An evapotranspiration estimation method and system based on a physical-data hybrid enhanced HS improved model

Abstract

The present application relates to a kind of evapotranspiration estimation method and system based on physical-data hybrid enhancement's HS improved model, belong to evapotranspiration estimation technical field.The method is aimed at the problem of missing physical mechanism, poor spatial adaptability and high system module coupling in the estimation of high-cold complex region, by constructing multidimensional feature system, climate self-adaptive parameter optimization and physical data layered hybrid architecture, combined with modular system design, the advantages of physical mechanism and data-driven are complementary.The system has multi-source data compatibility, real-time physical verification and adaptive configuration function, significantly improves the estimation accuracy, physical rationality and cross-region application ability, and provides reliable technical support for cold region hydrological ecological research.

Description

Technical Field

[0001] This invention belongs to the field of evapotranspiration estimation technology, and relates to an evapotranspiration estimation method and system based on a physical-data hybrid enhancement HS model. Background Technology

[0002] Reference Evaporation Accurate estimation of water resources is a crucial foundation for hydrological and ecological research and water resource allocation and management. The HS model, a classic temperature radiation model, is widely used due to its low parameter requirements and ease of calculation. However, traditional HS models and their existing improvements reveal several key technical shortcomings under complex high-altitude and cold climates and underlying surface conditions.

[0003] Existing improvements to the HS model can be broadly categorized into two types. The first type is the traditional statistical optimization method, which uses linear or nonlinear regression to locally calibrate the model's core parameters. This type of method has significant limitations: firstly, it only fits parameters for a single region, making it difficult to adapt to spatial heterogeneity across climate zones; secondly, it uses single-factor or simple multi-factor regression, failing to effectively capture the complex nonlinear coupling effects between multiple driving factors; and finally, the parameter optimization process lacks physical constraints, potentially leading to optimization results that contradict the physical laws of evapotranspiration, resulting in poor model interpretability.

[0004] The second category is pure machine learning (ML) optimization methods, which use the HS model estimates as input features and combine them with other meteorological and remote sensing data to construct an ML model. While this type of method can improve estimation accuracy, it has three major drawbacks: First, the model lacks physical meaning, and the output relies entirely on data fitting, failing to explain the energy balance of evapotranspiration and the water transport mechanism, easily leading to black-box prediction bias; second, the problem of systematic underestimation or overestimation is prominent, and due to the lack of physical constraints, the generalization ability drops sharply under extreme climate conditions outside the data distribution; third, the feature selection process is highly blind, failing to incorporate physical mechanisms, resulting in high model redundancy and low training efficiency.

[0005] Besides inherent methodological flaws, existing evapotranspiration estimation systems also suffer from numerous shortcomings. The data acquisition module is functionally limited, typically supporting only data input from conventional meteorological stations and failing to effectively accommodate multi-source data such as remote sensing products and reanalysis datasets, resulting in insufficient data coverage. Internal module coupling is too high; feature processing, parameter optimization, and model calculation are not decoupled, posing significant challenges to subsequent maintenance and functional expansion. Systems generally lack physical constraint verification modules, failing to verify the physical rationality of output results in real time, easily leading to outliers that do not conform to energy balance laws. System adaptability is poor; the lack of adaptive parameter configuration modules for different climatic regions results in a significant drop in accuracy when applied across regions. System visualization and output capabilities are weak, failing to intuitively display the model operation process, parameter optimization process, and the spatiotemporal distribution characteristics of estimation results, hindering the analysis and application of the results.

[0006] In addition, existing technologies are insufficient in characterizing key temporal dynamic processes such as seasonal freeze-thaw cycles and vegetation growth cycles, and exhibit significant response lag.

[0007] In summary, existing technologies are suitable for high-altitude, cold, and complex regions. ET In 0-dimensional estimation, core challenges include the lack of physical meaning, insufficient adaptation to spatial heterogeneity, and poor generalization ability under extreme conditions. Existing systems also suffer from systemic bottlenecks such as high module coupling, poor data compatibility, and insufficient adaptability. Therefore, there is an urgent need in this field for an improved HS model method and supporting system that can balance the rationality of physical mechanisms, estimation accuracy, and spatial adaptability to solve the aforementioned key technical problems. Summary of the Invention

[0008] In view of this, the purpose of the present invention is to provide a method and system for estimating evapotranspiration based on a physical-data hybrid enhancement HS model.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] An improved HS model evapotranspiration estimation method based on physical-data hybrid enhancement includes the following steps: Step 1: Construct a multi-dimensional feature system, which includes basic features, derived features, and constraint features; Step 2: Optimize the HS model parameters using a climate-adaptive parameter optimization strategy; Step 3: Calculate reference evapotranspiration using a physical-data layered hybrid architecture ET The physical-data layered hybrid architecture, with a value of 0, includes a physical model layer, a residual correction layer, and a dynamic weighted fusion layer. Step 4: Verify the estimation results from multiple dimensions.

[0011] Furthermore, in step one, constructing a multi-dimensional feature system specifically includes: screening basic features based on the evapotranspiration energy balance equation and water transport mechanism; generating derived features based on physical mechanisms, including energy-water constraint features, multi-factor coupling interaction features, and temporal rhythm coding features; and using a multi-level feature selection framework to screen features, including physical constraint forced retention, multi-method integrated scoring, and cross-validation optimization.

[0012] Furthermore, the physical constraint forced retention in the multi-level feature selection framework refers to the forced retention of the average daily temperature. Solar radiation from the top of the atmosphere and daily temperature range Three core physical characteristics; The multi-method ensemble scoring employs a weighted ensemble scoring based on Pearson Correlation Coefficient (PCC), XGBoost Feature Importance (FI), and Mutual Information (MI). The weighting formula is as follows: ;in The Pearson correlation coefficient score, XGBoost feature importance score, The mutual information score is used; the cross-validation optimization uses the validation set determination coefficient R. 2 The objective is to minimize the maximum sum and root mean square error (RMSE).

[0013] Furthermore, in step two, the climate adaptive parameter optimization strategy specifically includes: Climate units are divided based on core climate factors; The search space for HS model parameters is dynamically defined based on the statistical characteristics of climate units. The physically constrained differential evolution (DE) algorithm is used to globally optimize the parameters, where the HS model parameters include... C , a and E , C This is an empirical coefficient. a These are temperature-related parameters. E These are radiation-related parameters.

[0014] Furthermore, in step three, the computation of the physical-data layered hybrid architecture specifically includes: integrating four types of physical models at the physical model layer and outputting the result using a weighted average method. ,in The output values ​​are from the physical model layer; the residuals are predicted using the XGBoost model in the residual correction layer. Among them, residual , The measured evapotranspiration values ​​were used; the Sequential Least Squares Programming (SLSQP) algorithm was employed in the dynamic weighted fusion layer to solve for the weights and calculate the final value. The fusion formula is ,in and These are the weighting coefficients. Output for the physical model layer. To predict the residuals, the constraints are as follows: + =1 and ∈[0.6,0.8].

[0015] Furthermore, in step four, the multi-dimensional verification includes statistical accuracy verification, physical rationality verification, and extreme condition stability verification. The statistical accuracy verification uses the coefficient of determination. Root mean square error RMSE Mean absolute error MAE Nash-Sutcliffe efficiency coefficient NSE index; The physical rationality verification includes energy balance closure test, parameter sensitivity analysis, and time series trend consistency test. The extreme condition stability verification targets evaluations under extreme low temperature, extreme high radiation, and persistent drought scenarios. RMSE Increase.

[0016] A system for estimating evapotranspiration using a HS improved model based on physical-data hybrid enhancement is provided to implement the aforementioned HS improved model evapotranspiration estimation method based on physical-data hybrid enhancement. The system includes a data acquisition module, a feature processing module, a parameter optimization module, a hybrid model calculation module, a physical verification module, a result visualization and output module, and a control module. The data acquisition module is used to access multi-source data; The feature processing module is used to construct and select features at multiple levels; The parameter optimization module is used to perform climate adaptive parameter optimization; The hybrid model computation module is used for computation via a physical-data hierarchical hybrid architecture. Value; the physical verification module is used to verify the physical reasonableness of the result; The result visualization and output module is used to output the results; the control module is used to coordinate the operation of each module; and each module is connected through a standardized data interface.

[0017] Furthermore, the feature processing module includes a basic feature extraction unit, a derived feature generation unit, and a feature selection unit; The parameter optimization module includes a climate unit partitioning unit, a parameter space delimitation unit, and a DE optimization algorithm unit; The hybrid model computation module includes a physical model unit, an XGBoost residual correction unit, and an SLSQP weighted fusion unit. The physical verification module includes an energy balance verification unit, a parameter sensitivity verification unit, and a time-series trend verification unit.

[0018] Furthermore, the data acquisition module supports multi-source data access, including meteorological station data, remote sensing products, and reanalysis datasets, and includes a format conversion unit and a quality control unit; the result visualization and output module supports visualization of time-series variation curves, spatial distribution heat maps, and statistical accuracy comparison charts, and outputs the results in CSV, Excel, and PDF formats.

[0019] Furthermore, the control module includes a parameter configuration unit and a status monitoring unit. The parameter configuration unit is used to set DE algorithm parameters, XGBoost hyperparameters and verification index thresholds. The status monitoring unit is used to monitor the module's operating status in real time and record operating logs.

[0020] The beneficial effects of this invention are as follows: (1) This invention significantly improves the accuracy and reliability of reference evapotranspiration estimation in high-altitude and complex regions. By constructing a hybrid architecture that deeply integrates physical mechanisms and data-driven approaches, it effectively resolves the contradiction between the insufficient accuracy of traditional methods and the lack of physical meaning in machine learning methods. Specifically, the consistency between the estimation results and measured values ​​reaches an excellent level, the coefficient of determination is significantly improved, and the root mean square error and mean absolute error are greatly reduced. The model maintains stable performance under extreme climatic conditions and has strong generalization ability. More importantly, this invention successfully ensures the physical rationality and interpretability of the model. Through a feature selection framework that prioritizes physical constraints and a climate-adaptive parameter optimization strategy, it ensures that all parameter changes strictly follow the evapotranspiration energy balance and water transport mechanism. The model output results can be traced back to explicit physical processes, completely avoiding black-box prediction bias and greatly enhancing the academic research value and engineering application credibility.

[0021] (2) The modular evapotranspiration estimation system designed in this invention demonstrates outstanding engineering application value. The system possesses strong multi-source data compatibility, seamlessly integrating meteorological station observations, remote sensing inversion products, and reanalysis datasets, fundamentally solving the data coverage and quality issues. Adopting a highly decoupled modular design, each functional module is highly independent, supporting flexible upgrades and customized expansions, greatly reducing system maintenance costs. The system innovatively embeds a real-time physical verification module, which can automatically identify and eliminate abnormal results that do not conform to physical laws, ensuring the reliability of the output data. Through intelligent adaptive configuration, the system can quickly adapt to the unique conditions of different climate zones, significantly improving the accuracy and efficiency of cross-regional applications. The system's visualization and output functions have been comprehensively enhanced, supporting multi-dimensional dynamic display of estimation results and diverse format output, greatly reducing the user's learning curve. Simultaneously, the system possesses a comprehensive operational status monitoring and early warning mechanism, ensuring long-term operational stability and reliability, providing solid support for large-scale operational applications.

[0022] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein: Figure 1 for Model-physical-data layered hybrid architecture diagram; Figure 2 This is a diagram showing the overall architecture of the evapotranspiration estimation system. Figure 3 for Comparison of statistical accuracy verification between the model and the HS-M3 model; Figure 4 This is a flowchart of the evapotranspiration estimation system. Detailed Implementation

[0024] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0025] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0026] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0027] Figure 1 for The model physics-data hierarchical hybrid architecture diagram illustrates the hierarchical relationship between the physics model layer (four basic HS physics models and their improved integration), the residual correction layer (XGBoost model), and the dynamic weighted fusion layer (SLSQP algorithm for weight calculation). It labels the inputs, outputs, and core parameters of each layer, intuitively presenting the complementary advantages of the hybrid architecture, and includes annotations of data interaction relationships between each layer. Figure 2 This diagram shows the overall architecture of the evapotranspiration estimation system. It illustrates the composition and interrelationships of the system's seven core modules (data acquisition module, feature processing module, parameter optimization module, hybrid model calculation module, physical verification module, result visualization and output module, and control module), and marks the core functions and data interface types of each module, clearly presenting the system's modular design concept and data flow path.

[0028] Figure 3 for The statistical accuracy verification comparison charts of the model and the HS-M3 model (the best among typical HS models) include a performance comparison chart of the basic model and the hybrid model, a scatter plot of the prediction results of the basic model and the hybrid model, a residual analysis chart of the basic model and the hybrid model, and a weight distribution chart of the hybrid model.

[0029] Figure 4 This is a flowchart of the evapotranspiration estimation system. The flowchart shows the entire process of the system from multi-source data access to result output. It includes the triggering conditions, operation sequence and exception handling process of each module, and marks the key parameter configuration nodes and quality control thresholds, intuitively presenting the fully automated operation logic of the system.

[0030] The core technical solution of this invention is to construct a physical-data hybrid enhancement HS improved model ( The model and its supporting evapotranspiration estimation system achieve the accuracy and generalization improvement of the HS model through the technical link of "construction of multi-dimensional feature system → optimization of climate adaptive parameters → fusion of hierarchical hybrid architecture → multi-dimensional verification". The system achieves efficient and stable operation through "modular design → multi-source data compatibility → physical verification embedding → adaptive configuration → visualization output".

[0031] I. Evapotranspiration Estimation Method of HS Improved Model Based on Physical-Data Hybrid Enhancement (I) The construction of the multi-dimensional feature system follows the principle of "prioritizing physical mechanisms and comprehensively representing data", and constructs a three-dimensional feature system covering basic features, derived features, and constraint features. 1. Construction of basic feature library: Based on the evapotranspiration energy balance equation and water transport mechanism, 18 core basic characteristics were selected, including meteorological characteristics (daily maximum temperature, daily minimum temperature, daily average temperature, 2m wind speed, relative humidity, etc.), radiation characteristics (solar radiation at the top of the atmosphere, net surface radiation), underlying surface characteristics (normalized difference vegetation index, soil temperature, soil volumetric water content, freeze-thaw state of permafrost, etc.), geographical characteristics (altitude, latitude, slope), and temporal characteristics (annual accumulated days, seasonal type, snow cover, precipitation). The definitions, units, data sources, and acquisition methods of each characteristic were clarified to ensure the physical correlation of the characteristics and the reliability of the data. See Table 1 for details.

[0032] Table 1 List of basic features

[0033] 2. Derived Feature Generation (Physical Mechanism Guided): Based on the physical mechanism and temporal dynamics of evapotranspiration, three types of derived features are constructed to quantify core environmental stress, multi-factor coupling effects, and temporal rhythms: Fundamental Derivative Features: Based on the physical mechanisms of key evapotranspiration driving processes, quantifying core environmental stresses and energy supply states: 2.1 Diurnal temperature range : = This characterizes the potential impact of diurnal thermal condition differences on evapotranspiration. 2.2 Relative temperature amplitude: ( Standardize and quantify the intensity of intraday temperature fluctuations to avoid interference from the absolute value of average temperature; 2.3 Saturated water vapor pressure difference ( ): Calculated using Tetens' empirical formula. , The saturated vapor pressure, Calculated using actual water vapor pressure, it characterizes the degree of atmospheric drought and the driving force of transpiration.

[0034] Multi-factor coupling and interaction characteristics: Based on the synergistic driving mechanism of evapotranspiration, capturing the composite influence mechanism of core factors: : Characterizes the synergistic effect of air temperature and radiation, reflecting the coupling efficiency of energy supply and thermal drive; It characterizes the combined effect of air temperature and relative humidity, reflecting the potential for atmospheric water vapor exchange; Temporal rhythm coding features: Sine-cosine coding is used to transform the accumulated days of the year into continuous numerical features, representing the annual periodic rhythm of evapotranspiration.

[0035] in Using the accumulated days of the year (1~365), this coding method avoids the orderliness deviation of category coding and accurately depicts the seasonal time sequence dynamics.

[0036] 3. Multi-level feature selection framework: A three-layer feature selection framework is adopted, which prioritizes physical constraints, integrates multiple methods for screening, and optimizes through cross-validation. 3.1 Physically Constrained Retention: Based on the core mechanism of evapotranspiration energy balance, forced retention is achieved. (Thermal-driven fundamentals) (Energy Supply Core) (Key to diurnal thermal fluctuations) Three core physical characteristics ensure that the model output conforms to the basic physical laws of evapotranspiration; 3.2 Multi-method integrated scoring: The Pearson correlation coefficient (PCC) was used. XGBoost Feature Importance (FI) The ensemble method of "Mutual Information (MI)" quantifies and scores the remaining features, and the weighted summation formula is as follows: Weighting criteria: PCC focuses on linear correlation (30%), FI enhances the capture of nonlinear and interaction effects (40%), and MI supplements linear / nonlinear statistical dependence (30%). 3.3 Cross-validation selection: using the XGBoost model validation set corresponding to the feature subset. With the goal of "maximum and minimum RMSE", 5-fold cross-validation is used to progressively filter the top N features in terms of comprehensive score to determine the optimal feature subset.

[0037] (II) Climate Adaptive Global Parameter Optimization Strategy: A three-level parameter optimization strategy is proposed, consisting of "climate unit division - dynamic parameter spatial delimitation - physical constraint global optimization". 1. Division of climate units: Based on five core climate factors (annual average) of the study area average annual average annual average annual After standardizing the data using Z-scores, the K-means clustering algorithm was used to divide the stations into climate units. The number of clusters was determined by silhouette coefficients (optimal silhouette coefficients). 1. Output a list of stations and climate characteristic statistics for each climate unit; 2. Spatial delineation of dynamic parameters: Based on the statistical characteristics (mean, standard deviation, extreme values) of the core climate factors of each climate unit, the core parameters of the HS model are dynamically defined. C, a, E A reasonable search space is needed to avoid physical parameter failure caused by a globally uniform search space. 3. Physically Constrained Differential Evolution (DE) Optimization: An improved DE algorithm is used to globally optimize the HS model parameters of each climate unit. A physical constraint penalty term is introduced to construct the objective function to ensure that the parameter optimization conforms to the physical mechanism of evapotranspiration. Algorithm parameters such as population size NP=50 and iteration number G=100 are set to ensure optimization efficiency and convergence.

[0038] (III) Physical-Data Layered Hybrid Architecture Design: A three-layer hybrid architecture consisting of a physical model layer, a residual correction layer, and a dynamic weighted fusion layer is adopted to achieve complementary advantages between physical mechanisms and data-driven approaches. 1. Physical Model Layer (Basic Estimation Framework): By integrating four types of physics models, a weighted average method (with weights determined by normalizing the NSE values ​​of each model's training set) is used to obtain the output of the physics model layer. : 1.1 Original HS Model: Classic parameter configuration is used, as shown in Table 2; 1.2 Climate Adaptive Parameter HS Model: Loading the optimized parameters of this strategy; 1.3 Simplified Temperature-Radiation Coupled Energy Balance Model: Focusing on the core energy exchange process and simplifying redundant parameters; Table 2 Typical HS Model and Improved Model

[0039] 2. Residual Correction Layer (Core of Accuracy Improvement): Estimating residuals using physical model layers Using the selected optimal feature subset as input and taking it as the target variable, an XGBoost residual correction model is constructed; the hyperparameters (learning rate) are optimized through grid search. Maximum tree depth Minimum sample weights and The model employs 5-fold cross-validation to ensure its generalization ability and outputs the predicted residuals. ; 3. Dynamically Weighted Fusion Layer (Key to Complementary Advantages): The Sequential Least Squares Programming (SLSQP) algorithm is used to solve for the optimal weight coefficients of the physical model layer and the residual correction layer, and a fusion formula is constructed: ) Constraints: and The weighting interval is determined based on the dominant physical mechanism of evapotranspiration in high-altitude and cold regions and is verified through preliminary experiments to ensure the core dominance of the physical model.

[0040] (iv) Construction of a multi-dimensional model verification system: a three-dimensional verification system consisting of "statistical accuracy verification - physical rationality verification - extreme condition stability verification". 1. Statistical accuracy verification: Four core statistical indicators are used to quantify and evaluate the model's accuracy: 1.1 Coefficient of Determination : Characterizes the degree of linear fit between the estimated value and the measured value; 1.2 Root Mean Square Error RMSE : Characterizes the degree of dispersion of the estimation error; 1.3 Mean Absolute Error MAE Characterizes the average level of estimation error; 1.4 Nash-Sutcliffe efficiency coefficient NSE : Characterizes the degree of improvement of the model relative to the mean of the measured values ​​(0.65) Good & Excellent, 0.5 .65: Medium .5: Poor); 2. Physical rationality verification: 2.1 Energy balance closure test: Calculation , (The latent heat of vaporization of water is taken as 2.45 MJ / kg) and The correlation between (net radiation and soil heat flux) requires... ; 2.2 Parameter Sensitivity Analysis: The Morris screening method was used to analyze the core parameters ( C, a, E To perform ±10% perturbation, the parameter change trend must be consistent with the physical laws of evapotranspiration (when the parameter increases...). The estimated value shows an increasing trend). 2.3 Consistency Test of Time Series Trend: The Mann-Kendall trend test is used, requiring that the Z-statistics of the estimated and measured values ​​have the same sign. ; 3. Stability verification under extreme conditions: Three extreme climate scenarios were selected to evaluate the model stability (RMSE increase). For stability): 3.1 Extreme low temperature scenarios: ; 3.2 Extremely Strong Radiation Scenarios: ; 3.3 Persistent drought scenario: 15 consecutive days without precipitation and RH <30%.

[0041] II. Evapotranspiration Estimation System Based on the Improved HS Model with Physical-Data Hybrid Enhancement This system serves as the hardware and software implementation platform for the aforementioned estimation methods. Adopting a modular design concept, it achieves fully automated operation across the entire process, from multi-source data access, feature processing, parameter optimization, model calculation, physical verification, to result output. The system comprises seven core modules, each independent yet collaboratively linked through standardized data interfaces: 1. Data Acquisition Module Function: Enables the integration and access of multi-source data, format standardization, and quality control, providing a reliable data source for subsequent processing; Composition: Data interface unit, format conversion unit, quality control unit; Specific implementation: 1.1 Data Interface Unit: Supports batch uploading and real-time data stream access of meteorological station observation data (HTTP / HTTPS interface), remote sensing products (FTP interface, compatible with MODIS, AMSR2 and other series of products), and reanalysis datasets (API interface, compatible with GLDAS-2.1 and other datasets); 1.2 Format Conversion Unit: Converts the incoming multi-format data (CSV, NetCDF, HDF, etc.) into a standardized CSV format, and unifies the data timestamp (UTC time) and spatial projection (WGS-84 coordinate system). 1.3 Quality Control Unit: Outlier Handling: Grubbs Test (significance level) Identify outliers, especially those exceeding physical thresholds (such as...). ) or logical contradictions (such as For the data, the moving median of the nearest 5 days was used as a substitute or statistical repair was performed based on historical data from the same period; missing value completion: the inverse distance weighting method was used to complete a small number of missing data (missing rate). ), regarding missing rate The site data was marked and removed.

[0042] 2. Feature Processing Module Function: Based on standardized data, it completes basic feature extraction, derived feature generation, and multi-level feature selection, and outputs the optimal feature subset; Composition: Basic feature extraction unit, derived feature generation unit, feature selection unit; Specific implementation: 2.1 Basic Feature Extraction Unit: Extracts 18 basic features from standardized data, generating a time series data. Feature type The basic feature matrix of "number of sites"; 2.2 Derived Feature Generation Unit: Based on preset physical formulas and coding rules, it automatically calculates basic derived features, multi-factor coupled interaction features, and temporal rhythm coding features to generate a derived feature matrix; 2.3 Feature Selection Unit: Embedded with a multi-level feature selection algorithm, first forcibly retaining... , , The system uses three core physical features, and then selects the optimal feature subset through integrated scoring and cross-validation, outputting a feature selection report (including the scores of each feature and the selection logic) and the optimal feature subset data.

[0043] 3. Parameter optimization module Function: Enables adaptive optimization of climate unit division and core parameters of the HS model, and outputs the optimal parameter combination for each climate unit; Composition: Climate unit division unit, parameter space delimitation unit, DE optimization algorithm unit; Specific implementation: 3.1 Climate Unit Division: Embedding K-means clustering algorithm, loading core climate factor data (annual average) average annual average annual average annual After Z-score standardization, clustering is completed, and the clustering results, a list of sites in each unit, and a silhouette coefficient verification report are output. 3.2 Spatial Delineation of Parameter Units: Based on the climate unit division results, and using the statistical values ​​of climate characteristics of each unit, the corresponding parameters are automatically matched ( C, a, E Search space range; 3.3 DE Optimization Algorithm Unit: Embeds an improved DE optimization algorithm, loads physical constraint penalty terms and objective functions, inputs training set data of each climate unit to complete parameter optimization, and outputs the optimal parameter combination and parameter optimization report (including optimization curve and convergence analysis) for each climate unit.

[0044] 4. Hybrid Model Calculation Module Function: Enables collaborative computation of the physical model layer, residual correction layer, and dynamic weighted fusion layer, outputting the final result. Estimation results; Composition: Physical model unit, XGBoost residual correction unit, SLSQP weighted fusion unit; Specific implementation: 4.1 Physical Model Unit: Four types of physical models are deployed in parallel, and the original or optimized parameters are automatically loaded to complete the calculation. The results of the four types of models are integrated using a weighted average method to obtain ET0. phy ; 4.2 XGBoost Residual Correction Unit: Load the trained XGBoost model (optimized hyperparameters: , , ), input the optimal feature subset to calculate ; 4.3 SLSQP Weighted Fusion Unit: Embedded with the SLSQP algorithm, using " RMSE With the objective of "minimizing", the optimal weight coefficients are solved under constraints, and then substituted into the fusion formula to obtain the result. .

[0045] 5. Physical verification module Function: To The estimation results are physically validated, and outlier results are eliminated. Composition: Energy balance verification unit, parameter sensitivity verification unit, and time series trend verification unit; Specific implementation: 5.1 Energy Balance Verification Unit: Calculation and Rn G The correlation, if Then mark it as an abnormal result; 5.2 Parameter Sensitivity Verification Unit: Performs verification on core parameters. Disturbances whose parameter change trends contradict the physical laws of evapotranspiration are marked as abnormal results. 5.3 Time Series Trend Verification Unit: The Mann-Kendall trend test is used. If the time series trends of the estimated and measured values ​​are inconsistent (e.g., the Z statistic has opposite signs or...), the unit will be verified. If ) is an abnormal result; 5.4 Handling of abnormal results: For marked abnormal results, the results are automatically fed back to the hybrid model calculation module for recalculation (up to 3 iterations). If the result is still abnormal, an early warning message is output and the cause of the abnormality is recorded.

[0046] 6. Results Visualization and Output Module Functionality: Enables multi-dimensional visualization and multi-format output of estimation results, supporting result analysis and application; Components: Visualization unit, output unit; Specific implementation: 6.1 Visualization Unit: Supports real-time display of time-series change curves (daily / monthly / yearly scales), spatial distribution heatmaps, statistical indicator comparison bar charts / box plots, and model operation flowcharts, and supports interactive operations (zooming, filtering, and exporting). 6.2 Output Unit: Supports outputting estimation results in formats such as CSV, Excel, and NetCDF, and automatically generates PDF statistical reports, including model parameter configuration, feature information, validation index values, anomaly statistics, and model accuracy comparison analysis.

[0047] 7. Control Module Function: As the core control center of the system, it enables coordinated operation of various modules, parameter configuration, and monitoring of operating status; Composition: Module scheduling unit, parameter configuration unit, status monitoring unit; Specific implementation: 7.1 Module Scheduling Unit: Automatically schedules each module to run sequentially according to the preset process (data acquisition → feature processing → parameter optimization → model calculation → physical verification → result output), and supports custom process configuration; 7.2 Parameter Configuration Unit: Supports user-defined settings of core parameters (DE algorithm parameters, XGBoost hyperparameters, validation metric thresholds, data interface parameters, etc.), and saves the parameter configuration scheme for later use; 7.3 Status Monitoring Unit: Monitors the running status of each module in real time, records the running log (including running time, data volume, algorithm execution progress, error information, etc.), and automatically alarms and triggers backup plans (such as restarting the module, switching the backup data source, etc.) if a module failure occurs.

[0048] III. Implementation Area and Data Preparation This embodiment selects the high-altitude and cold region of the Qinghai-Tibet Plateau as the typical implementation area, which has significant climatic complexity and underlying surface heterogeneity. The system is deployed on a server equipped with an Intel Xeon Gold 6230R processor, 128GB of memory, and a 2TB SSD. The operating system is Ubuntu 20.04 LTS, and the software environment is built on Python 3.9. The main dependent libraries include Pandas 1.4.2, Scikit-learn 1.0.2, SciPy 1.8.0, and XGBoost 1.5.1.

[0049] Data acquisition encompassed multiple datasets: ground observation data came from the TORP observation network (including six stations: Nyingchi, Nagqu, Namtso, Mount Everest, Ali, and Muztagh Ata) and the China Meteorological Administration; soil moisture data used China's 1km daily soil moisture product; and vegetation data was based on the 0.05° gapless NDVI product for China developed from NOAA's daily NDVI product from 1981 to 2023. All data were obtained through official channels such as the National Tibetan Plateau Data Center to ensure traceability and authority.

[0050] 1. Implementation area: This study selected measured data from the TORP observation network, including surface radiation and latent heat flux data. The stations from east to west are Nyingchi (SETORS), Nagqu (BJ), Namtso (NAMORS), QOMS, Ngari (NADORS), and Muztagh Ata (MAWORS).

[0051] 2. Data Sources: TORP data is obtained from the National Tibetan Plateau Data Center (https: / / data.tpdc.ac.cn). Meteorological data is from the National Meteorological Science Data Center of the China Meteorological Administration (http: / / data.cma.cn / ). Soil moisture data is based on daily 1km soil moisture data from various stations in China (https: / / data.tpdc.ac.cn / zh-hans / data / 49b22de9-5d85-44f2-a7d5-a1ccd17086d2). The Normalized Difference Vegetation Index (NDVI) in the vegetation data is a daily 0.05° gapless NDVI product for China from 1981 to 2023, developed based on the daily NDVI product from the National Oceanic and Atmospheric Administration (NOAA) (https: / / figshare.com / s / 16f1fbaff259272249f1 / articles / 24922725?file=43865235) and combined with effective data identification and spatiotemporal filling technology.

[0052] 3. System Environment Configuration: This system is deployed on a server (CPU: Intel Xeon Gold 6230R, Memory: 128GB, Hard Disk: 2TB SSD), the operating system is Ubuntu 20.04 LTS, the software environment is based on Python 3.9, and the dependent library versions are Pandas 1.4.2, Scikit-learn 1.0.2, SciPy 1.8.0, XGBoost 1.5.1, Matplotlib 3.5.1, and Flask 2.0.1 (for system visualization interface development).

[0053] IV. Implementation Steps of the Estimation Method (I) Construction and Implementation of Multi-Dimensional Feature System 1. Construction of basic feature library: Multi-source data is accessed through the system data acquisition module. After format conversion and quality control, the basic feature extraction unit of the feature processing module extracts 18 basic features to form a basic feature matrix.

[0054] 2. Derived Feature Generation: The derived feature generation unit of the feature processing module calculates the derived feature according to a preset formula. Relative temperature amplitude Three types of basic derived features, , Two types of interaction features, ;in For accumulated days within the year (1 365) Two types of time-series coding features are used to form a derived feature matrix.

[0055] 3. Multi-level feature selection: 3.1 Feature selection units in the feature processing module are forcibly retained. , , Core physical characteristics; 3.2 Calculate PCC, FI, and MI scores for the remaining 25 feature classes, and then sum them using weighted averages. Receive a comprehensive score; 3.3 This paper expands the 15 basic features into 28 derived features, and after screening, 8 core features are obtained. This process reduces redundant features by dimensionality reduction and eliminates the risk of overfitting, while retaining key physical driving features, thereby improving the model's interpretability and predictive reliability.

[0056] (II) Implementation of Climate Adaptive Parameter Optimization 1. Climate Unit Division: The climate unit division unit of the parameter optimization module loads the annual average values ​​of 5 core climate factors from 6 stations. After Z-score standardization, K-means clustering algorithm is used for clustering. The number of clusters is determined to be 3 through silhouette coefficient verification (optimal silhouette coefficient = 0.72), and the list of stations for each climate unit is output.

[0057] 2. Dynamic parameter space definition: The parameter space definition unit of the parameter optimization module automatically matches the parameter search space of each climate unit based on the clustering results.

[0058] 3. Physical constraint DE optimization: 3.1 Input Data: Training sets for each climate unit 、 、 、 and actual measurement data; 3.2 Algorithm Execution: The DE optimization algorithm unit of the parameter optimization module loads the preset algorithm parameters and substitutes them into the objective function for global optimization; 3) Output results: Obtain the optimal parameter combination of the HS model, construct and store the climate adaptive parameter library.

[0059] (III) Implementation of Physical-Data Layered Hybrid Architecture 1. Physical Model Layer Implementation: The physical model unit of the hybrid model calculation module runs four types of physical models in parallel. After completing the calculation by inputting the corresponding parameters, the results are arithmetically averaged to obtain the result. .

[0060] 2. Implementation of residual correction layer: 2.1 Residual Calculation: Based on the Training Set and Calculate residuals ; 2.2 Model Training: The XGBoost residual correction unit of the hybrid model computation module takes 12 optimal feature subsets as input and Res as output, loads the optimized hyperparameters for model training, and uses 5-fold cross-validation for optimization; 2.3 Residual Prediction: Input the validation set features into the trained model to obtain... .

[0061] 3. Dynamic weighted fusion implementation: The SLSQP weighted fusion unit of the hybrid model calculation module loads the validation set. and ,by" RMSE With the objective of "minimizing", the optimal weights are solved under constraints, and then substituted into the fusion formula to obtain the result. .

[0062] (iv) Implementation of multi-dimensional model verification 1. Statistical accuracy verification: Calculate the validation set. Compared with actual measurement Four types of statistical indicators RMSE and NSE are all within the expected range.

[0063] 2. Physical rationality verification: 2.1 Energy balance closure test: Calculation and The correlation, This satisfies the energy balance constraint; 2.2 Parameter Sensitivity Analysis: The Morris screening method was used to analyze the core parameters. The results showed that the parameter sensitivity was ranked as follows: And when the parameter increases The estimated values ​​all show an increasing trend, which is consistent with physical laws; 2.3 Consistency test of time series trend: The monthly trend test shows that the estimated value and the measured value Z statistic are both positive and consistent in trend.

[0064] 3. Stability verification under extreme conditions: 3.1 Extracting data from three types of extreme scenarios (extreme low temperature: 128 samples; extreme strong radiation: 96 samples; persistent drought: no rainfall for 15 consecutive days and (72 samples) 3.2 The RMSE of each scenario was calculated, and the increase was less than 20%, indicating excellent stability.

[0065] V. System Implementation Steps (I) System Deployment and Parameter Configuration 1. Deploy the evapotranspiration estimation system on the server and install the required software environment and dependency libraries; 2. The hybrid model computation module implements a three-layer collaborative computing architecture: the physical model unit runs four types of HS physical models in parallel (including the original HS model and three improved versions), and integrates the output using a weighted average method. Set core parameters: DE algorithm parameters (NP=50, G=100, etc.), XGBoost hyperparameters ( , etc.), verification indicator threshold ( RMSE increase in extreme scenarios Record parameters such as FTP server address and API key, and save the parameter configuration scheme.

[0066] (II) Data Acquisition and Preprocessing 1. Start the system data acquisition module and access multi-source data (meteorological station data, remote sensing products, GLDAS dataset) through the data interface unit. 2. The format conversion unit converts the incoming multi-format data into a standardized CSV format, unifying the timestamps and spatial projections; 3. The quality control unit diagnoses and corrects outliers and fills in missing values ​​in the data, outputting standardized data.

[0067] (III) Feature Processing and Parameter Optimization 1. The control module schedules the feature processing module to start, and completes basic feature extraction, derived feature generation and multi-level feature selection based on standardized data, and outputs the optimal feature subset; 2. The control module schedules the parameter optimization module, which is started. Based on the core climate factor data, it completes the division of climate units, the spatial definition of dynamic parameters and the optimization of physical constraints (DE), and outputs the optimal parameter combination for each climate unit.

[0068] (iv) Model calculation and physical verification 1. The control module schedules the startup of the hybrid model calculation module, loads the optimal feature subset and optimal parameters, completes the calculations of the physical model layer, residual correction layer, and dynamic weighted fusion layer, and outputs the results. ; 2. The control module schedules the physical verification module to start, and... Physical rationality verification is performed. In this embodiment, there are no abnormal results, so proceed directly to the next step. If there are abnormal results, the system will automatically send feedback to the hybrid model calculation module for recalculation.

[0069] (v) Results visualization and output 1. The control module scheduling result visualization and output module is started, and the visualization unit generates time-series change curves, spatial distribution heat maps, statistical accuracy comparison charts, etc.; 2. The output unit outputs the estimation results in CSV and Excel formats and automatically generates a PDF statistical report, which includes model parameters, feature information, validation metrics, etc.

[0070] (vi) System operation monitoring system The status monitoring unit of the control module monitors the operating status of each module in real time and records the operation log (running time, data volume, no error information). In this embodiment, the system operates stably without fault alarms.

[0071] VI. Implementation Safeguards and Repeatability Explanation 1. Data traceability: All data used in this embodiment comes from official public channels and can be obtained through the corresponding official websites, ensuring the authenticity and repeatability of the data; 2. Environment Configuration: The hardware and software environment deployed in the system are all general configurations, and the dependent libraries are all publicly available versions that can be easily replicated; 3. Step Reproduction: All parameter settings, calculation processes, and system operation steps in this embodiment have been clearly quantified. Those skilled in the art can completely reproduce the implementation process and results of this invention after obtaining the same data source and configuring the same hardware and software environment.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A method for estimating evapotranspiration using an improved HS model based on physical-data hybrid enhancement, characterized in that: Includes the following steps: Step 1: Construct a multi-dimensional feature system, which includes basic features, derived features, and constraint features; Step 2: Optimize the HS model parameters using a climate-adaptive parameter optimization strategy; Step 3: Calculate reference evapotranspiration using a physical-data layered hybrid architecture ET The physical-data layered hybrid architecture, with a value of 0, includes a physical model layer, a residual correction layer, and a dynamic weighted fusion layer. Step 4: Verify the estimation results from multiple dimensions.

2. The evapotranspiration estimation method based on the HS improved model with physical-data hybrid enhancement according to claim 1, characterized in that: In step one, the construction of a multi-dimensional feature system specifically includes: screening basic features based on the evapotranspiration energy balance equation and the water transport mechanism; generating derived features based on physical mechanisms, including energy-water constraint features, multi-factor coupling interaction features, and temporal rhythm coding features; and using a multi-level feature selection framework to screen features, including physical constraint forced retention, multi-method integrated scoring, and cross-validation optimization.

3. The evapotranspiration estimation method based on the HS improved model with physical-data hybrid enhancement as described in claim 2, characterized in that: In the multi-level feature selection framework, the physical constraint-driven retention refers to the mandatory retention of the average daily temperature. Solar radiation from the top of the atmosphere and daily temperature range Three core physical characteristics; The multi-method ensemble scoring employs a weighted ensemble scoring based on Pearson correlation coefficient (PCC), XGBoost feature importance (FI), and mutual information (MI). The weighting formula is as follows: ;in The Pearson correlation coefficient score, XGBoost feature importance score, The mutual information score is used; the cross-validation optimization uses the validation set determination coefficient. Maximum and root mean square errors RMSE Minimum is the target.

4. The evapotranspiration estimation method based on the HS improved model with physical-data hybrid enhancement according to claim 1, characterized in that: In step two, the climate adaptive parameter optimization strategy specifically includes: Climate units are divided based on core climate factors; The search space for HS model parameters is dynamically defined based on the statistical characteristics of climate units. The physically constrained differential evolution (DE) algorithm is used to globally optimize the parameters, where the HS model parameters include... C , a and E , C This is an empirical coefficient. a These are temperature-related parameters. E These are radiation-related parameters.

5. The evapotranspiration estimation method based on the HS improved model with physical-data hybrid enhancement according to claim 1, characterized in that: In step three, the computation of the physical-data layered hybrid architecture specifically includes: integrating four types of physical models at the physical model layer and outputting the result using a weighted average method. ,in The output values ​​are from the physical model layer; the residuals are predicted using the XGBoost model in the residual correction layer. Among them, residual , The measured evapotranspiration values ​​were used; the Sequential Least Squares (SLSQP) algorithm was employed in the dynamic weighted fusion layer to solve for the weights and calculate the final values. The fusion formula is ,in and These are the weighting coefficients. Output for the physical model layer. To predict the residuals, the constraints are as follows: + =1 and ∈[0.6,0.8].

6. The evapotranspiration estimation method based on the HS improved model with physical-data hybrid enhancement according to claim 1, characterized in that: In step four, the multi-dimensional verification includes statistical accuracy verification, physical rationality verification, and extreme condition stability verification. The statistical accuracy verification uses the coefficient of determination. Root mean square error RMSE Mean absolute error MAE Nash-Sutcliffe efficiency coefficient NSE index; The physical rationality verification includes energy balance closure test, parameter sensitivity analysis, and time series trend consistency test. The extreme condition stability verification assesses the increase in RMSE under extreme low temperature, extreme high radiation, and persistent drought scenarios.

7. A system for estimating evapotranspiration using a HS improved model based on physical-data hybrid enhancement, for implementing the HS improved model evapotranspiration estimation method based on physical-data hybrid enhancement as described in any one of claims 1 to 6, characterized in that: It includes a data acquisition module, a feature processing module, a parameter optimization module, a hybrid model calculation module, a physical verification module, a result visualization and output module, and a control module; The data acquisition module is used to access multi-source data; The feature processing module is used to construct and select features at multiple levels; The parameter optimization module is used to perform climate adaptive parameter optimization; The hybrid model computation module is used for computation via a physical-data hierarchical hybrid architecture. Value; the physical verification module is used to verify the physical reasonableness of the result; The result visualization and output module is used to output the results; the control module is used to coordinate the operation of each module; and each module is connected through a standardized data interface.

8. The evapotranspiration estimation system based on the HS improved model with physical-data hybrid enhancement according to claim 7, characterized in that: The feature processing module includes a basic feature extraction unit, a derived feature generation unit, and a feature selection unit; The parameter optimization module includes a climate unit partitioning unit, a parameter space delimitation unit, and a DE optimization algorithm unit; The hybrid model computation module includes a physical model unit, an XGBoost residual correction unit, and an SLSQP weighted fusion unit. The physical verification module includes an energy balance verification unit, a parameter sensitivity verification unit, and a time-series trend verification unit.

9. The evapotranspiration estimation system based on the HS improved model with physical-data hybrid enhancement according to claim 7, characterized in that: The data acquisition module supports multi-source data access, including meteorological station data, remote sensing products, and reanalysis datasets, and includes a format conversion unit and a quality control unit. The result visualization and output module supports visualization of time-series variation curves, spatial distribution heat maps, and statistical accuracy comparison charts, and outputs the results in CSV, Excel, and PDF formats.

10. The evapotranspiration estimation system based on the HS improved model with physical-data hybrid enhancement according to claim 7, characterized in that: The control module includes a parameter configuration unit and a status monitoring unit. The parameter configuration unit is used to set DE algorithm parameters, XGBoost hyperparameters and verification index thresholds. The status monitoring unit is used to monitor the module's operating status in real time and record operating logs.

Images