A method for rapid climate prediction of future scenarios in small watersheds that integrates deep learning and dynamic downscaling

Abstract

This invention relates to the field of climate prediction and meteorological data downscaling technology, providing a rapid method for predicting future climate scenarios in small watersheds by integrating deep learning and dynamic downscaling. The method includes: Step 1: Target area delineation and input data preprocessing; Step 2: Dynamic downscaling simulation and annotation, and benchmark data generation; Step 3: Training sample construction and TA-UNet downscaling surrogate model training; Step 4: Future climate downscaling and prediction dataset generation. This invention addresses the core contradiction in existing technologies that cannot simultaneously achieve prediction accuracy, physical consistency, and computational efficiency. It provides a future climate prediction scheme for high-altitude small watersheds, constructs a standardized, end-to-end technical system adapted to high-altitude cold mountain scenarios, and enables the rapid and stable generation of high-resolution future climate prediction data for high-altitude glacial small watersheds. This provides reliable and refined meteorological data support for geological and hydrological disaster risk assessment and major engineering construction in high-altitude cold mountain areas.

Description

Technical Field

[0001] This invention relates to the field of climate prediction and meteorological data downscaling technology, and in particular to a method for rapid climate prediction of future scenarios in small watersheds that integrates deep learning and dynamic downscaling. Background Technology

[0002] Long-term global climate prediction mainly relies on coarse-resolution data output from global coupled climate models. However, the spatial scale of small watersheds in high-altitude mountainous areas, where disasters such as flash floods and debris flows are frequent, is much smaller than the grid resolution of global climate models. Coarse-resolution climate data cannot directly support disaster hazard analysis, risk assessment, and engineering applications in small watersheds in mountainous areas. Therefore, downscaling is necessary to achieve refined adaptation of climate data. Currently, the mainstream climate data downscaling techniques are mainly divided into three categories: statistical downscaling, dynamic downscaling, and machine learning downscaling.

[0003] Statistical downscaling is currently the most widely used downscaling method. Its core is to achieve spatial downscaling of low-resolution climate data through bias correction and linear interpolation. It has the advantages of simple calculation process and high computational efficiency. However, this method relies entirely on the linear transformation of low-resolution data and completely ignores the nonlinear effects of complex terrain and differentiated underlying surfaces on local climate change. The prediction error is large in complex terrain areas such as high-altitude glacier-covered areas and river valleys, which cannot meet the accuracy requirements of small watershed-scale engineering applications.

[0004] Dynamic downscaling methods, using regional climate models at the convective permissible scale, simulate atmospheric and land surface processes based on atmospheric physics equations. This results in high-resolution meteorological data with clearly defined physical mechanisms, accurately capturing local climate characteristics under complex terrain, and demonstrating significant advantages in physical interpretability and prediction accuracy. However, this method consumes enormous computational resources and has extremely long computation cycles. For climate prediction across multiple scenarios on a century-scale, its time and economic costs are too high to meet the timeliness and cost control requirements of engineering applications.

[0005] Machine learning downscaling is a new paradigm of downscaling technology that has emerged in recent years. Its core principle is to efficiently downscale climate data by learning the nonlinear mapping relationship between high-resolution observational data and low-resolution model outputs, offering significant efficiency advantages compared to traditional methods. However, existing machine learning downscaling techniques are mostly fragmented model optimization studies, lacking a standardized and implementable end-to-end technical system adapted to high-altitude glacier watershed scenarios. Furthermore, existing deep learning downscaling models do not fully consider the differentiated effects of complex mountain terrain and underlying surfaces, failing to guarantee the rationality of key atmospheric physical processes such as water vapor and energy. This results in insufficient accuracy in representing core meteorological elements such as temperature and precipitation in high-altitude mountainous areas, directly affecting the reliability of subsequent hydrological simulations and disaster risk assessments.

[0006] Overall, existing climate downscaling techniques cannot simultaneously achieve prediction accuracy, physical consistency, and computational efficiency. For the special application scenario of small watersheds in high-altitude glaciers, there is a lack of a mature, standardized, and engineering-practical climate prediction technology solution. Summary of the Invention

[0007] This invention provides a rapid climate prediction method for small watershed future scenarios that integrates deep learning and dynamic downscaling. It addresses the core contradiction in existing technologies that cannot balance prediction accuracy, physical consistency, and computational efficiency. It offers a climate prediction scheme for high-altitude small watersheds, constructs a standardized full-process technical system adapted to high-altitude cold mountain scenarios, and enables the rapid and stable generation of high-resolution future climate prediction data for high-altitude glacial small watersheds. This provides reliable and refined meteorological data support for geological and hydrological disaster risk assessment and major engineering construction in high-altitude cold mountain areas.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: A method for rapid climate prediction of future scenarios in small watersheds that integrates deep learning and dynamic downscaling includes the following steps: Step 1: Define the latitude and longitude range of the target high mountain watershed, and acquire low-resolution historical and future scenario meteorological data, atmospheric reanalysis driving field data, and high-resolution static geographic data from the global climate model output. Perform spatial grid alignment and normalization preprocessing on all acquired data to obtain input data with a uniform format. Step 2: Using the WRF-CPM convection-allowable scale regional climate model optimized for ice and snow processes, input preprocessed atmospheric reanalysis driving field data and high-resolution static geographic data, perform dynamic downscaling simulation on the target area to generate high-resolution meteorological field data for historical periods. This data is used as high-resolution labeled data for training the deep learning model, and a high-resolution baseline meteorological field is calculated based on this data. Step 3: The preprocessed low-resolution historical meteorological data is interpolated and aligned to the grid of high-resolution labeled data, forming a training sample pair with the high-resolution labeled data and the preprocessed high-resolution static geographic data; a TA-UNet network model with a terrain attention mechanism is constructed, in which a terrain attention gate is set at the jump connection between the encoder and decoder of the TA-UNet network model, and the terrain attention gate uses high-resolution static geographic data as the guiding signal; a total loss function composed of data loss and physical constraint loss is designed; the TA-UNet network model is trained using the training sample pair and based on the total loss function until the model converges, resulting in a trained downscaled surrogate model; Step 4: Input the preprocessed low-resolution future scenario meteorological data into the trained downscaling surrogate model and output the high-resolution future climate downscaling results; use the high-resolution labeled data as a benchmark to correct the bias of the high-resolution future climate downscaling results; superimpose the daily climate change amount of the corrected downscaling results onto the high-resolution benchmark meteorological field to generate a high-resolution daily future climate prediction dataset for the target high-altitude small watershed.

[0009] In this specification, the low-resolution historical and future scenario meteorological data output by the global climate model include daily-scale data of temperature, precipitation, relative humidity, specific humidity, wind field, longwave radiation, and shortwave radiation. The low-resolution historical and future scenario meteorological data are derived from the original CMIP6 global coupled climate model data or from publicly available CMIP6 derivative datasets that have undergone bias correction and preliminary downscaling.

[0010] In this specification, the WRF-CPM convection-allowable scale regional climate model optimized for snow and ice processes adopts the Noah-MP land surface model specifically optimized by snow decay curves and snow albedo schemes. During the simulation, the topographic turbulence drag scheme is enabled, and the cumulus parameterization scheme is disabled for the inner high-resolution grid.

[0011] In this specification, the historical period of the dynamic downscaling simulation is no less than 20 years, and a model warm-up period of no less than one month is set during the simulation process. The output high-resolution meteorological field data includes daily meteorological variables corresponding to the low-resolution historical period and future scenario meteorological data, as well as sensible heat flux, latent heat flux, surface upward radiation and surface temperature.

[0012] In this specification, the high-resolution static geographic data includes digital elevation models, slope, aspect, topographic shadow, land use type, and vegetation cover data; the topographic attention gate uses the above-mentioned high-resolution static geographic data as a guiding signal to dynamically calculate attention weights and adaptively enhance the model's learning of meteorological characteristics of topographically sensitive areas and different land cover areas.

[0013] In this specification, the physical constraint loss in the total loss function includes surface energy balance constraint, radiation-temperature correspondence constraint, inherent consistency constraint of temperature and humidity variables, and coupling constraint of precipitation and relative humidity. The weight of each constraint can be adjusted according to the meteorological characteristics of the target area.

[0014] In this manual, during the model training process, a seasonal independent training strategy is adopted for precipitation variables to capture the precipitation variation characteristics of the target area in different seasons. The model training adopts a phased training strategy. In the first phase, only the data loss is used to train the backbone network of the model. In the second phase, physical constraint loss is added to optimize the overall parameters of the model.

[0015] In this specification, the bias correction adopts the quantile mapping method or the linear translation-variance scaling method. Based on the high-resolution labeled data, the high-resolution future climate downscaling results are corrected month by month to align the statistical characteristics of the corrected data with the high-resolution labeled data.

[0016] In this manual, after generating a high-resolution daily future climate prediction dataset for the target high-altitude small watershed, the downscaling results obtained are input again into the trained downscaling surrogate model for the target key work area, and a second downscaling is performed according to a fixed scaling ratio to obtain higher resolution local climate prediction results.

[0017] In this specification, the high-resolution reference meteorological field is the historical daily average of high-resolution meteorological field data generated by dynamic downscaling, or a publicly available high-resolution reference meteorological dataset that matches the simulation period.

[0018] In summary, the present invention has at least the following beneficial effects: This method achieves a synergistic improvement in prediction accuracy, physical consistency, and computational efficiency during climate downscaling, addressing the core pain point that existing technologies cannot simultaneously achieve all three. It avoids the shortcomings of traditional statistical downscaling in its ability to characterize the nonlinear features of local climate under complex terrain, and also overcomes the engineering application bottlenecks of high computational cost and poor timeliness of pure dynamic downscaling. It can efficiently generate refined climate prediction data adapted to small watershed scales.

[0019] A standardized, full-process climate prediction technology system adapted to the high-altitude glacier watershed scenario has been constructed. It has achieved deep integration and scenario-specific optimization of multiple disciplines such as atmospheric science, geography, and computer science, overcoming interdisciplinary technical barriers and forming a replicable, portable, and engineering-applicable technical solution, which has significantly reduced the technical threshold for refined climate prediction in high-altitude and cold mountainous areas.

[0020] This approach overcomes the technical bias in the field that complex atmospheric physical mapping relationships are difficult for deep learning models to stably converge and effectively learn. Through the collaborative design of a terrain attention mechanism and a loss function constrained by multi-dimensional physical information, it not only enhances the model's ability to learn meteorological characteristics of complex mountainous terrain and differentiated underlying surface regions, but also ensures the physical rationality and numerical stability of long-term time-series prediction data at the algorithm level, significantly improving the reliability of climate prediction results for engineering applications.

[0021] It provides efficient, reliable, and refined meteorological data support for predicting future risk trends of disaster chains such as ice-rockfall, glacial debris flow, and glacial lake outburst in high-altitude and cold mountainous areas, and can comprehensively serve the meteorological support and disaster risk prevention and control work of major projects throughout the entire life cycle in high-altitude and cold mountainous areas. Attached Figure Description

[0022] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the 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.

[0023] Figure 1 This is a schematic diagram of the rapid climate prediction method for small watershed future scenarios that integrates deep learning and dynamic downscaling involved in this invention.

[0024] Figure 2 This is a schematic diagram illustrating the different downscaling techniques involved in this invention in glacier-covered areas and debris flow basins.

[0025] Figure 3 This is a schematic diagram of the method flow involved in this invention.

[0026] Figure 4 This is a schematic diagram illustrating the changes in the verification parameters involved in this invention. Detailed Implementation

[0027] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0028] The following disclosure provides many different implementations or examples for carrying out different structures of the embodiments of the present invention. To simplify the disclosure of the embodiments of the present invention, specific examples of components and arrangements are described below. Of course, these are merely examples and are not intended to limit the embodiments of the present invention. Furthermore, reference numerals and / or reference letters may be repeated in different examples of the embodiments of the present invention; such repetition is for simplification and clarity and does not in itself indicate a relationship between the various implementations and / or arrangements discussed.

[0029] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0030] like Figure 1 As shown, this embodiment provides a method for rapid climate prediction of future scenarios in small watersheds that integrates deep learning and dynamic downscaling, including the following steps: Step 1: Define the latitude and longitude range of the target high mountain watershed, and acquire low-resolution historical and future scenario meteorological data, atmospheric reanalysis driving field data, and high-resolution static geographic data from the global climate model output. Perform spatial grid alignment and normalization preprocessing on all acquired data to obtain input data with a uniform format. Step 2: Using the WRF-CPM convection-allowable scale regional climate model optimized for ice and snow processes, input preprocessed atmospheric reanalysis driving field data and high-resolution static geographic data, perform dynamic downscaling simulation on the target area to generate high-resolution meteorological field data for historical periods. This data is used as high-resolution labeled data for training the deep learning model, and a high-resolution baseline meteorological field is calculated based on this data. Step 3: The preprocessed low-resolution historical meteorological data is interpolated and aligned to the grid of high-resolution labeled data, forming a training sample pair with the high-resolution labeled data and the preprocessed high-resolution static geographic data; a TA-UNet network model with a terrain attention mechanism is constructed, and a terrain attention gate is set at the jump connection between the encoder and decoder of the model, with the high-resolution static geographic data as the guiding signal; a total loss function composed of data loss and physical constraint loss is designed; the TA-UNet network model is trained using the training sample pair and the total loss function until the model converges, resulting in a trained downscaled surrogate model; Step 4: Input the preprocessed low-resolution future scenario meteorological data into the trained downscaling surrogate model and output the high-resolution future climate downscaling results; use the high-resolution labeled data as a benchmark to correct the bias of the high-resolution future climate downscaling results; superimpose the daily climate change amount of the corrected downscaling results onto the high-resolution benchmark meteorological field to generate a high-resolution daily future climate prediction dataset for the target high-altitude small watershed.

[0031] In some embodiments, the low-resolution historical and future scenario meteorological data output by the global climate model include daily-scale data of temperature, precipitation, relative humidity, specific humidity, wind field, longwave radiation, and shortwave radiation; the low-resolution historical and future scenario meteorological data adopt the original CMIP6 global coupled climate model data, or adopt the CMIP6 derived public dataset after bias correction and preliminary downscaling.

[0032] In some embodiments, the WRF-CPM convection-allowable scale regional climate model optimized for snow and ice processes adopts the Noah-MP land surface model specifically optimized by snow decay curves and snow albedo schemes. During the simulation process, the topographic turbulence drag scheme is enabled, and the cumulus parameterization scheme is disabled in the inner high-resolution grid.

[0033] In some embodiments, the historical period of the dynamic downscaling simulation is no less than 20 years, and a model warm-up period of no less than one month is set during the simulation process. The output high-resolution meteorological field data includes daily meteorological variables corresponding to the low-resolution historical period and future scenario meteorological data, as well as diagnostic quantities such as sensible heat flux, latent heat flux, surface upward radiation, and surface temperature.

[0034] In some embodiments, the high-resolution static geographic data includes digital elevation models, slope, aspect, topographic shadow, land use type, and vegetation cover data; the topographic attention gate uses the aforementioned high-resolution static geographic data as a guiding signal to dynamically calculate attention weights and adaptively enhance the model's learning of meteorological characteristics of topographically sensitive areas and different land cover areas.

[0035] In some embodiments, the physical constraint loss in the total loss function includes a surface energy balance constraint term, a radiation-temperature correspondence constraint term, an intrinsic consistency constraint term for temperature and humidity variables, and a coupling constraint term for precipitation and relative humidity. The weight of each constraint term can be adjusted according to the meteorological characteristics of the target area.

[0036] In some embodiments, during model training, a seasonal independent training strategy is adopted for precipitation variables to capture the precipitation variation characteristics of the target area in different seasons; the model training adopts a phased training strategy, in which the first phase trains the model backbone network only through data loss, and in the second phase, physical constraint loss is added to optimize the overall parameters of the model.

[0037] In some embodiments, the bias correction employs quantile mapping or linear translation-variance scaling, using high-resolution labeled data as a benchmark, and corrects the high-resolution future climate downscaling results month by month, so that the statistical characteristics of the corrected data are aligned with the high-resolution labeled data.

[0038] In some embodiments, after generating a high-resolution daily future climate prediction dataset for the target high-altitude small watershed, the downscaling results obtained are input again into the trained downscaling surrogate model for the target key work area, and a second downscaling is performed according to a fixed scaling ratio to obtain higher resolution local climate prediction results.

[0039] In some embodiments, the high-resolution benchmark meteorological field is a historical daily average of high-resolution meteorological field data generated by dynamic downscaling, or a publicly available high-resolution benchmark meteorological dataset that matches the simulation period.

[0040] The technical concept of this invention is as follows: This invention employs a hybrid technology framework that deeply integrates dynamic downscaling and deep learning, and completes full-process technical specialization and optimization around the scenario of high-altitude glaciers and small watersheds, referencing... Figure 3 The core implementation process is summarized as follows: Data preparation and preprocessing: Delineate the target study area, collect low-resolution output data from global climate models, atmospheric reanalysis driving field data, and high-resolution static geographic data, and complete spatial grid alignment and normalization preprocessing of the data to form standardized input data.

[0041] Dynamic downscaling generates supervised labeled data: The WRF-CPM convection-allowable climate model optimized for high-altitude and cold snow and ice regions is used to complete the special optimization of snow and ice process parameters such as snow decay curve and snow albedo. Through dynamic downscaling simulation, long-term, high- and low-resolution historical period and future scenario meteorological datasets with complete physical consistency are generated as supervised labels for deep learning model training and reference fields for subsequent product generation.

[0042] Constructing and training a TA-UNet surrogate model with physical constraints: The TA-UNet network architecture, which integrates terrain attention mechanism, is adopted. Attention gates guided by high-resolution terrain and underlying surface data are added at the jump connection of encoder-decoder to enhance the model's feature learning ability for complex terrain areas. A loss function that integrates data fitting terms and multi-dimensional physical constraint terms is designed to ensure the physical rationality of prediction results from the algorithm level. The model is trained, validated and optimized by paired datasets to obtain a downscaled surrogate model that can be quickly used for inference.

[0043] Future climate prediction product generation: Input low-resolution global climate model data of future scenarios into the trained surrogate model, quickly output high-resolution downscaling results, and after bias correction and reference field fusion processing, generate a high-resolution daily future climate prediction dataset that can be directly used for engineering applications and disaster risk assessment.

[0044] This invention aims to achieve rapid and accurate downscaling prediction of high-resolution future climate data from small watersheds in high-altitude glaciers. The overall solution consists of three core modules: a dynamic downscaling module (generating physically continuous labeled data), a deep learning downscaling module (training an efficient proxy model), and an application module (rapidly generating the final product). The implementation process of the technical solution is described in detail below: 1: Data Preparation and Input 1) Define the latitude and longitude range of the region; 2) Obtain low-resolution output data (horizontal resolution typically >100 km) from the global climate model (CMIP6), including daily meteorological variables such as temperature, precipitation, humidity, and radiation for historical periods (1950-2014) and future scenarios (SSP1-2.6, SSP2-4.5, SSP5-8.5, 2015-2100); [Optional: Use "semi-finished" CMIP6 forecast data, such as CLIMEA-BCUD, BC-CMIP6, NEX-GDDP-CMIP6, etc. These data are processed and publicly released by other independent researchers or research institutions based on the original CMIP6 data. Compared with the original CMIP6 data, these "semi-finished" data have undergone model screening, bias correction, and even preliminary downscaling, and have higher accuracy and resolution, and can be directly used as input for further downscaling]; 3) Prepare atmospheric reanalysis driving fields (such as ERA5) as initial and boundary conditions for driving the WRF dynamic downscaling model; 4) Prepare high-resolution static geographic data (such as terrain elevation and land use type), and adjust the grid resolution to match the target output for use as boundary condition input for dynamic simulation and terrain feature fusion for deep learning models; 5) Perform spatial grid alignment and normalization preprocessing on all data to ensure a consistent data format.

[0045] 2: Dynamic downscaling generates physically consistent high-resolution data, which serves as supervision labels for deep learning models. 1) The WRF-CPM convection-allowed regional climate model was used, with the model and physical process parameter scheme configured (Table 1). Data from the European Centre for Medium-Range Weather Forecasts (ERA5) was input as the driving field for dynamic downscaling simulation. Among them, the coupled land surface model adopted the Noah-MP model with improved snow and ice processes, which greatly corrected the negative bias of the simulated temperature data in high-altitude glacial regions.

[0046] The land surface model optimizes the snow decay curve (SCF) and snow albedo. In mountainous areas with complex and steep terrain, topographic variations should be considered in the parameterization of the snow accumulation-decay process; therefore, a topographic scaling factor is introduced to correct the default SCF scheme, as shown in the following formula: ; Where swe is the snow water equivalent (in millimeters) and SD is the snow depth (in centimeters). and These are the snow density of fresh snow and the average snow density per grid point, respectively, which can be provided based on existing datasets. It is the rough length, and tanh is the hyperbolic tangent operator. This is the standard deviation of the subgrid topography. The effect of black coal and dust on snow albedo is considered. The effect of aging snow albedo ( The formula is as follows: ; Where t is time (in seconds). Let be the snow albedo at time t-1. For time step, The lower limit of snow albedo is calculated using the following SNICAR model: ; in, The albedo of bare soil is represented by a threshold of SD=2.0cm, which takes into account that when the snow depth is less than 2.0cm, the effect of soil on snow albedo can be considered linear, and increases with snow depth. Albedo of fresh snow. The calculation is as follows: ; Where z is the depth of the new snow (cm). The albedo of snow at infinity. The snow albedo attenuation coefficient is... and These are the intercept and slope parameters for linear fitting, respectively. and These are the multiplier and exponential parameter for the exponential fit, respectively. (Direct radiation is 1.1132, diffuse radiation is 1.1465). (Direct radiation is -0.059, diffuse radiation is -0.052). (Direct radiation is 23.939, diffuse radiation is 10.361). (Direct radiation is -0.629, diffuse radiation is -0.484). D is the snow grain size (m), through... The calculation yielded: ; Snow albedo at current time The calculation is as follows: ; in, Rainfall intensity (mm / s) Snowfall intensity (mm / s). This represents the proportion of snowfall in the total precipitation.

[0047] 2) Run the WRF-CPM model, solve the atmospheric physical equations, and integrate year by year. The first month is for model warm-up. Generate high-resolution meteorological field data (such as temperature, precipitation, air pressure, wind field, specific humidity, long wave and short wave radiation) for at least 20 years (1994-2014). By running only 20 years of simulation, the completion time can be controlled within 3 months and the calculation cost can be controlled within 30,000 yuan.

[0048] 3) The output results are used as high-resolution labeled data (Y_HR), and the 20-year daily average is used as the baseline field (Y_BASE); the CMIP6 low-resolution data (X_LR) is aligned to the WRF output grid through nearest neighbor interpolation to form a paired dataset.

[0049] Table 1. WRF-CPM Downscaling Model and Main Physics Parameterization Scheme Settings

[0050] Optional: Dynamic downscaling simulations at a 4km daily scale over the past 20-30 years have been completed in the Tibetan Plateau and surrounding high-altitude areas of Asia. This data can be directly used as high-resolution annotation data for downscaling in this region in the future, thus skipping the simulation step. In addition, there are other independently developed high-resolution meteorological downscaling data available, such as TPMFD, which may also be used directly. Using existing data as annotations can further improve efficiency and save costs.

[0051] 3: Training a TA-UNet deep learning WRF surrogate model with physical information constraints 1) Model Construction: The TA-UNet (Terrain Attention U-Net) architecture based on a dynamic weight adaptive mechanism is adopted. Its encoder-decoder structure includes skip connections and a terrain attention gate. Downsampling and upsampling are performed through depthwise convolution and channel reorganization to reorganize low-resolution, high-channel feature maps into high-resolution, low-channel outputs. The attention gate uses high-resolution terrain (containing multi-scale terrain features such as slope, aspect, and terrain shadow), vegetation, and land data as guiding signals, and dynamically weights and fuses meteorological features to enhance the learning of complex areas of the underlying terrain. The input consists of three parts: low-resolution CMIP6 meteorological data X (including temperature tas, relative humidity hurs, specific humidity huss, precipitation pr, wind field wind_u & wind_v, longwave and shortwave radiation variables rss & rls), high-resolution labeled data Y obtained from WRF (containing the corresponding 8 variables), and static geographic data A (including topographic DEM, topographic features, land use, and vegetation cover); the output is a high-resolution prediction field Y_hat (containing the corresponding 8 target variables and 5 calculated diagnostic quantities: sensible heat flux hfss, latent heat flux hfls, surface upward radiation rlus & rsus, and surface temperature ts). A fixed scaling ratio is set. The mathematical expression for the TA gate is as follows: ; ; ; ; Attention feature layer; : Input meteorological variable field; Terrain guidance signal from the decoder; Trainable weights and biases; Convolution operators and their biases; Normalization via the sigmoid function The calculated attention coefficient (between 0 and 1) is used to... The feature layer is weighted to make the model pay more attention to areas that are strongly correlated with terrain, and the attention weights are adaptively adjusted according to the type of meteorological variables. This is the set of all learnable parameters in the attention module.

[0052] 2) Loss Function Design: The total loss function is the sum of data loss and physical constraint loss. ; , which is the weighting coefficient (hyperparameter) used to adjust the proportion of physical constraint loss in the total loss; For regularization parameters. Data loss ( ): The mean squared error (MSE) is used to measure the predicted value. With WRF annotation value Differences: ; Physical constraint loss ( ): Introduce the residuals of the physical equations as regularization terms.

[0053] i. Introduce energy balance constraints to penalize predictions of surface energy not being closed: ; It is the net radiation, which is the sum of RSS and RLS; It is the sensible heat flux, which can be obtained from the difference between ts and tas through thermodynamic relations; ts, in turn, can be obtained from tas, 、 Wind speed is obtained through empirical formulas; Soil heat flux, generally A fixed ratio; This is the latent heat flux, obtained from the energy balance residual.

[0054] ii. Introduce a radiation-temperature constraint (Stephen-Boltzmann formula) to penalize predictions that do not conform to the temperature-energy relationship: ; This is the radiation-temperature constraint loss, used to constrain the thermodynamic relationship between Earth's surface temperature and long-wave radiation; It is the long-wave radiation rising from the earth's surface, which can be calculated using RSS and albedo parameters. Albedo is related to vegetation and land cover characteristics. The absorption rate is set to a fixed value based on empirical relationships; The blackbody radiation constant; For ts.

[0055] iii. Combining the variables of specific humidity (q), relative humidity (RH), and temperature (T), RH (RH_mag) is calculated from T and q using the saturated vapor pressure formula (Magnus formula). The RH_mag calculated from T and q is compared with the RH directly predicted by the model. A constraint loss term is added to penalize predictions that are inconsistent with the relationship between temperature, specific humidity, and relative humidity, ensuring the inherent consistency of the humidity field. Humidity consistency constraint loss. : ; iv. In monsoon summers, heavy rainfall typically occurs in high-humidity environments. Therefore, a constraint can be imposed to prevent the model from predicting heavy rainfall in low-relative-humidity environments. This is the precipitation-humidity constraint loss. : ; in The model predicts precipitation; c is a threshold, set to 50% (which can be adjusted according to the actual situation), indicating that if the model still predicts precipitation when the relative humidity is below 50%, it will be penalized, and the greater the precipitation, the heavier the penalty.

[0056] In summary, the total physical constraint loss is: ; This represents the strength of the physical constraint (here, n=1 / 2 / 3 / 4, corresponding to the loss of each constraint), with a value range of 0-1. When =0, this constraint is not applied.

[0057] In addition, the precipitation model can be chosen as follows: the model does not directly predict specific precipitation values, but rather predicts a parameter that describes the probability and intensity of precipitation (Bernoulli-Gamma distribution), where the precipitation probability distribution can be written as: ; Indicates precipitation. Represents probability density, These are the shape parameter and steepness parameter of the gamma distribution, respectively. The gamma function. The loss function at this point. It can be written as: ; in and This represents the actual observed precipitation and its probability.

[0058] 3) Model Training: The Adam optimizer was used with a fixed random seed, a learning rate of 0.0001, and an ExponentialLR learning rate decay strategy (decay coefficient of 0.9). Batch size ranged from 32 to 64. For precipitation, a seasonal training strategy was adopted (spring, summer, autumn, and winter were trained independently) to capture the seasonal characteristics of precipitation in the monsoon region. The total loss was minimized through backpropagation to update the model parameters until convergence.

[0059] Optional: A phased training strategy is adopted. Phase 1: First, train the TA-UNet backbone, using only the data loss of 8 basic variable fields ( The first stage involves teaching the system to downscale directly input variables; the second stage involves freezing the core parameters, adding a physical diagnostic module, and using diagnostic variable field data loss (…). Training diagnostic module parameters; Third stage: Adding physical constraint loss ( This allows the model to optimize physical consistency.

[0060] 4) Model Validation: The validation parameters include MSE, mean absolute error (MAE), and peak signal-to-noise ratio (PSNR). The formulas for calculating MAE and PSNR are as follows: ; ; This represents the maximum signal power, where m is the bit depth, typically 8 bits. For example... Figure 4 (The validation parameters for the bilinear model and the trained machine learning downscaling model are: loss function (Loss), mean squared error (MSE), mean absolute error (MAE), and peak signal-to-noise ratio (PSNR)). As shown, the TA-UNet results are significantly better than the results of traditional statistical downscaling bilinear interpolation.

[0061] 5) After training, a deployable TA-UNet proxy model is obtained. This model inherits the characteristics of the WRF output, but has extremely high computational efficiency (one test case can be completed in about 1-2 days); the model supports multivariate collaborative downscaling and maintains the physical consistency between variables.

[0062] 4: Product Development and Business Application The daily-scale CMIP6 low-resolution data (Xpred_LR) for the future period (2015-2100) is input into the trained TA-UNet surrogate model, which outputs downscaled future climate predictions (Xpred_HR). Further, bias correction is performed using quantile mapping or linear translation-variance scaling, with the correction based on the labeled data and performed monthly. The probability density distribution or mean and variance of daily temperature / precipitation are aligned with the labeled data. The calculation formulas for the two methods are as follows: ; ; in, To correct the output of the TA-UNet model before, This is the result after bias correction, where F is the cumulative distribution function. The mean, The standard deviation is represented by the subscripts wrf and pred, which indicate the labeled data obtained by WRF and the output results of TA-UNet, respectively.

[0063] The daily variation (delta) of the downscaling result after bias correction is superimposed onto the reference field. )superior[ The result can be obtained from the WRF output, or other source datasets can be used, such as the commonly used WorldClim and MSWX datasets, but it is necessary to ensure that the time period of the baseline data and the time period for calculating delta are consistent. The formula is: ; and These are the future predicted values ​​of the downscaled variables after bias correction and the average value of the baseline period, respectively. The final product generated is a high-resolution daily future climate prediction dataset, containing basic variables such as temperature and precipitation, which can be directly used for disaster simulation and risk assessment in mountainous areas.

[0064] Furthermore, for some key work areas, we can input the downscaling results back into the trained TA-UNet surrogate model for a second downscaling at the same scaling ratio to obtain higher-resolution climate prediction results. This step relies on the assumption that the dependency between climate change and topography and underlying surface conditions is consistent with the previous step. Since this step involves unlabeled data for supervised learning, the results cannot be used for further downscaling to prevent exponential error growth. We consider two downscaling paths: one at a 1:5 ratio: 25km => 5km => 1km; and the other at a 1:3 ratio: 10km => 3.33km => 1km.

[0065] like Figure 2 Performance of Delta statistical downscaling, TA-UNet downscaling, and WRF dynamic downscaling in glacier-covered areas and debris flow basins: mean temperature variation from 1994 to 2013 compared to 2014-2023 (altitude). A scatter plot of .

[0066] The purpose of this invention is to address the challenges posed by the advancement of major engineering projects in glacier-covered high-altitude mountainous regions in recent years, which have brought various new geological hazards (such as ice-rockfall, avalanches, glacial debris flows, and glacial lake outburst disasters). Meteorological conditions are a crucial driver of these disasters. This invention provides a high-resolution, high-precision meteorological-driven dataset for rapidly predicting the future risk trends of these hazards. By training a lightweight TA-UNet model with a small number of WRF simulations as annotations, the computation time for downscaling meteorological data for the next century can be reduced from several years to approximately three months, meeting the timeliness and cost control requirements of engineering applications. This solution introduces an attention mechanism as guidance to adaptively focus on the influence of key topographic features such as slopes, valleys, and glaciers. Simultaneously, a physical information-constrained loss function is added to achieve stable and efficient learning of complex physical mappings like WRF simulations.

[0067] Key technical innovations of this application: First, a standardized process for rapid climate prediction in alpine glacial watersheds was established: First, a high-resolution dynamic downscaling simulation of the target region was performed using a WRF-CPM model optimized for snow cover and albedo parameters, generating a high-resolution dataset with continuous physical mechanisms. Then, the WRF simulation results were used as the "ground truth" for supervised learning, forming a training sample pair with the corresponding coarse-resolution GCM (General Circulation Model) prediction data to train a specific deep learning model (TA-UNet). Finally, the target data was generated based on the trained model. This entire approach integrates research techniques from atmospheric science, geography, and computer science, and its application scenarios involve hydrology and geological engineering. The models used underwent a series of comparisons and optimizations for the target application scenarios. This proposed approach integrates previously fragmented, multidisciplinary technologies to achieve rapid and standardized generation of alpine glacial watershed climate prediction data with both reliable physical mechanisms and high spatiotemporal resolution.

[0068] Second, existing technologies generally believe that the physical process of WRF simulation is too complex, and the mapping relationship between its output and low-resolution input is extremely nonlinear and changes drastically with terrain, making it difficult for ordinary deep learning models to converge stably and learn effectively. This scheme integrates two core mechanisms within the deep learning model to overcome the technical bias of meteorological downscaling in small watersheds of high mountains and glaciers. (1) Terrain attention mechanism: At the jump connection between the encoder and decoder of the TA-UNet network, an attention gate with high-resolution terrain feature maps and soil and vegetation cover classification maps as guiding signals is added. This mechanism calculates dynamic weight maps, enabling the model to adaptively strengthen the learning of meteorological features of key terrain-sensitive areas such as slopes and valleys, as well as different land cover areas such as glaciers and forests during feature fusion and reconstruction, thereby significantly improving the downscaling precision of complex terrain and meteorological variable fields under the underlying surface. (2) Physical information constraint loss function: In the loss function of model training, in addition to data fitting terms (such as mean square error), a physical constraint term based on simplified atmospheric physical information residuals is introduced. This term, acting as a regularizer, explicitly penalizes the model for outputs that produce physically unreasonable relationships between variables such as water vapor, temperature, precipitation, and energy during the optimization process. This ensures the physical rationality and numerical stability of the generated long-term climate data from an algorithmic perspective.

[0069] The WRF model used in this application for optimizing ice and snow feedback parameter schemes has been described in the literature (Zhou et al., 2023). The basic structure of the TA-UNet model in this application has been described in the literature (Jian, He, et al., 2025) and the patent (Jian Jihao, He Siming, et al., 2025).

[0070] However, the innovation of this application lies in the first-time deep integration of TA-UNet and the WRF dynamic downscaling system, along with the addition of physical constraints for high-altitude glacier regions, constructing a complete, physically constrained, integrated "dynamic simulation-deep learning" climate prediction framework. The specific implementation of this framework includes a training method using long-term WRF-CPM simulation outputs as supervised labels, a physically constrained loss function designed for climate variables, and a complete technical solution for applying it to rapid future climate prediction under multiple scenarios in the CMIP6 system for small watersheds in mountainous areas.

[0071] The embodiments described above are for illustrative purposes only and are not intended to limit the invention. Therefore, any changes in numerical values ​​or substitutions of equivalent elements should still fall within the scope of this invention.

[0072] The above detailed description will enable those skilled in the art to understand that the present invention can indeed achieve the aforementioned objectives and has complied with the provisions of the Patent Law.

[0073] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the invention. The above descriptions are merely preferred embodiments of the invention and are not intended to limit the invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the invention should be included within the scope of protection of the invention.

[0074] It should be noted that the above description of the process is for illustrative purposes only and does not limit the scope of this specification. Those skilled in the art can make various modifications and changes to the process under the guidance of this specification. However, these modifications and changes remain within the scope of this specification.

[0075] The basic concepts have been described above. Obviously, for those skilled in the art who have read this application, the above disclosure is merely illustrative and does not constitute a limitation of this application. Although not explicitly stated herein, those skilled in the art may make various modifications, improvements, and corrections to this application. Such modifications, improvements, and corrections are suggested in this application, and therefore, such modifications, improvements, and corrections still fall within the spirit and scope of the exemplary embodiments of this application.

[0076] Furthermore, this application uses specific terms to describe its embodiments. For example, "an embodiment," "one embodiment," and / or "some embodiments" refer to a particular feature, structure, or characteristic related to at least one embodiment of this application. Therefore, it should be emphasized and noted that "an embodiment," "one embodiment," or "an alternative embodiment" mentioned twice or more in different positions in this specification do not necessarily refer to the same embodiment. In addition, certain features, structures, or characteristics in one or more embodiments of this application can be appropriately combined.

[0077] Furthermore, those skilled in the art will understand that aspects of this application can be described and illustrated through several patentable types or situations, including any new and useful combination of processes, machines, products, or substances, or any new and useful improvements thereof. Therefore, aspects of this application can be implemented entirely in hardware, entirely in software (including firmware, resident software, microcode, etc.), or a combination of hardware and software. All of the above hardware or software can be referred to as a “unit,” “module,” or “system.” Furthermore, aspects of this application can take the form of a computer program product embodied in one or more computer-readable media, wherein computer-readable program code is contained therein.

[0078] The computer program code required for the operation of each part of this application can be written in any one or more programming languages, including object-oriented programming languages ​​such as Java, Scala, Smalltalk, Eiffel, JADE, Emerald, C++, C#, VB.NET, and Python; general programming languages ​​such as C; Visual Basic, Fortran2103, Perl, COBOL2102, PHP, and ABAP; dynamic programming languages ​​such as Python, Ruby, and Groovy; or other programming languages. This program code can run entirely on the user's computer, or as a standalone software package on the user's computer, or partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In the latter case, the remote computer can be connected to the user's computer via any network, such as a local area network (LAN) or wide area network (WAN), or connected to an external computer (e.g., via the Internet), or in a cloud computing environment, or used as a service such as Software as a Service (SaaS).

[0079] Furthermore, unless expressly stated in the claims, the order of processing elements and sequences, the use of numbers and letters, or other names described in this application are not intended to limit the order of the processes and methods of this application. Although some currently considered useful embodiments of the invention have been discussed in the foregoing disclosure by way of various examples, it should be understood that such details are for illustrative purposes only, and the appended claims are not limited to the disclosed embodiments; rather, the claims are intended to cover all modifications and equivalent combinations that conform to the substance and scope of the embodiments of this application. For example, although the implementation of the various components described above can be embodied in a hardware device, it can also be implemented as a purely software solution, such as an installation on an existing server or mobile device.

[0080] Similarly, it should be noted that, in order to simplify the description of the present application and thus aid in the understanding of one or more embodiments of the invention, the foregoing description of the embodiments of the present application sometimes combines multiple features into a single embodiment, drawing, or description thereof. However, this approach of the present application should not be construed as reflecting an intention that the claimed subject matter requires more features than expressly recited in each claim. Rather, the subject of the invention should possess fewer features than in any single embodiment described above.

Claims

1. A method for rapid climate prediction of future scenarios in small watersheds that integrates deep learning and dynamic downscaling, characterized in that, include: Step 1: Define the latitude and longitude range of the target high mountain watershed, and acquire low-resolution historical meteorological data and future scenario meteorological data, atmospheric reanalysis driving field data and high-resolution static geographic data output by the global climate model. Perform spatial grid alignment and normalization preprocessing on all acquired data to obtain input data with a uniform format. Step 2: Using the WRF-CPM convection-allowable scale regional climate model optimized for ice and snow processes, input preprocessed atmospheric reanalysis driving field data and high-resolution static geographic data, perform dynamic downscaling simulation on the target area to generate high-resolution meteorological field data for historical periods. This high-resolution meteorological field data is used as high-resolution labeled data for training the deep learning model. At the same time, a high-resolution baseline meteorological field is calculated based on this high-resolution meteorological field data. Step 3: The preprocessed low-resolution historical meteorological data is interpolated and aligned to the grid of the high-resolution labeled data, and together with the high-resolution labeled data and the preprocessed high-resolution static geographic data, they form a training sample pair. A TA-UNet network model integrating a terrain attention mechanism is constructed. A terrain attention gate is set at the jump connection between the encoder and decoder of the TA-UNet network model. The terrain attention gate uses high-resolution static geographic data as a guiding signal. A total loss function composed of data loss and physical constraint loss is designed. The TA-UNet network model is trained using training sample pairs based on the total loss function until the model converges, and a trained downscaled surrogate model is obtained. Step 4: Input the preprocessed low-resolution future weather scenario data into the trained downscaling surrogate model and output high-resolution future climate downscaling results. Based on high-resolution labeled data, bias corrections are performed on the high-resolution future climate downscaling results; the daily climate change amounts of the corrected downscaling results are then superimposed onto the high-resolution baseline meteorological field to generate a high-resolution daily future climate prediction dataset for the target high-altitude small watershed.

2. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, The low-resolution historical and future scenario meteorological data output by the global climate model include daily-scale data of temperature, precipitation, relative humidity, specific humidity, wind field, longwave radiation, and shortwave radiation. The low-resolution historical and future scenario meteorological data are derived from the original CMIP6 global coupled climate model data or from publicly available CMIP6 derivative datasets that have undergone bias correction and preliminary downscaling.

3. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, The WRF-CPM convection-allowable scale regional climate model optimized for snow and ice processes adopts the Noah-MP land surface model specifically optimized by snow decay curves and snow albedo schemes. During the simulation, the topographic turbulence drag scheme is enabled, and the cumulus parameterization scheme is disabled in the inner high-resolution grid.

4. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, The historical period of the dynamic downscaling simulation is no less than 20 years, and a model warm-up period of no less than one month is set during the simulation. The output high-resolution meteorological field data includes daily meteorological variables corresponding to the low-resolution historical period and future scenario meteorological data, as well as sensible heat flux, latent heat flux, surface upward radiation and surface temperature.

5. The rapid climate prediction method for small watershed future scenarios that integrates deep learning and dynamic downscaling as described in claim 1, characterized in that, The high-resolution static geographic data includes digital elevation models, slope, aspect, topographic shadow, land use type, and vegetation cover data; based on the topographic attention gate, attention weights are dynamically calculated, and the adaptive reinforcement model learns the meteorological characteristics of topographically sensitive areas and different land cover areas.

6. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, The physical constraint loss in the total loss function includes surface energy balance constraint, radiation-temperature correspondence constraint, inherent consistency constraint of temperature and humidity variables, and coupling constraint of precipitation and relative humidity. The weight of each constraint is adjusted according to the meteorological characteristics of the target area.

7. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, During the training of the TA-UNet network model, a strategy of independent training for precipitation variables in different seasons is adopted to capture the precipitation variation characteristics of the target area in different seasons. The TA-UNet network model training adopts a phased training strategy. In the first phase, only the backbone network of the model is trained using data loss. In the second phase, physical constraint loss is added to optimize the overall parameters of the model.

8. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, The bias correction adopts the quantile mapping method or the linear translation-variance scaling method. Based on the high-resolution labeled data, the high-resolution future climate downscaling results are corrected month by month to align the statistical characteristics of the corrected data with the high-resolution labeled data.

9. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, After generating a high-resolution daily climate prediction dataset for the target high-altitude small watershed, the downscaling results obtained are input again into the trained downscaling surrogate model for the target key work area. The model is then downscaled a second time according to a fixed scaling ratio to obtain higher-resolution local climate prediction results.

10. The method for rapid climate prediction of future scenarios in small watersheds by integrating deep learning and dynamic downscaling as described in claim 1, characterized in that, The high-resolution benchmark meteorological field is the historical daily average of high-resolution meteorological field data generated by dynamic downscaling, or a publicly available high-resolution benchmark meteorological dataset that matches the simulation period.

Images