Intelligent climate downscaling method and system based on deep learning

Abstract

The application discloses an intelligent climate downscaling method and system based on deep learning, and relates to the technical field of climate downscaling data processing.The intelligent climate downscaling method based on deep learning comprises the following steps: auxiliary feature redundancy monitoring, climate downscaling data load monitoring, model adaptability monitoring and scenario adaptability monitoring.The application determines whether auxiliary features are needed for redundancy processing by performing climate downscaling preprocessing data redundancy discrimination, then determines whether light preprocessing is needed by performing climate downscaling preprocessing data load analysis, and finally determines whether scenario commonality analysis is needed by performing model adaptability analysis, so that the timeliness of high-resolution climate change scenario estimation data is improved, and the problem of low timeliness of high-resolution climate change scenario estimation data in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of climate downscaling data processing technology, and in particular to an intelligent climate downscaling method and system based on deep learning. Background Technology

[0002] In the process of intelligent climate downscaling based on deep learning, the following steps are first taken: First, regional climate models, such as the WRF (Weather Research and Forecasting Model), are collected, outputting low-resolution meteorological forecast data (e.g., monthly gridded data) and high-resolution reference data (e.g., ERA5-Land reanalysis data, station observation data). Next, the data is standardized (e.g., bilinear interpolation). Then, multiple types of features are extracted from the time domain (e.g., Fourier transform to extract diurnal variation, seasonality, etc.), frequency domain (e.g., fast Fourier transform to obtain frequency distribution), spatial domain (e.g., calculating regional correlation coefficients to identify spatial autocorrelation), and nonlinear dimensions (e.g., nonlinear regression, polynomial transform, etc.). These features are then combined to form a hybrid feature library. Finally, meteorological features are weighted based on a spatiotemporal attention mechanism to identify key spatiotemporal regions, and an adaptive downsampling window is designed. The importance score is weighted and averaged to reduce data dimensionality while preserving core features. Next, model training is performed, involving designing model architectures (such as convolutional neural network architectures or generative adversarial network architectures), selecting activation functions (such as ReLU) and loss functions (such as mean squared error or adversarial loss) to introduce nonlinearity, and then using cross-validation to divide the dataset. Optimization algorithms (such as Adam) are used to fine-tune hyperparameters, minimizing downscaling loss and bias correction loss until the loss converges. Finally, independent test sets are used to evaluate model performance (e.g., calculating root mean square error, mean absolute error, etc.) to verify the accuracy of the downscaled data. The trained model is then applied to real-world scenarios, inputting low-resolution climate data to generate high-resolution meteorological fields (such as temperature and precipitation distribution).

[0003] In the process of intelligent climate downscaling, existing technologies first collect low-resolution hydro-meteorological forecast data (such as precipitation, temperature, and other variables) and high-resolution reference data (such as ERA5-Land reanalysis data and station observation data) from regional climate models, and preprocess the data (including standardization and missing value imputation) to construct a training dataset. Next, a model architecture based on generative adversarial networks is built, in which the generator is responsible for learning the mapping relationship from low-resolution data to high-resolution data, and the discriminator is used to distinguish between generated high-resolution data and real high-resolution reference data. Then, through unsupervised adversarial training, the generator and discriminator compete and iteratively optimize each other: the generator continuously adjusts its parameters to generate high-resolution data that is closer to reality, and the discriminator continuously improves its ability to distinguish between real data and generated data. In the process, the adversarial loss is continuously reduced, so that the trained generative deep learning network has the ability to handle complex nonlinear relationships of climate variables. Finally, the trained model is used to downscale and correct biases in the hydro-meteorological variable forecast results of the multi-model ensemble.

[0004] For example, Chinese invention patent application CN116701839A discloses a spatial downscaling method, apparatus, device, and storage medium for climate variable data, comprising: first, acquiring climate variable data for a target region within a target time period, and constructing a first climate variable time series and a first climatological time series for the region based on the data; then, using the same first resolution grid as a unified reference, calculating first climate variable anomalous data corresponding to several time units based on the aforementioned first climate variable time series and first climatological time series, thereby constructing a first climate variable anomalous time series for the target region; then, acquiring a second climatological time series for the target region; subsequently, combining the second climatological time series with the first climate variable anomalous time series, and performing spatial downscaling processing using a spatial interpolation method; finally, obtaining the second climate variable time series for the target region through spatial downscaling processing.

[0005] For example, Chinese invention patent application CN119003975B discloses an integrated method for spatial reconstruction and downscaling of all-sky surface temperature. The method includes: first, acquiring multi-source remote sensing data and atmospheric driving data, and performing preprocessing operations on these data, including data cleaning, radiometric correction, geometric correction, and spatiotemporal benchmark unification; then, spatial reconstruction of all-sky surface temperature data based on the preprocessed dataset, by integrating effective data from different observation conditions and time periods to fill in missing data areas and construct a complete and continuous all-sky surface temperature dataset covering the target area; finally, spatial downscaling processing is applied to the reconstructed all-sky surface temperature data, combining statistical analysis and other data processing techniques with surface temperature monitoring methods to establish a correlation model between the data and high-resolution covariates, thereby upgrading low spatial resolution data to higher resolution.

[0006] The above-mentioned technology has at least the following technical problems:

[0007] With the deepening expansion of practical application scenarios (such as regional climate adaptation planning, disaster prevention and mitigation in small and medium-sized watersheds, and precision agricultural production), the demand for high-resolution climate change scenario prediction data at the kilometer level and below is becoming increasingly urgent. However, in the process of intelligent climate downscaling, traditional statistical downscaling methods (such as delta quantile mapping) require upsampling the low-resolution regional climate model simulation results to the target high resolution through spatial interpolation methods (such as bilinear interpolation), and then constructing a transfer function with the high-resolution observation data time-by-time. This process needs to cover massive amounts of raster data for all simulation periods, which places extremely high demands on computing resources. As a result, the stage of constructing the transfer function requires processing several terabytes of interpolated data. The interpolation error correction in complex terrain areas also requires additional parameter tuning and iterative verification steps. The amount of data computation and verification costs under multiple scenarios in a single region have increased significantly, ultimately resulting in low timeliness of high-resolution climate change scenario prediction data. Summary of the Invention

[0008] To address the technical problem of low timeliness in high-resolution climate change scenario prediction data in existing technologies, this invention provides an intelligent climate downscaling method and system based on deep learning. The technical solution is as follows:

[0009] On the one hand, a deep learning-based intelligent climate downscaling method is provided. This method includes: in the process of intelligent climate downscaling based on deep learning, performing data redundancy judgment in climate downscaling preprocessing to determine whether auxiliary feature de-redundancy processing is needed. Auxiliary feature de-redundancy processing is used to reduce data dimensionality and storage load, and improve the feature effectiveness of input data. After the data redundancy judgment in climate downscaling preprocessing is qualified, performing data load analysis in climate downscaling preprocessing to determine whether lightweight preprocessing is needed. Lightweight preprocessing is used to balance data processing efficiency and downscaling result reliability. After the data load analysis in climate downscaling preprocessing is qualified, performing model adaptability analysis to determine whether scenario commonality analysis is needed. If scenario commonality analysis is not needed, a model training failure warning is sent. Otherwise, based on the results of scenario commonality analysis, it is determined whether cross-scenario transfer optimization is needed. Cross-scenario transfer optimization is used to improve the downscaling adaptation accuracy and training efficiency of deep learning models for multiple types of climate scenarios.

[0010] On the other hand, a deep learning-based intelligent climate downscaling method system is provided. This system, employing a deep learning-based intelligent climate downscaling method, includes: an auxiliary feature redundancy monitoring module, a climate downscaling data load monitoring module, a model adaptability monitoring module, and a scenario adaptability monitoring module. Specifically, the auxiliary feature redundancy monitoring module performs data redundancy assessment during the deep learning-based intelligent climate downscaling process to determine whether auxiliary feature de-redundancy processing is required. The climate downscaling data load monitoring module performs data load analysis after passing the data redundancy assessment to determine whether lightweight preprocessing is required. The model adaptability monitoring module performs model adaptability analysis after passing the data load analysis to determine whether scenario commonality analysis is required. The scenario adaptability monitoring module sends a model training failure warning if scenario commonality analysis is not required; otherwise, it determines whether cross-scenario transfer optimization is required based on the results of the scenario commonality analysis.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0012] 1. By performing redundancy assessment on climate downscaling preprocessed data to determine whether auxiliary feature de-redundancy processing is needed, it helps filter out invalid data interference at the source, providing a high-quality data foundation for subsequent climate downscaling task data preprocessing and model training. This reduces the computational resources consumed by redundant features or misleads the model in learning core climate laws. After the climate downscaling preprocessed data redundancy assessment is passed, climate downscaling preprocessed data load analysis is performed to determine whether lightweight preprocessing is needed. This helps to accurately match data processing resources with data scale and reduce redundant consumption of computational resources. After the climate downscaling preprocessed data load analysis is passed, model adaptability analysis is performed to determine whether scenario commonality analysis is needed. This helps to identify the adaptation defects between the model and the preprocessed data in advance and reduce subsequent meaningless model training processes. If scenario commonality analysis is not needed, a model training failure warning is sent. Otherwise, based on the results of scenario commonality analysis, it is determined whether cross-scenario transfer optimization is needed, which helps to enhance the adaptability of deep learning models to complex climate scenarios.

[0013] 2. By selectively choosing model training convergence metrics, model prediction error metrics, and data fit goodness metrics as core multi-dimensional evaluation metrics, we can comprehensively capture key adaptation characteristics of deep learning models in climate downscaling scenarios, such as training stability, prediction accuracy, and data fitting quality. The result of weighted coupling of these core multi-dimensional evaluation metrics and their corresponding multi-dimensional adaptability weight parameters serves as the multi-dimensional adaptability metric. This comprehensively considers the combined impact of the three types of metrics on the model-data fit, reducing bias caused by single-metric evaluation. The multi-dimensional adaptability metric more objectively reflects whether the preset deep learning model meets the adaptation requirements of qualified climate downscaling preprocessed data. Determining whether the multi-dimensional adaptability metric exceeds the preset model adaptability threshold helps identify model-data fit defects in advance, achieving accurate matching between model processing strategies and actual model adaptation characteristics. This reduces problems such as ineffective training and insufficient predictive reliability of downscaling models caused by inaccurate analysis of model-data fit characteristics.

[0014] 3. By selectively choosing baseline climate scenarios with high data integrity and completed model training, along with other climate scenarios requiring adaptation, and combining the baseline and adaptation feature vectors extracted by merging the shallow feature vectors output from the shallow feature extraction layer and the residual feature vectors output from the residual group, the core climate feature correlation logic between deep and shallow layers under different climate scenarios can be comprehensively captured. Then, by quantifying the ratio of the dot product of the baseline and adaptation feature vectors with the product of their magnitudes, a climate scenario parameter migration adaptation index can be obtained. This can accurately assess the similarity of cross-scenario feature vectors, reduce adaptation bias caused by subjective judgment, and make the climate scenario parameter migration adaptation index more objectively reflect the feasibility of sharing model parameters between different climate scenarios. Determining whether the climate scenario parameter migration adaptation index is greater than the preset parameter sharing threshold helps to achieve accurate decision-making for cross-scenario migration optimization and model adaptation failure warning, reducing the waste of computing power and time caused by blindly training new scenario models, and reducing the problems of low model downscaling accuracy and poor adaptation efficiency caused by insufficient analysis of cross-scenario feature commonalities. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of the intelligent climate downscaling method based on deep learning provided in an embodiment of the present invention;

[0017] Figure 2 This is a flowchart outlining the intelligent climate downscaling method based on deep learning provided in this embodiment of the invention.

[0018] Figure 3 This is a schematic diagram of the lightweight preprocessing logic of the intelligent climate downscaling method based on deep learning provided in an embodiment of the present invention.

[0019] Figure 4 This is a schematic diagram of the structure of the intelligent climate downscaling system based on deep learning provided in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0021] like Figure 1 The diagram shown is a flowchart of a deep learning-based intelligent climate downscaling method provided in an embodiment of the present invention. Figure 1It can be seen that: First, auxiliary feature redundancy monitoring is performed during the intelligent climate downscaling process based on deep learning. Redundancy judgment of climate downscaling preprocessing data is performed to determine whether auxiliary feature redundancy removal is needed. Auxiliary feature redundancy removal is used to remove worthless auxiliary features with low correlation to core climate variables, reduce data dimensionality and storage load, and improve the feature effectiveness of input data. Auxiliary feature redundancy monitoring helps the preset deep learning model to focus more on the correlation patterns of core climate variables during training, reduce the interference of low-value features on model parameter learning, and lay the feature foundation for subsequent high-precision downscaling.

[0022] Secondly, climate downscaling data load monitoring involves analyzing the redundancy of climate downscaling preprocessed data after it has passed the redundancy check to determine whether lightweight preprocessing is necessary. Lightweight preprocessing reduces the computational load of preprocessing while ensuring the accuracy of core scenario data, balancing data processing efficiency with the reliability of downscaling results. Climate downscaling data load monitoring helps reduce excessive computational resources invested in non-critical data, while ensuring that core climate information is not lost and reducing the time spent on preprocessing climate downscaling task data.

[0023] Next, model fit monitoring is conducted after the climate downscaling preprocessed data load analysis is deemed satisfactory. This model fit analysis determines whether scenario commonality analysis is necessary. Model fit monitoring helps to quantify the degree of matching between the model and the preprocessed data from multiple dimensions, promptly identify potential problems in model convergence, prediction accuracy, etc., and reduce the problem of low timeliness of high-resolution climate change scenario prediction data caused by excessively long model training time.

[0024] Finally, scenario adaptability monitoring sends a model training failure warning if scenario commonality analysis is not required. Otherwise, it determines whether cross-scenario transfer optimization is needed based on the results of scenario commonality analysis. Cross-scenario transfer optimization is used to improve the downscaling adaptation accuracy and training efficiency of deep learning models for multiple types of climate scenarios. Scenario adaptability monitoring helps to accurately determine the commonalities and differences between different climate scenarios, avoiding the accuracy loss caused by blindly performing cross-scenario model transfer. At the same time, the failure warning mechanism can terminate inapplicable training processes in a timely manner, improving the system's response flexibility and adaptation reliability to complex and variable climate scenarios.

[0025] It should be noted that before designing the deep learning-based intelligent climate downscaling method proposed in this application, a database for storing various preset data has been pre-built. This database covers, but is not limited to, preset values ​​such as preset correlation threshold, preset feature redundancy threshold, preset absolute redundancy threshold, preset data volume threshold, preset key time period discrimination threshold, preset training time threshold, and preset parameter sharing threshold. All of the above preset values ​​are directly configured and set by professional technicians.

[0026] In this embodiment, a complete intelligent climate downscaling end-to-end support system is constructed through the coordinated linkage of auxiliary feature redundancy monitoring, climate downscaling data load monitoring, model adaptability monitoring, and scenario adaptability monitoring. This system ensures the feature validity and processing efficiency of input data, as well as the model-data adaptation accuracy. It also improves the flexibility of the preset deep learning model in responding to multiple climate scenarios and the reliability of downscaling results. This effectively solves the problems of low climate downscaling accuracy, poor timeliness, and weak cross-scenario adaptability caused by data redundancy, inefficient processing, and insufficient model adaptation.

[0027] Specifically, auxiliary feature redundancy monitoring provides high-quality, low-redundancy basic data support for subsequent climate downscaling data load monitoring. If the auxiliary feature redundancy monitoring fails to accurately remove invalid features, the data load analysis will be based on redundant data, increasing the computational pressure of lightweight preprocessing and thus affecting the overall data processing efficiency. The results of climate downscaling data load monitoring directly determine the input data quality and processing timeliness of model adaptability monitoring. If climate downscaling preprocessing data load analysis is missing, the model adaptability analysis will be based on inefficient data or data lacking core information, directly causing the timeliness of the preset deep learning model to be substandard during training. The results of model adaptability monitoring also provide key model basis judgment for scenario adaptability monitoring. If the model adaptability analysis is inaccurate, the cross-scenario transfer optimization effect will be poor due to the substandard timeliness of model training under the baseline condition, ultimately affecting the overall accuracy and output efficiency of multi-scenario downscaling.

[0028] like Figure 2 The diagram shown is a flowchart outlining the intelligent climate downscaling method based on deep learning provided in this embodiment of the invention. Figure 2It can be seen that in the process of intelligent climate downscaling based on deep learning, redundancy judgment is performed on the preprocessed climate downscaling data, and the redundancy index of the preprocessed data is obtained. It is then determined whether the redundancy index of the preprocessed data is less than a preset feature redundancy threshold. If it is, climate downscaling preprocessed data load analysis is performed; otherwise, auxiliary feature de-redundancy processing is carried out, and the redundancy index of the preprocessed data is determined whether it is less than a preset absolute redundancy threshold. If it is, the absolute redundant features and associated storage modules are deleted; otherwise, dimension fusion is performed on the non-redundant auxiliary features. After the auxiliary feature de-redundancy processing is completed, it is determined whether the redundancy judgment of the preprocessed climate downscaling data is qualified. If not, a redundancy removal failure warning is sent; otherwise, climate downscaling preprocessed data load analysis is performed. The system acquires the total amount of climate downscaling preprocessed data and determines whether it exceeds a preset data volume threshold. If not, conventional climate downscaling preprocessing is performed, followed by model fit analysis. Otherwise, lightweight preprocessing is adopted. After lightweight preprocessing, model fit analysis is performed, and the preset deep learning model training time is acquired. The system determines whether the preset deep learning model training time is less than a preset training time threshold. If not, a model training failure warning is issued. Otherwise, scenario commonality analysis is performed, and the climate scenario parameter migration fit index is acquired. The system determines whether the climate scenario parameter migration fit index exceeds a preset parameter sharing threshold. If so, cross-scenario migration optimization is adopted. Otherwise, a model fit failure warning is issued.

[0029] Furthermore, the specific process for redundancy assessment of climate downscaling preprocessed data is as follows: A preprocessed data redundancy index is obtained to assess the degree of redundancy in the preprocessed data. This index is represented by the quantification result of the redundancy of auxiliary features in the climate downscaling preprocessed data. The quantification result of the auxiliary feature redundancy is obtained by quantifying the proportion of the total amount of low-climate-related auxiliary features (such as minor topographical indicators and low-resolution vegetation subtypes) monitored by the auxiliary feature quantity monitor to the total number of auxiliary features. Low-climate-related auxiliary features refer to those whose correlation with core climate variables (such as temperature and precipitation) is less than a preset correlation threshold. The proportion quantification is represented by a ratio calculation. The preset correlation threshold is represented by the average correlation between auxiliary features and core climate variables over a historical time period. The correlation of core climate variables is assessed using a preset correlation algorithm (such as Pearson correlation coefficient, Spycorrelation coefficient, etc.). The results of linear correlation between auxiliary feature time series data (such as ten-day slope change rate, monthly soil texture classification, quarterly low-resolution vegetation cover, etc.) and core climate variable time series data (such as hourly temperature observations, daily precipitation records, monthly average humidity data, etc.) are represented by the correlation coefficient (e.g., Mann correlation coefficient). Auxiliary features represent non-core data features that assist in the downscaling simulation of core climate variables and reflect the characteristics of local environmental elements (such as slope change rate, rough soil texture classification, etc.). It is determined whether the preprocessed data redundancy index is less than the preset feature redundancy threshold. If so, climate downscaling preprocessed data load analysis is performed; otherwise, auxiliary feature redundancy removal is performed. The preset feature redundancy threshold is represented by the average value of the preprocessed data redundancy index over historical time periods. Climate downscaling preprocessed data load analysis is used to reduce the computational power consumption and preprocessing time caused by excessive data volume in climate downscaling preprocessed data.

[0030] Specifically, the auxiliary feature redundancy removal process is as follows: When the preprocessed data redundancy index is detected to be greater than the preset absolute redundancy threshold, the corresponding auxiliary feature field is marked as an absolute redundancy feature, and a prompt is sent to the designated personnel to delete the corresponding absolute redundancy feature and associated storage module. The associated storage module refers to a storage unit or storage sub-module used to store auxiliary information (such as the feature's acquisition log, preprocessing records, historical association analysis results, etc.) that has a direct data reference, dependency relationship, or logical association with the data field marked as an absolute redundancy feature. The preset absolute redundancy threshold is determined by the average value of the preprocessed data redundancy index over a historical time period that is not greater than the preset feature redundancy threshold. The process involves the following steps: When the preprocessed data redundancy index is detected to be no greater than the preset absolute redundancy threshold, the corresponding auxiliary feature fields are marked as non-redundant auxiliary features. A preset feature dimensionality reduction algorithm (such as Principal Component Analysis (PCA) or Linear Discriminant Analysis (LDA) is used to perform dimensionality fusion on the non-redundant auxiliary features, transforming the high-dimensional feature matrix into a low-dimensional core feature vector to reduce data dimensionality redundancy. After the auxiliary feature redundancy removal process is completed, the preprocessed data redundancy index is re-acquired. If the preprocessed data redundancy index is still no less than the preset feature redundancy threshold, a redundancy removal failure warning is sent; otherwise, climate downscaling preprocessed data load analysis is performed. The preset absolute redundancy threshold is greater than the preset feature redundancy threshold.

[0031] In this embodiment, by redundancy identification and auxiliary feature deredundancy processing of climate downscaling preprocessed data, accurate identification and hierarchical optimization of redundant features in climate downscaling preprocessed data are achieved. The redundancy identification step relies on the correlation degree of core climate variables, effectively reducing the subjective judgment bias of redundant features and ensuring the objectivity of redundancy identification. The auxiliary feature deredundancy processing reduces the loss of effective information caused by single processing through a hierarchical strategy, taking into account both data simplification and feature effectiveness, while ensuring the effect of redundancy processing. It removes valueless data from the source, compresses data dimensions, significantly reduces the computing power consumption and preprocessing time of subsequent data load analysis, improves the feature purity of input climate downscaling task data, reduces interference from low-correlation features, and lays a high-quality data foundation for subsequent high-precision climate downscaling simulation and model adaptation.

[0032] Furthermore, the specific process of climate downscaling preprocessing data load analysis is as follows: The total amount of climate downscaling preprocessed data monitored by a data statistics monitoring instrument is obtained to reflect the potential impact of the data scale of the climate downscaling preprocessing stage on the consumption pressure of computing resources; it is determined whether the total amount of climate downscaling preprocessed data exceeds a preset data volume threshold. If so, lightweight preprocessing is adopted; if the total amount of climate downscaling preprocessed data does not exceed the preset data volume threshold, conventional climate downscaling preprocessing is adopted. The climate downscaling preprocessed data obtained after conventional climate downscaling preprocessing is marked as qualified climate downscaling preprocessed data, and model adaptability analysis is performed. Model adaptability analysis is used to evaluate whether the preset deep learning model has effective downscaling and efficient processing capabilities. The preset data volume threshold is set in advance by preset personnel. Conventional climate downscaling preprocessing refers to performing standardized basic preprocessing procedures on the climate downscaling preprocessed data, including data cleaning, radiometric correction, and full-time bilinear interpolation. Conventional climate downscaling preprocessing is used to reduce data anomaly bias, unify data accuracy standards, ensure the spatial continuity of data throughout the time period, and reduce the risk of model training errors.

[0033] In this embodiment, by analyzing the data load of climate downscaling preprocessing, the preprocessing time and redundant computing power consumption caused by the excessive amount of data in climate downscaling preprocessing are reduced. Furthermore, by relying on the standardized process of conventional preprocessing, the elimination of data anomalies, consistency of accuracy, and spatial continuity are ensured. This provides efficient and high-quality input data for subsequent model adaptability analysis and supports the accurate evaluation of the downscaling processing capability of the preset deep learning model.

[0034] like Figure 3 The diagram shown illustrates the lightweight preprocessing logic of the deep learning-based intelligent climate downscaling method provided in this embodiment of the invention. Figure 3 It can be seen that: lightweight preprocessing is performed and the climate element fluctuation index is obtained. It is determined whether the climate element fluctuation index is greater than the preset key period discrimination threshold. If it is, high-precision processing is performed; otherwise, simplified interpolation processing is performed. After lightweight preprocessing is completed, it is determined whether the climate downscaling preprocessing data load analysis is qualified. If not, a lightweight processing failure warning is sent; otherwise, secondary preprocessing is performed and model adaptability analysis is performed.

[0035] Furthermore, the specific process of lightweight preprocessing is as follows: Lightweight preprocessing evaluation data items are obtained, including the total amount of climate downscaling preprocessed data, the proportion of climate variable fluctuations, and the proportion of climate variable extreme biases. The proportion of climate variable fluctuations is used to assess the intensity of spatial fluctuations in climate variables and to assist in quantifying the severity of climate characteristic changes. It is represented by the proportion of the standard deviation of the climate variable for each date within the preset data processing time period to the preset climate variable value. The standard deviation of the climate variable is obtained by taking the arithmetic mean of the squared differences between the observed climate variable values ​​and the preset climate variable values ​​in all raster grids within the target area on a single date within the preset data processing time period. The square root of the mean is used to represent the standard deviation of climate variables, which reflects the spatial dispersion of daily climate variables. Preset climate variable values ​​are represented by the average of historical climate variable observations over a given period. These observations are derived from the training set of climate variable downscaling data collected beforehand to train the preset deep learning model; for example, temperature observations are represented by the daily average temperature in degrees Celsius. The extreme deviation ratio of climate variables is used to assess the degree to which extreme values ​​deviate from the baseline. It is calculated by comparing the absolute difference between the extreme values ​​of climate variables (such as daily maximum temperature, daily maximum precipitation, etc.) for each date within a preset data processing period and the preset climate variable values. The results of the value quantification by proportion are represented; the results of weighted coupling processing of the lightweight preprocessing assessment data items and the corresponding preprocessing assessment parameter influence values ​​are used as the climate element fluctuation index to quantify the severity of climate characteristic changes within the preset data processing time period; the larger the total amount of climate downscaling preprocessed data, the wider the data coverage or time span, and the higher the contribution to the index under the same fluctuation characteristics. When the proportion of climate variable fluctuation increases, it indicates that the spatial dispersion of climate variables in the target area is intensified, directly driving the increase of the proportion of extreme deviation of climate variables. The increase of the proportion of extreme deviation of climate variables indicates that the deviation of extreme climate values ​​from the baseline level is more significant. Any one of the three Individual increases in any indicator will lead to an increase in the climate element volatility index, weighted by the preprocessing assessment parameters. The weighted coupling process involves performing multiplication followed by summation to transform scattered, multi-dimensional data into a single, comprehensive quantitative value. The preprocessing assessment parameters include the total amount of climate downscaling preprocessed data (reflecting its influence on the climate element volatility index), the proportion of climate variable fluctuations (reflecting its influence on the climate element volatility index), and the proportion of extreme deviations in climate variables (reflecting its influence on the climate element volatility index).The system determines whether the climate element fluctuation index exceeds a preset critical period discrimination threshold. If so, the climate-sensitive period is marked as a climate critical period, and high-precision processing is performed on the climate critical period using bilinear interpolation. A climate-sensitive period refers to the time period corresponding to a single date extracted from the climate downscaling preprocessed data within a preset data processing period. Conversely, if the climate element fluctuation index is not greater than a preset critical period, the time period corresponding to the single date is marked as a climate non-critical period, and simplified interpolation processing is performed on the climate non-critical period using nearest neighbor interpolation to reduce the computational load of non-critical period preprocessing and balance data processing efficiency with core accuracy requirements. The preset critical period discrimination threshold is represented by the average value of the climate element fluctuation index over historical time periods. After lightweight preprocessing, the total amount of climate downscaling preprocessed data is reacquired. If the total amount of climate downscaling preprocessed data is still greater than a preset data volume threshold, a lightweight processing failure warning is sent. Otherwise, the lightweight preprocessed climate downscaling data undergoes secondary preprocessing to obtain qualified climate downscaling preprocessed data, and model adaptability analysis is performed. Secondary preprocessing refers to other conventional preprocessing besides interpolation preprocessing.

[0036] It is important to note that the lightweight preprocessing and model fit analysis involve a set of weighted parameters used to quantify the influence of each parameter on the corresponding assessment results. These weighted parameters are pre-set by technical personnel and stored in a database to provide a basis for matching the weighted coefficients between the climate element fluctuation index and the model's multi-dimensional core assessment indicators.

[0037] Specifically, for example, first extract a large amount of actual historical processing data under climate downscaling scenarios, covering parameter combinations for two core assessment scenarios (such as, in the lightweight preprocessing scenario, the numerical combinations of the total amount of climate downscaling preprocessed data and the corresponding total amount of climate downscaling preprocessed data, the influence of the proportion of climate variable fluctuations and the corresponding proportion of climate variable fluctuations, and the influence of the proportion of extreme deviations of climate variables and the corresponding proportion of extreme deviations of climate variables; in the model fit analysis scenario, the value combinations of model training convergence indicators and the corresponding model training convergence weights, model prediction error indicators and the corresponding prediction accuracy weights, and data fit goodness of fit indicators and the corresponding data fit weights), and provide input parameters for each scenario. The data is assigned a weighted quantitative value based on its influence on the corresponding assessment results (climate element fluctuation index, multi-dimensional adaptability index). The actual effective values ​​of the total influence value of climate downscaling preprocessed data, the influence value of the proportion of climate variable fluctuation, the influence value of the proportion of extreme deviation of climate variables, the model training convergence weight, the prediction accuracy weight, and the data fitting weight are recorded simultaneously under each historical scenario. Then, through correlation analysis (such as Spearman correlation coefficient), abnormal correlation data caused by sudden climate events, monitoring equipment failures, and temporary adaptation bias of preprocessing algorithms are removed. The correspondence between statistically significant parameter combinations and weight coefficients is retained. Finally, all effective data are integrated to form a mapping group containing multiple mapping sets.

[0038] Specifically, the association rules within the mapping group represent the weight coefficient proportions using a numerical range of 0 to 1, achieving a one-to-one mapping or many-to-one adaptation between input parameter combinations and corresponding weight coefficients in different assessment scenarios. When the system performs lightweight preprocessing and model adaptability analysis, it can quickly retrieve the matching weight coefficients from the mapping group, accurately quantify the degree of influence of each parameter on the assessment conclusion, and thus ensure the accuracy and reliability of the climate element fluctuation index and multi-dimensional adaptability indicators.

[0039] In this embodiment, lightweight preprocessing enables precise and differentiated simplification of climate downscaling preprocessed data, effectively resolving the contradiction between computational power consumption and accuracy assurance in large-scale data processing. While ensuring data quality in core scenarios such as drastic climate change, it reduces the computational power consumption and time consumption of preprocessing non-critical data, effectively alleviating the computational pressure and preprocessing time caused by excessive data volume. This provides crucial support that combines efficiency and quality for subsequent model adaptation and high-precision climate downscaling simulation.

[0040] Furthermore, the specific process of model adaptability analysis is as follows: Qualified climate downscaling preprocessed data is input into a preset deep learning model (such as a residual channel attention network model, an efficient sub-pixel convolutional network model, etc.) for model training, and the training time of the preset deep learning model is obtained. It is determined whether the training time of the preset deep learning model is less than a preset training time threshold. If so, the corresponding preset deep learning model is marked as a qualified deep learning model. Qualified deep learning models are used to adapt to climate downscaling tasks, and scenario commonality analysis is performed based on qualified deep learning models. Otherwise, a model training failure warning is sent. Scenario commonality analysis is used to identify the core climate variable correlation logic and model parameter configuration experience that can be reused across scenarios, improving the downscaling adaptability of deep learning models to multiple types of climate scenarios. The preset training time threshold is represented by the average training time of the preset deep learning model over a historical time period.

[0041] In this embodiment, model adaptability analysis enables a rapid and effective determination of the adaptability of a preset deep learning model to the climate downscaling task. By screening qualified deep learning models that meet the training time requirements, it ensures that subsequent scenario commonality analysis can be based on a reliable model foundation, reducing the impact of low model training efficiency on the timeliness of the downscaling task. Furthermore, by sending failure warnings to models that exceed the training time requirements, it promptly terminates subsequent processes for incompatible models, reducing unnecessary computing power consumption and time consumption. From the model level, this lays the foundation for improving the downscaling adaptability of various climate scenarios. Overall, while ensuring model training efficiency, it enhances the timeliness of response and the reliability of adaptability to the climate downscaling task.

[0042] As an embodiment of the second aspect, the specific process of model fit analysis is as follows: Qualified climate downscaling preprocessed data is input into a pre-set deep learning model for model training, and multi-dimensional core evaluation indicators of the model are obtained. These multi-dimensional core evaluation indicators include a model training convergence indicator to measure the stability and convergence effect of the pre-set deep learning model training process; a model prediction error indicator to quantify the deviation between the predicted climate variable values ​​and the actual climate variable values; and a data fit goodness indicator to reflect the linear correlation between the predicted climate variable values ​​and the actual climate variable values. The influence of each evaluation indicator on the multi-dimensional fit indicators is then obtained. The multi-dimensional adaptability weight parameters include model training convergence weight, prediction accuracy weight, and data fitting weight. The result of weighted coupling of the model's multi-dimensional core evaluation indicators and corresponding multi-dimensional adaptability weight parameters is used as a multi-dimensional adaptability index to comprehensively evaluate the adaptability of the preset deep learning model to qualified climate downscaling preprocessed data. It is determined whether the model's multi-dimensional adaptability index is greater than the preset model adaptability threshold. If it is, a scenario commonality analysis is performed; otherwise, a model training failure alarm is sent. The preset model adaptability threshold is represented by the average value of the model's multi-dimensional adaptability index over a historical time period.

[0043] Specifically, the formula for the model training convergence metric is as follows:

[0044]

[0045] Where A represents the model training convergence index, L final L represents the loss function value for a single training iteration. stable The default loss function value is represented by L, and the default baseline constant is represented by Q. The single-cycle training loss function value is the output value of the default deep learning model at the end of training, used to quantify the final fit and error level at the end of training. The default loss function value is the average loss function value after the default deep learning model enters the convergence and stabilization phase, used to quantify the convergence of the model training. The default baseline constant is set in advance by a designated person to avoid errors caused by L. final equal to L stable When the denominator is in the ideal convergence state, it is meaningless.

[0046] Specifically, the formula for the model prediction error index is as follows:

[0047]

[0048] Where B represents the model prediction error index, RMSE represents the root mean square error of the actual values ​​of climate variables, MAE represents the mean absolute error of climate variables, and max(V) obs) represents the maximum value of the climate variable, min(V obs The minimum value of a climate variable is represented by the root mean square error of the actual value of the climate variable, which is expressed as the root mean square error between the predicted climate variable value output during the training of the pre-set deep learning model and the actual value of the climate variable extracted from the test set of qualified climate downscaling preprocessed data. The mean absolute error of the climate variable is expressed as the mean absolute error calculated between the model's predicted climate variable value and the actual value of the climate variable, and is used to quantify the average deviation between the model's predicted value and the actual value. The maximum value of the climate variable is expressed as the maximum value of the actual value of the climate variable, and the minimum value of the climate variable is expressed as the minimum value of the actual value of the climate variable, and is used to normalize the accuracy index.

[0049] Specifically, the process for determining the goodness-of-fit index is as follows:

[0050]

[0051] Where C represents the goodness-of-fit index, R0 2 The climate variable correlation index is represented by the result of linear fitting between the model-predicted climate variable values ​​and the actual climate variable values ​​based on a linear regression analysis algorithm, and is used to reflect the goodness of fit of the data.

[0052] Specifically, the formula for the multi-dimensional adaptability index is as follows:

[0053]

[0054] Wherein, W represents the multi-dimensional fit index, with a value ranging from 0 to 1. The closer to 1, the better the fit. w1 represents the model training convergence weight, w2 represents the prediction accuracy weight, and w3 represents the data fitting weight. w1, w2, and w3 are all preset by a designated person. The model training convergence weight is used to adjust the coefficient of influence of the preset deep learning model training convergence index among the multi-dimensional fit index. The prediction accuracy weight is used to adjust the coefficient of influence of the preset deep learning model prediction error index among the multi-dimensional fit index. The data fitting weight is used to adjust the coefficient of influence of the data fitting goodness index among the multi-dimensional fit index.

[0055] As an embodiment of the second aspect, in this example, by constructing a multi-dimensional and quantitative model adaptability analysis system, a comprehensive and accurate assessment of the adaptability of deep learning models to climate downscaling data is achieved. The constructed multi-dimensional adaptability index captures model performance characteristics from three key dimensions: training stability, prediction accuracy, and data matching degree, intuitively reflecting the strength of data correlation. It can not only screen out models with excellent adaptability to support subsequent scenario commonality analysis and provide a high-quality foundation for cross-scenario parameter reuse and model optimization, but also terminate the training process of models with insufficient adaptability in a timely manner through failure alarms, reducing the ineffective consumption of computing power and time, improving the objectivity and accuracy of model adaptability judgment, and laying a solid model foundation for subsequent high-precision climate downscaling simulation and multi-scenario adaptation.

[0056] Further, the specific process of scenario commonality analysis is as follows: First, obtain the baseline climate scenario and the climate scenario to be adapted. The baseline climate scenario represents the reference scenario in the climate downscaling task that has completed the training of the preset deep learning model (e.g., SSP2-4.5, a medium greenhouse gas emission scenario). The climate scenario to be adapted represents other scenarios that need to be adapted based on the baseline scenario model (e.g., SSP1-2.6, i.e., a low greenhouse gas emission scenario, SSP5-8.5, i.e., a high greenhouse gas emission scenario, etc.). Second, perform feature vector extraction: Input the qualified climate downscaling preprocessed data under the baseline scenario and the qualified climate downscaling preprocessed data under the scenario to be adapted into the qualified deep learning model under the baseline scenario, respectively, to obtain the baseline feature vector and the feature vector to be adapted. The baseline feature vector represents the shallow feature vector output by the shallow feature extraction layer of the model under the baseline scenario. The multi-dimensional vector is obtained by merging the residual feature vectors output by the quantity and the residual group; the feature vector to be adapted represents the multi-dimensional vector obtained by merging the shallow feature vectors output by the shallow feature extraction layer of the model under the scenario to be adapted with the residual feature vectors output by the residual group; the climate scenario parameter transfer adaptation index is obtained to evaluate the similarity of the feature vectors output by the feature extraction layer of the qualified deep learning model under different scenarios. The climate scenario parameter transfer adaptation index is represented by the ratio quantification result of the dot product of the baseline feature vector and the feature vector to be adapted and the product of the magnitudes of the two vectors; it is determined whether the climate scenario parameter transfer adaptation index is greater than the preset parameter sharing threshold. If it is, cross-scenario transfer optimization is adopted; otherwise, a model adaptation failure warning is sent. The preset parameter sharing threshold is represented by the average value of the climate scenario parameter transfer adaptation index over a historical period.

[0057] In this example, a cross-climate scenario adaptation evaluation system based on feature vector similarity was constructed through scenario commonality analysis. This system accurately judges the feasibility of model parameter reuse between the baseline scenario and the scenario to be adapted, effectively improving the efficiency and adaptation accuracy of multi-scenario climate downscaling tasks. By extracting multi-dimensional vectors that integrate shallow and residual features through a qualified deep learning model, the system captures the basic representation of climate data while retaining deep key features, reducing the omission of common information caused by single feature extraction, and enhancing the system's adaptation flexibility and downscaling reliability for complex climate scenarios.

[0058] Furthermore, the specific process of cross-scenario transfer optimization is as follows: The climate scenario parameter transfer adaptation index and the mean value of the climate variables in the scenario to be adapted are input into the linear mapping layer weight mapping set for querying to obtain the adjustment coefficient of the linear mapping layer weight matrix. When the mean value of the climate variables in the scenario to be adapted is greater than the mean value of the climate variables in the baseline scenario, the adjustment magnitude corresponding to the adjustment coefficient of the linear mapping layer weight matrix is ​​used as the adjustment step size, and the linear mapping layer weight values ​​of all elements in the linear mapping layer weight matrix are increased step by step (after each adjustment of the weight values ​​of all elements in the weight matrix, the mean square error of the test set of the scenario to be adapted and the ratio of the shortening of the training period are re-monitored, and cross-scenario transfer optimization is judged. If the cross-scenario transfer optimization is qualified, the adjusted weight values ​​are used as the initial values ​​for the next adjustment). This helps to reduce the impact of a single large adjustment. To mitigate model parameter oscillations or overfitting, the model is made more suitable for the overall high distribution of climate variables in the scenario to be adapted. Conversely, the adjustment step size is set by using the adjustment magnitude corresponding to the adjustment coefficient of the linear mapping layer weight matrix, and the weight values ​​of all elements in the weight matrix are gradually reduced (after each adjustment of the weight values ​​of all elements in the weight matrix, the mean square error of the test set of the scenario to be adapted and the ratio of the shortened training period are re-monitored, and cross-scenario transfer optimization is performed. If the cross-scenario transfer optimization is qualified, the adjusted weight values ​​are used as the initial values ​​for the next adjustment). This helps to weaken the model's excessive focus on high-value features, guides the model to adapt to the overall low level of climate variables in the scenario to be adapted, and prevents the loss of feature capture ability caused by a sudden drop in weights through step-by-step reduction of adjustment, thus ensuring the prediction accuracy under low-value climate scenarios.

[0059] Specifically, the mean value of climate variables in the scenario to be adapted is represented by the arithmetic mean of all core climate variables in the qualified climate downscaling preprocessed data under the scenario to be adapted, after standardization, to reflect the overall numerical level of climate variables under that scenario. The mean value of climate variables in the baseline scenario is represented by the arithmetic mean of all core climate variables in the qualified climate downscaling preprocessed data under the baseline scenario, after standardization, to serve as a benchmark reference level for scenario numerical comparison, to quantify the direction and degree of deviation between the scenario to be adapted and the baseline scenario in terms of the overall level of climate variables, ensuring the pertinence and rationality of cross-scenario parameter transfer. The mean square error of the test set and the proportion of training cycle shortening in the scenario to be adapted are continuously monitored, and cross-scenario transfer optimization is judged. When the cross-scenario transfer optimization is qualified, the qualified deep learning model corresponding to the qualified cross-scenario transfer optimization is marked as the qualified deep learning model of the scenario to be adapted, and high-resolution climate change scenario prediction is performed. Otherwise, cross-scenario transfer optimization continues to be performed. When the number of cross-scenario transfer optimization executions exceeds the preset maximum number of scenario transfer optimizations, if the cross-scenario transfer optimization judgment is unqualified, a cross-scenario transfer failure warning is sent. The preset maximum number of scenario transfer optimizations is set in advance by preset personnel.

[0060] Specifically, the cross-scenario transfer optimization discrimination means that if the mean square error of the test set of the scenario to be adapted is less than the preset transfer accuracy threshold and the training cycle reduction ratio is greater than the preset transfer efficiency threshold, then the cross-scenario transfer optimization is deemed qualified; otherwise, it is deemed unqualified. The preset transfer accuracy threshold is represented by the average mean square error of the test set of the scenario to be adapted over a historical period, and the preset transfer efficiency threshold is represented by the average training cycle reduction ratio over a historical period. The training cycle reduction ratio is calculated by subtracting the time taken to complete the training of the preset deep learning model under the baseline scenario from the time taken to complete the fine-tuning training of the preset deep learning model under the scenario to be adapted, and then quantifying the result by comparing this difference with the time taken to complete the training of the preset deep learning model under the baseline scenario. The high-resolution climate change scenario prediction means that the process of inputting low-resolution climate simulation data under the scenario to be adapted into a qualified deep learning model under the scenario to be adapted, and outputting high-resolution climate change scenario prediction data.

[0061] It should be noted that before designing the deep learning-based intelligent climate downscaling system provided in this application, a mapping set is established that associates climate scenario characteristic parameters with the linear mapping layer weight adjustment coefficients. This set is pre-set by designated personnel and stored in a database. It is used to reflect the precise correspondence between the input parameters (climate scenario parameter migration adaptation index, mean of climate variables in the scenario to be adapted) and the linear mapping layer weight matrix adjustment coefficients related to cross-scenario parameter migration under different climate scenarios. This provides a directly queryable decision basis for the dynamic adjustment of weights during the cross-scenario migration optimization process.

[0062] In this embodiment, the established mapping set is a linear mapping layer weight mapping set. This mapping set is gradually constructed through statistical analysis and parameter verification of historical climate downscaling cross-scenario migration optimization scenarios: First, a large amount of specific combination data of input parameters and output adjustment coefficients corresponding to the linear mapping layer weight mapping set in actual optimization scenarios are extracted, including the climate scenario parameter migration adaptation index, the mean of climate variables in the scenario to be adapted, and the combination of adjustment coefficients of the corresponding linear mapping layer weight matrix. Each input parameter is assigned a weight value based on its influence on the output adjustment coefficient. The actual effective weight adjustment results under each scenario are recorded simultaneously. Then, through correlation analysis (such as Spearman rank correlation coefficient analysis), abnormal correlation data caused by temporary failure of monitoring equipment and sudden abnormality of climate data are removed. The correspondence between statistically significant input parameter combinations and output adjustment coefficients is retained. Finally, all effective data are integrated to form a linear mapping layer weight mapping set in which the corresponding adjustment coefficient can be obtained directly by querying the input parameters.

[0063] In this embodiment, cross-scenario transfer optimization achieves accurate transfer and dynamic optimization of model parameters between the baseline scenario and the scenario to be adapted. This provides a reasonable solution to the problem of low timeliness of high-resolution climate change scenario prediction data caused by excessive computational load in traditional statistical downscaling methods. By adjusting the weight matrix of the linear mapping layer in a stepwise manner and verifying it in real time, only targeted fine-tuning of the model parameters is required, which greatly reduces the amount of data computation and verification cost in multiple scenarios in a single region. It avoids repetitive and computationally intensive steps such as full-time raster data processing and interpolation error correction in traditional methods, significantly reducing the demand for computing resources and ensuring a significant improvement in the training efficiency of the model for the scenario to be adapted. It effectively compresses the entire process time from data input to high-resolution prediction result output, improves the problem of poor timeliness of prediction data caused by the surge in computational load in traditional methods, and provides timely and accurate high-resolution climate change scenario support for various practical application scenarios.

[0064] like Figure 4The diagram shown is a structural schematic of the intelligent climate downscaling system based on deep learning provided in this application embodiment. This intelligent climate downscaling system based on deep learning applies an intelligent climate downscaling method based on deep learning and is characterized by including: an auxiliary feature redundancy monitoring module, a climate downscaling data load monitoring module, a model adaptability monitoring module, and a scenario adaptability monitoring module. Specifically, the auxiliary feature redundancy monitoring module is used to determine whether auxiliary feature de-redundancy processing is needed during the intelligent climate downscaling process based on deep learning; the climate downscaling data load monitoring module is used to analyze the climate downscaling data load after the climate downscaling data redundancy determination is passed, to determine whether lightweight preprocessing is needed; the model adaptability monitoring module is used to perform model adaptability analysis after the climate downscaling data load analysis is passed, to determine whether scenario commonality analysis is needed; and the scenario adaptability monitoring module is used to send a model training failure warning if scenario commonality analysis is not needed, and otherwise, determine whether cross-scenario transfer optimization is needed based on the results of scenario commonality analysis.

[0065] In this embodiment, the auxiliary feature redundancy monitoring module, the climate downscaling data load monitoring module, the model adaptability monitoring module, and the scenario adaptability monitoring module are progressively linked and work in synergy to construct a dynamic monitoring and control system for the entire intelligent climate downscaling process. The auxiliary feature redundancy monitoring module provides low-redundancy, high-value basic data for subsequent modules by eliminating low-correlation auxiliary features. Its judgment result directly determines the quality of the analysis object of the climate downscaling data load monitoring module. If redundancy elimination is incomplete, it can lead to the data load analysis falling into the trap of ineffective computational power assessment. The climate downscaling data load monitoring module undertakes... The preceding climate downscaling preprocessing data balances data efficiency and accuracy through differentiated preprocessing strategies, providing lightweight and high-quality input data for the model adaptability monitoring module. The model adaptability monitoring module then conducts adaptation assessments based on the high-quality preprocessed data. As the terminal monitoring link, the scenario adaptability monitoring module not only connects the model adaptability results to achieve accurate early warnings, but also provides direction for cross-scenario migration optimization through scenario commonality analysis. The four modules support each other and progress in a layered manner, improving the stability, efficiency, and adaptability flexibility of the intelligent climate downscaling system, and solving the problem of low timeliness of high-resolution climate change scenario prediction data output.

[0066] In summary, redundancy assessment of climate downscaling preprocessed data to determine whether auxiliary feature de-redundancy processing is needed helps filter out invalid data interference at the source, providing a high-quality data foundation for subsequent climate downscaling task data preprocessing and model training. This reduces the computational resources consumed by redundant features or misleading the model in learning core climate laws. After passing the climate downscaling preprocessed data redundancy assessment, climate downscaling preprocessed data load analysis is performed to determine whether lightweight preprocessing is needed. This helps to accurately match data processing resources with data scale and reduce redundant consumption of computational resources. After passing the climate downscaling preprocessed data load analysis, model adaptability analysis is performed to determine whether scenario commonality analysis is needed. This helps to identify model and preprocessed data compatibility defects in advance and reduce subsequent meaningless model training processes. If scenario commonality analysis is not needed, a model training failure warning is sent. Otherwise, based on the results of scenario commonality analysis, it is determined whether cross-scenario transfer optimization is needed, which helps to enhance the adaptability of deep learning models to complex climate scenarios.

[0067] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based intelligent climate downscaling method, characterized in that, The method includes: In the process of intelligent climate downscaling based on deep learning, redundancy judgment is performed on climate downscaling preprocessing data to determine whether auxiliary feature redundancy removal processing is needed. The auxiliary feature redundancy removal processing is used to reduce data dimensionality and storage load, and improve the feature effectiveness of input data. After the redundancy of the climate downscaling preprocessed data passes the redundancy check, a climate downscaling preprocessed data load analysis is performed to determine whether lightweight preprocessing is required. The lightweight preprocessing is used to balance data processing efficiency and the reliability of downscaling results. After the climate downscaling preprocessed data load analysis is qualified, model fit analysis is performed to determine whether scenario commonality analysis is needed. If scenario commonality analysis is not required, a model training failure warning is sent; otherwise, a decision is made based on the results of scenario commonality analysis as to whether cross-scenario transfer optimization is needed. The cross-scenario transfer optimization is used to improve the downscaling adaptation accuracy and training efficiency of deep learning models for multiple types of climate scenarios. The specific process of the model fit analysis is as follows: Input qualified climate downscaling preprocessed data into a preset deep learning model for model training, and obtain the training time of the preset deep learning model; Determine whether the training time of the preset deep learning model is less than the preset training time threshold. If so, mark the corresponding preset deep learning model as a qualified deep learning model and perform a common scenario analysis based on the qualified deep learning model. Otherwise, send a model training failure warning. The scenario commonality analysis is used to identify the core climate variable correlation logic and model parameter configuration experience that can be reused across scenarios, thereby improving the downscaling adaptation capability of deep learning models to multiple types of climate scenarios. The qualified deep learning model is used to adapt to climate downscaling tasks; The specific process of the scenario commonality analysis is as follows: The qualified climate downscaling preprocessed data under the baseline scenario and the qualified climate downscaling preprocessed data under the scenario to be adapted are respectively input into the qualified deep learning model under the baseline scenario to obtain the baseline feature vector and the feature vector to be adapted. The baseline feature vector represents a multi-dimensional vector obtained by merging the shallow feature vector output by the shallow feature extraction layer of the model under the baseline scenario with the residual feature vector output by the residual group. The feature vector to be adapted represents a multi-dimensional vector obtained by merging the shallow feature vector output by the shallow feature extraction layer of the model under the scenario to be adapted with the residual feature vector output by the residual group. Obtain the climate scenario parameter transfer adaptation index, which is used to evaluate the similarity of the output feature vectors of the feature extraction layer of a qualified deep learning model under different scenarios. Determine whether the climate scenario parameter migration adaptation index is greater than the preset parameter sharing threshold. If it is, then take cross-scenario migration optimization measures; otherwise, send a model adaptation failure warning.

2. The intelligent climate downscaling method based on deep learning according to claim 1, characterized in that, The specific process for redundancy determination of the climate downscaling preprocessed data is as follows: Obtain a preprocessing data redundancy index to determine the degree of redundancy in climate downscaling preprocessed data; The preprocessed data redundancy index is represented by the quantitative result of the redundancy of auxiliary features in the climate downscaling preprocessed data; Determine whether the redundancy index of the preprocessed data is less than the preset feature redundancy threshold; If so, then perform climate downscaling preprocessing data load analysis; otherwise, take auxiliary feature redundancy removal processing. The climate downscaling preprocessing data load analysis is used to reduce the computational power consumption and preprocessing time caused by the large amount of data in climate downscaling preprocessing data.

3. The intelligent climate downscaling method based on deep learning according to claim 2, characterized in that, The specific process of redundancy removal for auxiliary features is as follows: When the redundancy index of preprocessed data is detected to be greater than the preset absolute redundancy threshold, the corresponding auxiliary feature field is marked as an absolute redundancy feature, and a prompt is sent to the preset personnel to delete the corresponding absolute redundancy feature and associated storage module. When the preprocessed data redundancy index is detected to be no greater than the preset absolute redundancy threshold, the corresponding auxiliary feature field is marked as a non-redundant auxiliary feature, and the non-redundant auxiliary feature is dimensionally fused to reduce data dimensional redundancy. After the auxiliary feature redundancy removal process is completed, the redundancy index of the preprocessed data is reacquired. If the redundancy index of the preprocessed data is still not greater than the preset feature redundancy threshold, a redundancy removal failure warning is sent; otherwise, climate downscaling preprocessed data load analysis is performed.

4. The intelligent climate downscaling method based on deep learning according to claim 3, characterized in that, The specific process of the climate downscaling preprocessing data load analysis is as follows: Obtain the total amount of climate downscaling preprocessed data to reflect the potential impact of the data size in the climate downscaling preprocessing stage on the pressure on computing resources. Determine whether the total amount of climate downscaling preprocessed data exceeds the preset data volume threshold; if so, adopt lightweight preprocessing. If the total amount of climate downscaling preprocessed data is not greater than the preset data volume threshold, then conventional climate downscaling preprocessing is adopted. The climate downscaling preprocessed data obtained after conventional climate downscaling preprocessing is marked as qualified climate downscaling preprocessed data, and model adaptability analysis is performed to evaluate whether the preset deep learning model has the ability to effectively downscale and process efficiently. The conventional climate downscaling preprocessing refers to performing a standardized basic preprocessing procedure on climate downscaling preprocessed data. The conventional climate downscaling preprocessing is used to reduce data anomaly bias, unify data accuracy standards, ensure the spatial continuity of data throughout the time period, and reduce the risk of model training errors.

5. The intelligent climate downscaling method based on deep learning according to claim 4, characterized in that, The specific process of the lightweight preprocessing is as follows: Obtain lightweight preprocessing assessment data items, which include the total amount of climate downscaling preprocessed data, the proportion of climate variable fluctuations, and the proportion of climate variable extreme biases; The percentage of climate variable fluctuations is used to assess the intensity of spatial fluctuations in climate variables and to help quantify the severity of changes in climate characteristics. The percentage of extreme deviations of climate variables is used to assess the degree to which extreme values ​​of climate variables deviate from the baseline. The result of weighted coupling of lightweight preprocessing assessment data items and corresponding preprocessing assessment parameter influence values ​​is used as a climate element fluctuation index to quantify the severity of climate characteristic changes within a preset data processing time period. The preprocessing assessment parameters include the total amount of climate downscaling preprocessed data, which reflects the influence of the total amount of climate downscaling preprocessed data on the climate element fluctuation index; the influence of the proportion of climate variable fluctuation, which reflects the influence of the proportion of climate variable fluctuation on the climate element fluctuation index; and the influence of the proportion of climate variable extreme deviation, which reflects the influence of the proportion of climate variable extreme deviation on the climate element fluctuation index. Determine whether the climate element fluctuation index is greater than the preset critical period discrimination threshold. If so, mark the climate highly sensitive period as the climate critical period and perform high-precision processing on the climate critical period. Conversely, the time period corresponding to the single date is marked as a non-critical climate period, and simplified interpolation processing is performed on the non-critical climate period based on nearest neighbor interpolation to reduce the preprocessing computation load of the non-critical period and balance data processing efficiency and core accuracy requirements. After the lightweight preprocessing is completed, the total amount of climate downscaling preprocessed data is reacquired. If the total amount of climate downscaling preprocessed data is still greater than the preset data volume threshold, a lightweight processing failure warning is sent. Otherwise, the climate downscaling data after lightweight preprocessing is subjected to secondary preprocessing to obtain qualified climate downscaling preprocessed data, and model fit analysis is performed.

6. The intelligent climate downscaling method based on deep learning according to claim 5, characterized in that, The specific process of the model fit analysis is as follows: Qualified climate downscaling preprocessed data is input into a preset deep learning model for model training, and multi-dimensional core evaluation indicators of the model are obtained. The multi-dimensional core evaluation indicators of the model include the model training convergence index, which measures the stability and convergence effect of the training process of the preset deep learning model; the model prediction error index, which quantifies the degree of deviation between the predicted climate variable values ​​and the actual climate variable values ​​of the preset deep learning model; and the data fit index, which reflects the strength of the linear correlation between the predicted climate variable values ​​and the actual climate variable values. Obtain multi-dimensional adaptability weight parameters; The result of weighting and coupling the multi-dimensional core evaluation indicators of the model with the corresponding multi-dimensional adaptability weight parameters is used as a multi-dimensional adaptability indicator for comprehensively evaluating the adaptability of the preset deep learning model with qualified climate downscaling preprocessed data. If the multi-dimensional adaptability index of the model is greater than the preset model adaptability threshold, a scenario commonality analysis is performed; otherwise, a model training failure alarm is sent.

7. The intelligent climate downscaling method based on deep learning according to claim 1, characterized in that, The specific process of cross-scenario migration optimization is as follows: The climate scenario parameter migration adaptation index and the mean value of the climate variable of the scenario to be adapted are input into the linear mapping layer weight mapping set for querying to obtain the adjustment coefficient of the linear mapping layer weight matrix. When the mean value of the climate variables in the scenario to be adapted is greater than the mean value of the climate variables in the baseline scenario, the adjustment range corresponding to the adjustment coefficient of the linear mapping layer weight matrix is ​​used as the adjustment step size, and the weight values ​​of all elements in the weight matrix are increased step by step. Conversely, the adjustment step size is taken as the adjustment magnitude corresponding to the adjustment coefficient of the linear mapping layer weight matrix, and the weight values ​​of all elements in the weight matrix are gradually reduced. The mean values ​​of climate variables under the scenario to be adapted are used to reflect the overall numerical level of climate variables under that scenario. The mean value of the climate variables in the baseline scenario is represented by the arithmetic mean of all core climate variables in the qualified climate downscaling preprocessed data under the baseline scenario. It serves as a benchmark reference level for scenario numerical comparison and is used to quantify the direction and degree of deviation between the scenario to be adapted and the baseline scenario at the overall level of climate variables, ensuring the pertinence and rationality of cross-scenario parameter transfer. The mean square error of the test set for the scenario to be adapted and the ratio of the shortened training cycle are continuously monitored, and cross-scenario transfer optimization is judged. When the cross-scenario transfer optimization is qualified, the qualified deep learning model corresponding to the cross-scenario transfer optimization is marked as a qualified deep learning model for the scenario to be adapted, and high-resolution climate change scenario prediction is performed. Otherwise, cross-scenario transfer optimization is continued. When the number of cross-scenario transfer optimization executions exceeds the preset maximum number of scenario transfer optimizations, if the cross-scenario transfer optimization judgment is unqualified, a cross-scenario transfer failure warning is sent. If the mean square error of the test set of the scenario to be adapted is less than the preset transfer accuracy threshold and the shortening ratio of the training cycle is greater than the preset transfer efficiency threshold, then the cross-scenario transfer optimization is deemed qualified; otherwise, the cross-scenario transfer optimization is deemed unqualified. The high-resolution climate change scenario prediction refers to the process of inputting low-resolution climate simulation data under the scenario to be adapted into a qualified deep learning model of the scenario to be adapted, and outputting high-resolution climate change scenario prediction data.

8. A deep learning-based intelligent climate downscaling system, employing the deep learning-based intelligent climate downscaling method as described in any one of claims 1-7, characterized in that, include: The modules include: auxiliary feature redundancy monitoring module, climate downscaling data load monitoring module, model adaptability monitoring module, and scenario adaptability monitoring module. Among them, the auxiliary feature redundancy monitoring module is used to determine the redundancy of climate downscaling preprocessing data in the process of intelligent climate downscaling based on deep learning, so as to determine whether auxiliary feature de-redundancy processing is needed. The climate downscaling data load monitoring module is used to perform climate downscaling preprocessing data load analysis after the redundancy judgment of climate downscaling preprocessing data is qualified, in order to determine whether lightweight preprocessing is required. The model fit monitoring module is used to perform model fit analysis after the climate downscaling preprocessed data load analysis is qualified, in order to determine whether scenario commonality analysis is needed. The scenario adaptability monitoring module is used to send a warning of model training failure if scenario commonality analysis is not required, and otherwise, it determines whether cross-scenario transfer optimization is needed based on the results of scenario commonality analysis.

Images