Method, device and medium for climate prediction based on warm start and background constraints

Abstract

The application discloses a climate prediction method and device based on a hot start and background constraint, and a medium. The method comprises the following steps: training a self-encoder by using actual climate data of a first historical annual interval, and then decoding the self-encoder as an initial climate prediction model decoder; obtaining first climate data of a current annual interval and a second historical annual interval predicted by a specified climate prediction model; for each month in the second historical annual interval: taking the first climate data of the month and a set number of months before the month as input features, and taking actual climate data of the month as a label, to generate a sample pair of the month; training the climate prediction model trained in the last year by using a training sample library, with the joint loss being minimized as the target, wherein the joint loss comprises an accuracy constraint representing the prediction performance of the model, a climate background constraint and a prediction score constraint; and the trained model is used for predicting the climate of the current year. The application realizes continuous evolution and stable prediction of the prediction model, and improves the convergence speed of the prediction model.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and climate prediction, and more specifically, relates to a climate prediction method, device and medium based on hot start and background constraints. Background Technology

[0002] Climate prediction is a crucial decision-making tool for disaster prevention and mitigation, water resource management, and agricultural production planning. Deep learning technology, with its powerful nonlinear fitting capabilities, offers new insights into climate prediction. However, climate systems are non-stationary, and weather forecasting data typically has small sample sizes. Traditional deep learning downscaling methods often suffer from the following limitations.

[0003] On the one hand, model updates are lagging: traditional deep learning methods retain unchanged model parameters after annual training, failing to evolve using newly added observational data each year. Training from scratch annually (cold start) wastes computational resources and may lead to unstable local optima due to scarce sample size. On the other hand, long-term climate patterns are missing: existing models often train using paired data from recent decades (e.g., 2001-2020), ignoring earlier (e.g., 1951-2000) long-sequence observational data, which, while lacking corresponding dynamical patterns, contains valuable climatological distribution information. This results in insufficient predictive ability for extreme weather events (e.g., once-in-a-century rainstorms). Furthermore, physical consistency is insufficient: simply minimizing the root mean square error (RMSE) can lead to smoothed predictions, losing the extreme distribution characteristics of elements like precipitation. Therefore, a method is urgently needed that can continuously integrate newly added observational data each year and ensure the stability and timeliness of the prediction model through smooth parameter updates. Summary of the Invention

[0004] The main objective of this invention is to provide a climate prediction method, device, and medium based on hot start and background constraints, so as to overcome the shortcomings of the prior art.

[0005] The first aspect of this invention provides a climate prediction method based on hot start and background constraints, comprising: training an autoencoder using actual climate data from a first historical year interval to enable the autoencoder to learn the dependencies between sites, and using the weights of the trained autoencoder as the initial weights or regularization terms of the initial climate prediction model decoder; acquiring first climate data for the current year and a second historical year interval predicted by a specified climate prediction model, wherein the first historical year interval is earlier than the second historical year interval; for each month in the second historical year interval: generating sample pairs for that month using the first climate data of that month and a predetermined number of previous months as input features and the actual climate data of that month as labels; combining the sample pairs of each month to form a training sample library; training a climate prediction model trained in the previous year using the training sample library with the goal of minimizing the joint loss, wherein the joint loss includes accuracy constraints, climate background constraints, and prediction score constraints characterizing the model's prediction performance; for any month to be predicted in the current year: inputting the first climate data of the month to be predicted and a predetermined number of previous months into the trained climate prediction model, and outputting the climate prediction result for the month to be predicted.

[0006] Preferably, the joint loss is:

[0007] ;

[0008] in, For the joint loss; This is the first hyperparameter; This is the second hyperparameter; and The orders of magnitude of the accuracy constraint, the climate background constraint, and the prediction score constraint are used to control them. The accuracy constraint characterizes the error between the predicted value and the actual value; As a constraint of the climate background, it characterizes whether the fluctuation range and extreme value probability of the predicted value conform to the climate patterns of the first and second historical year intervals of the station; The prediction score constraint characterizes the degree of closeness between the predicted value and the actual value.

[0009] Preferably, the climate background constraint is:

[0010] ;

[0011] Alternatively, the aforementioned climate background constraint may be:

[0012] ;

[0013] in, Constrained by the aforementioned climate background; Let KL divergence function be used. The Wasserstein distance function; This represents the distribution of the current batch of prediction results from the climate prediction model. This represents the cumulative distribution or probability density of actual climate data for the first historical year interval.

[0014] Preferably, the prediction score constraint is:

[0015] ;

[0016] in, For the predicted score constraints, The total number of observation stations, This represents the number of observation stations where the predicted percentage anomaly and the actual percentage anomaly have the same sign. The absolute values ​​of both the predicted and actual anomaly percentages are within [a certain range]. The number of observation stations within the interval, The absolute values ​​of both the predicted and actual anomaly percentages are within [a certain range]. The number of observation stations within the interval, For the absolute value of the predicted value's percentage anomaly to be no less than 1 and the absolute value of the actual value's percentage anomaly to be within 1, the following conditions must be met: The number of observation stations within the interval, The first preset parameter, This is the second preset parameter. , where a is the first coefficient, b is the second coefficient, and c is the third coefficient.

[0017] Preferably, if there is no climate prediction model trained in the previous year, the initial climate prediction model is used as the climate prediction model trained in the previous year.

[0018] Preferably, obtaining the first climate data for the current year and the second historical year interval predicted by the specified climate system model specifically includes: obtaining the data of each set member under the current year and the second historical year interval predicted by the specified climate system model; for each year: performing a set average on the data of each set member under that year to obtain the climate factor field for that year, interpolating and resampling the climate factor field to a preset standard spatial grid to generate the first climate data for that year.

[0019] Preferably, the climate prediction model includes: an input adaptation layer for receiving multi-channel sliding time window data input to the climate prediction model; a deep feature extraction module, including alternating residual blocks and max pooling layers, for extracting nonlinear spatial features from the multi-channel sliding time window data to obtain a feature map; a spatial feature aggregation module, including an adaptive average pooling layer, for compressing the feature map to a specified size; and a decoder, including alternating fully connected layers and a nonlinear regression layer, wherein the nonlinear regression layer uses an activation function and a random deactivation function for nonlinear regression processing, and the decoder is used to map the feature map of the specified size to the predicted values ​​of each station.

[0020] Preferably, the method further includes: the climate prediction model trained for the current year outputs the climate prediction results for each month of the current year, and calculates the quarterly climate prediction results based on the climate prediction results for the corresponding months.

[0021] A second aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the climate prediction method based on hot start and background constraints as described above.

[0022] A third aspect of the present invention provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the climate prediction method based on hot start and background constraints as described above.

[0023] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention provides a climate prediction method, device, and medium based on hot start and background constraints; the constructed training sample library is a dynamically sliding training set, ensuring that the data of the year to be predicted does not participate in any learning or optimization process, completely eliminating the risk of future information leakage and improving the objectivity and credibility of the prediction model; the introduction of climate background constraints and prediction score constraints into the joint loss, the long-sequence climate background constraints compensate for the deficiency of recent training data possibly lacking extreme samples, making the prediction results more consistent with the long-term climate statistical laws of the site, thereby effectively improving the spatial correlation and physical rationality of the prediction results, and making the spatial distribution of the model output more consistent with the long-term climate statistical laws of the site. It more closely reflects actual atmospheric processes; it is directly optimized and trained on the climate prediction model trained in the previous year, so that the model can inherit long-term climate patterns and absorb the latest climate change characteristics in real time, realizing continuous evolution and stable prediction; the model can output climate prediction results for each month, achieving a breakthrough in refined prediction from seasonal to monthly scales, providing richer and more decision-making guidance information; it has a high degree of automation and scalability, supporting automatic updates of the training set and model iteration based on new observation data every year, realizing continuous learning and performance evolution of the prediction model, ensuring that the model can adapt to climate change, and guaranteeing the long-term stable and efficient operation of the operational forecasting system. Attached Figure Description

[0024] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0025] Figure 1 A flowchart of a climate prediction method based on hot start and background constraints provided in an embodiment of the present invention.

[0026] Figure 2 This diagram illustrates the implementation process of the climate prediction method based on hot start and background constraints provided in this embodiment of the invention.

[0027] Figure 3 This is a schematic diagram of the structure of the climate prediction model provided in an embodiment of the present invention.

[0028] Figure 4 This is a timeline diagram of the rolling hot start update mechanism provided in an embodiment of the present invention.

[0029] Figure 5 This is a schematic diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0030] In view of the shortcomings of the prior art, the inventors of this invention, through long-term research and extensive practice, have proposed the technical solution of this invention. The following will further explain and illustrate the technical solution, its implementation process, and its principles.

[0031] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0032] Furthermore, in the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "horizontal," "vertical," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0033] In the description of this specification, the references to terms such as "an embodiment," "a particular embodiment," or "the embodiment" indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0034] Figure 1 The flowchart of the climate prediction method based on hot start and background constraints provided in the embodiments of the present invention is shown below. Figure 1 The method includes the following operations S100-S500.

[0035] Operate S100 to train an autoencoder using actual climate data from the first historical year interval, so that the autoencoder learns the dependencies between stations, and use the weights of the trained autoencoder as the initial weights or regularization terms of the initial climate prediction model decoder.

[0036] Operation S200 retrieves the first climate data for the current year and the second historical year interval predicted by the specified climate prediction model. The first historical year interval is earlier than the second historical year interval.

[0037] Operation S300: For each month in the second historical year interval: use the first climate data of the month and the previous set number of months as input features, and use the actual climate data of the month as labels to generate sample pairs for that month; combine the sample pairs of each month to form a training sample library.

[0038] The S400 operation aims to minimize the joint loss by training the climate prediction model trained in the previous year using the training sample library. The joint loss includes accuracy constraints, climate background constraints, and prediction score constraints that characterize the model's prediction performance.

[0039] Operation S500: For any month to be predicted in the current year: Input the first climate data of the month to be predicted and the previous set number of months into the trained climate prediction model, and output the climate prediction result of the month to be predicted.

[0040] The current year is, for example, 2021, 2022, 2023, 2024, or 2025; the first historical year range is, for example, 1951 to 2000; and the second historical year range is, for example, 2001 to 2020. Climate predictions include, for example, rainfall, snowfall, temperature, wind speed, wind direction, relative humidity, solar radiation, and visibility.

[0041] This invention achieves accurate spatial downscaling from grid points to discrete stations using a grid-station heterogeneous mapping architecture. The dynamic model (i.e., the specified climate model) outputs regular two-dimensional grid data, while the prediction target is discrete, unevenly distributed station data. Traditional methods typically interpolate to stations before regression, or train a model separately for each station. The former loses spatial structure, and the latter ignores the spatial correlation between stations. This method designs an "encoder-decoder" heterogeneous mapping network. The encoder preserves the two-dimensional grid spatial structure and extracts high-dimensional features through convolution and pooling. The decoder directly maps the extracted regional prediction field features to all stations (e.g., 160).

[0042] Compared to single-station models, the multi-station joint output structure of this method allows the model to learn the spatial covariance between stations, meaning that climate (e.g., precipitation) variations among adjacent stations are usually consistent. A single model can predict all stations within a region simultaneously, significantly reducing computational complexity.

[0043] There is spatial correlation between stations (e.g., if it rains at station A, it usually rains at the nearby station B as well). This spatial relationship is physically fixed and does not change drastically with the year. The data from the first historical year interval reflects this spatial covariance structure.

[0044] In operation S100, the model's decoder is pre-trained or regularized using actual climate data from the first historical year interval. Specifically, before formal training, a simple autoencoder (e.g., capable of inputting 160 station values ​​→ compression → outputting 160 station values) is trained to allow the network to learn the dependencies between stations. The weights of this autoencoder are used as the initial weights of the climate prediction model's decoder, or as a regularization term to lock in the physical associations between stations. That is, a spatial topological prior is constructed using long-sequence unpaired historical observation data to initialize or constrain the output layer of the climate prediction model.

[0045] Preferably, operation S200 specifically includes the following sub-operations S210-S220.

[0046] In suboperation S210, data of each set member is obtained for the current year and the second historical year interval predicted by the specified climate system model.

[0047] Specifically, based on a preset file path mapping, it can automatically read data from each ensemble member of a climate system model (e.g., BCC-CSM) within a specified year range. The specified year could be, for example, the current year and the 20 years preceding it.

[0048] In sub-operation S220, for each year: the data of each set member under that year are ensembled to obtain the climate factor field of that year, and the climate factor field is interpolated and resampled to a preset standard spatial grid to generate the first climate data of that year.

[0049] Specifically, the data of each set member is parsed, and the ensemble average of all set member data for the same year is performed to obtain the climate factor field representing the climate state of that year. This climate factor field is then interpolated and resampled to a preset standard spatial grid (e.g., 1°×1°) to eliminate the differences in spatial resolution of the original data and ensure that all input data are spatially aligned.

[0050] The method also includes reading and filtering observational data (actual climate data). Specifically, it reads station observational data files, parses time information and climate data from each station, and filters out actual climate data for the first historical year interval (e.g., 70 to 20 years before the current year) for training the autoencoder; and filters out the second historical year interval (e.g., 20 years before the current year) as labels for training the climate prediction model.

[0051] Taking the current year as 2026, the second historical year interval as 2016-2025, and a set number of samples of 3 as an example, this explains the sample pairs generated in operation S300. For each month in the 2016-2025 period, such as June 2024, the first climate data of June, May, April, and March 2024 are used as input features, and the actual climate data of June 2024 are used as labels to generate "input feature-label" sample pairs for June 2024; and so on, to build a training sample library covering sample pairs from January 2016 to December 2025.

[0052] This method improves the ability to capture precursory signals of precipitation. Traditional statistical downscaling often uses only contemporaneous model data to predict precipitation for the current month. However, precipitation processes are often influenced by the cumulative effects of previous circulation (e.g., the lag in El Niño-Southern Oscillation signals, and the memory effect of soil moisture). Single-time input leads to information loss. This method constructs a "time-delay multi-channel input tensor." For the target prediction month, dynamic model data from several months in advance (e.g., 3 months) are selected and stacked along the channel dimension. This forces the model to learn the evolution trend and cumulative characteristics of the physical field over the past few months.

[0053] Furthermore, the feature tensors, observation labels, and corresponding metadata (such as year) of all generated sample pairs can be saved as files in a standardized format, completing the automated pipeline processing from raw data to a model-ready dataset. In this embodiment, the sample pairs in the training sample library can be randomly divided into training and validation sets, and the sample pairs of the current year can be used as the test set for the optimization training of the climate prediction model.

[0054] When optimizing the climate prediction model for the current year (year T), the mean is calculated using only historical data from year T-1 and earlier. and standard deviation And further generate sample pairs :

[0055] ;

[0056] ;

[0057] in, , The original values ​​before standardization. , The result after standardization It should be a very small positive number to prevent division by zero. This is the primary climate data. For actual climate data, The standard deviation of the first climate data for year T-1 and earlier. The standard deviation of actual climate data for year T-1 and earlier. This represents the average of the first climate data for year T-1 and earlier. This represents the average of actual climate data from year T-1 and earlier.

[0058] Preferred, such as Figure 3 As shown, the climate prediction model includes: an input adaptation layer for receiving multi-channel sliding time window data input to the climate prediction model; a deep feature extraction module, including alternating connected residual blocks and max pooling layers, for extracting nonlinear spatial features from the multi-channel sliding time window data to obtain a feature map; a spatial feature aggregation module, including an adaptive average pooling layer, for compressing the feature map to a specified size; and a decoder, including alternating connected fully connected layers and nonlinear regression layers. The nonlinear regression layer uses activation functions and random deactivation functions for nonlinear regression processing, and the decoder maps the feature map of the specified size to the predicted values ​​of each station.

[0059] Residual blocks are used to extract nonlinear spatial features from dynamic model data (such as trough-ridge patterns in pressure fields and water vapor transport channels), which can accelerate model convergence, prevent gradient vanishing, allow gradients to flow directly to shallow layers, and support the construction of deeper networks without degradation. Adaptive average pooling layers force feature maps of arbitrary sizes to be compressed to a fixed size (e.g., 4). 4) This allows the model to adapt to dynamic pattern data input at different resolutions without frequent adjustments to the network structure. The spatial feature aggregation module also includes a feature flattening layer, which transforms two-dimensional spatial feature maps into one-dimensional feature vectors, achieving a data format conversion from "field" to "point." The random deactivation layer randomly sets the output of some neurons to zero during training to simulate model ensemble effects, significantly enhancing the model's resistance to overfitting. The fully connected layer maps the extracted abstract atmospheric circulation features to specific climate values ​​from various meteorological stations, completing the final observation constraint correction.

[0060] Preferably, the joint loss is:

[0061] ;

[0062] in, For joint losses; This is the first hyperparameter; This is the second hyperparameter; and The control accuracy constraints, climate background constraints, and prediction score constraints are of the same order of magnitude; For accuracy constraints, it characterizes the error between the predicted value and the actual value; As a constraint of climate background, it characterizes whether the fluctuation range and extreme value probability of the predicted values ​​conform to the climate patterns of the first and second historical year intervals of the station; The prediction score constraint characterizes the degree of closeness between the predicted value and the actual value.

[0063] The precision constraint is calculated using labeled data from the second historical year interval and can be expressed as:

[0064] ;

[0065] in, The total number of weather stations. Let be the actual observed precipitation value (or its standardized value) at the i-th station. The predicted precipitation value (or its standardized value) for the i-th station output by the model.

[0066] The climate background constraint-forcing model not only provides accurate predictions (low root mean square error), but also ensures that the fluctuation range and extreme value probability of the predicted values ​​conform to the historical climate patterns of the site (the first historical year interval and the second historical year interval), thus solving the problem of "predictions tending to be smooth" caused by simple root mean square error. Preferably, the climate background constraint is:

[0067] ;

[0068] Alternatively, the climate background constraint is:

[0069] ;

[0070] in, Constrained by climate background; Let KL divergence function be used. The Wasserstein distance function; This represents the distribution of the current batch of prediction results from the climate prediction model. This represents the cumulative distribution or probability density of actual climate data for the first historical year interval.

[0071] Preferably, the prediction score constraint is:

[0072] ;

[0073] in, For predicting score constraints, The total number of observation stations, This represents the number of observation stations where the predicted percentage anomaly and the actual percentage anomaly have the same sign. The absolute values ​​of both the predicted and actual anomaly percentages are within [a certain range]. The number of observation stations within the interval, The absolute values ​​of both the predicted and actual anomaly percentages are within [a certain range]. The number of observation stations within the interval, For the absolute value of the predicted value's percentage anomaly to be no less than 1 and the absolute value of the actual value's percentage anomaly to be within 1, the following conditions must be met: The number of observation stations within the interval, The first preset parameter, This is the second preset parameter. , where 'a' is the first coefficient, 'b' is the second coefficient, and 'c' is the third coefficient. For example... , a=2, b=2, c=4. The predicted percentage anomaly and the actual percentage anomaly may be positive, negative, or 0.

[0074] The implementation details of the optimized training and prediction process in operating S400 to S500 are as follows: Figure 2 As shown. During training, the joint loss is monitored in real time. If the joint loss does not decrease for several consecutive (e.g., 50) epochs (an epoch is defined as the process of a complete dataset passing through the neural network once and returning once), an early stopping mechanism is triggered, and the system rolls back to the weight state with the minimum joint loss, which is then used as the final weights for the climate prediction model for the current year. The next year's model is then optimized and trained based on the final climate prediction model for the current year.

[0075] Preferably, if there is no climate prediction model trained in the previous year, the initial climate prediction model is used as the climate prediction model trained in the previous year.

[0076] See Figure 4 This illustrates a cold start / hot start update strategy according to an embodiment of the present invention. In the first year (e.g., 2021), there is no climate prediction model trained in the previous year, so the initial climate prediction model is directly optimized and trained. In the years following the first year (e.g., 2022, 2023, 2024, and 2025), there is a climate prediction model trained in the previous year, so the climate prediction model trained in the previous year is optimized and trained again.

[0077] Preferably, the method further includes: the climate prediction model trained for the current year outputs the climate prediction results for each month of the current year, and calculates the quarterly climate prediction results based on the climate prediction results for the corresponding months.

[0078] Taking the current year as 2026 as an example, the trained climate prediction model directly outputs monthly climate prediction results. , , , , , , , , , , , ,in, This represents the climate prediction result for the i-th site in the m-th month of 2026. For example, As part of the first quarter's climate forecasts, This is a result of summer climate forecasts.

[0079] The climate prediction method based on hot start and background constraints provided by this invention has the following advantages.

[0080] (1) Observational constraints effectively correct systematic biases in climate models. This method uses real observed climate data from meteorological stations as the sole supervisory signal (observational constraints) for the deep learning model, constructing a nonlinear mapping relationship between the large-scale background field of the dynamic model and station-level climate. Compared with traditional statistical downscaling methods, this method utilizes the powerful feature extraction capabilities of deep residual networks to more accurately capture and correct systematic biases in dynamic models under complex terrain, significantly improving the accuracy and score of station-level climate predictions.

[0081] (2) Hot start solves catastrophic forgetting and model aging. It abandons the traditional static mode of "train once and use for life" and loads the optimal model parameters of the previous year as the initial weights of the current year. The model can inherit long-term climate patterns and absorb the latest climate change characteristics in real time, so as to achieve continuous evolution and stable prediction of the model.

[0082] (3) Maximize the utilization efficiency of small sample climate data. The full historical data training strategy and hot start mechanism maximize the use of limited sample resources and improve the convergence speed and stability of the model.

[0083] (4) Enhance the ability to predict extreme events. By introducing the distribution constraints of long-series climate background data, the shortcomings of recent training data that may lack extreme samples are compensated for, making the prediction results more consistent with the long-term climate statistics of the site.

[0084] (5) Based on physical consistency and spatial correlation, a "climate distribution constraint term" is added to the loss function. The model is not required to reproduce a specific climate event in the first historical year interval (e.g., 1951 to 2000), but the statistical distribution of precipitation predicted by the model is required not to deviate too far from the long-term climate state of the first historical year interval of the station. The cumulative effect is captured by time-delayed multi-channel input, and the spatial covariance between stations is preserved by heterogeneous mapping architecture to ensure the consistency of the prediction results in both physics and space.

[0085] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the climate prediction method based on hot start and background constraints as described in any of the above embodiments.

[0086] Figure 5 This embodiment illustrates a more specific hardware structure of an electronic device, which may include a processor 510, a memory 520, an input / output interface 530, a communication interface 540, and a bus 550. The processor 510, memory 520, input / output interface 530, and communication interface 540 are interconnected internally via the bus 550.

[0087] The processor 510 can be implemented using a general-purpose CPU (Central Processing Unit), microprocessor, application-specific integrated circuit (ASIC), or one or more integrated circuits, and is used to execute relevant programs to implement the technical solutions provided in the embodiments of this specification.

[0088] The memory 520 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 520 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented by software or firmware, the relevant program code is stored in the memory 520 and is called and executed by the processor 510.

[0089] Input / output interface 530 is used to connect input / output modules to realize information input and output. Input / output modules can be configured as components in the device (not shown in the figure) or externally connected to the device to provide corresponding functions. Input devices may include keyboards, mice, touch screens, microphones, various sensors, etc., and output devices may include displays, speakers, vibrators, indicator lights, etc.

[0090] The communication interface 540 is used to connect a communication module (not shown in the figure) to enable communication between this device and other devices. The communication module can communicate via wired means (such as USB, Ethernet cable, etc.) or wireless means (such as mobile network, WIFI, Bluetooth, etc.).

[0091] Bus 550 includes a pathway for transmitting information between various components of the device, such as processor 510, memory 520, input / output interface 530, and communication interface 540.

[0092] It should be noted that although the above-described device only shows the processor 510, memory 520, input / output interface 530, communication interface 540, and bus 550, in specific implementations, the device may also include other components necessary for normal operation. Furthermore, those skilled in the art will understand that the above-described device may only include the components necessary for implementing the embodiments of this specification, and not necessarily all the components shown in the figures.

[0093] The electronic devices described above are used to implement the corresponding climate prediction methods based on hot start and background constraints in any of the foregoing embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0094] Based on the same inventive concept, corresponding to the methods of any of the above embodiments, the present invention also provides a non-transitory computer-readable storage medium storing computer instructions for causing the computer to execute the climate prediction method based on hot start and background constraints as described in any of the above embodiments.

[0095] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0096] The computer instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the climate prediction method based on hot start and background constraints as described in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0097] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of the invention (including the claims) is limited to these examples; within the framework of the invention, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of the invention as described above, which are not provided in the details for the sake of brevity.

[0098] Additionally, to simplify the description and discussion, and to avoid obscuring the embodiments of the invention, the well-known power / ground connections to integrated circuit (IC) chips and other components may or may not be shown in the provided drawings. Furthermore, the apparatus may be shown in block diagram form to avoid obscuring the embodiments of the invention, and this also takes into account the fact that the details of implementation of these block diagram apparatuses are highly dependent on the platform on which the embodiments of the invention will be implemented (i.e., these details should be fully understood by those skilled in the art). While specific details (e.g., circuits) have been set forth to describe exemplary embodiments of the invention, it will be apparent to those skilled in the art that the embodiments of the invention may be implemented without these specific details or with variations thereof. Therefore, these descriptions should be considered illustrative rather than restrictive.

[0099] Although the invention has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art from the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) may be used with the embodiments discussed.

[0100] It should be understood that the above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made in accordance with the spirit and essence of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A climate prediction method based on hot start and background constraints, characterized in that, include: The autoencoder was trained using actual climate data from the first historical year interval, so that the autoencoder could learn the dependencies between stations. The weights of the trained autoencoder were used as the initial weights or regularization terms of the initial climate prediction model decoder. Obtain the first climate data for the current year and the second historical year interval predicted by a specified climate system model, wherein the first historical year interval is earlier than the second historical year interval; For each month in the second historical year interval: use the first climate data of that month and the previous set number of months as input features, and the actual climate data of that month as labels to generate sample pairs for that month; combine the sample pairs of each month to form a training sample library; With the goal of minimizing the joint loss, the climate prediction model trained in the previous year is trained using the training sample library. The joint loss includes accuracy constraints, climate background constraints, and prediction score constraints that characterize the model's prediction performance. If there is no climate prediction model trained in the previous year, the initial climate prediction model is used as the climate prediction model trained in the previous year. For any month to be predicted in the current year: input the first climate data of the month to be predicted and the previous set number of months into the trained climate prediction model, and output the climate prediction result for the month to be predicted.

2. The climate prediction method based on hot start and background constraints according to claim 1, characterized in that, The joint loss is: ； in, For the joint loss; This is the first hyperparameter; This is the second hyperparameter; and The orders of magnitude of the control used for the accuracy constraint, the climate background constraint, and the prediction score constraint are the same. The accuracy constraint characterizes the error between the predicted value and the actual value; As a constraint of the climate background, it characterizes whether the fluctuation range and extreme value probability of the predicted value conform to the climate patterns of the first and second historical year intervals of the station; The prediction score constraint characterizes the degree of closeness between the predicted value and the actual value.

3. The climate prediction method based on hot start and background constraints according to claim 1 or 2, characterized in that, The climate background constraint is: ； Alternatively, the climate background constraint is: ； in, Constrained by the aforementioned climate background; Let KL divergence function be used. The Wasserstein distance function; This represents the distribution of the current batch of prediction results from the climate prediction model. This represents the cumulative distribution or probability density of actual climate data for the first historical year interval.

4. The climate prediction method based on hot start and background constraints according to claim 1 or 2, characterized in that, The prediction score constraint is: ； in, For the predicted score constraints, The total number of observation stations, This represents the number of observation stations where the predicted percentage anomaly and the actual percentage anomaly have the same sign. The absolute values ​​of both the predicted and actual anomaly percentages are within [a certain range]. The number of observation stations within the interval, The absolute values ​​of both the predicted and actual anomaly percentages are within [a certain range]. The number of observation stations within the interval, For the absolute value of the predicted value's percentage anomaly to be no less than 1 and the absolute value of the actual value's percentage anomaly to be within 1, the following conditions must be met: The number of observation stations within the interval, The first preset parameter, This is the second preset parameter. , where a is the first coefficient, b is the second coefficient, and c is the third coefficient.

5. The climate prediction method based on hot start and background constraints according to claim 1, characterized in that, Obtain the first climate data for the current year and the second historical year interval predicted by the specified climate system model, specifically including: Obtain the data of each set member in the current year and the second historical year interval predicted by the specified climate system model; For each year: the data of each set member in that year are averaged to obtain the climate factor field for that year. The climate factor field is then interpolated and resampled to a preset standard spatial grid to generate the first climate data for that year.

6. The climate prediction method based on hot start and background constraints according to claim 1, characterized in that, The climate prediction model includes: An input adaptation layer is used to receive multi-channel sliding time window data input to the climate prediction model; The deep feature extraction module, including alternating residual blocks and max pooling layers, is used to extract nonlinear spatial features from multi-channel sliding time window data to obtain feature maps. The spatial feature aggregation module includes an adaptive average pooling layer for compressing feature maps to a specified size; The decoder includes alternating fully connected layers and nonlinear regression layers, wherein the nonlinear regression layers perform nonlinear regression processing using activation functions and random deactivation functions, and the decoding is used to map feature maps of a specified size to predicted values ​​for each station.

7. The climate prediction method based on hot start and background constraints according to claim 1, characterized in that, The method further includes: the climate prediction model trained for the current year outputs the climate prediction results for each month of the current year, and calculates the quarterly climate prediction results based on the climate prediction results for the corresponding months.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the climate prediction method based on hot start and background constraints as described in any one of claims 1-7.

9. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the climate prediction method based on hot start and background constraints as described in any one of claims 1-7.

Images