anomaly field of the downscaled climate prediction target elements to accurately obtain the anomaly field of the downscaled climate prediction target elements.

Among them, the anomaly field of the downscaled climate prediction target element can represent the climate prediction situation of the downscaled climate prediction target element, so the anomaly field is used as the nonlinear quantitative prediction result of the downscaled climate prediction target element, which can accurately represent the regional refined climate situation. In this way, the artificial intelligence downscaling climate prediction method based on the large-scale optimal climate mode can further complete the efficient and accurate artificial intelligence downscaling nonlinear prediction based on the prediction results of the large-scale dynamic model with coarse resolution for the prediction target climate elements of different scales, which can effectively meet the needs of actual business applications and provide an efficient and reliable solution for improving the business level of downscaling climate prediction.

In steps 101, 104 and 105, any abnormal field of climate elements is decomposed into an abnormal relative tendency field and a recent background abnormal field. In this way, the prediction target and prediction factors are decomposed on a time scale through the above process, and the prediction is concentrated on the required time scale to reduce the interference of signals of other time scales on the prediction.

The artificial intelligence downscaling climate prediction method based on the large-scale optimal climate mode provided in the embodiment of the present application has a smaller demand for computing resources than the regional climate dynamic model, and at the same time increases the prediction ability for nonlinear systems compared to the downscaling scheme in the related art. The selection of prediction factors is based on the spatiotemporal coupling relationship between large-scale climate elements. There are strong physical constraints between the prediction factors and the prediction targets. At the same time, through the artificial intelligence model, the prediction factors of the climate elements affecting different regions are nonlinearly combined and weighted, so as to build a more complete nonlinear prediction model. Based on the large-scale optimal mode, the nonlinear downscaling prediction of the downscaled climate prediction target elements can effectively grasp the overall characteristics of the large-scale background of the regional climate, and perform efficient and accurate nonlinear quantitative intelligent prediction through the nonlinear prediction model, thereby improving the accuracy of the downscaled climate prediction.

In some embodiments, the climate system is a complex nonlinear system. The relationship between large-scale climate elements and small-scale climate elements cannot be accurately described by a simple linear model. Artificial intelligence technology has strong nonlinear modeling capabilities, so it has broad application prospects in downscaling climate prediction. Artificial intelligence solutions in related technologies usually need to be trained based on a large amount of observational data and dynamic model outputs, which not only consumes a lot of computing resources, but also lacks interpretability.

Based on this, the embodiment of the present application provides an artificial intelligence downscaling climate prediction method based on the large-scale optimal climate mode, which uses the clear physical relationship between the large-scale prediction output of the dynamic model and the large-scale climate mode elements established by the climate statistical method, and uses artificial intelligence technology to construct a nonlinear prediction model between the large-scale climate mode and the small-scale climate elements, thereby realizing nonlinear intelligent prediction of regional refined climate element anomalies.

In the embodiment of the present application, in order to solve the problem of regional refined climate prediction, especially to build an efficient nonlinear downscaling climate prediction model, an artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode is proposed. Based on the physical statistical relationship between key elements such as large-scale optimal climate mode and large-scale precipitation, an artificial intelligence solution is used to build a nonlinear prediction model between the optimal climate mode and regional refined climate elements, and the large-scale climate mode prediction results output by the dynamic model are used to finally realize the nonlinear intelligent prediction of the abnormal regional refined climate elements. The process of climate prediction artificial intelligence modeling based on the optimal mode can be realized through the following steps:

In the first step, climate elements of different scales are selected and the corresponding anomaly fields, abnormal relative tendency fields and recent background anomaly fields are calculated.

Here, according to the climate dynamics theory, the contemporaneous large-scale circulation field that can determine the large-scale prediction target climate elements corresponding to the downscaled climate prediction target elements is selected as the source for extracting and calculating the prediction factors in the subsequent steps.

As shown in FIG4 , in the first step, the historical observation data anomaly 401 of the large-scale climate element is determined through the historical observation data set 41 of the climate element, and the anomaly includes: the recent background anomaly 410 of the large-scale climate element and the relative tendency 411 of the large-scale climate element anomaly; the relative tendency 412 of the large-scale climate element anomaly in the same period that determines the prediction of the large-scale prediction target element anomaly is selected from the relative tendency of the large-scale climate element anomaly. And the relative tendency 413 of the large-scale prediction target climate element anomaly is selected from the relative tendency of the large-scale climate element anomaly. In the historical observation data set 42 of the regional refined climate element, the historical observation data anomaly 402 is determined. The relative tendency 47 of the regional refined climate element anomaly and the recent background anomaly 46 of the regional refined climate element can be obtained through the historical observation data anomaly 402. At the same time, the relative tendency 48 of the regional refined prediction target climate element anomaly is determined in the relative tendency of the regional refined climate element anomaly, and the relative tendency of the regional refined prediction target climate element anomaly corresponds to the relative tendency of the large-scale prediction target climate element anomaly.

In the second step, the large-scale optimal climate mode is extracted through the spatiotemporal coupling method and the corresponding time series are calculated.

Here, the method for extracting the large-scale optimal climate mode and the corresponding time series is: use the space-time coupling decomposition method (such as the singular value decomposition method) to decompose the large-scale circulation field and the large-scale prediction target climate elements of the same period, sort them based on the covariance contribution and obtain the large-scale optimal climate mode that determines the large-scale prediction target, and use the projection method to calculate the time series corresponding to the large-scale optimal climate mode.

As shown in Fig. 4, a spatiotemporal coupled mode decomposition method 403 (e.g., singular value decomposition method) is used to decompose the relative tendency of large-scale climate element anomalies and the relative tendency of target climate element anomalies obtained in the first step to obtain an optimal climate mode (SM) 404. The optimal mode time series 405 corresponding to the optimal climate model is determined using a projection method, and the optimal mode time series 405 is used as a prediction factor.

The third step is to use time series as the prediction factor and the relative tendency field of regional refined climate element anomalies as the prediction target, and to construct a downscaling nonlinear prediction model based on artificial intelligence methods.

Here, the large-scale optimal climate mode time series is used as the prediction factor, the downscaled climate prediction target element anomaly relative tendency field is used as the prediction target, and the artificial intelligence model is used to train the nonlinear climate prediction model. When building the model, a variety of artificial intelligence models can be selected, including classification tree schemes, support vector machines, recurrent neural networks, etc.

As shown in FIG4 , by using the regional refined prediction target climate element anomaly relative tendency and the optimal mode time series as the prediction target, an artificial intelligence method is used to build a downscaled nonlinear prediction model 406 .

The fourth step is to calculate the time coefficient corresponding to the large-scale optimal climate mode predicted by the global climate dynamic model during the same period, import the above-mentioned artificial intelligence downscaling prediction model, and make nonlinear predictions on the relative tendency field of regional refined climate element anomalies.

Here, the specific calculation method for calculating the time coefficient corresponding to the relative tendency field of the large-scale circulation field anomaly predicted by the global climate dynamic model in the same period is as follows: using the projection method, the large-scale circulation field anomaly relative tendency field predicted by the global climate dynamic model corresponding to the downscaled climate prediction target element is projected to the large-scale optimal climate mode obtained in the second step to obtain the corresponding time coefficient. This time coefficient is used as the actual prediction factor and brought into the artificial intelligence nonlinear climate prediction model constructed in the third step to obtain the prediction result of the relative tendency field of the downscaled climate prediction target element anomaly.

As shown in FIG4 , based on the historical return data set 43 of the climate dynamic model, the large-scale climate elements 407 predicted by the dynamic model of the same period are determined. Then, through the large-scale climate elements 407 predicted by the dynamic model of the same period, the relative tendency of the large-scale climate elements predicted by the dynamic model of the same period is determined 408. The relative tendency of the large-scale climate elements predicted by the dynamic model of the same period is projected onto the optimal climate mode to obtain the time coefficient of the prediction factor of the same period 409. The time coefficient is input into the downscaled nonlinear prediction model 406, and the prediction result 44 of the relative tendency of the abnormal climate elements of the target climate elements predicted by the regional refined prediction is output.

The fifth step is to obtain the downscaling prediction results of the regional refined climate element anomaly field based on the prediction results of the abnormal relative tendency field and the corresponding recent background anomaly field.

Here, the method for calculating the anomaly field of the downscaled climate prediction target element is: according to the corresponding relationship, the prediction results of the recent background anomaly field of the downscaled climate prediction target element in the first step and the anomaly relative tendency field of the downscaled climate prediction target element obtained in the fourth step are added together, and then the anomaly field of the downscaled climate prediction target element is obtained, and finally the artificial intelligence downscaling prediction based on the large-scale optimal climate mode is realized.

As shown in FIG4 , by combining the relative tendency prediction result 44 of the regional refined prediction target climate element anomaly with the recent background anomaly 46 of the regional refined climate element, the downscaled climate prediction target element anomaly field 45 can be obtained.

In some embodiments, taking the prediction of the 2019 summer precipitation anomaly field in the middle and lower reaches of the Yangtze River as an example, the artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode is described in detail.

First, the basic information of the embodiments of the present application is described:

In the embodiment of the present application, the regional refined climate prediction target is the 2019 summer precipitation anomaly field in the middle and lower reaches of the Yangtze River (the data is extracted from the meteorological information data of 2374 stations released by the CIPAS system of the Meteorological Bureau, and the middle and lower reaches of the Yangtze River are between 25°N~35°N, 110°E~125°E), and the corresponding large-scale element is selected as regional precipitation (the data is extracted from the meteorological information data of 160 stations released by the Meteorological Bureau), and the long-wave radiation outward from the tropical region (30°N~30°S) and the 500hPa potential height data of the mid- and high-latitude regions of the Northern Hemisphere (90°N~20°N) that determine the regional summer precipitation anomaly are used as the basis for extracting the large-scale optimal climate mode. As shown in Figure 6, spatial distribution 601 represents the spatial distribution of the precipitation anomaly field in the middle and lower reaches of the Yangtze River at 2374 stations in the summer of 2019, and spatial distribution 602 represents the spatial distribution of the precipitation anomaly field in the middle and lower reaches of the Yangtze River at 160 stations in the summer of 2019.

In an embodiment of the present application, the method of spatiotemporal coupling decomposition is singular value decomposition (SVD). The SVD method is used to extract the large-scale climate modes of tropical OLR and the relative tendency field of 500hPa potential height anomalies in the middle and high latitudes of the Northern Hemisphere and large-scale precipitation at 160 stations. According to the principle of maximum covariance, the top few modes with a total variance contribution of more than 90% are selected as the large-scale optimal climate modes. The time series corresponding to these modes are used as forecasting factors, and the refined station precipitation in the middle and lower reaches of the Yangtze River in the corresponding period is used as the prediction target for artificial intelligence modeling. The time dimension of the modeling training set is 1989~2018 (the first 30 years).

In the embodiment of the present application, the selected artificial intelligence model is a random forest regression model (RFRM), which is an integrated learning method based on decision trees. It improves the accuracy and robustness of the prediction by constructing multiple decision trees and combining all their prediction results for integrated prediction. The model first constructs multiple different sub-datasets by randomly sampling from the original training set through bootstrap aggregation, trains a decision tree with each feature data set, and finally calculates the mean of all decision tree predictions to generate prediction results. Compared with a single decision tree model, this model can effectively reduce the overfitting problem, and can provide an estimate of the importance of each predictor, which plays an important role in understanding the causal relationship between the predictor and the prediction target. In the downscale climate prediction application, the model can not only construct a nonlinear prediction model, but also analyze the contribution of the large-scale optimal climate mode affecting different refined sites, which helps to understand the causes of climate prediction results. As shown in Figure 7, the original training set 701 is randomly sampled to construct sets 1, 2, ···, N-1 and N. The corresponding decision trees are trained through set 1, set 2, ..., set N-1 and set N, and the training of each decision tree is completed through the first layer regression tree, the second layer regression tree and the third layer regression tree respectively, and the set voting 702 is performed to output the prediction result 703.

In the embodiment of the present application, the tropical OLR and the 500hPa potential height in the mid- and high-latitude regions of the Northern Hemisphere predicted by the dynamic model during the same period, as well as the prediction results of the summer precipitation anomaly in the middle and lower reaches of the Yangtze River as a control are given by the current main operational dynamic model BCC_CSM1.1 (m) of the National Climate Center (the data is released by the Beijing Climate Center). Taking into account the actual seasonal climate forecast business needs, the starting time of the dynamic model forecast in the embodiment of the present application is set to March 1, 2019.

As shown in Figure 5, the goal is to predict the precipitation anomaly field of the refined stations in the middle and lower reaches of the Yangtze River Basin in summer 2019, which can be achieved through the following steps:

In the first step, we selected the OLR field in the tropical region in the summer of 1989-2018, the Z500 field data in the middle and high latitudes of the Northern Hemisphere, and the precipitation data of 160 stations from the historical observation data set 51 of climate elements, and selected the precipitation data of 2374 stations from the historical observation data set 52 of regional refined climate elements, and calculated the anomaly field, the anomaly relative tendency field and the recent background anomaly field corresponding to each data. That is, we obtained the anomaly 501 of the historical observation data of large-scale climate elements, the recent background anomaly 502 of large-scale climate elements, the relative tendency of large-scale climate element anomalies from 1989 to 2018, the anomaly of historical observation data from 1989 to 2018, and the relative tendency of precipitation anomalies at 2374 stations from 1989 to 2018 as shown in Figure 5. Afterwards, the relative tendency of precipitation anomalies at 160 stations was determined to be 503 from the relative tendency of large-scale climate element anomalies from 1989 to 2018; and the relative tendency of precipitation anomalies at 2374 stations was determined to be 504 from the relative tendency of precipitation anomalies at 2374 stations from 1989 to 2018.

The tropical OLR data for the summer of the current year and the Z500 data for the mid- and high-latitudes in the Northern Hemisphere reported from March 1 each year are selected from the dynamic model historical report data set 53 to obtain the OLR and Z500 (2019) predicted by the dynamic model for the same period. The corresponding anomaly fields, abnormal relative tendency fields, and recent background anomaly fields are calculated respectively. The corresponding summer precipitation data are selected and interpolated to 2374 stations to obtain the precipitation data set for 2374 stations reported historically by the dynamic model, and the corresponding anomaly fields are calculated as the control group of this embodiment.

In the second step, the relative tendency field of tropical OLR anomalies and the relative tendency field of Z500 anomalies in the mid- and high-latitudes of the Northern Hemisphere in the summer from 1981 to 2010, that is, the relative tendency field of OLR and Z500 anomalies in the same period that determine the precipitation anomalies of the 160 stations, are decomposed by SVD using the spatiotemporal coupled mode decomposition method54; then, the large-scale optimal modes55 corresponding to OLR and Z500 and their corresponding time series56 are obtained by sorting them using the projection method and the maximum covariance principle, and the length of the time series is 30 (years).

In the third step, according to the principle that the sum of covariance contribution ratios exceeds 90%, the first 12 OLR modes and the first 13 Z500 modes calculated in the second step are selected, and the time series corresponding to the above modes are used as prediction factors. The relative tendency field of precipitation anomalies in the 2374 stations in the middle and lower reaches of the Yangtze River in the summer of 1989-2018 obtained in the first step is used as the prediction target, and the random forest regression model is used for modeling training to obtain the downscaling climate prediction model of refined precipitation artificial intelligence in the middle and lower reaches of the Yangtze River based on the optimal climate mode of large-scale precipitation 57. The main parameters of the model are set as follows: the number of decision trees is 10, the maximum depth of the decision tree is 3 layers, and the tree is constructed by sampling with replacement, and the rest of the parameters are set to the default values.

In the fourth step, based on the large-scale optimal climate mode obtained in the second step and the 2019 OLR and northern hemisphere mid- and high-latitude Z500 anomaly relative tendency fields calculated in the first step for the actual prediction target (for example, the large-scale climate element anomaly relative tendency field 506 predicted by the concurrent dynamic model in Figure 5), the time coefficient corresponding to the large-scale optimal climate mode in the summer of 2019 is calculated by the projection method, for example, the concurrent (2019) prediction factor time coefficient 507 in Figure 5; this time coefficient is used as the actual prediction factor for calculating the precipitation anomaly relative tendency field in the middle and lower reaches of the Yangtze River in the summer of 2019, and is input into the downscaled climate prediction model 57 obtained in the third step, so as to obtain the artificial intelligence downscaling prediction results of the precipitation anomaly relative tendency field at the middle and lower reaches of the Yangtze River in the summer of 2019; for example, the prediction result 508 of the precipitation anomaly relative tendency field at 2374 stations in 2019 in Figure 5.

The fifth step is to base on the recent background anomaly tendency field of precipitation in the middle and lower reaches of the Yangtze River in the summer of 2019 calculated in the first step (for example, the recent background anomaly tendency field of precipitation at station 2374 in 2019 is 510 in Figure 5) and the downscaling prediction results of the relative tendency field of precipitation anomaly in the middle and lower reaches of the Yangtze River in the summer of 2019 calculated in the fourth step. According to the corresponding relationship in the first step, the prediction results of the precipitation anomaly field at station 2374 in the middle and lower reaches of the Yangtze River in the summer of 2019 can be finally obtained; for example, the precipitation anomaly value of station 2374 in the summer of 2019 is 509 in Figure 5.

In the sixth step, after all calculations are completed, the prediction results of the precipitation anomaly field in the middle and lower reaches of the Yangtze River in the summer of 2019 obtained by the prediction method provided in the embodiment of the present application are compared with the historical report results of the precipitation anomaly field in the middle and lower reaches of the Yangtze River in the summer of 2019 of the BCC_CSM1.1 (m) dynamic model obtained in the first step. The spatial distribution is shown in Figure 8, wherein spatial distribution 801 represents the prediction results of the business model of the precipitation anomaly field in the middle and lower reaches of the Yangtze River in the summer of 2019, and spatial distribution 802 represents the prediction results of the scheme of the embodiment of the present application. It can be found that the artificial intelligence downscaling climate prediction method based on the large-scale optimal climate mode in the embodiment of the present application, in the prediction of the precipitation anomaly field in the middle and lower reaches of the Yangtze River in the summer of 2019, can more accurately grasp the overall characteristics of precipitation than the dynamic model, and at the same time, for the prediction of refined sites, especially in some areas of the lower reaches of the Yangtze River, it shows better prediction performance.

In the embodiment of the present application, the spatiotemporal coupled modal decomposition method is used to extract the corresponding large-scale optimal climate mode and time series of the same period that determine the relative tendency of the abnormal large-scale climate elements based on the large-scale climate elements corresponding to the target climate elements of regional refined prediction; the artificial intelligence model is used to train and construct the downscaling prediction model of the nonlinear relationship between the time series of the large-scale optimal climate mode and the relative tendency of the abnormal abnormal large-scale climate elements of the target climate elements of regional refined prediction; the time coefficient of the large-scale optimal climate mode predicted by the global climate dynamic model is brought into the nonlinear downscaling prediction model to realize the prediction of the relative tendency of the abnormal abnormal large-scale climate elements of the regional refined prediction; combined with the recent background anomalies of historical observations, the artificial intelligence downscaling climate prediction of the regional refined prediction target climate element anomaly is finally realized. Compared with the downscaling climate prediction method in the related art, the embodiment of the present application can make full use of the nonlinear prediction ability of artificial intelligence modeling and the prediction ability of the global climate dynamic model for large-scale climate modes according to the physical relationship between the large-scale optimal climate mode and the large-scale climate elements corresponding to the refined prediction target, and at the same time, the advantages of a variety of existing prediction schemes are integrated to establish an efficient and accurate downscaling climate prediction model, which can effectively improve the regional refined climate prediction ability.

The embodiment of the present application provides a climate prediction device, and FIG9 is a structural schematic diagram of an artificial intelligence downscaling climate prediction device based on a large-scale optimal climate mode provided by the embodiment of the present application. Exemplarily, as shown in FIG9 , the artificial intelligence downscaling climate prediction device 900 based on a large-scale optimal climate mode includes:

The first determination module 901 is used to select a large-scale circulation field for constructing a physical statistical relationship based on the large-scale prediction target element corresponding to the downscaled climate prediction target element, and respectively determine the corresponding climate anomaly field, anomaly relative tendency field and recent background anomaly field;

The first decomposition module 902 is used to perform spatiotemporal coupling decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale prediction target element and the time series corresponding to the large-scale optimal climate mode;

A training module 903 is used to train the nonlinear prediction model to be trained by taking the time series as a prediction factor and the abnormal relative tendency field of the downscaled climate prediction target element as a prediction target to obtain a nonlinear prediction model;

The second determination module 904 is used to determine the time coefficient corresponding to the large-scale circulation field in the same period according to the large-scale optimal climate mode and the large-scale circulation field in the same period, and import the time coefficient into the nonlinear prediction model to obtain the prediction result of the abnormal relative tendency field of the downscaled climate prediction target element;

The third determination module 905 is used to obtain the nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element based on the prediction results of the recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element.

In the above device, the large-scale optimal climate mode is extracted by a spatiotemporal coupling decomposition method based on the large-scale climate elements corresponding to the downscaled climate prediction target elements and the large-scale circulation elements that determine the anomalies of the large-scale climate elements.

In the above device, the first determination module is also used to add the abnormal relative tendency field and the recent background abnormal field to obtain the climate abnormal field.

In the above-mentioned device, the first determination module is also used to adopt climate dynamics theory to select the large-scale circulation field for constructing physical statistical relationships based on the large-scale prediction target elements corresponding to the downscaled climate prediction target elements.

In the above-mentioned device, the first decomposition module is also used to perform singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode; project the abnormal relative tendency field of the large-scale circulation field to the large-scale optimal climate mode to obtain the time series corresponding to the large-scale optimal climate mode.

In the above-mentioned device, the first decomposition module is also used to perform singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element; sort them according to the covariance corresponding to the singular value decomposition results to obtain the large-scale optimal climate mode.

In the above-mentioned device, the second determination module is also used to use the global climate dynamic model to predict the circulation field corresponding to the downscaled climate prediction target element to obtain the contemporaneous large-scale circulation field; project the contemporaneous large-scale circulation field to the large-scale optimal climate mode to obtain the time coefficient corresponding to the contemporaneous large-scale circulation field.

In the above device, the third determination module is also used to add the recent background anomaly field of the downscaled climate prediction target element and the abnormal relative tendency field of the downscaled climate prediction target element to obtain a nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element.

In the above-mentioned device, the training module is also used to obtain an artificial intelligence model; based on the artificial intelligence model, a nonlinear prediction model to be trained including multiple decision trees is constructed; wherein the artificial intelligence model includes: a random forest regression model.

In the above device, the first determination module is also used to decompose any abnormal field of climate elements into an abnormal relative tendency field and a recent background abnormal field.

It should be noted that the artificial intelligence downscaling climate prediction device based on the large-scale optimal climate mode provided in the above embodiment is only illustrated by the division of the above functional modules. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the computer device is divided into different functional modules to complete all or part of the functions described above. In addition, the category prediction device and the category prediction method embodiment provided in the above embodiment belong to the same concept. The specific implementation process is detailed in the method embodiment and will not be repeated here.

An embodiment of the present application further provides an electronic device. FIG10 is a schematic diagram of the structure of an electronic device provided in an embodiment of the present application.

Exemplarily, as shown in FIG10 , the electronic device 1000 includes: a memory 1001 and a processor 1002 , wherein the memory 1001 stores an executable program code 10011 , and the processor 1002 is used to call and execute the executable program code 10011 to perform an artificial intelligence downscaling climate prediction method based on large-scale optimal climate modes.

In addition, an embodiment of the present application also protects a device, which may include a memory and a processor, wherein the memory stores executable program code, and the processor is used to call and execute the executable program code to execute an artificial intelligence downscaling climate prediction method based on a large-scale optimal climate mode provided in an embodiment of the present application.

In this embodiment, the functional modules of the device can be divided according to the above method example. For example, each functional module can be corresponded, 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 modules in the embodiment of the present application is schematic and is only a logical function division. There may be other division methods in actual implementation.

In the case of dividing each module according to each function, the device may also include a signal uploading module, a determining module, an adjusting 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.

It should be understood that the device provided in this embodiment is used to execute the above-mentioned artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode, and thus can achieve the same effect as the above-mentioned implementation method.

In the case of an integrated unit, the device may include a processing module and a storage module. When the device is applied to a device, the processing module may be used to control and manage the actions of the device. The storage module may be used to support the device to execute mutual program codes, etc.

The processing module may be a processor or a controller, which may implement or execute various exemplary logic blocks, modules and circuits disclosed in the present application. The processor may also be a combination that implements computing functions, 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.

In addition, the device provided in the embodiments of the present application may specifically be a chip, a component or a module, and the chip may include a connected processor and a memory; wherein the memory is used to store instructions, and when the processor calls and executes the instructions, the chip can execute an artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode provided in the above embodiment.

This embodiment also provides a computer-readable storage medium, which stores computer program code. When the computer program code runs on a computer, the computer executes the above-mentioned related method steps to implement an artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode provided by the above embodiment.

This embodiment 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 an artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode provided by the above embodiment.

Among them, the device, computer-readable storage medium, computer program product or chip provided in this embodiment 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.

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.

In the embodiments provided in the present application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic, for example, the division of modules or units is only a logical function division, and there may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another device, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

The above contents are only specific implementation methods of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.

Claims (10)

1. An artificial intelligence downscaling climate prediction method based on large-scale optimal climate mode, characterized in that the method comprises: Based on the large-scale prediction target elements corresponding to the downscaled climate prediction target elements, the large-scale circulation field for constructing the physical statistical relationship is selected, and the corresponding climate anomaly field, anomaly relative tendency field and recent background anomaly field are determined respectively; Performing spatiotemporal coupling decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale prediction target element and the time series corresponding to the large-scale optimal climate mode; Taking the time series as a prediction factor and the abnormal relative tendency field of the downscaled climate prediction target element as a prediction target, training the artificial intelligence prediction model to be trained to obtain a nonlinear prediction model based on an artificial intelligence method; According to the large-scale optimal climate mode and the large-scale circulation field of the same period, the time coefficient corresponding to the large-scale circulation field of the same period is determined, and the time coefficient is introduced into the artificial intelligence prediction model to obtain the prediction result of the abnormal relative tendency field of the downscaled climate prediction target element; Based on the prediction results of the recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element, a nonlinear quantitative artificial intelligence prediction result of the anomaly field of the downscaled climate prediction target element is obtained. 2. The method according to claim 1 is characterized in that the large-scale optimal climate mode is extracted by a spatiotemporal coupling decomposition method based on the large-scale climate elements corresponding to the downscaled climate prediction target elements and the large-scale circulation elements that determine the anomalies of the large-scale climate elements. 3. The method according to claim 1, characterized in that the method further comprises: The abnormal relative tendency field and the recent background abnormal field are added together to obtain the climate abnormal field. 4. The method according to claim 1, characterized in that the large-scale prediction target element corresponding to the downscaled climate prediction target element is selected to construct a large-scale circulation field for physical statistical relationship, comprising: The large-scale circulation field for constructing the physical statistical relationship is selected by using the climate dynamics theory based on the large-scale prediction target elements corresponding to the downscaled climate prediction target elements. 5. The method according to claim 1, characterized in that the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element are subjected to spatiotemporal coupling decomposition to obtain the large-scale optimal climate mode of the abnormal relative tendency field of the large-scale prediction target element and the time series corresponding to the large-scale optimal climate mode, including: Performing singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode; The abnormal relative tendency field of the large-scale circulation field is projected onto the large-scale optimal climate mode to obtain a time series corresponding to the large-scale optimal climate mode. 6. The method according to claim 5, characterized in that the step of performing singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element to obtain the large-scale optimal climate mode comprises: Performing singular value decomposition on the abnormal relative tendency field of the large-scale circulation field and the abnormal relative tendency field of the large-scale prediction target element; The large-scale optimal climate mode is obtained by sorting according to the covariance corresponding to the singular value decomposition results. 7. The method according to claim 1, characterized in that the determining of the time coefficient corresponding to the large-scale circulation field in the same period according to the large-scale optimal climate mode and the large-scale circulation field in the same period comprises: Using a global climate dynamic model to predict the circulation field corresponding to the downscaled climate prediction target element, to obtain the large-scale circulation field in the same period; The large-scale circulation field of the same period is projected onto the large-scale optimal climate mode to obtain the time coefficient corresponding to the large-scale circulation field of the same period. 8. The method according to claim 1, characterized in that the prediction results of the recent background anomaly field of the downscaled climate prediction target element and the abnormal relative tendency field of the downscaled climate prediction target element are used to obtain the nonlinear quantitative prediction results of the anomaly field of the downscaled climate prediction target element, including: The recent background anomaly field of the downscaled climate prediction target element and the anomaly relative tendency field of the downscaled climate prediction target element are added to obtain a nonlinear quantitative prediction result of the anomaly field of the downscaled climate prediction target element. 9. The method according to claim 1, characterized in that the time series is used as a prediction factor, the abnormal relative tendency field of the downscaled climate prediction target element is used as a prediction target, and the nonlinear prediction model to be trained is trained to obtain a nonlinear prediction model, comprising: Get AI models; Based on the artificial intelligence model, a nonlinear prediction model to be trained including multiple decision trees is constructed; wherein the artificial intelligence model includes: a random forest regression model. 10. The method according to claim 1, characterized in that the method further comprises: The anomaly field of any climate element is decomposed into the anomaly relative tendency field and the recent background anomaly field.