A lightgbm-based precipitation bias correction method

Abstract

The application discloses a precipitation deviation correction method based on LightGBM and belongs to the technical field of hydrological prediction and analysis, and the method comprises the following steps: step 1, data collection and preprocessing; step 2, establishment of multi-scale precipitation characteristics; step 3, training of a classification model and a regression model; and step 4, deviation correction of precipitation. According to the method, time, space and other multi-scale precipitation characteristics are introduced, whether precipitation occurs is determined through the trained classification model, then the deviation correction of precipitation is carried out under the condition of wet days, the generalized Pareto distribution is used to correct the high-quantile extreme value, high-precision deviation correction of precipitation is realized, and the accuracy of hydrological prediction and analysis is improved.

Description

Technical Field

[0001] This invention belongs to the field of hydrological prediction and analysis technology, and in particular relates to a precipitation deviation correction method based on LightGBM. Background Technology

[0002] Due to the shortcomings of current numerical weather prediction models based on physical processes (such as uncertainties in the initial and variable fields), high-precision weather forecasts, especially extended-range precipitation forecasts to seasonal scales, have significant systematic biases in their direct precipitation outputs, which greatly limits their direct application value in forecasting and scientific decision-making.

[0003] Currently, the Cumulative Transfer Distribution Function (CDF-t) method is an important tool for post-processing forecast data and has been applied in the meteorological and hydrological fields. However, the CDF-t method still has some drawbacks, including: 1) The CDF-t method is based on the Empirical Distribution Function (ECDF) to establish relationships. If extreme events that have not occurred in the historical period occur in the future, the CDF-t method cannot correct for extreme events exceeding the historical period range. 2) The estimation quality of ECDF is highly dependent on the length of the historical sample size. For areas with short historical records or sparse data, the ECDF estimation, especially at the distribution tail, will be very uncertain and unstable, leading to unreliable transformation function estimates and increased variance in the correction results. 3) CDF-t is a univariate correction method that operates independently at each grid point (or station), relying only on the statistical distribution of the precipitation sequence at that point, while completely ignoring other valuable information that may explain the cause of bias. Especially under the context of global climate change, the above-mentioned shortcomings of the CDF-t method become more prominent. Therefore, how to achieve higher accuracy precipitation bias correction is a technical problem that needs to be solved in this field. Summary of the Invention

[0004] The purpose of this invention is to provide a precipitation deviation correction method based on LightGBM to solve the above-mentioned technical problems.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention discloses a precipitation deviation correction method based on LightGBM, the method comprising the following steps:

[0007] Step 1: Data Collection and Preprocessing: Collect historical long-series CN05.1 grid-based meteorological data for the study area, including measured precipitation data; collect historical long-series ECMWF model precipitation forecast data for the study area; and then use bilinear interpolation to unify the spatial resolution of the measured precipitation data and the precipitation forecast data.

[0008] Step 2: Establish multi-scale precipitation characteristics: Read the basic information of historical precipitation forecast data and establish multi-scale precipitation characteristics grid by grid, including latitude, longitude, current day's model precipitation, whether the current day is a wet day, previous day's model precipitation, previous two days' model precipitation, previous three days' model precipitation, previous three days' total precipitation, previous seven days' total precipitation, previous thirty days' total precipitation, previous three days' maximum precipitation, consecutive wet day length up to yesterday, consecutive dry day length up to yesterday, annual total precipitation, monthly total precipitation, seasonal quantile position, and logarithmic anomaly with respect to the seasonal median.

[0009] Step 3: Train the classification and regression models: Input multi-scale precipitation features at daily scales, labeled as measured wet / dry days, and train the LightGBM classifier to obtain the trained classification model. Under wet day conditions, using the multi-scale characteristics of wet day samples from precipitation forecast data as input and the corresponding measured precipitation as labels, a quantile regression model based on LightGBM is trained. The specific process is as follows:

[0010] 1) Training a classification model based on log loss:

[0011] (2)

[0012] In the formula: The loss function for the classification model; This represents the number of training samples; The true label of the i-th sample ( ); The predicted probability output by the model;

[0013] Using a trained classification model The system outputs the probability of daily precipitation, performs probability calibration, and determines the optimal threshold for wet day discrimination based on the historical wet day frequency to determine whether precipitation has occurred. The specific process is as follows:

[0014] First, using the trained classification model Mapping historical multi-scale precipitation characteristics to the probability of precipitation occurrence:

[0015] (3)

[0016] Secondly, isotonic regression is used for probability calibration:

[0017] (4)

[0018] Then determine the optimal threshold for wet day discrimination after calibration:

[0019] (5)

[0020] In the formula: For a certain grid point; For a certain processing time; For feature vectors; For the Sigmoid function, specifically... ; This represents the probability of a wet day occurring, as output by the model. The calibrated probability of wet days; It is a monotonic calibration function; This indicates the search for parameter T that minimizes the following expression; For indicator functions; This represents the total number of training samples, which is the sum of all days for all grid points. To observe the frequency of wet days; For the threshold, ; The optimal threshold for wet day discrimination after calibration;

[0021] 2) Training a quantile regression model based on minimizing risk:

[0022] (6)

[0023] in:

[0024] (7)

[0025] In the formula: This is the loss function for quantile regression; For all wet day samples; The quantile loss function; For quantiles, ; This represents the actual measured precipitation. For features Precipitation at quantile points; For residuals;

[0026] Step 4: Correct for precipitation bias: For future precipitation forecast data, establish multi-scale precipitation characteristics according to the same rules, first by using a classification model. Determine whether it's a dry or wet day; in the case of a wet day, use the characteristics of the wet day samples as input and apply a quantile regression model. The precipitation quantiles are corrected, and the high-quantile extreme values ​​are corrected using a generalized Pareto distribution. The tails are then extrapolated to obtain the final corrected precipitation. The specific process is as follows:

[0027] 1) If we set wet days as 1 and dry days as 0, we can determine whether precipitation occurs at a specific location and time:

[0028] (8)

[0029] In the formula: Used to indicate whether precipitation has occurred at a specific location and time;

[0030] 2) Under wet day conditions, use the trained regression model The quantile for predicting precipitation is given by the following formula:

[0031] (9)

[0032] In the formula: The quantile regression model predicts the 1st Precipitation at quantile points; This is a quantile regression model;

[0033] 3) Improve numerical stability using logarithmic transformation:

[0034] (10)

[0035] (11)

[0036] (12)

[0037] In the formula: This represents the value of the observed precipitation after adding 1 and taking the logarithm. Indicates the prediction of the th in the logarithmic space Precipitation at quantile points; It is a logarithmic space model;

[0038] 4) Apply generalized Pareto distribution (GPD) correction to high quantile extreme values:

[0039] (13)

[0040] Push the tail outward:

[0041] (14)

[0042] In the formula: The cumulative distribution function of the generalized Pareto distribution; The precipitation value corresponding to the 95th percentile; It is a scaling function; The shape parameters are obtained using maximum likelihood estimation; The first one is obtained by extrapolation based on the generalized Pareto distribution. ( ) Quantile correction value;

[0043] 5) The precipitation at each quantile is obtained as follows:

[0044] (15)

[0045] Interpolation yields the final precipitation as follows:

[0046] (16)

[0047] In the formula: This is the corrected quantile precipitation series; , for The adjacent quantiles, and ; , quantiles The corresponding quantile precipitation value.

[0048] Furthermore, the measured precipitation data mentioned in step 1 was obtained from the official website of the National Climate Change Research Center; the precipitation forecast data was obtained from the official website of ECMWF; and the historical long-series data consisted of data from at least ten years.

[0049] Furthermore, the formula for unifying the spatial resolution of measured precipitation data and forecast precipitation data using bilinear interpolation is as follows:

[0050] (1)

[0051] In the formula: The coordinates of the points to be interpolated and their corresponding values; , , , These are the coordinates of four known points and their corresponding values.

[0052] The beneficial effects of the present invention are as follows: The method of the present invention introduces precipitation characteristics at multiple scales such as time and space. First, it determines whether precipitation has occurred through a trained classification model. Then, it performs precipitation deviation correction under wet day conditions. For high quantile extreme values, it uses generalized Pareto distribution correction, which achieves high-precision precipitation deviation correction and improves the accuracy of hydrological prediction and analysis.

[0053] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. Attached Figure Description

[0054] Figure 1 This is a schematic diagram of the method flow described in Example 1;

[0055] Figure 2 This is a schematic diagram illustrating the classification results of whether it rained in a local area on a certain day in Example 1.

[0056] Figure 3 This is a Taylor plot showing the spatial correlation between simulated and measured data in Example 1.

[0057] Figure 4 The figure shows the monthly average precipitation curves for China and the eight study areas in Example 1. Detailed Implementation

[0058] This invention discloses a precipitation deviation correction method based on LightGBM, such as... Figure 1 As shown, the method includes the following steps:

[0059] Step 1: Data Collection and Preprocessing: Collect historical long-series CN05.1 grid observation meteorological data for the study area, including measured precipitation data, which can be obtained from the official website of the National Climate Change Research Center; collect historical long-series ECMWF model precipitation forecast data for the study area, which can be obtained from the ECMWF official website; the historical long-series data should be at least ten years long.

[0060] Then, the spatial resolution of the measured precipitation data and the precipitation forecast data is unified using bilinear interpolation, as shown in the following formula:

[0061] (1)

[0062] In the formula: The coordinates of the points to be interpolated and their corresponding values; , , , These are the coordinates of four known points and their corresponding values.

[0063] Step 2: Establish Multi-Scale Precipitation Features: Establish spatial and temporal multi-scale precipitation features grid-by-grid. Specifically: Read basic information from historical precipitation forecast data (including latitude and longitude, precipitation values, etc.). Read spatial location features, including latitude (lat) and longitude (lon). Based on the current day's precipitation value, determine whether it rained that day and obtain instantaneous features, including the current day's model precipitation (pr_mod) and whether the day is a wet day (is_wet_mod). Read time lag features, including the previous day's model precipitation (lag1_pr_mod), the previous two days' model precipitation (lag2_pr_mod), and the previous three days' model precipitation (lag3_pr_mod). Set time windows to calculate temporal statistical features, including the total precipitation of the previous three days (roll3_sum_mod), the total precipitation of the previous seven days (roll7_sum_mod), the total precipitation of the previous thirty days (roll30_sum_mod), and the maximum precipitation of the previous three days (roll3_max_mod). Obtain persistence features, including the length of consecutive wet days up to yesterday (runlen_wet_mod) and the length of consecutive dry days up to yesterday (runlen_dry_mod). That is, if yesterday was a wet day, trace back to the first dry day; if yesterday was a dry day, trace back to the first wet day. Calculate cumulative features, including total annual precipitation (ytd_sum_mod) and total monthly precipitation (mtd_sum_mod). Sort the daily precipitation data for the entire season (e.g., spring, summer, etc.). Then, calculate the position of the current date's precipitation in this sort, i.e., its quantile position in the seasonal precipitation distribution, which is the feature's seasonal quantile position (pr_mod_quantile_doywin). First, calculate the median precipitation for the season, then compare the current date's precipitation with this median, and calculate the difference between them through a logarithmic transformation, which is the feature's logarithmic anomaly with respect to the seasonal median (pr_mod_anom_logp50). The calculation formula is: pr_mod_anom_logp50 = log (current precipitation value / seasonal median precipitation value).

[0064] The names and units of each feature are shown in Table 1:

[0065] Table 1. Multiscale precipitation characteristics

[0066]

[0067] Step 3: Train the classification and regression models: Input multi-scale precipitation features at the daily scale from historical periods, labeled as measured wet / dry days, and train the LightGBM classifier to obtain the trained classification model. Under wet day conditions, a quantile regression model based on LightGBM is trained using the multi-scale features of wet day samples from precipitation forecast data as input and the corresponding measured precipitation as labels. The specific process is as follows:

[0068] 1) Training a classification model based on log loss:

[0069] (2)

[0070] In the formula: The loss function for the classification model; This represents the number of training samples; The true label of the i-th sample ( ); This represents the predicted probability output by the model.

[0071] Using a trained classification model The system outputs the daily probability of precipitation, performs probability calibration, and determines the optimal threshold for wet day discrimination based on the historical wet day frequency to determine whether precipitation has occurred. The specific process is as follows:

[0072] First, using the trained classification model Mapping historical multi-scale precipitation characteristics to the probability of precipitation occurrence:

[0073] (3)

[0074] Secondly, isotonic regression is used for probability calibration:

[0075] (4)

[0076] Then determine the optimal threshold for wet day discrimination after calibration:

[0077] (5)

[0078] In the formula: For a certain grid point; For a certain processing time; For feature vectors; For the Sigmoid function, specifically... ; This represents the probability of a wet day occurring, as output by the model. The calibrated probability of wet days; The calibration function is monotonic to ensure that the model's predicted precipitation probability is close to the actual precipitation frequency. This indicates the search for parameter T that minimizes the following expression; For indicator functions; This represents the total number of training samples, which is the sum of all days for all grid points. To observe the frequency of wet days; For the threshold, ; The optimal threshold for wet day discrimination after calibration.

[0079] 2) Training a quantile regression model based on minimizing risk:

[0080] (6)

[0081] in:

[0082] (7)

[0083] In the formula: This is the loss function for quantile regression; For all wet day samples; The quantile loss function; For quantiles, ; These are measured precipitation values. For features Precipitation values ​​at quantile points; It represents the residual.

[0084] Step 4: Correct for precipitation bias: For future precipitation forecast data, establish multi-scale precipitation characteristics according to the same rules. First, use a classification model to determine dry / wet days; in the case of wet days, input the features of the wet day samples into the trained quantile regression model. This yields corrected quantile values ​​consistent with the observed distribution. For extreme values ​​in the higher quantiles, a generalized Pareto distribution is used for correction, and the tails are extrapolated to obtain the final corrected precipitation. The specific process is as follows:

[0085] 1) If we set wet days as 1 and dry days as 0, we can determine whether precipitation occurs at a specific location and time:

[0086] (8)

[0087] In the formula: Used to indicate whether precipitation has occurred at a specific location and time.

[0088] 2) Utilizing the characteristics of the future period Using a trained regression model The quantile for predicting precipitation is given by the following formula:

[0089] (9)

[0090] In the formula: The quantile regression model predicts the 1st Precipitation at quantile points; This is a quantile regression model.

[0091] 3) Improve numerical stability using logarithmic transformation:

[0092] (10)

[0093] (11)

[0094] (12)

[0095] In the formula: This represents the value of the observed precipitation after adding 1 and taking the logarithm. Indicates the prediction of the th in the logarithmic space Precipitation at quantile points; It is a logarithmic space model.

[0096] 4) Apply generalized Pareto distribution (GPD) correction to high quantile extreme values:

[0097] (13)

[0098] Push the tail outward:

[0099] (14)

[0100] In the formula: The cumulative distribution function of the generalized Pareto distribution; The measured precipitation value corresponds to the 95th percentile. It is a scaling function; The shape parameters are obtained using maximum likelihood estimation; The first one is obtained by extrapolation based on the generalized Pareto distribution. ( ) Quantile correction value.

[0101] 5) The precipitation at each quantile is obtained as follows:

[0102] (15)

[0103] Interpolation yields the final precipitation as follows:

[0104] (16)

[0105] In the formula: This is the corrected quantile precipitation series; , for The adjacent quantiles, and ; , quantiles The corresponding quantile precipitation value.

[0106] Example 1

[0107] This embodiment is an application example of the above method, using China as an example to illustrate and verify the method. Based on China's climate, topography, and administrative divisions, China is divided into eight major regions: North China (NC), Northeast China (NE), East China (EC), South China (SC), Central China (CC), Southwest China (SW), Northwest China (NW), and the Qinghai-Tibet Plateau (TP) for research.

[0108] This embodiment discloses a precipitation deviation correction method based on LightGBM, such as... Figure 1 As shown, the method includes the following steps:

[0109] Step 1: Data Collection and Preprocessing: Collect historical long-series CN05.1 grid-based meteorological data for the study area, including measured precipitation data; collect historical long-series ECMWF model precipitation forecast data for the study area; then use bilinear interpolation to unify the spatial resolution of the measured precipitation data and the precipitation forecast data. In this embodiment, the time scale for data collection is from January 1, 1981 to December 31, 2022, with data from January 1, 1981 to December 31, 2017 considered as the historical period, and data from January 1, 2018 to December 31, 2022 considered as the future period (i.e., the validation period).

[0110] Step 2: Establish multi-scale precipitation characteristics: Based on historical forecast precipitation data, establish multi-scale precipitation characteristics for each grid, including latitude, longitude, current day's model precipitation, whether the current day is a wet day, previous day's model precipitation, previous two days' model precipitation, previous three days' model precipitation, previous three days' total precipitation, previous seven days' total precipitation, previous thirty days' total precipitation, previous three days' maximum precipitation, consecutive wet day length up to yesterday, consecutive dry day length up to yesterday, annual total precipitation, monthly total precipitation, seasonal quantile position, and logarithmic anomaly with respect to the seasonal median.

[0111] Step 3: Train the classification and regression models: Input multi-scale precipitation features at the daily scale from historical periods, labeled as measured wet / dry days, and train the LightGBM classifier model to obtain the trained classification model. Using a trained classification model The system outputs the daily probability of precipitation, performs probability calibration, and determines the optimal threshold for wet day discrimination based on the historical wet day frequency to determine whether precipitation has occurred. In the case of wet days, the features of the wet day samples are used as input, and the corresponding measured precipitation is used as the label to train a quantile regression model based on LightGBM. This embodiment determines whether it rained that day by checking each grid cell. Figure 2 The classification results show whether it rained in a local area on a certain day.

[0112] Step 4: Correct for precipitation bias: For future precipitation forecast data, establish multi-scale features using the same rules as historical data. First, use a classification model to determine dry / wet days. In the case of wet days, use the features of the wet day samples as input and utilize a quantile regression model. The precipitation quantiles are corrected, and the extreme values ​​at the high quantiles are corrected using a generalized Pareto distribution. The tail is extrapolated to obtain the final corrected precipitation.

[0113] To verify the advantages of the method described in this invention, this embodiment compares the corrected precipitation in the future period (verification period) obtained by the method described in this invention with the precipitation obtained by other methods. Five precipitation indices that can reflect the average and extreme conditions of precipitation are selected to evaluate the method described in this invention, as shown in Table 2.

[0114] Table 2 Precipitation Index

[0115]

[0116] Figure 3 The Taylor plots show the spatial correlation between ECMWF model precipitation forecast data, simulated data after correction using the CDF-t method, and simulated data after correction using the LightGBM method.

[0117] Depend on Figure 3 It can be seen that the data released by ECMWF can capture the spatial distribution pattern of precipitation in China, which is "more in the east and less in the west." However, due to the deviation of other indices, it indicates that it significantly overestimates the intensity of precipitation in China. After bias correction, the simulation ability of the spatial distribution characteristics of each precipitation index is improved. However, the LightGBM method shows better overall performance: for the four indices PRCPTOT, SDII, RZ, and P95, the spatial Corr after correction by the LightGBM method is higher than that after correction by the CDF-t method, increasing from 0.90, 0.93, 0.90, and 0.96 to 0.92, 0.96, 0.91, and 0.98 respectively, and the RMSE is significantly reduced.

[0118] Figure 4The monthly average precipitation curves for China and the eight study areas are shown, with OBS representing measured precipitation data. Table 3 shows the mean absolute deviation for China and the eight study areas.

[0119] Table 3 Mean Absolute Deviation of the Region

[0120]

[0121] Table 3 shows that the LightGBM method exhibits the lowest mean absolute error (MAE) nationwide, reducing it by approximately 16.7% and 6.4% compared to the original model and the CDF-t method, respectively, indicating stronger applicability and stability across the country. The improvement is particularly significant in regions with concentrated and highly variable precipitation, such as North China, Northeast China, and East China, where the MAE reduction exceeds 14%. In Central and South China, precipitation is significantly influenced by strong monsoons, typhoons, and topography, resulting in greater uncertainty; therefore, the bias correction method's improvement on the data is less pronounced. In regions with complex topography and low precipitation, such as Southwest China, Northwest China, and the Qinghai-Tibet Plateau, the LightGBM method demonstrates superiority, indicating that it effectively avoids "false wet days" and the misrepresentation of "low precipitation" events.

[0122] In terms of precipitation curves, the LightGBM method shows better improvement in peak precipitation and duration than the CDF-t method. In North China, Northeast China, East China, South China, Central China, Southwest China, and Northwest China, the correlation of monthly average precipitation curves using the LightGBM method is superior to that using the CDF-t method. Particularly in North China, East China, and Southwest China, the CDF-t method exhibits a "negative correction," which aligns with the aforementioned shortcomings of the CDF-t method. However, it is worth noting that in the Tibetan Plateau region, although the LightGBM method reduces MAE better than the CDF-t method, the correlation coefficient of its monthly average precipitation curve is slightly lower than that of the CDF-t method.

[0123] Finally, it should be noted that the above description is only used to illustrate the technical solution of the present invention and not to limit it. Although the present invention has been described in detail with reference to the preferred arrangement, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solution of the present invention without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A precipitation deviation correction method based on LightGBM, characterized in that, The method includes the following steps: Step 1: Data Collection and Preprocessing: Collect historical long-series CN05.1 grid-based meteorological data for the study area, including measured precipitation data; collect historical long-series ECMWF model precipitation forecast data for the study area; and then use bilinear interpolation to unify the spatial resolution of the measured precipitation data and the precipitation forecast data. Step 2: Establish multi-scale precipitation characteristics: Read the basic information of historical precipitation forecast data and establish multi-scale precipitation characteristics grid by grid, including latitude, longitude, current day's model precipitation, whether the current day is a wet day, previous day's model precipitation, previous two days' model precipitation, previous three days' model precipitation, previous three days' total precipitation, previous seven days' total precipitation, previous thirty days' total precipitation, previous three days' maximum precipitation, consecutive wet day length up to yesterday, consecutive dry day length up to yesterday, annual total precipitation, monthly total precipitation, seasonal quantile position, and logarithmic anomaly with respect to the seasonal median. Step 3: Train the classification and regression models: Input multi-scale precipitation features at daily scales, labeled as measured wet / dry days, and train the LightGBM classifier to obtain the trained classification model. Under wet day conditions, using the multi-scale characteristics of wet day samples from precipitation forecast data as input and the corresponding measured precipitation as labels, a quantile regression model based on LightGBM is trained. The specific process is as follows: 1) Training a classification model based on log loss: （2） In the formula: The loss function for the classification model; This represents the number of training samples; Let i be the true label of the i-th sample. ; The predicted probability output by the model; Using a trained classification model The system outputs the probability of daily precipitation, performs probability calibration, and determines the optimal threshold for wet day discrimination based on the historical wet day frequency to determine whether precipitation has occurred. The specific process is as follows: First, using the trained classification model Mapping historical multi-scale precipitation characteristics to the probability of precipitation occurrence: （3） Secondly, isotonic regression is used for probability calibration: （4） Then determine the optimal threshold for wet day discrimination after calibration: （5） In the formula: For a certain grid point; For a certain processing time; For feature vectors; For the Sigmoid function, specifically... ; This represents the probability of a wet day occurring, as output by the model. The calibrated probability of wet days; It is a monotonic calibration function; This indicates the search for parameter T that minimizes the following expression; For indicator functions; This represents the total number of training samples, which is the sum of all days for all grid points. To observe the frequency of wet days; For the threshold, ; The optimal threshold for wet day discrimination after calibration; 2) Training a quantile regression model based on minimizing risk: （6） in: （7） In the formula: This is the loss function for quantile regression; For all wet day samples; The quantile loss function; For quantiles, ; This represents the actual measured precipitation. For features Precipitation at quantile points; For residuals; Step 4: Correct for precipitation bias: For future precipitation forecast data, establish multi-scale precipitation characteristics according to the same rules, first by using a classification model. Determine whether it's a dry or wet day; in the case of a wet day, use the characteristics of the wet day samples as input and apply a quantile regression model. The precipitation quantiles are corrected, and the high-quantile extreme values ​​are corrected using a generalized Pareto distribution. The tails are then extrapolated to obtain the final corrected precipitation. The specific process is as follows: 1) If we set wet days as 1 and dry days as 0, we can determine whether precipitation occurs at a specific location and time: （8） In the formula: Used to indicate whether precipitation has occurred at a specific location and time; 2) Under wet day conditions, use the trained regression model. The quantile for predicting precipitation is given by the following formula: （9） In the formula: The quantile regression model predicts the 1st Precipitation at quantile points; This is a quantile regression model; 3) Improve numerical stability using logarithmic transformation: （10） （11） （12） In the formula: This represents the logarithm of the observed precipitation plus 1. Indicates the prediction of the th in the logarithmic space Precipitation at quantile points; It is a logarithmic space model; 4) Apply generalized Pareto distribution (GPD) correction to high quantile extreme values: （13） Push the tail outward: （14） In the formula: The cumulative distribution function of the generalized Pareto distribution; The precipitation value corresponding to the 95th percentile; It is a scaling function; The shape parameters are obtained using maximum likelihood estimation; The first one is obtained by extrapolation based on the generalized Pareto distribution. quantile correction value ; 5) The precipitation at each quantile is obtained as follows: （15） Interpolation yields the final precipitation as follows: （16） In the formula: This is the corrected quantile precipitation series; , for The adjacent quantiles, and ; , quantiles The corresponding quantile precipitation values.

2. The precipitation deviation correction method based on LightGBM according to claim 1, characterized in that, The measured precipitation data in step 1 was obtained from the official website of the National Climate Change Research Center; the precipitation forecast data was obtained from the official website of ECMWF; and the historical long-series data consisted of data from at least ten years.

3. The precipitation deviation correction method based on LightGBM according to claim 1, characterized in that, The formula for unifying the spatial resolution of measured precipitation data and forecast precipitation data using bilinear interpolation is as follows: （1） In the formula: The coordinates of the points to be interpolated and their corresponding values; , , , These are the coordinates of four known points and their corresponding values.

Images