Evaporation pan evaporation uniformization sequence reconstruction method based on multi-source data fusion

Abstract

The application discloses an evaporation dish evaporation amount homogenization sequence reconstruction method based on multi-source data fusion, specifically, daily evaporation dish evaporation amount data of each station in a target area, meteorological data and station metadata are screened, automatic secondary quality control is performed on the screened data, and a high-quality basic data set is obtained; a random forest regression and recursive feature elimination cross-validation method is used for selecting features, and finally a partial least squares regression model is used for splicing and interpolation on the daily evaporation dish evaporation amount data, so that the evaporation dish evaporation amount reconstruction sequence is generated; the evaporation dish evaporation amount reconstruction sequence is subjected to homogenization test, breakpoints are identified and corrected, and the evaporation dish evaporation amount homogenization reconstruction sequence is obtained, so that the problem of discontinuous evaporation amount data sequence caused by non-climatic factors such as mixed use of large and small evaporation dishes is solved, and the limitations of traditional conversion coefficient method and multivariate regression equation, such as large difference between stations, insufficient consideration of meteorological conditions and the like, are overcome.

Description

Technical Field

[0001] This invention belongs to the field of meteorological data processing technology, specifically relating to a method for reconstructing a uniform sequence of evaporation amount in evaporation pans based on multi-source data fusion. Background Technology

[0002] Surface evaporation is crucial for understanding the patterns and causes of climate change. However, due to historical reasons, my country uses a mix of large and small evaporation pans for observation. Furthermore, the data is influenced by multiple factors, including observation techniques, station locations, observation times, and methods, resulting in unavoidable inconsistencies in historical evaporation data, which severely impacts the quality of surface meteorological data. Current technologies address the issue of mixed evaporation pan use primarily by employing conversion factor methods to unify observation data from different evaporation pans. Other methods utilize partial least squares to build models that estimate daily evaporation from large evaporation pans. Researchers have also used methods such as Random Forest (RF), Deep Neural Network (DNN), and Support Vector Regression (SVR) to model predicted evaporation from evaporation pans. Despite these advancements, nonlinear modeling methods based on machine learning still face challenges in handling nonlinear relationships in evaporation and splicing evaporation data from two-source evaporation pans, including low model interpretability, susceptibility to overfitting, and strong dependence on data quality.

[0003] In homogenization studies, the SNHT (Standard Normal Homogeneity Test), RHtests, and MASH (Multiple Analysis of Series for Homogenization) methods have been primarily used to homogenize normally distributed meteorological elements such as temperature and precipitation. However, in practical applications of homogenized evaporation, different methods have different limitations. First, the conversion factor method does not consider other factors that may affect evaporation, resulting in potentially incomplete results that cannot fully reflect complex realities. Linear multiple regression equations perform poorly in handling nonlinear relationships, are sensitive to multicollinearity, fail to adequately consider differences between stations, and are susceptible to outliers. Second, research on non-normally distributed evaporation using homogenization methods is still in its early stages. Existing non-homogeneity tests for other meteorological elements are mostly based on the MASH method, but its regional application is limited due to the low tolerance for missing data and the varying start and end times of evaporation observations among ground observers nationwide. Summary of the Invention

[0004] The purpose of this invention is to provide a method for reconstructing a uniform evaporation sequence of evaporation pans based on multi-source data fusion, so as to solve the problems of inconsistent evaporation data and discontinuous time series caused by differences in observation instruments.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for reconstructing the homogenization sequence of evaporation amount in evaporating dishes based on multi-source data fusion, comprising the following steps:

[0006] S1: Screen the daily evaporation data of small evaporation pans and large evaporation pans of ground stations in the target area, meteorological data and station metadata. Based on the station metadata, the stations are divided into overlapping observation stations, connecting observation stations and pure small observation stations. Secondary quality control is performed on the screened data to obtain a high-quality basic dataset.

[0007] S2: Based on random forest regression and recursive feature elimination cross-validation, the optimal feature subset of each station is selected, and a partial least squares regression model is constructed based on the optimal feature subset. The optimal hyperparameter of the model is searched using Grid Search CV cross-validation to obtain the daily data prediction model of evaporation of large evaporating pans.

[0008] S3: Using the partial least squares regression model, the daily evaporation data of the evaporating dish are spliced ​​and interpolated to generate a reconstructed evaporation sequence of the evaporating dish.

[0009] S4: Perform a homogenization test on the reconstructed evaporation amount sequence of the evaporating dish, identify non-homogeneous breakpoints, correct the breakpoints, and obtain the homogenized reconstructed evaporation amount sequence of the evaporating dish.

[0010] Furthermore, the specific steps of S1 are as follows:

[0011] S1.1: Obtain daily data on evaporation of small evaporating pans, daily data on evaporation of large evaporating pans, meteorological element data, and station metadata of ground stations within the target area. Based on the observation instrument records in the station metadata, classify each station into overlapping observation stations, connecting observation stations, and purely small observation stations.

[0012] S1.2: Organize the data by station number to form a long-sequence basic dataset containing year, month, day, evaporation data and observation type identification code;

[0013] S1.3: Perform secondary quality control on the evaporation data in the long sequence basic dataset to remove extreme outliers in the evaporation of the evaporating dish, and obtain a high-quality basic dataset.

[0014] Furthermore, the specific steps for secondary quality control of evaporation data in S1.3 are as follows:

[0015] S1.3.1: Perform boundary value tests on the daily evaporation data of small and large evaporating dishes in the long-sequence basic dataset to obtain the data that passes the boundary value test;

[0016] S1.3.2: Perform climate anomaly testing on the data that have passed the threshold test, and mark the data that exceed the range of 3 to 5 times the standard deviation as suspicious values;

[0017] S1.3.3: Perform adjacent-day variation analysis on the suspicious values, calculate the absolute difference between the current day's evaporation and the adjacent day's evaporation, and combine it with the same day's precipitation, relative humidity and sunshine duration for judgment and quality control;

[0018] S1.3.4: All suspicious values ​​are re-verified and corrected to obtain a high-quality basic dataset.

[0019] Furthermore, the calculation of the absolute difference between the current daily evaporation and the evaporation of adjacent days corresponding to the suspicious value, as described in S1.3.3, for judgment and quality control purposes specifically includes:

[0020] Let PE i Let PE be the evaporation on day i. i-1 PE represents the evaporation on day i-1. i+1 Let PE be the evaporation on day i+1. diff PE is the evaporation amount on day i-1. i-1 Evaporation PE on day i+1 i+1 Absolute difference, n is the year of evaporation observation, QC is the evaporation after quality control, PRE is the precipitation on the same day, RH is the relative humidity, and SSD is the sunshine duration;

[0021] Calculate the evaporation PE on day i. i Evaporation PE on day i-1 i-1 Or the evaporation PE on day i+1 i+1 absolute difference PE diff :

[0022] If PE is satisfied diff If the thickness is ≥5mm and simultaneously meets the following conditions: PRE≥10mm, RH≥80%, and SSD≤2h, then PE is determined. i This is a non-extreme outlier, and the quality control value is 0.0.

[0023] Furthermore, the specific steps of S2 are as follows:

[0024] S2.1: Using the data of overlapping observation periods of overlapping observation stations as training samples and time series coding, extract the daily data of small evaporation pan evaporation and related meteorological elements of each station as initial feature variables. The meteorological elements include temperature, wind speed, precipitation, relative humidity, water vapor pressure and sunshine duration. The evaporation of large evaporation pans in the same period is used as the target variable.

[0025] S2.2: The importance of each feature to the target variable is evaluated based on the random forest regression method, and the optimal feature subset of each site is selected by combining the recursive feature elimination cross-validation method.

[0026] S2.3: Based on the optimal feature subset, construct a partial least squares regression model, and use Grid Search CV cross-validation to search for the optimal hyperparameters of the model to obtain a daily data prediction model for the evaporation of large evaporating pans.

[0027] Furthermore, the specific method of S3 is as follows:

[0028] For purely small-scale observation stations, the original small-scale evaporation pan observation sequence is extracted from the high-quality basic dataset after S1 quality control processing, and this original observation sequence is directly used as the evaporation reconstruction sequence of the purely small-scale observation station.

[0029] For the connecting observation station, its observation history is divided into an early small evaporation pan observation stage and a later large evaporation pan observation stage. In the early stage, n qualified overlapping observation stations around the connecting observation station are selected, and the partial least squares regression model trained on them is used to convert the small evaporation pan observation data into equivalent predicted values ​​for large evaporation pans, and then interpolation is performed to generate the missing data estimate for the connecting observation station in the early stage. In the later stage, the measured observation value sequence of large evaporation pans is directly used, and the estimated value of the connecting observation station in the early stage is spliced ​​with the measured observation value sequence to form the evaporation reconstruction sequence of the connecting observation station.

[0030] For overlapping observation stations, the partial least squares regression model trained during the overlapping observation period is applied to convert all available small evaporation pan observation data of the station into equivalent predicted values ​​of large evaporation pans, and this conversion result is used as the evaporation reconstruction sequence of the overlapping observation station.

[0031] The evaporation reconstruction sequences of pure small observation stations, connected observation stations, and overlapping observation stations are fused to obtain the evaporation reconstruction sequence of the evaporation pan.

[0032] Furthermore, the specific steps of S4 are as follows:

[0033] S4.1: Based on the evaporation rate reconstruction sequence of the evaporating dish generated in S3, a regional reference sequence is constructed using the first-order difference method and the correlation coefficient weighted average method;

[0034] S4.2: Apply the penalized maximum T-test and penalized maximum F-test in the RHtests algorithm to perform a homogenization test on the reconstructed sequence and identify breakpoints;

[0035] S4.3: The identified breakpoints are corrected using the quantile mapping correction method to obtain a uniform reconstruction sequence of the evaporation amount of the evaporating dish.

[0036] 8. The method for reconstructing the homogenized evaporation sequence of evaporation pans based on multi-source data fusion according to claim 6, characterized in that the calculation of the estimated value of missing data from the previous observation station includes the following steps:

[0037] Step a: Select n qualified overlapping meteorological stations from the surrounding meteorological stations connecting the observation station;

[0038] Step b: For qualified overlapping observation stations, apply the partial least squares regression model obtained during the overlapping observation period to convert all available small evaporating pan observation data of the station into equivalent predicted values ​​for large evaporating pans.

[0039] Step c: The equivalent predicted values ​​of the large evaporating dish are fused using an interpolation algorithm to generate estimated values ​​of the missing data from the previous observation station. The interpolation algorithm is shown in equation (1):

[0040]

[0041] Where: E e,i It is the estimated value of the missing data from observation station i, EF ij It is a linear regression estimate of the missing data of overlapping observation station j on the connecting observation station i, RMSE j It is the root mean square error of the regression model evaluation for the overlapping observation station j.

[0042] Compared with the prior art, the beneficial effects of the present invention are:

[0043] 1) This invention innovatively constructs a machine learning-based dynamic feature selection PLS (Partial Least Squares) regression model. By using methods such as random forest regression, recursive feature elimination, and cross-validation, features are dynamically selected for each ground observation station. This machine learning-based dynamic feature selection PLS regression model enables efficient and accurate prediction and interpolation, completes the splicing of evaporation sequences of large and small evaporation pans, and can achieve localized correction for each station, effectively reducing errors.

[0044] 2) This invention solves the historical problem of discontinuous evaporation data sequences caused by the mixed use of large and small evaporation pans. At the same time, it overcomes the limitations of traditional conversion factor method and multiple regression equation, such as large differences between stations and insufficient consideration of meteorological conditions, and helps to accurately grasp the spatiotemporal variation trend of evaporation.

[0045] 3) This invention effectively improves the continuity and uniformity of evaporation data by integrating machine learning and homogenization methods, overcoming the limitations of traditional methods in terms of station adaptability and meteorological factor considerations. This method fully utilizes the advantages of random forest models in capturing nonlinear relationships and complex feature interactions, combined with partial least squares regression's ability to handle multicollinearity of variables, forming a complementary modeling mechanism. Taking Hunan Province as an example, by dynamically selecting features station by station and constructing a prediction model, high-precision reconstruction of long-term evaporation sequences based on large evaporation pans is achieved. The generated homogenized dataset can truly reflect the historical climate change process of regional evaporation, possessing significant scientific value and application significance. Attached Figure Description

[0046] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0047] Figure 1 This is a diagram illustrating the architecture of the method for reconstructing the evaporation amount homogenization sequence of evaporating dishes based on multi-source data fusion proposed in this invention.

[0048] Figure 2 A schematic diagram illustrating the overall technical approach and process for climate sequence homogeneity testing and correction;

[0049] Figure 3 This diagram illustrates the relationship between the number of features and model performance during feature selection using RFECV combined with random forest regression.

[0050] Figure 4 The diagram shows the prediction results of the daily data prediction model for the evaporation of a large evaporating dish proposed in this invention, where (a) is the prediction result of the evaporation of the evaporating dish at station 57554 (overlapping observation station), and (b) is the prediction result of the evaporation of the evaporating dish at station 57574 (connected observation station).

[0051] Figure 5 The results of the performance analysis of the daily data prediction model for the evaporation rate of the large evaporating dish proposed in this invention are as follows: (a) is the coefficient of determination (R²). 2 (a) Experimental results of Score, correlation coefficient (R), upper confidence interval (CI Upper) and lower confidence interval (CI Lower), and (b) experimental results of root mean square error (RMSE) and mean square error (MSE);

[0052] Figure 6A radar chart comparing the standard deviation, coefficient of variation, and interquartile range of the evaluation indicators between the raw evaporation data and the homogenized evaporation data;

[0053] Figure 7 The diagram shows the spatial distribution of the basic dataset of Hunan Province and the reconstructed evaporation sequence of the evaporation pan. (a) is the spatial distribution of the evaporation change rate of the basic dataset of Hunan Province, (b) is the spatial distribution of the evaporation change rate of the reconstructed evaporation sequence of the evaporation pan, (c) is the spatial distribution of the climate trend rate of the basic dataset of Hunan Province, and (d) is the spatial distribution of the climate trend rate of the reconstructed evaporation sequence of the evaporation pan.

[0054] Figure 8 This diagram illustrates the spatial distribution of climate tendency rates in the basic dataset and the homogenized reconstruction sequence of evaporation from evaporation pans for each season in Hunan Province. (a) shows the spatial distribution of climate tendency rates in the basic dataset for spring in Hunan Province; (b) shows the spatial distribution of climate tendency rates in the homogenized reconstruction sequence of evaporation from evaporation pans for spring in Hunan Province; (c) shows the spatial distribution of climate tendency rates in the basic dataset for summer in Hunan Province; (d) shows the spatial distribution of climate tendency rates in the homogenized reconstruction sequence of evaporation from evaporation pans for summer in Hunan Province; (e) shows the spatial distribution of climate tendency rates in the basic dataset for autumn in Hunan Province; (f) shows the spatial distribution of climate tendency rates in the homogenized reconstruction sequence of evaporation from evaporation pans for autumn in Hunan Province; (g) shows the spatial distribution of climate tendency rates in the basic dataset for winter in Hunan Province; and (h) shows the spatial distribution of climate tendency rates in the homogenized reconstruction sequence of evaporation from evaporation pans for winter in Hunan Province.

[0055] Figure 9 The data are normalized evaporation data, where (a) is the frequency distribution of the month at the breakpoint, (b) is the frequency distribution of the year at the breakpoint, (c) is the original and normalized daily evaporation values, and (d) is the frequency distribution of the monthly data.

[0056] Figure 10 The data shows the correction magnitude for the annual evaporation value, where (a) is the frequency distribution of the correction magnitude, (b) is the frequency distribution of the average correction magnitude for a single station, and (c) is the spatial distribution of the correction magnitude.

[0057] Figure 11 This is a diagram comparing evaporation from August to November 2022 with uniformized evaporation.

[0058] Figure 12 This is a diagram showing the comparison of evaporation in September 2024, where (a) is a diagram of the basic evaporation climate value and (b) is a diagram of the homogenized evaporation climate value. Detailed Implementation

[0059] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] like Figure 1 As shown, the method for reconstructing the homogenization sequence of evaporation amount in evaporating dishes based on multi-source data fusion proposed in this invention is implemented according to the following steps:

[0061] S1: Screen the daily evaporation data of small evaporation pans and large evaporation pans of ground stations in the target area, meteorological data and station metadata. Based on the station metadata, the stations are divided into overlapping observation stations, connecting observation stations and pure small observation stations. Secondary quality control is performed on the screened data to obtain a high-quality basic dataset.

[0062] The specific steps for S1 are as follows:

[0063] S1.1: Obtain daily data on evaporation of small evaporating pans, daily data on evaporation of large evaporating pans, meteorological element data, and station metadata of ground stations within the target area. Based on the observation instrument records in the station metadata, classify each station into overlapping observation stations, connecting observation stations, and purely small observation stations.

[0064] In the specific implementation of this application, the data used in S1.1 specifically comes from the daily observation data of 97 national-level ground meteorological stations in Hunan Province, which are extracted from the daily value dataset of basic meteorological elements of national-level ground meteorological stations in China (V3.0) released by the National Meteorological Information Center. At the same time, the daily evaporation data of small evaporating pans and large evaporating pans are obtained by combining the monthly report A and J files of each national-level ground meteorological station.

[0065] In the specific implementation of this application, based on the instrument replacement records and observation period information in the station metadata, each station is divided into three types: overlapping observation stations (the period when both large and small evaporating pan observation data are available), connecting observation stations (small and large evaporating pan observations are conducted successively), and pure small observation stations (small evaporating pan observations are conducted only).

[0066] S1.2: Organize the data by station number to form a long-sequence basic dataset containing year, month, day, evaporation data and observation type identification code;

[0067] In a specific implementation of this application, the data processing in S1.2 includes: compiling all observation data since the station's establishment according to station number to form a long-sequence basic dataset based on the evaporation of large evaporating pans. This dataset contains five fields: year, month, day, evaporation PE (mm), and type identifier code. The type identifier code is defined as follows: 1 represents a large evaporating pan observation, 2 represents a small evaporating pan observation, and 0 represents a missing value. Data quality statistics show that the total missing rate for daily evaporation values ​​from 97 stations is 2.62%. The following statistical rules are used to ensure data continuity: when the number of missing days in a month is less than 10, the statistical value for that month is considered valid, with a monthly missing rate of 1.93%; when the number of missing months in a year is less than 3, the statistical value for that year is considered valid, with an annual missing rate of 0.80%.

[0068] S1.3: Perform secondary quality control on the evaporation data in the long-sequence basic dataset to remove extreme outliers in the evaporation pan caused by precipitation interference, and obtain a high-quality basic dataset.

[0069] The specific steps for secondary quality control of evaporation data in S1.3 are as follows:

[0070] S1.3.1: Perform boundary value tests on the daily evaporation data of small and large evaporating dishes in the long-sequence basic dataset to obtain the long-sequence data that passes the boundary value test;

[0071] In a specific embodiment of this application, the method for verifying the daily data boundary value of evaporation amount for small and large evaporating dishes is as follows: the reasonable range of the daily evaporation amount of the evaporating dish is set to 0 mm to 30 mm, and data exceeding this range is considered erroneous data and is discarded.

[0072] S1.3.2: Perform climate anomaly testing on the data that have passed the threshold test, and mark the data that exceed the range of 3 to 5 times the standard deviation as suspicious values;

[0073] S1.3.3: Based on the observation specifications and the principle of water vapor pressure balance, calculate the absolute difference between the current daily evaporation and the evaporation of adjacent days corresponding to the suspicious value, and make judgments and quality control.

[0074] Calculate the absolute difference between the current daily evaporation and the evaporation of adjacent days corresponding to the suspicious value, and make judgments and quality control decisions. Specifically, this includes:

[0075] Let PE i Let PE be the evaporation on day i. i-1 PE represents the evaporation on day i-1. i+1 Let PE be the evaporation on day i+1. diff PE is the evaporation amount on day i-1. i-1 Evaporation PE on day i+1 i+1Absolute difference, n is the year of evaporation observation, QC is the evaporation after quality control, PRE is the precipitation on the same day, RH is the relative humidity, and SSD is the sunshine duration.

[0076] Calculate the evaporation PE on day i. i Evaporation PE on day i-1 i-1 Or the evaporation PE on day i+1 i+1 absolute difference PE diff :

[0077] If PE is satisfied diff If the thickness is ≥5mm and simultaneously meets the following conditions: PRE≥10mm, RH≥80%, and SSD≤2h, then PE is determined. i This is a non-extreme outlier, and based on the evaporation process theory, the water vapor balance principle, and the quality control of the observation operation specifications, it is set to 0.0.

[0078] S1.3.4: All suspicious values ​​are re-verified and corrected to obtain a high-quality basic dataset.

[0079] S2: Based on random forest regression and recursive feature elimination cross-validation, the optimal feature subset of each station is selected, and a partial least squares regression model is constructed based on the optimal feature subset. The optimal hyperparameter of the model is searched using Grid Search CV cross-validation to obtain the daily data prediction model of evaporation of large evaporating pans.

[0080] S2.1: Using the data of overlapping observation periods of overlapping observation stations as training samples and time series coding, extract the daily data of small evaporation pan evaporation and related meteorological elements of each station as initial feature variables. The meteorological elements include temperature, wind speed, precipitation, relative humidity, water vapor pressure and sunshine duration. The evaporation of large evaporation pans in the same period is used as the target variable.

[0081] In the specific implementation of this application, based on the historical data of 97 national-level ground meteorological stations in Hunan Province, the start and end times of evaporation observation data and the replacement of evaporation pan instruments were statistically analyzed for each station. Accordingly, the observation data of each station can be divided into three categories according to instrument type: ① pure small-scale observation (59 stations); ② connected observation (13 stations); ③ overlapping observation (25 stations). Then, the effective data of 25 meteorological stations in the third category (overlapping observation) during the overlapping observation period of large and small evaporation pans were further statistically analyzed.

[0082] Daily evaporation data of small evaporation pans from each overlapping observation station were obtained during the overlapping observation period. Simultaneously, meteorological elements such as temperature (TEM), wind speed (WIN), precipitation (PRE), relative humidity (RHU), vapor pressure (E), and sunshine duration (SSD) were also acquired. Based on this, ΔT was further calculated. m(i.e., the difference between the highest temperature of the day and the previous day), and introduce time series coding to form a candidate feature set. The time series coding is used to convert time information into a numerical format that the model can recognize, so as to help the model effectively capture seasonal changes and identify meteorological characteristics, thereby improving the accuracy of evaporation prediction.

[0083] S2.2: Using the daily evaporation data of large evaporating pans as the prediction target, a random forest regression model is used to evaluate the importance of each feature in the candidate feature set. Then, a recursive feature elimination cross-validation method is used to select the optimal feature subset for predicting the daily evaporation data of large evaporating pans for each overlapping observation station.

[0084] To achieve the objective described in S2.2—namely, using daily evaporation data from large evaporation pans as the prediction target—this application employs a Random Forest Regression (RFR) model to evaluate the importance of each feature in the candidate feature set. Combined with Recursive Feature Elimination with Cross-Validation (RFECV), the optimal feature subset is selected for each overlapping observation station. RFR, by integrating the prediction results of multiple decision trees, improves prediction accuracy, and its built-in feature importance evaluation mechanism effectively handles high-dimensional data, revealing the contribution of each meteorological factor to evaporation prediction. Furthermore, RFECV is used for recursive feature selection, eliminating features with lower contributions round by round. Cross-validation rigorously evaluates model performance changes, ultimately retaining the most critical feature combinations for prediction. This combined strategy not only enhances the model's predictive ability but also improves its interpretability, achieving accurate identification of key meteorological factors influencing evaporation in large evaporation pans. Figure 3 To illustrate the relationship between the number of features and model performance in feature selection using RFECV combined with random forest regression, from... Figure 3 It can be seen that blindly increasing the number of features not only failed to improve model performance, but also led to a significant increase in model prediction error and a deterioration in model performance.

[0085] S2.3: Based on the optimal feature subset, construct a partial least squares regression model, and use Grid SearchCV cross-validation to search for the optimal hyperparameters of the model to obtain a daily data prediction model for the evaporation of large evaporating pans.

[0086] It needs to be explained that, in order to convert the daily evaporation data (PES) of small evaporating dishes into the daily evaporation data (PEL) of large evaporating dishes, this application introduces a generalized function for the daily evaporation data of large evaporating dishes as shown in equation (1):

[0087] PEL=f(PES,TEM,ΔTm PRE,WIN,SSD,E,RHU) (1)

[0088] Where: PEL represents daily evaporation data for large evaporating dishes, PES represents daily evaporation data for small evaporating dishes, TEM represents the corresponding air temperature during the overlapping observation period, and ΔT m The value is the difference between the highest temperature of the day and the previous day, PRE is the precipitation, WIN is the wind speed, SSD is the sunshine duration, E is the water vapor pressure, and RHU is the relative humidity.

[0089] Although the generalized functional relationship of daily evaporation data from large evaporation pans initially appears solvable using linear multiple regression, the widespread correlation among the selected meteorological variables easily leads to multicollinearity, resulting in unstable regression coefficient estimations and increased errors. Furthermore, the limited number of stations in China capable of simultaneously observing both large and small evaporation pans, coupled with a small sample size, further restricts the robustness of traditional regression models. Therefore, this study employs partial least squares regression (PLR) based on machine learning to construct a transformation model. PLS, by extracting the most explanatory component for the target variable, effectively overcomes the limitations of high correlation between variables and small sample size, more reliably capturing the complex relationships between features, thereby improving the prediction accuracy and model stability of evaporation transformation.

[0090] In the specific implementation of this application, a partial least squares regression model is first established. Grid search cross-validation is used to automatically optimize the model in the full parameter space to determine the optimal combination of hyperparameters, thereby training a model applicable to daily prediction of evaporation rates in large evaporating pans. In the model application phase, for each site, after data cleaning, evaporation rate data of small evaporating pans coded as "S" and their corresponding optimal feature subset are selected from the basic dataset. The data is then divided into training and test sets according to time sequence. The trained partial least squares regression model is then used to predict the test set sample by sample to obtain the predicted values ​​of evaporation rates in large evaporating pans for each site. Figure 4The diagram illustrates the prediction results of the daily evaporation data prediction model for large evaporating pans proposed in this invention. (a) shows the predicted evaporation results for station 57554, which is an overlapping observation station; (b) shows the predicted evaporation results for station 57574, which is a connecting observation station. Overall, the model effectively captures the actual changing trends of the PES index at both stations (57554 and 57574), demonstrating good predictive ability. At station 57554, the predicted value generally follows the same trend as the actual value, and the model successfully identifies the main peaks and fluctuation cycles. At station 57574, the prediction effect is even better, with the predicted curve almost completely overlapping the actual value curve, showing extremely high fitting accuracy. Further analysis shows that the model has significant advantages in time series prediction. It can not only accurately learn the long-term changing patterns of data but also effectively identify its fluctuation characteristics, a capability verified by the consistency performance at both stations. Although there is a slight lag at some turning points at station 57554, the model successfully predicted the key peak changes overall. This stable performance across different sites demonstrates that the model has fully grasped the inherent patterns of environmental data and has the potential to be deployed and applied in actual monitoring systems.

[0091] In the specific implementation of this application, the partial least squares regression model technically integrates the ideas of principal component analysis and canonical correlation analysis, and its core process is as follows:

[0092] Suppose there are p standardized independent variables X and a dependent variable Y. Partial least squares regression first extracts independent variable components t1 and dependent variable components u1 from independent variables X and dependent variable Y, respectively. The independent variable component t1 is a linear combination of independent variables X. Partial least squares regression requires that the extracted components t1 and u1 can carry as much variation information as possible from the original data (i.e., maximize the variance Var(t1) and Var(u1)) while also ensuring that the correlation r(t1,u1) between them is maximized. After the first component is established, the independent variable component t1 is used to regress the dependent variable Y. If the preset accuracy is not achieved, the residuals of the independent variable X are used to extract components in the next round until the extracted m independent variable components can fully explain the variation of the dependent variable Y. Ultimately, the dependent variable Y is represented as a linear combination of the independent variable components t. Since each independent variable component t is a linear function of the standardized independent variable X, the partial least squares regression model is also represented as a linear regression equation of the dependent variable Y with respect to the standardized independent variable X. During the modeling process, Grid Search Cross-Validation is used to search for the optimal hyperparameters of the model, thereby training a prediction model for daily evaporation data of large evaporating pans. To determine the optimal hyperparameters for predicting daily evaporation data of large evaporating pans, the following steps are taken: When determining the number of components extracted to achieve the desired regression level, cross-validation is applied to x in the partial least squares regression method to define the possible sum of squared errors, and Grid Search Cross-Validation is used to determine its optimal hyperparameters to construct a model for predicting daily evaporation data of large evaporating pans.

[0093] First, construct the parameter grid param_grid, which contains key hyperparameters to be optimized, such as n_components. This parameter specifies the number of latent components extracted from the PLS model, directly affecting the model complexity and feature extraction capability.

[0094] Grid SearchCV is used to exhaustively search the grid of preset parameters. K-fold cross-validation is used to evaluate the performance of each set of hyperparameters. Specifically, the training data is divided into K subsets each time. K-1 subsets are used for training in sequence, and the remaining 1 subset is used for validation. This is repeated K times to ensure that each fold of data participates in the validation, thereby comprehensively evaluating the model's generalization ability.

[0095] Record the performance score of each set of hyperparameters in cross-validation, and finally select the hyperparameter combination with the best average performance as the final modeling basis. GridSearchCV not only returns the optimal parameters, but also outputs the evaluation results of the corresponding model, ensuring that the selected configuration achieves a balance between bias and variance, thereby improving the model's predictive robustness on new data.

[0096] This application uses a learning curve and a coefficient of determination (R²). 2The model's performance at various stations is evaluated from multiple dimensions, including goodness of fit, error level, and forecast stability, by using quantitative indicators such as correlation coefficient (R), root mean square error (RMSE), mean square error (MSE), and confidence interval. This allows the system to determine the model's forecasting capability and reliability.

[0097] S3: Using the partial least squares regression model, the daily evaporation data of the evaporating dish are spliced ​​and interpolated to generate a reconstructed evaporation sequence of the evaporating dish.

[0098] The specific method for S3 is as follows:

[0099] For purely small-scale observation stations, the original small-scale evaporation pan observation sequence is extracted from the high-quality basic dataset after S1 quality control processing, and this original observation sequence is directly used as the evaporation reconstruction sequence of the purely small-scale observation station.

[0100] For the connecting observation station, its observation history is divided into an early small evaporation pan observation stage and a later large evaporation pan observation stage. In the early stage, n qualified overlapping observation stations around the connecting observation station are selected, and the partial least squares regression model trained on them is used to convert the small evaporation pan observation data into equivalent predicted values ​​for large evaporation pans, and then interpolation is performed to generate the missing data estimate for the connecting observation station in the early stage. In the later stage, the measured observation value sequence of large evaporation pans is directly used, and the estimated value of the connecting observation station in the early stage is spliced ​​with the measured observation value sequence to form the evaporation reconstruction sequence of the connecting observation station.

[0101] The calculation of the estimated values ​​of missing data from previous observation stations includes the following steps:

[0102] Select n qualified overlapping observation stations from the surrounding meteorological stations of the connecting observation station;

[0103] Select three overlapping observation stations (nearby stations) that have the smallest distance between the connecting observation station (target station), the best correlation, and an acceptable significance level (0.05). If the number of qualified overlapping observation stations is less than three, n in formula (2) will be equal to the actual number of qualified overlapping observation stations.

[0104] For qualified overlapping observation stations, the partial least squares regression model obtained during the overlapping observation period is used to convert all available small evaporating pan observation data of the station into equivalent predicted values ​​for large evaporating pans.

[0105] An interpolation algorithm is used to fuse the equivalent predicted values ​​of the large evaporating dish to generate estimated values ​​of missing data from the previous observation station. The interpolation algorithm is shown in equation (2):

[0106]

[0107] Where: E e,iIt is the estimated value of the missing data from observation station i, EF ij It is a linear regression estimate of the missing data of overlapping observation station j on the connecting observation station i, RMSE j It is the root mean square error of the regression model evaluation for the overlapping observation station j.

[0108] For overlapping observation stations, the partial least squares regression model trained during the overlapping observation period is applied to convert all available small evaporation pan observation data of the station into equivalent predicted values ​​of large evaporation pans, and this conversion result is used as the evaporation reconstruction sequence of the overlapping observation station.

[0109] The evaporation reconstruction sequences of pure small observation stations, connected observation stations, and overlapping observation stations are fused to obtain the evaporation reconstruction sequence of the evaporation pan.

[0110] S4: Perform a homogenization test on the reconstructed evaporation amount sequence of the evaporating dish, identify non-homogeneous breakpoints, correct the breakpoints, and obtain the homogenized reconstructed evaporation amount sequence of the evaporating dish.

[0111] like Figure 2 As shown, the specific steps of S4 are as follows:

[0112] S4.1: Based on the evaporation rate reconstruction sequence of the evaporating dish generated in S3, a regional reference sequence is constructed using the first-order difference method and the correlation coefficient weighted average method;

[0113] S4.2: Apply the penalized maximum T-test and penalized maximum F-test in the RHtests algorithm to perform a homogenization test on the reconstructed sequence and identify breakpoints;

[0114] S4.3: The identified breakpoints are corrected using the quantile mapping correction method to obtain a uniform reconstruction sequence of the evaporation amount of the evaporating dish.

[0115] In the specific embodiments of this application, the present invention utilizes two testing methods provided by RHtest: the penalized maximum T test and the penalized maximum F test, to conduct testing on the monthly and daily values ​​of spliced ​​evaporation data from large and small evaporation pans. Based on methods such as the First Difference Method (FDM) and Weighted Average based on correlation coefficients (WAA), a regional average reference sequence is constructed station by station. Metadata is incorporated into the algorithm, and the MASH (Multiple Analysis of Series for Homogenization) homogenization algorithm is used to assist in the judgment. The number and occurrence time of discontinuities at each station are compared and analyzed. Taking into full account the influence of regional climate change factors in Hunan Province, the test results are analyzed for climate rationality, and quantile matching adjustments (QM Adjustments) are performed to correct discontinuities. This establishes a homogenized reconstructed evaporation sequence from evaporation pans, reflecting the true historical climate change in evaporation in Hunan Province.

[0116] This application conducted simulation experiments to verify the performance of the large-scale evaporation pan evaporation rate prediction model proposed in this application. This study selected observation data from 25 stations in Hunan Province, randomly allocating 75% for training the hybrid model and 25% for testing. The performance of the large-scale evaporation pan evaporation rate prediction model is as follows: Figure 5 As shown.

[0117] from Figure 5 (b) It can be seen that the model proposed in this application has a small prediction error, with a minimum MSE of 0.48 and excellent RMSE performance. Figure 5 (a) It can be seen that the model has strong explanatory power (R). 2 The predicted values ​​range from 0.70 to 0.88, the predicted values ​​are highly correlated with the measured values ​​(R = 0.84 to 0.94), and the confidence interval is narrow (0.02 to 0.08). These results indicate that the model has high accuracy and reliability.

[0118] Figure 6 To compare the standard deviation, coefficient of variation, and interquartile range of the raw evaporation data with the homogenized evaporation data, from... Figure 6As can be seen, this application employed multiple RHtest methods, combined with regional reference sequences and metadata analysis, and found a significant breakpoint in 2013. The homogenized data were generally lowered compared to the original data, especially the large evaporation pan data, which underwent a significant correction (annual values ​​decreased from 739-1578 mm to 557-1113 mm). After QM correction, the data stability was significantly improved (mean standard deviation decreased from 0.77 to 0.73, and coefficient of variation decreased from 3.86 to 3.47), and the spurious downward trend in the original data was corrected (the trend slope at the large evaporation pan sites was adjusted from -1.47 to 2.79), more accurately reflecting the historical evaporation variation characteristics.

[0119] from Figure 7 (a) This paper presents the spatial distribution characteristics of the decadal rate of change of evaporation in the basic dataset. Statistical results show that 26% of the stations in the dataset show an increasing trend, mainly distributed in western and southern Hunan. However, the spatial consistency is poor, and the trends of change between adjacent stations differ significantly, especially in southern Hunan. Figure 7 (c) The spatial distribution pattern of climate trend rates in the basic dataset is presented. Data shows that 24% of the stations exhibit a positive climate trend rate, whose spatial distribution characteristics are basically consistent with the decadal change rate: positive trend rates are predominant in parts of western and southern Hunan, while most of northeastern Hunan shows a negative trend rate. After normalization correction, the dataset shows significantly different change characteristics. In the revised data, the proportion of stations with an increasing decadal evaporation trend rises to 46%, and the proportion of stations with positive climate trend rates reaches 39%. Figure 7 (b) shows the spatial distribution of the decadal rate of change in the homogenized dataset. Compared to the base dataset, the revised data better reflects the true trend of increasing evaporation in central and northern Hunan. Figure 7 (d) shows the spatial distribution of climate trend rates in the homogenized dataset, revealing that the magnitude of the trend in areas of reduced evaporation has also weakened. This result demonstrates that dynamic feature selection and systematic homogenization correction using the PLS regression model effectively reduce systematic errors introduced by factors such as inconsistent data splicing, station migration, and instrument replacement, making the identification of evaporation change trends more accurate and reliable, and providing a higher-quality data foundation for in-depth research on regional climate change.

[0120] from Figure 8 Looking at the spatial distribution of climate tendency rate across the four seasons, the basic data shows a downward trend in most areas during spring, summer, and autumn. This phenomenon is mainly due to the systematic overestimation of evaporation by small evaporation pans, which leads to an overall downward trend during the replacement of observation instruments. Figure 8(a) shows that the percentage of stations with a positive trend in the spring baseline data is 39.2%, and the climate tendency rate varies from -37.65 to 14.94 mm / 10a. Figure 8 (c) indicates that the proportion of stations with positive trends is the lowest in summer, at only 29.9% (29 stations), but the climate trend rate varies the most, with the negative extreme value reaching -71.54 mm / 10a and the positive extreme value reaching 53.63 mm / 10a. The stations with positive trends are mainly scattered in the southern Hunan region. Figure 8 (e) shows that only 5% of the stations showed a positive trend in the fall, but the overall negative growth was relatively small, with the largest negative value being -42.8 mm / 10a. Figure 8 (g) indicates that the variation trend was smallest at all stations in winter, ranging from -15.59 to 15.61 mm / 10a, with 37.1% of stations showing a positive trend. After normalization correction, the proportion of stations with positive trends in each season significantly increased, reaching 58.8% (spring), 50.5% (summer), 30.9% (autumn), and 60.8% (winter), respectively, indicating a more balanced distribution of stations with positive and negative trends. Simultaneously, the ten-year variation amplitude was significantly reduced after correction, with the variation ranges for each season being: spring -15.37–16.96 mm / 10a, summer -28.85–41.0 mm / 10a, autumn -32.62–19.19 mm / 10a, and winter -11.44–9.8 mm / 10a, and the spatial distribution characteristics became clearer. Figure 8 (b) shows that the revised spring data all show a stable growth trend in western, central and northern Hunan. Figure 8 (d) indicates that the summer growth trend is mainly observed in southwestern and southern Hunan, while the range of variation in the declining areas has decreased. Figure 8 (f) and Figure 8 (h) shows that the changes in autumn and winter are relatively small, but the number of observation stations showing an increasing trend has increased significantly. Overall, the normalized and corrected dataset is significantly better than the basic dataset in terms of spatial distribution characteristics, and can more accurately reflect the regional variation patterns of evaporation in each season, providing more reliable data support for climate evolution research on seasonal scales.

[0121] Figure 9 The results of RHtest method detection of evaporation sequences from 97 stations in Hunan Province are presented. Figure 9 (a) The seasonal characteristics of the breakpoint distribution show that it mainly occurs in winter (56.5%), with January having the highest frequency of occurrence throughout the year (42.8%). This coincides with the timing of instrument replacement events. Figure 9 (b) shows that the frequency of breakpoints is significantly higher in 1963 and 2013, mainly corresponding to major changes such as the cessation of observation or relocation of stations. Figure 9 (c) and Figure 9 (d) shows that after the normalization correction, the frequency of monthly evaporation data greater than 100 mm decreased, the values ​​were more concentrated in the 30–80 mm range, the extreme high values ​​were reduced, the daily value distribution was smoother, the distribution in the 0–4 mm range was more uniform, and the extreme values ​​greater than 6 mm were reduced, effectively correcting the overestimation bias caused by small evaporation pan observations.

[0122] Figure 10 To correct the experimental results for homogenization of evaporation, from Figure 10 The statistical results in (a) show that the maximum positive correction is 0.41, and the maximum negative correction is -0.42. The distribution frequency is mainly concentrated between -0.4 and 0.1, exhibiting a bimodal distribution with peaks around -0.3 and 0, respectively. 69% of the stations showed negative corrections. This phenomenon is directly related to the historical widespread use of small evaporation pans in Hunan Province, which tend to overestimate evaporation measurements. Figure 10 As can be seen from (a) to (b), the evaporation sequence correction results exhibit a unimodal distribution, with the peak value located at -0.1. Negative correction sites dominate (71.6%, 68 sites), while positive correction sites account for 28.4% (27 sites). This systematic negative bias mainly stems from corrections made based on the alternation of large and small evaporating dishes. Figure 10 (c) It can be seen that, in terms of the correction magnitude, 45.3% (43 stations) of the stations had a magnitude less than -0.1, which was concentrated in the eastern and southwestern Hunan regions; 26.3% (25 stations) of the stations had a magnitude between -0.1 and 0, which was widely distributed in most areas except for southern Hunan; 18.9% (18 stations) of the stations had a magnitude between 0 and 0.1, and their number decreased from southeast to northwest; only 9.5% (9 stations) of the stations had a magnitude greater than 0.1, all of which were pure small evaporation pan stations, and were concentrated in southern Hunan. The evaporation sequence was significantly smaller in these areas due to non-uniform factors.

[0123] refer to Figure 11 The National Meteorological Center used the homogenized reconstructed evaporation sequence of Hunan Province constructed in this application to conduct a case study of a severe drought in Hunan Province in 2022. By comparing the actual evaporation data and distribution characteristics with the homogenized long-series historical evaporation data for the same period, the center assessed and analyzed the causes of this prolonged and large-scale drought. This provides data support and a clear visual representation of the historical drought, enabling analysts to conduct more in-depth research on the occurrence patterns, impact range, and severity of droughts, thus providing a scientific basis and decision support for future disaster prevention and mitigation efforts. The homogenized long-series evaporation data and actual evaporation provide reliable meteorological data support and climate data assurance for drought condition analysis.

[0124] In practical applications, raw data, due to its inherent limitations, often underestimates historical evaporation anomalies to some extent, leading to biases in the assessment of drought severity. In contrast, climate data after normalization can more timely reflect the development trend of drought, and its anomaly change trend is highly consistent with the actual drought evolution process. In particular, for the period from August 1 to October 7, the historical evaporation anomaly was positive for 68 consecutive days, with the maximum anomaly reaching 4.2 mm (August 21). During this period, the drought area increased rapidly, with the severely drought area reaching 100% and the extreme drought area reaching over 90%.

[0125] The Meteorological Bureau of Xiangxi Autonomous Prefecture, Hunan Province, used the historical homogenized data of monthly and daily evaporation values ​​in the Hunan Province evaporation pan homogenization reconstruction sequence constructed in this application. It calculated the comparison between the historical homogenized evaporation climate values ​​and the actual evaporation for the same period, thereby assessing the degree of meteorological drought development. It produced service reports such as the "Agricultural Meteorology Monthly Report" and "Climate Monthly Report", providing scientific agricultural advice to the 8 counties (cities) of Xiangxi Autonomous Prefecture, grain and oil stations, vegetable offices, plant protection stations, forestry bureaus, etc., with significant service benefits.

[0126] During the meteorological drought period in September 2024, historical homogenized data of daily evaporation values ​​in Hunan Province, constructed using the homogenized evaporation reconstruction sequence of evaporation pans in this application, were used to calculate the daily climatological values ​​of evaporation in Jishou (57649) and Baojing (57642) in September, and to analyze the anomalies between actual evaporation and historical homogenized evaporation for the same period. "The average evaporation for the entire prefecture was 119.9 mm, which is 32.1 mm more than the same period in previous years. The average daily evaporation for this month was 4.0 mm, which is 1.07 mm more than the same period in previous years." Drought development occurred from September 1-6 and 18-28, with the maximum daily evaporation occurring on September 22 (7.4 mm), indicating particularly severe drought development. The historical anomaly was 4.6 mm, reflecting the drought situation for this month.

[0127] refer to Figure 12 (a) and Figure 12 (b) Correlation analysis of monthly evaporation data from 57,640 stations in Xiangxi Prefecture that underwent homogenization with those that were considered homogenized but not homogenized showed correlation coefficients greater than 0.9 for all stations, indicating high reliability of the homogenized evaporation data. Furthermore, this homogenized evaporation dataset demonstrates significant improvement. Taking September 2024 as an example, the anomaly between the homogenized evaporation and historical averages for the same period ranged from -0.9 to 4.6 mm, an improvement compared to the historical anomalies of -1.6 to 3.9 mm for the same period in the original evaporation data. This result is consistent with the actual drought situation this month and is highly reliable. This dataset can be applied to the analysis of drought conditions in this region.

[0128] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0129] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A method for reconstructing a homogenized sequence of evaporation amount in an evaporating dish based on multi-source data fusion, characterized in that, Includes the following steps: S1: Screen the daily evaporation data of small evaporation pans and large evaporation pans of ground stations in the target area, meteorological data and station metadata. Based on the station metadata, the stations are divided into overlapping observation stations, connecting observation stations and pure small observation stations. Secondary quality control is performed on the screened data to obtain a high-quality basic dataset. S2: Based on random forest regression and recursive feature elimination cross-validation, the optimal feature subset of each station is selected, and a partial least squares regression model is constructed based on the optimal feature subset. The optimal hyperparameter of the model is searched using Grid Search CV cross-validation to obtain the daily data prediction model of evaporation of large evaporating pans. S3: Using the partial least squares regression model, the daily evaporation data of the evaporating dish are spliced ​​and interpolated to generate a reconstructed evaporation sequence of the evaporating dish. S4: Perform a homogenization test on the reconstructed evaporation amount sequence of the evaporating dish, identify non-homogeneous breakpoints, correct the breakpoints, and obtain the homogenized reconstructed evaporation amount sequence of the evaporating dish.

2. The method for reconstructing the homogenization sequence of evaporation amount in an evaporating dish based on multi-source data fusion according to claim 1, characterized in that, The specific steps for S1 are as follows: S1.1: Obtain daily data on evaporation of small evaporating pans, daily data on evaporation of large evaporating pans, meteorological element data, and station metadata of ground stations within the target area. Based on the observation instrument records in the station metadata, classify each station into overlapping observation stations, connecting observation stations, and purely small observation stations. S1.2: Organize the data by station number to form a long-sequence basic dataset containing year, month, day, evaporation data and observation type identification code; S1.3: Perform secondary quality control on the evaporation data in the long sequence basic dataset to remove extreme outliers in the evaporation of the evaporating dish, and obtain a high-quality basic dataset.

3. The method for reconstructing the homogenization sequence of evaporation amount in an evaporating dish based on multi-source data fusion according to claim 2, characterized in that, The specific steps for secondary quality control of evaporation data in S1.3 are as follows: S1.3.1: Perform boundary value tests on the daily evaporation data of small and large evaporating dishes in the long-sequence basic dataset to obtain the data that passes the boundary value test; S1.3.2: Perform climate anomaly testing on the data that have passed the threshold test, and mark the data that exceed the range of 3 to 5 times the standard deviation as suspicious values; S1.3.3: Perform adjacent-day variation analysis on the suspicious values, calculate the absolute difference between the current day's evaporation and the adjacent day's evaporation, and combine it with the same day's precipitation, relative humidity and sunshine duration for judgment and quality control; S1.3.4: All suspicious values ​​are re-verified and corrected to obtain a high-quality basic dataset.

4. The method for reconstructing the homogenization sequence of evaporation amount in an evaporating dish based on multi-source data fusion according to claim 3, characterized in that, The calculation of the absolute difference between the current daily evaporation and the evaporation of adjacent days corresponding to the suspicious value, as described in S1.3.3, for judgment and quality control purposes specifically includes: Let PE i Let PE be the evaporation on day i. i-1 PE represents the evaporation on day i-1. i+1 Let PE be the evaporation on day i+1. diff PE is the evaporation amount on day i-1. i-1 Evaporation PE on day i+1 i+1 Absolute difference, n is the year of evaporation observation, QC is the evaporation after quality control, PRE is the precipitation on the same day, RH is the relative humidity, and SSD is the sunshine duration; Calculate the evaporation PE on day i. i Evaporation PE on day i-1 i-1 Or the evaporation PE on day i+1 i+1 absolute difference PE diff : If PE is satisfied diff If the thickness is ≥5mm and simultaneously meets the following conditions: PRE≥10mm, RH≥80%, and SSD≤2h, then PE is determined. i This is a non-extreme outlier, and the quality control is 0.

0.

5. The method for reconstructing the homogenization sequence of evaporation amount in an evaporating dish based on multi-source data fusion according to claim 1, characterized in that, The specific steps for S2 are as follows: S2.1: Using the data of overlapping observation periods of overlapping observation stations as training samples and time series coding, extract the daily data of small evaporation pan evaporation and related meteorological elements of each station as initial feature variables. The meteorological elements include temperature, wind speed, precipitation, relative humidity, water vapor pressure and sunshine duration. The evaporation of large evaporation pans in the same period is used as the target variable. S2.2: The importance of each feature to the target variable is evaluated based on the random forest regression method, and the optimal feature subset of each site is selected by combining the recursive feature elimination cross-validation method. S2.3: Based on the optimal feature subset, construct a partial least squares regression model, and use Grid Search CV cross-validation to search for the optimal hyperparameters of the model to obtain a daily data prediction model for the evaporation of large evaporating pans.

6. The method for reconstructing the homogenization sequence of evaporation amount in an evaporating dish based on multi-source data fusion according to claim 1, characterized in that, The specific method for S3 is as follows: For purely small-scale observation stations, the original small-scale evaporation pan observation sequence is extracted from the high-quality basic dataset after S1 quality control processing, and this original observation sequence is directly used as the evaporation reconstruction sequence of the purely small-scale observation station. For the connecting observation station, its observation history is divided into an early small evaporation pan observation stage and a later large evaporation pan observation stage. In the early stage, n qualified overlapping observation stations around the connecting observation station are selected, and the partial least squares regression model trained on them is used to convert the small evaporation pan observation data into equivalent predicted values ​​for large evaporation pans, and then interpolation is performed to generate the missing data estimate for the connecting observation station in the early stage. In the later stage, the measured observation value sequence of large evaporation pans is directly used, and the estimated value of the connecting observation station in the early stage is spliced ​​with the measured observation value sequence to form the evaporation reconstruction sequence of the connecting observation station. For overlapping observation stations, the partial least squares regression model trained during the overlapping observation period is applied to convert all available small evaporation pan observation data of the station into equivalent predicted values ​​of large evaporation pans, and this conversion result is used as the evaporation reconstruction sequence of the overlapping observation station. The evaporation reconstruction sequences of pure small observation stations, connected observation stations, and overlapping observation stations are fused to obtain the evaporation reconstruction sequence of the evaporation pan.

7. The method for reconstructing the homogenization sequence of evaporation amount in an evaporating dish based on multi-source data fusion according to claim 1, characterized in that, The specific steps for S4 are as follows: S4.1: Based on the evaporation rate reconstruction sequence of the evaporating dish generated in S3, a regional reference sequence is constructed using the first-order difference method and the correlation coefficient weighted average method; S4.2: Apply the penalized maximum T-test and penalized maximum F-test in the RHtests algorithm to perform a homogenization test on the reconstructed sequence and identify breakpoints; S4.3: The identified breakpoints are corrected using the quantile mapping correction method to obtain a uniform reconstruction sequence of the evaporation amount of the evaporating dish.

8. The method for reconstructing the homogenization sequence of evaporation amount in an evaporating dish based on multi-source data fusion according to claim 6, characterized in that, The calculation of the estimated missing data from the previous phase of the connecting observation station includes the following steps: Step a: Select n qualified overlapping meteorological stations from the surrounding meteorological stations connecting the observation station; Step b: For qualified overlapping observation stations, apply the partial least squares regression model obtained during the overlapping observation period to convert all available small evaporating pan observation data of the station into equivalent predicted values ​​for large evaporating pans. Step c: The equivalent predicted values ​​of the large evaporating dish are fused using an interpolation algorithm to generate estimated values ​​of the missing data from the previous observation station. The interpolation algorithm is shown in equation (1): Among them: E e,i It is the estimated value of the missing data from observation station i, EF ij It is a linear regression estimate of the missing data of overlapping observation station j on the connecting observation station i, RMSE j It is the root mean square error of the regression model evaluation for the overlapping observation station j.

Images