Climate set prediction method and system based on artificial intelligence technology

Abstract

The invention provides a climate set prediction method and system based on an artificial intelligence technology, and the method is applied to the technical field of artificial intelligence, and comprises the steps: carrying out the space-time decomposition of a prediction target climate element, and obtaining an actual prediction target; dimensionality reduction and feature extraction are carried out on the early-stage large-scale circulation field influencing the principal components of the plurality of climatic elements to obtain an actual prediction factor; constructing a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target; determining a set prediction result of the main components of the plurality of climate elements; and reconstructing a complete prediction result for predicting the target climate elements. According to the method, predictable information can be extracted from climate big data based on an artificial intelligence technology, climate dynamics and climate predictability research, a plurality of artificial intelligence nonlinear fitting models are constructed through different parameter schemes, a prediction model with higher stability and generalization is further constructed by using a set prediction scheme, and the prediction efficiency is improved. Therefore, the climate elements are objectively and quantitatively predicted.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a climate ensemble prediction method and system based on artificial intelligence technology in the field of artificial intelligence technology. Background Art

[0002] Affected by non-human factors such as weather and environment, there are many natural disasters that occur around the world every year. Relevant statistics show that climate disasters cause huge economic losses every year and seriously affect the safety of people's lives and property. At the same time, agriculture, water conservancy, new energy and other fields are closely related to climate. Therefore, building an efficient climate prediction system is of great significance to economic and social development.

[0003] At present, there are "three major problems" in climate prediction: complex time scales, multiple influencing factors, and the "chaotic" variability of the system itself. Among them, complex time scales refer to the fact that the climate system involves changes in multiple time scales, and processes on different time scales interact with each other; multiple influencing factors refer to the fact that the climate system is affected by multiple factors; chaotic variability refers to the inherent nonlinear characteristics of the climate system, that is, small changes in initial conditions may lead to huge differences in results. The mutual influence between these three parts requires climate prediction to face this complex nonlinear system. Therefore, it is an important solution to improve climate prediction capabilities by extracting predictable information from climate big data in a targeted manner and building an efficient and stable nonlinear prediction model.

[0004] Artificial intelligence technology, with its excellent feature extraction and nonlinear fitting performance, has shown amazing results and potential in the fields of computer vision and time series prediction. Based on this, how to make full use of the feature extraction and nonlinear prediction capabilities of artificial intelligence technology to predict the climate has become an urgent problem to be solved. Summary of the invention

[0005] The present application provides a climate ensemble prediction method and system based on artificial intelligence technology. The method can extract predictable information from climate big data based on artificial intelligence technology and climate dynamics and climate predictability research, construct multiple artificial intelligence nonlinear fitting models through different parameter schemes, and use ensemble prediction schemes to further construct a more stable and generalized prediction model, thereby making objective and quantitative predictions of climate elements.

[0006] In a first aspect, a climate ensemble prediction method based on artificial intelligence technology is provided, the method comprising: performing spatiotemporal decomposition on the predicted target climate element to obtain a plurality of climate element principal components with different spatiotemporal characteristics, and using the plurality of climate element principal components as actual prediction targets; based on the study of climate dynamics and climate predictability, extracting the previous large-scale circulation field that affects the principal components of the plurality of climate elements from the large data set of historical observations of the earth system, performing dimensionality reduction and feature extraction on the previous large-scale circulation field through a convolutional neural network, obtaining a time series set of predictable circulation feature vectors, and using the time series set of predictable circulation feature vectors as actual prediction factors; using a recurrent neural network, selecting the historical climate data of 30 years before the prediction target period as a training set, and constructing a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target; training a plurality of nonlinear artificial intelligence fitting models for the plurality of climate element principal components through a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method, and obtaining the ensemble prediction results of the plurality of climate element principal components through an ensemble averaging scheme; based on the ensemble prediction results of the plurality of climate element principal components, reconstructing the complete prediction results of the prediction target climate element through a principal component synthesis method.

[0007] Compared with the prior art, the above technical solution has the following advantages: (1) This application aims to solve the nonlinear prediction problems existing in the climate system and climate prediction by using artificial intelligence methods and selecting appropriate artificial intelligence models to extract nonlinear prediction factors and build prediction models, thereby achieving nonlinear quantitative and objective prediction of climate elements.

[0008] (2) Based on the idea of ​​ensemble prediction, this application designs a variety of schemes for constructing different ensemble members, and makes comprehensive use of the above-mentioned schemes to quantitatively describe the possible uncertainties in the prediction model, and gives an ensemble prediction result containing more prediction information through an ensemble averaging scheme. Compared with the prediction of a single model, this method can effectively reduce the forecast error and improve the prediction effect of the prediction model.

[0009] (3) This application uses artificial intelligence methods to perform spatiotemporal decomposition of the target climate elements for prediction, focusing the prediction on the dominant principal components, thereby reducing the impact of other components such as noise in the climate elements on the prediction and improving the predictability of the model.

[0010] (4) Based on the principles of climate dynamics and climate predictability, this application selects the large-scale circulation field that determines the target climate elements to be predicted as the source of prediction factors, and uses artificial intelligence technology to reduce the dimension and extract features, thereby reducing the complexity of the prediction model and building a dedicated artificial intelligence prediction model that focuses on the prediction target. This greatly improves the training and reasoning efficiency, effectively reduces the consumption of computing resources and energy, and at the same time has stronger portability and simpler expansion and upgradeability.

[0011] In combination with the first aspect, in some possible implementations, the predicted target climate element is decomposed in time and space to obtain multiple climate element principal components with different time and space characteristics, including: decomposing the time series of the predicted target climate element from the time dimension and the space dimension through the multivariate variational mode decomposition method, and extracting the multiple climate element principal components.

[0012] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, based on the study of climate dynamics and climate predictability, the previous large-scale circulation field that affects the main components of the multiple climate elements is extracted from the big data set of historical observations of the earth system, including: based on the study of climate dynamics and climate predictability, the physical mechanism of variation of the predicted target climate element is analyzed, and different previous large-scale circulation fields are selected to determine the predicted target climate element.

[0013] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, the use of a recurrent neural network selects the historical climate data of the 30 years before the prediction target period as a training set, and constructs a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target, including: using a long short-term memory neural network to model the actual prediction target and the actual prediction factor, selecting the historical climate data of the 30 years before the prediction target period as a training set, and constructing a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target.

[0014] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method are used to train multiple nonlinear artificial intelligence fitting models for the multiple climate element main components, and the artificial intelligence ensemble prediction model of the actual prediction target is constructed through the ensemble averaging scheme, including: constructing the multiple nonlinear artificial intelligence fitting models for the multiple climate element main components through the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method, training and introducing the actual training factors to determine the multiple prediction results of the multiple climate element main components, and obtaining the ensemble prediction results of the multiple climate element main components through the ensemble averaging scheme.

[0015] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method are used to construct the multiple nonlinear artificial intelligence fitting models for the multiple climate element main components, train and introduce the actual training factor to determine the multiple prediction results of the multiple climate element main components, and obtain the collective prediction results of the multiple climate element main components through the ensemble average scheme, including: adding an initial random perturbation to the training set through the random initial perturbation scheme, setting different initialization parameters for the long short-term memory neural network through the random initialization parameter scheme, randomly setting the cyclic weights in the iterative cycle process of the long short-term memory network through the random cyclic weight scheme, selecting the learning rate and number of iterations of the long short-term memory network through the training method, constructing the multiple nonlinear artificial intelligence fitting models, training and determining the multiple prediction results based on the actual training factor, and obtaining the collective prediction results of the multiple climate element main components through the ensemble average scheme.

[0016] In combination with the first aspect and the above-mentioned implementation methods, in some possible implementation methods, the ensemble prediction results based on the principal components of the multiple climate elements are used to reconstruct the complete prediction results of the predicted target climate elements through a principal component synthesis method, including: reconstructing the complete prediction results of the predicted target climate elements through an inverse operation of the multivariate variational mode decomposition method, based on the ensemble prediction results of the principal components of the multiple climate elements, through the principal component synthesis method.

[0017] In a second aspect, another climate ensemble prediction method based on artificial intelligence technology is provided, the method comprising: extracting a target early stage large-scale circulation field that affects multiple target climate element principal components from a large data set of historical observations of the earth system, the multiple target climate element principal components being obtained by performing spatiotemporal decomposition of the target climate elements to be predicted; inputting the target early stage large-scale circulation field into multiple nonlinear artificial intelligence ensemble prediction models to obtain multiple prediction results for each of the multiple target climate element principal components, the multiple nonlinear artificial intelligence ensemble prediction models being constructed based on the method in the first aspect or any possible implementation of the first aspect; based on an ensemble averaging scheme, ensemble averaging the multiple prediction results for each target climate element principal component to obtain an ensemble prediction result of the multiple target climate element principal components; based on the ensemble prediction result of the multiple target climate elements, obtaining a complete prediction result of the target climate element to be predicted.

[0018] The third aspect provides a climate ensemble prediction system based on artificial intelligence technology, which includes: a spatiotemporal decomposition module for performing spatiotemporal decomposition of the predicted target climate elements to obtain multiple climate element principal components with different spatiotemporal characteristics, and using the multiple climate element principal components as actual prediction targets; a feature extraction module for extracting the previous large-scale circulation field that affects the multiple climate element principal components from the Earth system historical observation big data set based on climate dynamics and climate predictability research, performing dimensionality reduction and feature extraction on the previous large-scale circulation field through a convolutional neural network, and obtaining a time series set of predictable circulation feature vectors, and using the time series set of predictable circulation feature vectors as actual prediction factors; a fitting module A model construction module is used to use a recurrent neural network to select the historical climate data of the 30 years before the prediction target period as a training set to construct a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target; an ensemble prediction result determination module is used to train multiple nonlinear artificial intelligence fitting models for the multiple climate element principal components through a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method, and obtain the ensemble prediction results of the multiple climate element principal components through an ensemble averaging scheme; a prediction result reconstruction module is used to reconstruct the complete prediction result of the prediction target climate element through a principal component synthesis method based on the ensemble prediction results of the multiple climate element principal components.

[0019] In combination with the third aspect, in some possible implementations, the spatiotemporal decomposition module is specifically used to: decompose the time series of the predicted target climate element from the time dimension and the space dimension through the multivariate variational mode decomposition method, and extract the multiple principal components of the climate elements.

[0020] In combination with the third aspect and the above-mentioned implementation methods, in some possible implementation methods, the feature extraction module is specifically used to: based on climate dynamics and climate predictability research, analyze the physical mechanism of variation of the predicted target climate element, and select different early large-scale circulation fields for determining the predicted target climate element.

[0021] In combination with the third aspect and the above-mentioned implementation methods, in some possible implementation methods, the fitting model construction module is specifically used to: use a long short-term memory neural network to model the actual prediction target and the actual prediction factor, select the historical climate data for the 30 years before the prediction target period as a training set, and construct a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target.

[0022] In combination with the third aspect and the above-mentioned implementation methods, in some possible implementation methods, the collective prediction result determination module is specifically used to: construct the multiple nonlinear artificial intelligence fitting models for the multiple climate element main components through the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method, train and introduce the actual training factors to determine the multiple prediction results of the multiple climate element main components, and obtain the collective prediction results of the multiple climate element main components through the collective averaging scheme.

[0023] In combination with the third aspect and the above-mentioned implementation methods, in some possible implementation methods, the ensemble prediction result determination module is also used to: add initial random disturbance to the training set through the random initial disturbance scheme, set different initialization parameters for the long short-term memory neural network through the random initialization parameter scheme, randomly set the cyclic weights in the iterative cycle process of the long short-term memory network through the random cyclic weight scheme, select the learning rate and number of iterations of the long short-term memory network through the training method, construct the multiple nonlinear artificial intelligence fitting models, train and determine the multiple prediction results based on the actual training factors, and obtain the ensemble prediction results of the multiple climate element principal components through the ensemble averaging scheme.

[0024] In combination with the third aspect and the above-mentioned implementation methods, in some possible implementation methods, the prediction result reconstruction module is specifically used to: reconstruct the complete prediction result of the predicted target climate element through the inverse operation of the multivariate variational mode decomposition method, based on the collective prediction results of the principal components of the multiple climate elements, through the principal component synthesis method.

[0025] In a fourth aspect, a computer system is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the computer system implements the first aspect or any possible implementation method of the first aspect or the climate ensemble prediction method based on artificial intelligence technology in the second aspect. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 is a schematic flow chart of a climate ensemble prediction method based on artificial intelligence technology provided in an embodiment of the present application; Figure 2 It is a schematic diagram of a scenario for determining a collective prediction result of a principal component of any climate element provided in an embodiment of the present application; Figure 3 This is a schematic diagram of a scenario for performing spatiotemporal decomposition of precipitation at site B in country A, provided in an embodiment of the present application; Figure 4 It is a structural schematic diagram of a nonlinear artificial intelligence fitting model corresponding to a single training member provided in an embodiment of the present application; Figure 5 This is a schematic diagram of a scenario for evaluating a multi-member set prediction result provided in an embodiment of the present application; Figure 6 This is a schematic diagram of a scenario for comparing prediction effects of different prediction methods provided in an embodiment of the present application; Figure 7 It is a structural schematic diagram of a climate ensemble prediction system based on artificial intelligence technology provided in an embodiment of the present application; Figure 8 It is a structural diagram of a computer system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0027] 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.

[0028] 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.

[0029] At present, there are "three major problems" in climate prediction: complex time scales, multiple influencing factors, and the "chaotic" variability of the system itself. Among them, complex time scales refer to the fact that the climate system involves changes in multiple time scales, and processes on different time scales interact with each other; multiple influencing factors refer to the fact that the climate system is affected by multiple factors; chaotic variability refers to the inherent nonlinear characteristics of the climate system, that is, small changes in initial conditions may lead to huge differences in results. The mutual influence between these three parts requires climate prediction to face this complex nonlinear system. Therefore, it is an important solution to improve climate prediction capabilities by extracting predictable information from climate big data in a targeted manner and building an efficient and stable nonlinear prediction model.

[0030] There are two main types of traditional climate prediction schemes: statistical prediction schemes and climate dynamic model prediction schemes. Among them, the statistical prediction scheme is mainly based on historical observation data, using statistical methods to find the temporal and spatial correlation between different atmospheric signals, and then predict climate elements through statistical models. The climate dynamic model is mainly based on physical laws, and through numerical solutions, it obtains the mathematical and physical equations of the earth system, thereby simulating the earth system by numerical means to achieve the purpose of prediction.

[0031] However, the statistical prediction scheme model is simple and does not have the ability to predict situations that have never occurred in history. In addition, the statistical prediction scheme is mostly linear and lacks nonlinear prediction capabilities. The climate dynamic model prediction scheme has a clear physical mechanism, but is limited by theoretical research, computing power and observation basis. The climate dynamic model prediction scheme still lacks descriptions of some key physical processes and is affected by initial values, iteration errors and approximate errors, which can easily lead to prediction bias.

[0032] Artificial intelligence technology, with its excellent feature extraction and nonlinear fitting performance, has shown amazing results and potential in the fields of computer vision and time series prediction. Based on this, how to make full use of the feature extraction and nonlinear prediction capabilities of artificial intelligence technology to predict the climate has become an urgent problem to be solved.

[0033] The following is an introduction to the climate ensemble prediction method based on artificial intelligence technology provided in an embodiment of the present application.

[0034] Figure 1 It is a schematic flow chart of a climate ensemble prediction method based on artificial intelligence technology provided in an embodiment of the present application. It should be understood that the method can be applied to any electronic device integrated with the artificial intelligence ensemble prediction model provided in an embodiment of the present application.

[0035] For example, Figure 1 As shown, the method 100 includes: 101. Decompose the predicted target climate elements in time and space to obtain multiple climate element principal components with different time and space characteristics, and use the multiple climate element principal components as actual prediction targets.

[0036] 102. Based on the study of climate dynamics and climate predictability, the previous large-scale circulation field that affects the main components of multiple climate elements is extracted from the large data set of historical observations of the Earth system. The dimension reduction and feature extraction of the previous large-scale circulation field are carried out through convolutional neural networks to obtain a time series set of predictable circulation feature vectors. The time series set of predictable circulation feature vectors is used as the actual prediction factor.

[0037] 103. Using a recurrent neural network, we selected the historical climate data of the 30 years before the prediction target period as the training set and constructed a nonlinear artificial intelligence fitting model of the actual prediction factors and the actual prediction targets.

[0038] It should be understood that the core of the climate ensemble prediction method based on artificial intelligence technology provided by the embodiment of the present application is to obtain an artificial intelligence ensemble prediction model by constructing multiple nonlinear artificial intelligence fitting models, predict the main components of the climate elements of the target period to be predicted, and then obtain the final complete prediction result based on the prediction results of the main components of the climate elements. Therefore, the premise for the implementation of the embodiment of the present application is to first construct the nonlinear artificial intelligence fitting model required for the prediction.

[0039] Specifically, when constructing multiple nonlinear artificial intelligence fitting models, the embodiments of the present application are roughly divided into the following steps: data collection and preprocessing, spatiotemporal decomposition, feature extraction, and model construction. The construction of the nonlinear artificial intelligence fitting model is described below according to these steps.

[0040] (1) Data collection and preprocessing: First, during the model construction process, based on the actual prediction needs and prediction scenarios, the embodiments of the present application can first determine the prediction target period and the climate elements that need to be predicted.

[0041] Climate elements refer to the basic meteorological parameters that describe and define the climate characteristics of a region.

[0042] Optionally, climate elements include temperature, precipitation, wind speed, air pressure, sunshine hours, air pressure, visibility, cloud cover, evaporation, etc.

[0043] In one scenario, after determining the forecast target period and the climate elements to be forecasted, the electronic device can obtain the historical climate data of 30 years before the forecast target period as a training set for building a nonlinear artificial intelligence fitting model. The historical climate data includes various climate elements.

[0044] According to climate research, the climate changes periodically, and every 30 years as a decadal background, the climate changes. Therefore, the historical climate data 30 years before the target period can be used as a training set to pre-train multiple artificial intelligence models.

[0045] For example, assuming that the forecast target period is the summer of 2022, and the forecast target climate element is the precipitation at site A, the historical climate data may be the climate data of site A from 1992 to 2021.

[0046] After obtaining the historical climate data of the 30 years before the target forecast period, since the historical climate data includes various climate elements, the electronic device can extract the precipitation of the 30 years before the target forecast period from the historical climate data of the 30 years before the target forecast period according to the forecast requirements, that is, the "target forecast climate element" in the embodiment of the present application.

[0047] In another scenario, after determining the forecast target period and the climate elements to be forecasted, the electronic device can also directly obtain the climate elements of the 30 years before the forecast target period as historical climate data, which is also the forecast target climate element. For example, in the above example, the electronic device can directly collect the precipitation of site A in summer from 1992 to 2021, that is, obtain the forecast target climate element, without first obtaining the climate data of site A in summer from 1992 to 2021, and extracting the precipitation of site A in summer from 1992 to 2021.

[0048] After obtaining the predicted target climate element, the electronic device can perform spatiotemporal decomposition on the predicted target climate element to obtain a plurality of principal components of the climate element.

[0049] Principal components refer to new orthogonal variables extracted from the original data set. These variables can capture the maximum variance in the original data, help understand the main features and structure of the original data, and remove noise and redundant information in the original data.

[0050] In one possible implementation, the predicted target climate element is decomposed in time and space to obtain multiple principal components of climate elements with different time and space characteristics, including: Through the multivariate variational mode decomposition method, the time series of the predicted target climate elements are decomposed from the time dimension and the spatial dimension, and multiple principal components of climate elements are extracted.

[0051] Among them, Multiple Variational Mode Decomposition (MVMD) is a method for decomposing multivariate time series signals, and its main purpose is to decompose multivariate time series into several Intrinsic Mode Functions (IMFs). Each IMF represents a specific frequency component.

[0052] Before performing spatiotemporal decomposition on the time series of the predicted target climate element by using the MVMD method, the time series of the predicted target climate element can be converted into anomaly data.

[0053] The calculation formula of the anomaly data is shown in the following formula (1): Anomaly data = observed value - long-term average value formula (1) For example, assuming that the predicted target climate element is the summer precipitation at station A from 1992 to 2021, the summer precipitation at station A in 1992 is 500 mm, the summer precipitation at station A in 1993 is 480 mm, and so on. When the summer precipitation at station A from 1992 to 2021 is converted into anomaly data, the average value of the summer precipitation at station A from 1992 to 2021 can be calculated first, which is the long-term average value in formula (1).

[0054] After obtaining the long-term average value, the precipitation at site A in summer each year is subtracted from the long-term average value to obtain the anomaly data of the summer precipitation at site A from 1992 to 2021.

[0055] Furthermore, the MVMD method is used to perform spatiotemporal decomposition of the anomaly data of the predicted target climate elements.

[0056] When the MVMD method performs time-space decomposition, the specific steps include: 1. Initialization: Initialization specifically includes initialization modal function U k , center frequency ω k , and initialize the Lagrange multipliers λ .

[0057] 2. Iterative update: For each modal function U k , calculate the frequency domain representation of the current mode function .renew To focus it on the center frequency ω k Nearby. Through inverse Fourier transform Convert back to the time domain. Update the center frequency ω k To make it better reflect the frequency characteristics of the current mode function. Update the Lagrange multiplier λ .

[0058] 3. Convergence check: Check whether the residual is less than the set threshold. If the condition is met, stop the iteration; otherwise, continue iterating until the maximum number of iterations is reached.

[0059] Thus, through the above process, the electronic device can obtain multiple climate element principal components of the anomaly data of the predicted target climate element as the actual prediction target.

[0060] In the above technical solution, the MVMD method is used to perform spatiotemporal decomposition of the predicted target climate elements, which can separate the noise components, reduce the impact of noise on the prediction results, reduce the dimension of the data, effectively retain the most important information in the original data, and reduce the complexity of model training and calculation.

[0061] After obtaining the output results of the model, for the input of the model, in an embodiment of the present application, the electronic device can extract the previous large-scale circulation field that can affect the main components of the above-mentioned multiple climate elements from the large data set of historical observations of the earth system.

[0062] The Earth system historical observation big data set refers to a collection of long-term observation data of various components of the Earth system (such as the atmosphere, ocean, land surface, etc.). These data sets usually include a variety of meteorological and climate variables, such as temperature, precipitation, wind speed, sea level pressure, winter tropical outgoing longwave radiation (OLR), potential height field, etc. These data come from various observation platforms, such as ground meteorological stations, satellite remote sensing, radar, ocean buoys, etc.

[0063] The previous period refers to a period of preset duration before the period corresponding to the main components of multiple climate elements. For example, if the period corresponding to the main components of multiple climate elements is the summer of 1992-2021, then the previous period may refer to the winter period before 1992-2021.

[0064] Large-scale circulation field refers to the air flow pattern in the atmosphere on a global or regional scale, which includes air movement in the horizontal and vertical directions. The main components of large-scale circulation field include high-pressure system, low-pressure system, jet stream, monsoon system, etc.

[0065] It should be understood that the reason why the embodiment of the present application extracts the previous large-scale circulation field that affects the main components of multiple climate elements from the Earth system historical observation big data set after determining the main components of multiple climate elements of the predicted target climate elements is that the large-scale circulation field directly or indirectly controls multiple key variables in the climate system through dynamic processes, water vapor transport, temperature and humidity distribution, and climate phenomena in the atmosphere, thereby affecting the formation of climate. Therefore, when implementing climate prediction, the embodiment of the present application can achieve the prediction of the main components in the future by capturing the correlation between the previous large-scale circulation field of the main components of multiple climate elements and the main components.

[0066] The specific extraction process of the early large-scale circulation field is as follows.

[0067] In one possible implementation, based on the study of climate dynamics and climate predictability, the previous large-scale circulation fields that affect the main components of multiple climate elements are extracted from the large data set of historical observations of the Earth system, including: Based on the research on climate dynamics and climate predictability, the physical mechanism of variation of the predicted target climate elements is analyzed, and different early large-scale circulation fields are selected to determine the predicted target climate elements.

[0068] The physical mechanism of variation of the predicted target climate elements refers to the basic physical processes and interactions that cause specific climate elements (such as temperature, precipitation, wind speed, etc.) to change in time and space.

[0069] Optionally, the physical mechanisms of variation include atmospheric dynamics, ocean circulation, land surface processes, cryosphere changes, human activities, etc. Atmospheric dynamics include atmospheric circulation (such as Hadley cell, Ferrel cell, etc.), jet stream and wind system, pressure system, El Niño-Southern Oscillation (ENSO). Ocean circulation includes ocean heat transport, sea surface temperature (SST) changes, and ocean-atmosphere interactions. Land surface processes include soil temperature, vegetation cover, and ice and snow cover. Cryosphere changes include glaciers and ice sheets, Arctic sea ice. Human activities include greenhouse gas emissions, aerosol emissions, land use changes, etc.

[0070] Specifically, in the embodiments of the present application, the main physical mechanism affecting the predicted target climate element, i.e., the variation physical mechanism, can be identified with the help of relevant literature and existing research. Based on the physical mechanism analysis, the large-scale circulation field closely related to the predicted target climate element is selected, and then the large-scale circulation field data of the previous time period closely related to the predicted target climate element is extracted from the large data set of historical observations of the earth system.

[0071] For example, if the period corresponding to the predicted target climate element is the summer of 1992-2021, the early period can be the winter before each year from 1992 to 2021 (generally from December of the previous year to February of the next year). Assuming that the large-scale circulation field is SST, the embodiment of the present application can be based on the large data set of historical observations of the earth system, from which the SST data of the winter before each year from 1992 to 2021 can be extracted, that is, the early large-scale circulation field of the embodiment of the present application can be obtained.

[0072] In the above technical scheme, by analyzing the physical mechanism of variation of climate elements, the early large-scale circulation field is obtained, which can more accurately reflect the internal dynamic process of the climate system, deeply understand the key processes and interactions in the climate system, and provide a more accurate and reliable climate prediction scheme.

[0073] After obtaining the early large-scale circulation field, it can be further subjected to dimensionality reduction and feature extraction.

[0074] Similarly, before reducing the dimension and extracting features of the previous large-scale circulation field, the previous large-scale circulation field can be converted into anomaly data in the manner shown in formula (1) to obtain the anomaly data of the previous large-scale circulation field. Then, the anomaly data of the previous large-scale circulation field can be reduced in dimension and extracted features through the convolutional neural network (CNN) model.

[0075] Specifically, when the CNN model is used to reduce the dimension and extract features of the anomaly data of the previous large-scale circulation field, the structure of the CNN model includes a convolution layer and a pooling layer, and the anomaly data of the previous large-scale circulation field is input into the CNN model as input data. The anomaly data of the previous large-scale circulation field first enters the convolution layer of the CNN model. The convolution layer performs a convolution operation on the input data through a sliding convolution kernel to extract the local features of the input data. Each convolution kernel is responsible for detecting a specific type of pattern or feature. Nonlinearity is introduced through activation functions such as the Rectified Linear Unit (ReLU) to enhance the expression ability of the CNN model. By stacking multiple convolution layers, higher-level features can be gradually extracted.

[0076] After extracting the features of the input data, it immediately enters the pooling layer of the CNN model. The pooling layer reduces the spatial dimension of the features, reduces the amount of calculation, and retains the main features. After multiple layers of convolution and pooling operations, the feature map is usually multi-dimensional. In order to input it into the fully connected layer, these multi-dimensional feature maps need to be flattened into one-dimensional vectors. The flattened one-dimensional vector is the predictable circulation feature vector. Arranging the flattened feature vectors in chronological order can obtain a time series set of predictable circulation feature vectors, and the time series set of predictable circulation feature vectors is regarded as the actual prediction factor.

[0077] In the above technical solution, by reducing the dimension and extracting features of the early large-scale circulation field, the dimension of the input data can be significantly reduced, the computing resource requirements during model training and prediction can be reduced, noise and redundant information can be reduced, and the generalization ability of the model can be improved.

[0078] After obtaining the actual prediction target and the actual prediction factor, the embodiment of the present application can construct a nonlinear artificial intelligence fitting model based on these two types of data.

[0079] In one possible implementation, a recurrent neural network is used to select the historical climate data of the 30 years before the prediction target period as a training set, and a nonlinear artificial intelligence fitting model of the actual prediction factors and the actual prediction target is constructed, including: The actual prediction target and actual prediction factors are modeled using long short-term memory neural networks. The historical climate data for the 30 years before the prediction target period are selected as the training set to construct a nonlinear artificial intelligence fitting model of the actual prediction factors and the actual prediction target.

[0080] Since the actual prediction factors and actual prediction targets obtained above are both time series data, when constructing the nonlinear artificial intelligence fitting model of the actual prediction factors and the actual prediction targets, the Long Short-Term Memory Neural Network (LSTM) model used for time series processing in the recurrent neural network can be used to implement it.

[0081] Among them, in the time series of predictable circulation feature vectors extracted by the CNN model, the predictable circulation feature vectors of each time step are mapped to a hidden unit of the LSTM model (also called LSTM unit or LSTM neuron). Each hidden unit includes the LSTM's unique gating mechanism, namely the input gate, forget gate, cell state and output gate.

[0082] By mapping the features extracted by the CNN model to the hidden units of the LSTM model, the conversion of features from high-dimensional space to low-dimensional space can be achieved. Hidden units usually use activation functions and introduce nonlinearity, allowing the LSTM model to capture complex linear relationships in features. In addition, each hidden unit can learn different local features of the input data. In the LSTM model, the hidden unit contains not only information at the current moment, but also information at the past moments. The hidden unit can control the flow of information through the gating mechanism and learn the dynamic change pattern in the time series.

[0083] It should be understood that the output of the CNN model (actual prediction factor) is directly used as the input of the LSTM model. Therefore, the nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target constructed actually includes a CNN model for feature extraction and an LSTM model for prediction. In other words, for the nonlinear artificial intelligence fitting model including the CNN model and the LSTM model, its input data is the anomaly data of the previous large-scale circulation field, which is first subjected to dimensionality reduction and feature extraction by the CNN model to obtain the actual prediction factor, and then the actual prediction factor is used as the input data of the LSTM model. The output data of the LSTM model is the prediction results of the principal components of multiple climate elements.

[0084] During the training process, the actual prediction target (i.e., multiple climate element principal components) is the observed value of the LSTM model. The LSTM model outputs the predicted values ​​of multiple climate element principal components. The LSTM model is trained by comparing the difference between the predicted values ​​and the observed values. Based on the input parameters and output parameters and its unique gating mechanism, the LSTM model can learn the mapping relationship between the input parameters and the output data, so that it can output the prediction results of multiple climate element principal components during the training process.

[0085] Model construction mainly includes defining the model structure, setting hyperparameters, and compiling the model. Defining the model structure includes using Keras or other deep learning frameworks to define the structure of the LSTM model, which involves how many LSTM layers, fully connected layers, and appropriate activation functions the LSTM model includes. Compiling the model refers to setting the loss function, optimizer, and evaluation indicators of the LSTM model.

[0086] 104. Through the random initial perturbation scheme, random initialization parameter scheme, random cyclic weight scheme and pre-training method, multiple nonlinear artificial intelligence fitting models are trained for multiple main components of climate elements, and the ensemble prediction results of multiple main components of climate elements are obtained through the ensemble averaging scheme.

[0087] For the multiple main components of climate elements in the embodiments of the present application, in the embodiments of the present application, for each main component of the climate element, based on the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method, multiple nonlinear artificial intelligence fitting models (i.e., a combination model of multiple LSTM models and CNN models) can be trained to obtain multiple prediction results of the main component of the climate element, and then based on the ensemble averaging scheme, the ensemble prediction result corresponding to the main component of the climate element can be obtained.

[0088] In a possible implementation, a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method are used to train multiple nonlinear artificial intelligence fitting models for multiple climate element principal components, and an ensemble average scheme is used to obtain ensemble prediction results of multiple climate element principal components, including: Through random initial perturbation scheme, random initialization parameter scheme, random cyclic weight scheme and pre-training method, multiple nonlinear artificial intelligence fitting models are constructed for multiple main components of multiple climate elements. The actual training factors are trained and introduced to determine the multiple prediction results of the main components of multiple climate elements. Through the ensemble averaging scheme, the ensemble prediction results of the main components of multiple climate elements are obtained.

[0089] Specifically, in the embodiment of the present application, 1000 nonlinear artificial intelligence fitting models are trained for each principal component of the climate element. During the training process, the nonlinear artificial intelligence fitting model is subjected to random initial perturbations, random initialization parameters and random cyclic weights.

[0090] Random Initial Perturbation is a technique commonly used in numerical simulation, machine learning model initialization, and various scientific and engineering applications. It explores the diversity of system behavior by adding random noise to initial conditions or parameters.

[0091] In a possible implementation, the random initialization parameter scheme in the embodiment of the present application is specifically Glorot random initialization, and the random cyclic weight scheme is specifically orthogonal random cyclic weight.

[0092] Glorot random initialization is a method for initializing neural network weights. This method aims to keep the activation values ​​and gradients of each layer of the network within a suitable range during forward propagation and back propagation, thereby accelerating the training process and improving the convergence of the model.

[0093] Orthogonal random cyclic weights are a scheme for initializing neural network weights, especially for networks with cyclic connections such as recurrent neural network (RNN) models and LSTM models. This method helps solve the problems of gradient vanishing and gradient exploding by ensuring the orthogonality of the weight matrix to maintain stable gradient propagation.

[0094] In a possible implementation, through a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method, multiple nonlinear artificial intelligence fitting models are constructed for multiple climate element principal components, and multiple prediction results of multiple climate element principal components are determined by training and introducing actual training factors. Through an ensemble average scheme, the ensemble prediction results of multiple climate element principal components are obtained, including: The initial random disturbance is added to the training set through the random initial perturbation scheme, different initialization parameters are set for the long short-term memory neural network through the random initialization parameter scheme, the cyclic weights in the iterative cycle of the long short-term memory network are randomized through the random cyclic weight scheme, the learning rate and number of iterations of the long short-term memory network are selected through the training method, multiple nonlinear artificial intelligence fitting models are constructed, and multiple prediction results are determined based on the actual training factors after training. The ensemble prediction results of the main components of multiple climate elements are obtained through the ensemble averaging scheme.

[0095] When training an artificial intelligence ensemble prediction model, random initial perturbations can be added to the input data during the feature extraction stage, so that the modes extracted by different CNN models are disturbed.

[0096] Specifically, we first define a random noise distribution, generate random disturbances for each input sample, and add the generated random disturbances to the input parameters (anomaly data of the previous large-scale circulation field).

[0097] Different initialization parameters can be set for the LSTM model by setting different initializers. In addition, during the LSTM model iteration cycle, the loop weights are randomly set. For the learning rate and number of iterations of the LSTM model, different combinations of learning rate and number of iterations can be tried through grid search or random search (for example, defining a function) to obtain the learning rate and number of iterations of the LSTM model.

[0098] Through the above steps, multiple nonlinear artificial intelligence fitting models with different initial disturbances, initialization parameters, loop weights and training hyperparameters can be generated.

[0099] For any principal component of a climate element, when training multiple nonlinear artificial intelligence fitting models corresponding to the principal component of the climate element, the actual training factors are used as the input data of the LSTM model, and the LSTM model is trained with the actual training factors and the principal component of the climate element, so that the LSTM model outputs the prediction results of the principal component of the climate element. 1000 prediction results of 1000 principal components of the climate element can be output by 1000 LSTM models.

[0100] The 1000 prediction results of the main component of the climate element are averaged through the ensemble average scheme to obtain the ensemble prediction result of the main component of the climate element. The ensemble average specifically refers to the weighted average of the 1000 prediction results of the main component of the climate element.

[0101] Thus, by performing the above operations on each of the above climate element principal components, 1000 prediction results of each climate element principal component are obtained, and by ensemble averaging, the ensemble prediction result of each climate element principal component is obtained.

[0102] In order to facilitate the understanding of the determination process of the ensemble prediction results of each climate element principal component, the following Figure 2 The above process is introduced in detail.

[0103] Figure 2 It is a schematic diagram of a scenario for determining the collective prediction result of the principal component of any climate element provided in an embodiment of the present application.

[0104] For example, Figure 2 As shown, it is assumed that any one of the principal components of the climate element is principal component 1, and the predicted target climate element is precipitation. When performing spatiotemporal decomposition of precipitation, the precipitation data is first converted into anomaly data, and then spatiotemporal decomposition is performed through MVMD to obtain four principal components of the climate element, namely principal component 1, principal component 2, principal component 3 and principal component 4. In an embodiment of the present application, the number of multiple nonlinear artificial intelligence fitting models corresponding to each principal component is 1000, including 1000 CNN models and 1000 LSTM models.

[0105] For the principal component 1, the actual prediction factor is obtained by reducing the dimension and extracting the features of the previous large-scale circulation field corresponding to the principal components of the above four climate elements, that is, Figure 2 The predictor elements in .

[0106] The predictor elements are converted into anomaly data, that is, Figure 2 The prediction factor elements in are input as the input data of 1000 nonlinear artificial intelligence fitting models. 1000 CNN models are used to reduce the dimension and extract features to obtain a time series set of predictable circulation feature vectors, which are used as the input of each LSTM unit in the LSTM model. 1000 LSTM models are used to obtain the prediction results of 1000 principal components 1.

[0107] Specifically, the 1000 nonlinear AI fitting models include 1000 CNN models and 1000 LSTM models. The 1000 CNN models are Figure 2 Convolutional neural network 1, convolutional neural network 2, convolutional neural network ..., convolutional neural network 999 and convolutional neural network 1000 in. The 1000 LSTM models are Figure 2 LSTM neural network 1, LSTM neural network 2, LSTM neural network..., LSTM neural network 999 and LSTM neural network 1000.

[0108] After the input data is processed by 1000 CNN models and 1000 LSTM models, each set of CNN models and LSTM models can obtain a prediction result for principal component 1. Therefore, 1000 sets of CNN models and LSTM models can process 1000 prediction results for principal component 1, that is, Figure 2 Prediction 1, Prediction 2, Prediction..., Prediction 999 and Prediction 1000.

[0109] After obtaining 1000 prediction results of principal component 1, the ensemble prediction results of principal component 1 are obtained by ensemble averaging, that is, Figure 2 The principal component 1 prediction in .

[0110] 105. Based on the ensemble prediction results of the principal components of multiple climate elements, the complete prediction results of the target climate element are reconstructed through the principal component synthesis method.

[0111] After obtaining the ensemble prediction results of the principal components of different climate elements, in order to restore the prediction results, it is necessary to further reconstruct the prediction results to obtain the complete prediction results of the predicted target climate elements.

[0112] In a possible implementation, based on the ensemble prediction results of the principal components of different climate elements, the complete prediction results of the target climate element are reconstructed through the principal component synthesis method, including: Through the inverse operation of the multivariate variational mode decomposition method, based on the ensemble prediction results of the principal components of different climate elements, the complete prediction results of the target climate elements are reconstructed through the principal component synthesis method.

[0113] Specifically, after obtaining the collective prediction results of each climate element principal component, since the aforementioned multiple climate element principal components are obtained by decomposition through the MVMD method, in the embodiment of the present application, the collective prediction results of the principal components of each climate element can be resynthesized into the prediction results of the complete prediction target climate element based on the inverse operation of the MVMD method.

[0114] To facilitate understanding of the implementation process of the embodiments of the present application, the following takes the prediction of summer precipitation in country A in 2022 as an embodiment of the present application, and introduces in detail the climate ensemble prediction method based on artificial intelligence technology provided by the present application.

[0115] Figure 3 This is a schematic diagram of a scenario for performing spatiotemporal decomposition of precipitation at site B in country A, provided in an embodiment of the present application.

[0116] For example, Figure 3 As shown, the forecast target period is the summer of 2022, and the forecasted climate element is precipitation. Taking site B in country A as an example, when constructing and training the model provided in the embodiment of the present application, the summer precipitation of site B in the 30 years before 2022, that is, the summer precipitation of site B from 1992 to 2021, can be collected.

[0117] After obtaining the summer precipitation of station B 30 years before the above-mentioned prediction target period, the summer precipitation of station B in the above-mentioned 30 years is first converted into anomaly data, and then the anomaly data is decomposed in time and space using the MVMD method, and the precipitation anomaly data is decomposed into 4 main modes to obtain multiple principal components. In the embodiment of the present application, the number of the multiple principal components is 4.

[0118] For example, in the embodiment of the present application, the decomposition result of the anomaly data of summer precipitation at station B is as follows: Figure 3 As shown, it includes Mode 1, Mode 2, Mode 3 and Mode 4, which correspond to the first principal component, the second principal component, the third principal component and the fourth principal component, respectively, and correspond to the actual prediction targets in the embodiments of the present application.

[0119] exist Figure 3 In the four figures corresponding to the decomposition results of the anomaly data of summer precipitation at station B in , the horizontal axis represents the year, and the vertical axis represents the principal components under different modes of the anomaly data.

[0120] Depend on Figure 3 It can be seen that the first principal component of the anomaly data of summer precipitation at station B has gradually increased from negative to positive since 1992, and reached a peak in 2020, corresponding to the record-breaking summer precipitation at station B in 2020, which is rainy. The second principal component of the anomaly data of summer precipitation at station B turned from negative to positive after 2000. Related studies have shown that the position of the main summer rain belt shifted around 2000, and summer precipitation in city B increased. The third principal component sequence of the anomaly data of summer precipitation at station B is basically the same as the change trend of the Nino3.4 index, indicating that summer precipitation in city B is modulated by El Niño. The fourth principal component of the anomaly data of summer precipitation at station B shows that there have been two positive anomaly periods in precipitation in city B in the past 30 years, around 1998 and around 2016.

[0121] In addition, if Figure 3 As shown, in the embodiment of the present application, the anomaly data of summer precipitation at station B can also be reconstructed based on the inverse operation of the MVMD method through the above four principal components to obtain a reconstruction graph of the anomaly data. Figure 3 The horizontal axis of the anomaly data reconstruction graph represents the year, and the vertical axis represents the precipitation anomaly (PRCano). The data corresponding to "Original" is the original anomaly data of summer precipitation at station B, that is, the anomaly data of summer precipitation at station B before the aforementioned time-space decomposition. The data corresponding to "Reconstructed" is the anomaly data of summer precipitation at station B after reconstruction based on the four principal components. Figure 3 It can be seen that the reconstructed anomaly data of summer precipitation at site B is highly consistent with the original anomaly data of summer precipitation at site B, indicating that the principal components extracted by the spatiotemporal decomposition in the embodiment of the present application can well represent the main features of the original anomaly data, verifying the accuracy and reliability of the spatiotemporal decomposition method.

[0122] In addition to obtaining the four principal components of the anomaly data of summer precipitation at station B from 1992 to 2021 through spatiotemporal decomposition, the embodiment of the present application can also select the outgoing longwave radiation (OLR) in the tropical region in the early winter of each year from 1992 to 2021 and the 500hPa potential height field (Z500) in the mid- and high-latitude regions of the Northern Hemisphere from the Earth system historical observation big data set according to the study of climate dynamics and climate predictability. For example, for 1992, the corresponding early large-scale circulation field data are OLR and Z500 from December 1991 to February 1992.

[0123] OLR refers to the longwave radiation emitted by the Earth's surface and atmosphere to space, which reflects the heat emitted by the Earth system to space. OLR can usually be observed through satellite remote sensing technology.

[0124] The 500hPa potential height field in the mid- and high-latitude regions of the Northern Hemisphere refers to the distribution of potential heights on the 500hPa isobaric surface of the atmosphere in the mid- and high-latitude regions of the Northern Hemisphere (usually 30°N to 60°N or higher dimensions).

[0125] After collecting the OLR and Z500 for each year from 1992 to 2021, similarly, in the embodiment of the present application, the above two types of early large-scale circulation fields can be converted into anomaly data in the manner shown in formula (1) to obtain the anomaly data of the early large-scale circulation field.

[0126] refer to Figure 2 In the manner shown, the electronic device can be based on Figure 2 The construction process of the nonlinear artificial intelligence fitting model shown in Figure 1 constructs 1,000 training members. The main processing flow of the nonlinear artificial intelligence fitting model includes the input of the anomaly data of the prediction factor element - CNN model - LSTM model - prediction model output. Among them, the prediction factor element is the previous large-scale circulation field mentioned above. Among them, each training member will go through the above processing flow.

[0127] Figure 4 It is a structural schematic diagram of a nonlinear artificial intelligence fitting model corresponding to a single training member provided in an embodiment of the present application.

[0128] For example, Figure 4As shown in the figure, the CNN model includes a convolution layer and a pooling layer. In the CNN model processing part, the deviation data of the prediction factor elements are first input into the CNN model, and the convolution operation is performed through the convolution kernel, and then the average pooling operation is performed through the pooling layer to obtain the characteristic value of the prediction factor element. The characteristic value of the prediction factor is further mapped to different hidden units to complete the dimension reduction and feature extraction process of the prediction factor element. In this part, random initial perturbations are added to the input deviation data through a random initial perturbation scheme, so that the modes extracted by different CNN models are perturbed.

[0129] Next, we enter the LSTM model, which mainly includes the input gate, forget gate, cell candidate and output gate. During the iteration of the LSTM model, the "memory" signal is transmitted in the LSTM model by passing the implicit state and cell state, thereby realizing nonlinear time series fitting and prediction. During the iteration process, the LSTM model has three main parameters to learn, namely input weight, loop weight and bias value. Therefore, to realize multi-member set prediction based on the LSTM model, different solutions can be introduced for the above three parameters.

[0130] For example, Figure 4 As shown in the figure, for each LSTM model corresponding to a training member, the LSTM model uses the Glorot random initialization parameter scheme to input different initial parameters to the LSTM model. The core of this scheme is to construct an initial loop weight array with a standard normal distribution and input it into the network to start the first round of iteration; the LSTM model uses the orthogonal random loop weight scheme to randomize the loop weights between each loop of the LSTM model. The core of this scheme is to design a random 0-1 orthogonal matrix with the same size as the number of hidden units as the loop weight. In each loop, different hidden units may be used, so that the loop itself is random.

[0131] For the LSTM model, the number of hidden units, learning rate, and number of iterations are hyperparameters for model training. The above three hyperparameters can be obtained through the pre-training scheme. That is, according to the designed scheme, different numbers of hidden units and learning rates are set, and the training set is freely trained. It is considered that the loss of model training reaches a value less than the threshold for 30 rounds, that is, the model tends to converge. According to the pre-training and convergence of the model, the appropriate number of hidden units, learning rate, and number of iterations are recorded as hyperparameters for all member training. Among them, setting a fixed number of training iterations can achieve early stopping of iterations in subsequent training, thereby obtaining different deviation values.

[0132] In order to prevent the model from overfitting, the LSTM model introduces the L2 regularization scheme to correct the training model. While improving the generalization of the model, it also increases the randomness of the model. Figure 1 The fitting model design process shown in 103.

[0133] For example, Figure 2 As shown, when using the method of the embodiment of the present application to predict whether the summer precipitation at site B in 2022 is abnormal, the summer precipitation at site B is decomposed into four principal components using the historical observation data set from 1992 to 2021. The four principal components are as follows: Figure 3 shown.

[0134] Further integration Figure 4 ,When realizing the prediction, 1000 training members were designed, and the anomaly data of OLR and North Z500 from 1992 to the early ,2021 were input respectively, and dimension reduction and feature extraction operations were ,performed on them, and then nonlinear fitting training was performed on the actual prediction factors and ,actual prediction targets until the model training was successful.

[0135] After the model training is successful, during the model application process, when it is necessary to predict the summer precipitation at site B in 2022, the OLR corresponding to December 2021-February 2022 and the anomaly data of 500hPa in the middle and high dimensional regions of the Northern Hemisphere can be used as the actual prediction factors of the model in the application stage, thereby obtaining 1000 prediction results of the principal component 1 of the summer precipitation at site B in the summer of 2022, and then obtaining the ensemble prediction results of the principal component 1 through the ensemble averaging scheme.

[0136] Similarly, for the other three principal components, the OLR corresponding to December 2021-February 2022 and the anomaly data of 500hPa in the middle and high dimensional regions of the Northern Hemisphere are used as actual prediction factors, and 1000 prediction results of the principal component n (n=2, 3, 4) of the summer precipitation at station B in the summer of 2022 are obtained, and then the ensemble prediction results of the principal component n are obtained through the ensemble averaging scheme.

[0137] After obtaining the ensemble prediction results of the four principal components, the electronic device reconstructs the true prediction results of the precipitation at site B in the summer of 2022 based on the inverse operation of MVMD, thereby obtaining an artificial intelligence ensemble prediction of the summer precipitation at site B in the summer of 2022.

[0138] Related research points out that ensemble prediction can give more stable prediction results, so that the expected prediction results are often higher than the actual prediction results. In order to examine how ensemble prediction can quantitatively describe the uncertainty of prediction results, thereby reducing prediction errors and improving the prediction effect, the embodiment of the present application also provides an evaluation scheme.

[0139] Figure 5 It is a schematic diagram of a scenario for evaluating a multi-member set prediction result provided in an embodiment of the present application.

[0140] For example, Figure 5 As shown in the figure, after obtaining the 1000 prediction member results corresponding to each principal component, the ensemble averaging is not performed. Instead, the prediction results of 1000 anomaly data of summer precipitation at station B in 2022 are reconstructed by random combination reconstruction method. Then the above 1000 prediction results are randomly grouped and ensemble averaging is performed iteratively, that is, random ensemble averaging of one group and random ensemble averaging of two groups until all 1000 members are averaged. Then the ensemble average prediction results of each iteration are compared with the actual observation results, and evaluated by the anomaly correlation coefficient (Anomaly Correlation Coefficient, ACC), mean squared error (Mean Squared Error, MSE), symbol accuracy (Accuracy Symbols, AS) and difference (DELTA).

[0141] exist Figure 5 When the above different evaluation indicators are used to evaluate the ensemble prediction results, the horizontal axis is the number of samples, and the vertical axis is the value of ACC, MSE, AS and DELTA. The unit of the MSE evaluation indicator is ((mm / day) 2 ,(mm / day) 2 ), the unit of DELTA evaluation index is (mm / day).

[0142] Among them, ACC is used to measure the linear relationship between the forecast value and the observed value, and its calculation formula is shown in the following formula (2).

[0143] Formula (2) In formula (2): F i : No. i The forecast value at a time point; O i : No. i Observed values ​​at each time point; : The average of all forecast values; : The average of all observations; n : Sample size.

[0144] The value range of ACC is between -1 and 1. ACC=1 means that the forecast value is completely positively correlated with the observed value; ACC=0 means that there is no linear relationship between the forecast value and the observed value; ACC=-1 means that the forecast value is completely negatively correlated with the observed value.

[0145] MSE measures the difference between the predicted value and the actual observed value. It calculates the average of the squares of the differences between all predicted values ​​and the actual values. The calculation formula is shown in the following formula (3).

[0146] Formula (3) In formula (3): y i : No. i The actual observed value of samples; : No. i The predicted value of samples; n : Sample size.

[0147] The value range of MSE is from 0 to positive infinity. When MSE=0, it means that all predicted values ​​are completely equal to the actual observed values; the larger the MSE, the larger the prediction error.

[0148] AS: It is used to measure the proportion of correct predictions in classification tasks or symbol recognition tasks, which refers to the proportion of correctly classified or recognized samples to the total number of samples.

[0149] DELTA: An important indicator for evaluating the difference between prediction and observation. Its calculation formula is shown in the following formula (4).

[0150] Formula (4) In formula (4): y i : No. i The actual observed value of samples; : No. i The predicted value of a sample.

[0151] like Figure 5 As shown, it can be seen that the ensemble prediction scheme can give a variety of random prediction results, thereby quantitatively describing the uncertainty of the prediction results. Through the ensemble average scheme, when the number of average samples increases, that is, Figure 5 When the prediction result is close to 1000, regardless of the evaluation index, the result of the ensemble prediction reaches the peak of the normal distribution, indicating that the result of the ensemble prediction tends to be stable and remains at a good prediction level overall. Therefore, the climate ensemble prediction scheme based on artificial intelligence technology can improve the climate prediction ability.

[0152] In order to further illustrate the accuracy of the prediction results of the embodiment of the present application, the embodiment of the present application uses the anomaly data of summer precipitation from 1993 to 2016 at multiple stations for prediction, and compares it with the prediction results obtained using the dynamic model in the same time period.

[0153] Figure 6 This is a schematic diagram of a scenario for comparing the prediction effects of different prediction methods provided in an embodiment of the present application.

[0154] For example, Figure 6 As shown in Figure 1, the winter circulation elements were used as prediction factors to report the summer precipitation anomaly, and compared with the prediction results reported by business units A and B starting from March each year. The evaluation index is the prediction skill (PS) score. The evaluation results are shown in Figure 1. Figure 6 shown. Figure 6 In each prediction result, the horizontal axis is the year and the vertical axis is the PS score.

[0155] It can be found that the ensemble prediction (ENS) model designed in the present invention has an average prediction score of 74.75 points for the summer precipitation anomaly from 1994 to 2016, which is significantly higher than 66.71 points of business unit A and 69.58 points of business unit B, indicating that the climate ensemble prediction method based on artificial intelligence technology and its prediction model proposed in this application can effectively improve the effect of precipitation prediction.

[0156] Figure 7 It is a structural diagram of a climate ensemble prediction system based on artificial intelligence technology provided in an embodiment of the present application.

[0157] For example, Figure 7 As shown, the system 700 includes: The spatiotemporal decomposition module 701 is used to perform spatiotemporal decomposition of the predicted target climate element to obtain a plurality of climate element principal components with different spatiotemporal characteristics, and use the plurality of climate element principal components as actual prediction targets; Feature extraction module 702, for extracting the previous large-scale circulation field that affects the main components of the multiple climate elements from the large data set of historical observations of the earth system based on the study of climate dynamics and climate predictability, performing dimension reduction and feature extraction on the previous large-scale circulation field through a convolutional neural network, obtaining a time series set of predictable circulation feature vectors, and using the time series set of predictable circulation feature vectors as an actual prediction factor; The fitting model building module 703 is used to use a recurrent neural network to select the historical climate data of the 30 years before the prediction target period as a training set to build a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target; The ensemble prediction result determination module 704 is used to train multiple nonlinear artificial intelligence fitting models for the multiple climate element principal components through a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method, and obtain the ensemble prediction results of the multiple climate element principal components through an ensemble averaging scheme; The prediction result reconstruction module 705 is used to reconstruct the complete prediction result of the prediction target climate element through the principal component synthesis method based on the collective prediction results of the plurality of climate element principal components.

[0158] In a possible implementation, the spatiotemporal decomposition module 701 is specifically used to decompose the time series of the predicted target climate element from the time dimension and the space dimension through a multivariate variational mode decomposition method, and extract the principal components of the multiple climate elements.

[0159] In one possible implementation, the feature extraction module 702 is specifically used to: analyze the physical mechanism of variation of the predicted target climate element based on the study of climate dynamics and climate predictability, and select different early large-scale circulation fields for determining the predicted target climate element.

[0160] In one possible implementation, the fitting model construction module 703 is specifically used to: use a long short-term memory neural network to model the actual prediction target and the actual prediction factor, select the historical climate data of the 30 years before the prediction target period as a training set, and construct a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target.

[0161] In one possible implementation, the collective prediction result determination module 704 is specifically used to: construct the multiple nonlinear artificial intelligence fitting models for the multiple climate element main components through the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method, train and introduce the actual training factors to determine the multiple prediction results of the multiple climate element main components, and obtain the collective prediction results of the multiple climate element main components through the collective averaging scheme.

[0162] In one possible implementation, the ensemble prediction result determination module 704 is also used to: add initial random disturbance to the training set through the random initial disturbance scheme, set different initialization parameters for the long short-term memory neural network through the random initialization parameter scheme, randomly set the cyclic weights in the iterative cycle process of the long short-term memory network through the random cyclic weight scheme, select the learning rate and number of iterations of the long short-term memory network through the training method, construct the multiple nonlinear artificial intelligence fitting models, train and determine the multiple prediction results based on the actual training factors, and obtain the ensemble prediction results of the main components of the multiple climate elements through the ensemble averaging scheme.

[0163] In one possible implementation, the prediction result reconstruction module 705 is specifically used to reconstruct the complete prediction result of the predicted target climate element based on the collective prediction results of the principal components of the multiple climate elements through the inverse operation of the multivariate variational mode decomposition method and through the principal component synthesis method.

[0164] Figure 8 It is a structural diagram of a computer system provided in an embodiment of the present application.

[0165] For example, Figure 8 As shown, the computer system 800 includes: a memory 801 and a processor 802, wherein the memory 801 stores a computer program 8011, and the processor 802 is used to call and execute the computer program 8011 to perform a climate ensemble prediction method based on artificial intelligence technology.

[0166] In this embodiment, the computer system can be divided into functional modules 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 this embodiment is schematic and is only a logical function division. There may be other division methods in actual implementation.

[0167] In the case of dividing each functional module according to each function, the computer system may include: a spatiotemporal decomposition module, a feature extraction module, a fitting model construction module, a set prediction result determination module, and a prediction result reconstruction 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, and will not be repeated here.

[0168] The computer system provided in this embodiment is used to execute the above-mentioned climate ensemble prediction method based on artificial intelligence technology, and thus can achieve the same effect as the above-mentioned implementation method.

[0169] In the case of an integrated unit, the computer system may include a processing module and a storage module. The processing module may be used to control and manage the actions of the computer system. The storage module may be used to support the computer system in executing related program codes and data.

[0170] The processing module may be a processor or a controller, which may implement or execute various exemplary logic blocks, modules and circuits shown in combination with the contents disclosed in this 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.

[0171] This embodiment also provides a computer-readable storage medium, in which a computer program code is stored. When the computer program code is executed on a computer, the computer executes the above-mentioned related method steps to implement a climate ensemble prediction method based on artificial intelligence technology in the above-mentioned embodiment.

[0172] 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 a climate ensemble prediction method based on artificial intelligence technology in the above-mentioned embodiment.

[0173] In addition, the computer system provided in the embodiments of the present application may specifically be a chip, component or module, and the computer system may include a connected processor and memory; wherein the memory is used to store instructions, and when the computer system is running, the processor may call and execute instructions so that the chip executes a climate ensemble prediction method based on artificial intelligence technology in the above-mentioned embodiments.

[0174] Among them, the computer system, computer-readable storage medium, computer program product or chip provided in this embodiment are all used to execute the corresponding methods provided above. Therefore, the beneficial effects that can be achieved can refer to the beneficial effects in the corresponding methods provided above and will not be repeated here.

[0175] 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.

[0176] 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.

[0177] 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

1. A climate ensemble prediction method based on artificial intelligence technology, characterized in that: The method comprises: Decomposing the predicted target climate element in time and space to obtain a plurality of climate element principal components with different time and space characteristics, and taking the plurality of climate element principal components as actual prediction targets; Based on the study of climate dynamics and climate predictability, the previous large-scale circulation field that affects the main components of the multiple climate elements is extracted from the large data set of historical observations of the earth system. The previous large-scale circulation field is subjected to dimensionality reduction and feature extraction through a convolutional neural network to obtain a time series set of predictable circulation feature vectors. The time series set of predictable circulation feature vectors is used as an actual prediction factor. Using a recurrent neural network, selecting the historical climate data of 30 years before the prediction target period as a training set, and constructing a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target; Through a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method, multiple nonlinear artificial intelligence fitting models are trained for the multiple climate element principal components, and through an ensemble averaging scheme, ensemble prediction results of the multiple climate element principal components are obtained; Based on the ensemble prediction results of the principal components of the plurality of climate elements, a complete prediction result of the prediction target climate element is reconstructed through a principal component synthesis method.

2. The method according to claim 1, characterized in that The predicted target climate element is subjected to temporal and spatial decomposition to obtain a plurality of climate element principal components with different temporal and spatial characteristics, including: By using the multivariate variational mode decomposition method, the time series of the predicted target climate element is decomposed from the time dimension and the space dimension, and the principal components of the multiple climate elements are extracted.

3. The method according to claim 1, characterized in that Based on the study of climate dynamics and climate predictability, the previous large-scale circulation field that affects the main components of the multiple climate elements is extracted from the large data set of historical observations of the earth system, including: Based on the study of climate dynamics and climate predictability, the physical mechanism of variation of the predicted target climate element is analyzed, and different early large-scale circulation fields are selected to determine the predicted target climate element.

4. The method according to claim 1, characterized in that: The method uses a recurrent neural network to select historical climate data for 30 years before the prediction target period as a training set, and constructs a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target, including: The actual prediction target and the actual prediction factor are modeled using a long short-term memory neural network, the historical climate data for 30 years before the prediction target period is selected as a training set, and a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target is constructed.

5. The method according to claim 4, characterized in that The method trains multiple nonlinear artificial intelligence fitting models for the multiple climate element principal components through a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method, and obtains the ensemble prediction results of the multiple climate element principal components through an ensemble averaging scheme, including: Through the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method, the multiple nonlinear artificial intelligence fitting models are constructed for the multiple climate element principal components, and the actual training factors are trained and introduced to determine the multiple prediction results of the multiple climate element principal components. Through the ensemble averaging scheme, the ensemble prediction results of the multiple climate element principal components are obtained.

6. The method according to claim 5, characterized in that The method comprises: constructing the plurality of nonlinear artificial intelligence fitting models for the plurality of climate element principal components through the random initial perturbation scheme, the random initialization parameter scheme, the random cyclic weight scheme and the pre-training method, training and introducing the actual training factors to determine the plurality of prediction results of the plurality of climate element principal components, and obtaining the ensemble prediction results of the plurality of climate element principal components through the ensemble averaging scheme, including: The random initial disturbance scheme is used to add initial random disturbance to the training set, different initialization parameters are set for the long short-term memory neural network through the random initialization parameter scheme, the random cyclic weight scheme is used to randomly set the cyclic weights in the iterative cycle of the long short-term memory network, the learning rate and the number of iterations of the long short-term memory network are selected through the training method, the multiple nonlinear artificial intelligence fitting models are constructed, the multiple prediction results are determined based on the actual training factors after training, and the ensemble prediction results of the principal components of the multiple climate elements are obtained through the ensemble averaging scheme.

7. The method according to claim 2, characterized in that The method of reconstructing the complete prediction result of the prediction target climate element based on the ensemble prediction results of the principal components of the plurality of climate elements by a principal component synthesis method comprises: Through the inverse operation of the multivariate variational mode decomposition method, based on the collective prediction results of the principal components of the multiple climate elements, the complete prediction results of the predicted target climate elements are reconstructed through the principal component synthesis method.

8. A climate ensemble prediction method based on artificial intelligence technology, characterized in that: The method comprises: Extracting the target pre-phase large-scale circulation field that affects the principal components of multiple target climate elements from the Earth system historical observation big data set, wherein the principal components of multiple target climate elements are obtained by temporal and spatial decomposition of the target climate elements to be predicted; Inputting the target early large-scale circulation field into a plurality of nonlinear artificial intelligence ensemble prediction models to obtain a plurality of prediction results for each of the plurality of target climate element principal components, wherein the plurality of nonlinear artificial intelligence ensemble prediction models are constructed based on the method described in any one of claims 1 to 7; Based on the ensemble averaging scheme, ensemble averaging the multiple prediction results of the principal components of each target climate element to obtain the ensemble prediction results of the principal components of the multiple target climate elements; Based on the collective prediction results of the multiple target climate elements, a complete prediction result of the target climate element to be predicted is obtained.

9. A climate ensemble prediction system based on artificial intelligence technology, characterized in that: The system comprises: A spatiotemporal decomposition module is used to perform spatiotemporal decomposition of the predicted target climate element to obtain a plurality of climate element principal components with different spatiotemporal characteristics, and use the plurality of climate element principal components as actual prediction targets; A feature extraction module is used to extract the previous large-scale circulation field that affects the main components of the multiple climate elements from the large data set of historical observations of the earth system based on the study of climate dynamics and climate predictability, perform dimension reduction and feature extraction on the previous large-scale circulation field through a convolutional neural network, obtain a time series set of predictable circulation feature vectors, and use the time series set of predictable circulation feature vectors as an actual prediction factor; A fitting model building module is used to use a recurrent neural network to select historical climate data for the 30 years before the prediction target period as a training set to build a nonlinear artificial intelligence fitting model of the actual prediction factor and the actual prediction target; An ensemble prediction result determination module is used to train multiple nonlinear artificial intelligence fitting models for the multiple climate element principal components through a random initial perturbation scheme, a random initialization parameter scheme, a random cyclic weight scheme and a pre-training method, and obtain the ensemble prediction results of the multiple climate element principal components through an ensemble averaging scheme; The prediction result reconstruction module is used to reconstruct the complete prediction result of the prediction target climate element based on the collective prediction results of the principal components of the plurality of climate elements through a principal component synthesis method.

10. A computer system, characterized in that: The invention comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the computer program is loaded into the processor, the climate ensemble prediction method based on artificial intelligence technology according to any one of claims 1 to 7 or claim 8 is implemented.

Images