CMIP daily scale rainfall data mode set generation method based on machine learning

Abstract

The invention discloses a CMIP daily scale rainfall data pattern set generation method based on machine learning. The method comprises the following steps: step 1, data collection: collecting historical daily scale rainfall data of a research area; 2, constructing a training sample: constructing the training sample by taking the multi-mode monthly scale rainfall data and rainfall annual distribution characteristics of the central grid and the screened peripheral grids as sample characteristics and taking the corresponding observation rainfall data as a sample label; 3, model training: training a random forest regression model grid by grid by using the training sample constructed in the step 2; training a random forest classification model grid by grid by taking historical daily scale rain data as a training sample; and step 4, generating a daily scale rainfall data mode set: based on the intra-month rainfall weight of the rainy day, distributing the monthly scale rainfall data to the daily scale rainfall data. According to the method, the performance difference between different modes is considered, the spatial-temporal heterogeneity can be reflected more accurately, and the simulation precision is effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of hydrological technical analysis, and in particular relates to a method for generating a CMIP daily-scale precipitation data pattern set based on machine learning. Background Art

[0002] Global change and climate change have become one of the most concerned topics in the world in recent years. The IPCC Sixth Assessment Report pointed out that as global warming further intensifies in the future, the spatial and temporal distribution pattern of precipitation and extreme precipitation events will change significantly. The International Coupled Model Intercomparison Project Phase (CMIP) is an important tool for studying future climate change. However, there are large differences in the results of models from different institutions. In application, model ensembles are needed to correct the deviations of individual models. Existing studies mostly use multi-model averaging methods to ensemble models. The simulation results between different models may have large differences in simulation effects at different times and spaces. This is particularly evident in precipitation data. For example, some models simulate better in the flood season, but worse in the dry season; some models simulate better in mountainous areas, but worse in plains. Although the multi-model averaging method is simple and can reduce some deviations, it cannot effectively improve the simulation accuracy of the ensemble results, and does not consider the spatial differences between different models. At the same time, the rainy days and rainy days of different models are different on a daily scale. The model averaging method will result in different degrees of precipitation every day, and it cannot distinguish between rainy and rainless situations. Summary of the invention

[0003] The purpose of the present invention is to provide a method for generating a CMIP daily-scale precipitation data pattern set based on machine learning, which can generate a CMIP daily-scale precipitation data set with higher simulation accuracy and spatiotemporal differences compared with a multi-model averaging method.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] The present invention discloses a method for generating a CMIP daily-scale precipitation data pattern set based on machine learning, the method comprising the following steps:

[0006] Step 1: Data collection: Collect historical CMIP multi-model daily precipitation data and daily observation precipitation data in the study area, and interpolate precipitation data of different spatial resolutions to a unified spatial resolution;

[0007] Step 2, construct training samples: bias correction is performed on CMIP multi-model precipitation data according to historical observed precipitation data, and precipitation data is sampled to the monthly scale; the annual distribution characteristics of precipitation are extracted grid by grid based on the monthly observed precipitation data; for each grid, it is regarded as the central grid, the correlation between the central grid and its surrounding grids is calculated, and the surrounding grids with correlation coefficients greater than 0.95 are selected; the CMIP multi-model monthly precipitation data and annual distribution characteristics of precipitation of the central grid and the selected surrounding grids are used as training sample features, and the corresponding monthly observed precipitation data are used as training sample labels to construct training samples;

[0008] Step 3, model training: The random forest regression model is trained grid by grid with the training samples constructed in step 2. Based on the trained model, the monthly precipitation under different scenarios in the future is calculated. The random forest classification model is trained grid by grid with the historical daily-scale rain data as training samples, and the daily rain data under different scenarios in the future are calculated.

[0009] Step 4, generate a daily-scale precipitation data pattern set: based on the daily rainy or not data under different scenarios in the future, calculate the monthly precipitation weight of each rainy day; based on the monthly-scale precipitation under different scenarios in the future and the monthly precipitation weight of rainy days, generate the daily-scale precipitation under different scenarios in the future, and then obtain the daily-scale precipitation data pattern set.

[0010] Furthermore, the step 1 of collecting historical CMIP multi-model daily precipitation data and daily observed precipitation data of the study area and interpolating precipitation data of different spatial resolutions to a uniform spatial resolution is as follows: the collected data include CMIP6 multi-model precipitation data and CN05.1 grid observed precipitation data;

[0011] The CMIP6 multi-model precipitation data is obtained from the Internet and interpolated to a uniform spatial resolution of 0.25° using a bilinear interpolation method. The formula for the bilinear interpolation method is:

[0012]

[0013] Where: f(x,y) is the coordinate point to be interpolated and its corresponding value; Q 11 (x1,y1),Q 21 (x2,y1),Q 21 (x2,y1),Q 12 (x1, y2) are the coordinates of four known points and their corresponding values;

[0014] The CN05.1 grid observation precipitation data is a daily observation data with a spatial resolution of 0.25° obtained by interpolation and superposition of more than 2,400 national stations of the National Meteorological Information Center using the thin disk spline function method and the angular distance weight method.

[0015] Furthermore, the bias correction of CMIP multi-model precipitation data according to historical observed precipitation data described in step 2 is specifically as follows: bias correction is performed using the transfer cumulative probability step-by-step method, assuming that there is a transfer function T in the historical period, which can establish a relationship between the cumulative probability distribution function of historical observed precipitation data and the cumulative probability distribution function of historical CMIP multi-model precipitation data, as shown in the following formula:

[0016] T(F MH (X))=F OH (X) (2)

[0017] First, define u = F MH (X), then The value range of u is [0,1]. Substituting X into the equation, the transfer function T is:

[0018]

[0019] Assuming that the transfer function still holds true in the future, the corrected result is:

[0020] T(F MH (X))=F OF (X) (4)

[0021]

[0022] Where: F OH (X) is the cumulative probability distribution function of historical observed precipitation data; F MH (X) is the cumulative probability distribution function of historical CMIP multi-model precipitation data; F OF (X) is the cumulative probability distribution function of future observed precipitation; F CF (X) is the cumulative probability distribution function of future CMIP multi-model precipitation data after bias correction.

[0023] Furthermore, the extraction of annual distribution characteristics of precipitation grid by grid based on monthly scale observed precipitation data in step 2 is specifically as follows: given monthly scale observed precipitation data P m (x, y), where m represents the month m = 1, 2, ..., 12, (x, y) represents the grid position, for each grid (x, y), calculate the multi-year average monthly precipitation To determine the month with the maximum precipitation, mark the annual distribution characteristics; define the feature labeling matrix T(x,y), whose size is the same as the grid data; extract and label the annual distribution characteristics of precipitation grid by grid according to the following rules: label the month with the maximum precipitation as 4, the second and third months as 3, the fourth and fifth months as 2, and the remaining months as 0;

[0024] Average monthly precipitation over the years The calculation formula is:

[0025]

[0026] Where: It represents the precipitation in the mth month of the nth year; N is the number of observation years.

[0027] Furthermore, the calculation of the correlation between the central grid and its surrounding grids in step 2 is specifically as follows: for each grid (x, y), it is regarded as the central grid, and its correlation with the surrounding 8 grids is calculated. The correlation calculation adopts the Pearson correlation coefficient PCC, and the calculation formula is:

[0028]

[0029] Where: X is the monthly precipitation data of the central grid; Y is the monthly precipitation data of the surrounding grids.

[0030] Furthermore, in step 3, the training samples constructed in step 2 are used to train the random forest regression model grid by grid, and based on the trained model, the monthly precipitation under different scenarios in the future is calculated as follows: the training samples constructed in step 2 are used, the CMIP multi-model monthly precipitation data and the annual distribution characteristics of precipitation of the central grid and the screened surrounding grids are used as training sample features, the corresponding monthly observed precipitation data are used as training sample labels, the random forest regression model is trained grid by grid, and the grid search algorithm is used for hyperparameter optimization;

[0031] The random forest regression model is:

[0032]

[0033] Where: represents the predicted value of the random forest model in the grid (x, y); f k (x, y; θ) represents the kth decision tree model; θ represents the model parameter set;

[0034] The grid search algorithm formula is:

[0035]

[0036] Where: θ *is the best hyperparameter combination; GridNum is the total number of grids; Represents the predicted value of the model at the grid (x, y); P obs (x,y) represents the actual observed value of the grid (x,y);

[0037] Based on the trained random forest regression model for each grid, the monthly precipitation under different scenarios in the future is calculated by taking the bias-corrected CMIP multi-model monthly precipitation data and annual distribution characteristics as input.

[0038] Furthermore, in step 3, the random forest classification model is trained grid by grid using the historical daily-scale rain data as training samples, and the daily rain data under different scenarios in the future are calculated as follows: the historical CMIP multi-model daily-scale precipitation data and the daily-scale observed precipitation data are classified into two types, rain and no rain, and marked as 0 and 1 respectively; the rain or no rain of the multi-model daily-scale precipitation data is used as the training sample feature, and the rain or no rain of the observed precipitation data is used as the training sample label, and the random forest classification model is trained grid by grid, and the grid search algorithm is used for hyperparameter optimization;

[0039] The two-category labeling formula is:

[0040]

[0041] Where: τ is the rainfall threshold (0.5 mm), which is used to distinguish between rain and no rain; is the precipitation of the kth mode on the dth day;

[0042] Based on the trained random forest classification model, the bias-corrected future CMIP multi-model daily-scale rain-or-no rain characteristics are used as input to calculate the daily rain-or-no rain data under different scenarios in the future.

[0043] Furthermore, the step 4 of calculating the monthly precipitation weight of each rainy day based on the daily rainy or not data under different scenarios in the future is specifically as follows: based on the calculated daily rainy or not data under different scenarios in the future, the daily precipitation of the rainy day is calculated by a multi-mode averaging method, and the monthly precipitation weight of each rainy day is calculated;

[0044] The calculation formula for precipitation weight is:

[0045]

[0046] Where: W d,m represents the precipitation weight of day d in month m, reflecting the proportion of the precipitation on that day in the total precipitation of the whole month; P d (x, y) is the model average precipitation on rainy day d; P m(x,y) is the precipitation on rainy days in the month P d The sum of (x,y);

[0047] Based on the monthly precipitation under different scenarios in the future and the monthly precipitation weights of rainy days, the daily precipitation under different scenarios in the future is generated, and then the daily precipitation data pattern set is obtained as follows:

[0048] Based on the calculated monthly precipitation under different scenarios in the future and the calculated monthly precipitation weights of rainy days, the monthly precipitation data is spread to the daily scale to generate the daily precipitation under different scenarios in the future, and then the daily precipitation data pattern set is obtained;

[0049] The daily precipitation calculation formula is:

[0050]

[0051] Where: W d,m (x,y) is the precipitation weight of the dth day in the mth month; is the monthly precipitation of the mth month on the grid (x, y); d,ensemble (x,y) is the daily precipitation on the dth day on the grid (x,y) under different scenarios in the future.

[0052] The beneficial effects of the present invention are as follows: a method for generating a CMIP daily-scale precipitation data pattern set based on machine learning is proposed in the present invention. By introducing the annual distribution characteristics of precipitation to train a random forest regression model on a monthly scale, a better data set result than a pattern averaging method is achieved. The method of the present invention takes into account the performance differences between different patterns, can more accurately reflect the temporal and spatial heterogeneity, and effectively improves the simulation accuracy of the results; secondly, by training a random forest classification model on a daily scale, it is judged whether there will be rain in the future, the distribution of rainy days in a month under future scenarios is calculated, and the daily-scale precipitation data is further deployed with the monthly-scale precipitation data as a constraint, thereby forming a high-temporal-resolution future scenario daily-scale precipitation data set.

[0053] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific implementation methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 is a flow chart of the method of the present invention;

[0055] Figure 2 This is an overview of the study area of ​​Example 1;

[0056] Figure 3 This is a schematic diagram of the interpolation process of CMIP multi-model precipitation data in Example 1;

[0057] Figure 4Box plots of annual precipitation before and after deviation correction for four watershed units in Example 1;

[0058] Figure 5 This is a tree structure diagram of the random forest regression model after training in Example 1;

[0059] Figure 6 Taylor diagrams of the CMIP models and model sets for precipitation in four watershed units in Example 1;

[0060] Figure 7 This is a binary classification result diagram of whether there is rain or not in random months in Example 1;

[0061] Figure 8 This is a precipitation weight map for a rainy day in Example 1. DETAILED DESCRIPTION

[0062] The present invention discloses a method for generating a CMIP daily-scale precipitation data pattern set based on machine learning, such as Figure 1 As shown, the method comprises the following steps:

[0063] Step 1: Collect data: Collect historical CMIP multi-model daily precipitation data and daily observed precipitation data in the study area, and interpolate precipitation data of different spatial resolutions to a unified spatial resolution.

[0064] Among them, the collected data include CMIP6 (Sixth International Coupled Model Intercomparison Project) multi-model precipitation data and CN05.1 grid observation precipitation data.

[0065] CMIP6 multi-model precipitation data can be obtained from the Internet at the specific URL: https: / / esgf-data.dkrz.de / search / cmip6-dkrz / . Since the spatial resolutions of the collected CMIP6 multi-model precipitation data are different, bilinear interpolation is required to interpolate to a uniform spatial resolution of 0.25° for subsequent calculations. The formula for the bilinear interpolation method is:

[0066]

[0067] Where: f(x,y) is the coordinate point to be interpolated and its corresponding value; Q 11 (x1,y1),Q 21 (x2,y1),Q 21 (x2,y1),Q 12 (x1, y2) are the coordinates of four known points and their corresponding values.

[0068] The CN05.1 grid precipitation data is a daily observation data with a spatial resolution of 0.25° obtained by interpolating and superimposing more than 2,400 national stations of the National Meteorological Information Center using the thin disk spline function method (ANUSPLIN) and the angular distance weighting method (ADW).

[0069] Step 2, construct training samples: bias-correct the CMIP multi-model precipitation data according to historical observed precipitation data, and sample the precipitation data to the monthly scale; extract the annual distribution characteristics of precipitation grid by grid based on the monthly observed precipitation data; for each grid, regard it as the central grid, calculate the correlation between the central grid and its surrounding grids, and select the surrounding grids with a correlation coefficient greater than 0.95; use the CMIP multi-model monthly precipitation data and annual distribution characteristics of precipitation of the central grid and the selected surrounding grids as training sample features, and the corresponding monthly observed precipitation data as training sample labels to construct training samples.

[0070] The specific steps include:

[0071] Step 21: perform bias correction on the CMIP multi-model precipitation data based on historical observed precipitation data, and sample the precipitation data to the monthly scale. Specifically, the transfer cumulative probability step-by-step method (CDF-t) is used to perform bias correction. The main idea of ​​CDF-t is that, assuming that there is a transfer function T in the historical period, the cumulative probability distribution function (CDF) of the historical observed precipitation data can be related to the cumulative probability distribution function of the historical CMIP multi-model precipitation data, as shown in the following formula:

[0072] T(F MH (X))=F OH (X) (2)

[0073] First, define u = F MH (X), then The value range of u is [0,1]. Substituting X into the equation, the transfer function T is:

[0074]

[0075] Assuming that the transfer function still holds true in the future, the corrected result is:

[0076] T(F MH (X))=F OF (X) (4)

[0077]

[0078] Where: F OH (X) is the cumulative probability distribution function of historical observed precipitation data; F MH(X) is the cumulative probability distribution function of historical CMIP multi-model precipitation data; F OF (X) is the cumulative probability distribution function of future observed precipitation; F CF (X) is the cumulative probability distribution function of future CMIP multi-model precipitation data after bias correction.

[0079] Step 22: Extract the annual distribution characteristics of precipitation grid by grid based on the monthly scale observed precipitation data. Specifically, given the monthly scale observed precipitation data P m (x, y), where m represents the month m = 1, 2, ..., 12, (x, y) represents the grid position, for each grid (x, y), calculate the multi-year average monthly precipitation To determine the month with the maximum precipitation, mark the annual distribution characteristics. Define the feature labeling matrix T(x,y), whose size is the same as the raster data. According to the following rules, extract and label the annual distribution characteristics of precipitation grid by grid: label the month with the maximum precipitation as 4, the second and third months as 3, the fourth and fifth months as 2, and the remaining months as 0.

[0080] Average monthly precipitation over the years The calculation formula is:

[0081]

[0082] Where: It represents the precipitation in the mth month of the nth year; N is the number of observation years.

[0083] Step 23: Form training samples. For each grid (x, y), regard it as the central grid and calculate its correlation with the surrounding 8 grids. The correlation calculation uses the Pearson correlation coefficient PCC, and the calculation formula is:

[0084]

[0085] Where: X is the monthly precipitation data of the central grid; Y is the monthly precipitation data of the surrounding grids.

[0086] For each central grid (x, y), if the correlation coefficient PCC of a surrounding grid with the central grid is greater than 0.95, the surrounding grid is added to the training sample. The annual distribution characteristics of precipitation refer to the corresponding values ​​(0, 2, 3 or 4) in the feature labeling matrix T(x, y) defined in step 22. Some historical CMIP6 precipitation data and historical observation data are used as training sets to construct the following training samples: feature X{CMIP multi-model monthly precipitation, monthly labeling features}, label Y{observed monthly precipitation}, and the training set contains the central grid and the filtered surrounding grid data. Repeat the above process until all grids are processed to form the final training sample set.

[0087] Step 3, model training: The random forest regression model is trained grid by grid with the training samples constructed in step 2. Based on the trained model, the monthly precipitation under different scenarios in the future is calculated. The random forest classification model is trained grid by grid with the historical daily-scale rain data as training samples, and the daily rain data under different scenarios in the future is calculated.

[0088] The specific steps include:

[0089] Step 31: Using the training samples constructed in step 2, the CMIP multi-model monthly precipitation data and annual distribution characteristics of precipitation of the central grid and the screened surrounding grids are used as training sample features, and the corresponding monthly observed precipitation data are used as training sample labels. The random forest regression model is trained grid by grid, and the grid search algorithm is used for hyperparameter optimization.

[0090] The random forest regression model is:

[0091]

[0092] Where: represents the predicted value of the random forest model in the grid (x, y); f k (x, y; θ) represents the kth decision tree model; θ represents the model parameter set.

[0093] The grid search algorithm formula is:

[0094]

[0095] Where: θ * is the best hyperparameter combination; GridNum is the total number of grids; Represents the predicted value of the model at the grid (x, y); P obs (x,y) represents the actual observed value of the grid (x,y).

[0096] Step 32: Based on the random forest regression model of each grid trained in step 31, the monthly precipitation data of the future CMIP multi-model after bias correction and the annual distribution characteristics are used as input to calculate the monthly precipitation under different scenarios in the future, that is,

[0097] Step 33: Classify the historical CMIP multi-model daily precipitation data and the daily observed precipitation data into two types: rain and no rain, and label them as 0 and 1 respectively. The multi-model daily precipitation data with or without rain is used as the training sample feature, and the observed precipitation data with or without rain is used as the training sample label. The random forest classification model is trained grid by grid, and the grid search algorithm is used for hyperparameter optimization.

[0098] The two-category labeling formula is:

[0099]

[0100] Where: τ is the rainfall threshold (0.5 mm), which is used to distinguish between rain and no rain; is the precipitation of the kth mode on the dth day.

[0101] Step 34: Based on the random forest classification model trained in step 33, the bias-corrected future CMIP multi-model daily-scale rain-or-no rain characteristics are used as input to calculate the daily rain-or-no rain data under different scenarios in the future.

[0102] Step 4, generate a daily-scale precipitation data pattern set: based on the daily rainy or not data under different scenarios in the future, calculate the monthly precipitation weight of each rainy day; based on the monthly-scale precipitation under different scenarios in the future and the monthly precipitation weight of rainy days, generate the daily-scale precipitation under different scenarios in the future, and then obtain the daily-scale precipitation data pattern set.

[0103] The specific steps include:

[0104] Step 41: Based on the daily rainy or not data under different future scenarios obtained in step 34, the daily precipitation on rainy days is calculated by a multi-mode averaging method, and the monthly precipitation weight of each rainy day is calculated.

[0105] The calculation formula for precipitation weight is:

[0106]

[0107] Where: W d,m represents the precipitation weight of day d in month m, reflecting the proportion of the precipitation on that day in the total precipitation of the whole month; P d (x, y) is the model average precipitation on rainy day d; P m (x,y) is the precipitation on rainy days in the month P d The sum of (x,y).

[0108] Step 42: Based on the monthly precipitation under different scenarios in the future calculated in step 32 and the monthly precipitation weight of the rainy day calculated in step 41, the monthly precipitation data is spread to the daily scale to generate the daily precipitation under different scenarios in the future, and then a daily precipitation data pattern set is obtained;

[0109] The daily precipitation calculation formula is:

[0110]

[0111] Where: W d,m (x,y) is the precipitation weight of the dth day in the mth month; is the monthly precipitation of the mth month on the grid (x, y); d,ensemble (x,y) is the daily precipitation on the dth day on the grid (x,y) under different scenarios in the future.

[0112] The present invention is based on the bias-corrected multi-mode CMIP precipitation data, takes the CN05.1 precipitation dataset as the observation data, and adopts the random forest method with the introduction of time features to perform regression analysis grid by grid, so that the annual distribution and spatial variation characteristics are considered in the model training process, and the daily-scale data are spread out by binary classification according to whether there is rain or not, so as to generate a CMIP precipitation daily-scale dataset with higher simulation accuracy and spatiotemporal differences compared with the multi-mode averaging method.

[0113] Embodiment 1

[0114] This embodiment is an application example of the above method.

[0115] This embodiment discloses a method for generating a CMIP daily-scale precipitation data pattern set based on machine learning. The Tianshan area is used as a research area to verify the above method. The research area range is as follows: Figure 2 The Tianshan area is subdivided into four third-level watershed units, namely the Eastern Tianshan (ETM), Southern Tianshan (STM), Central Tianshan (CTM) and Northern Tianshan (NTM).

[0116] The method described in this embodiment includes the following steps:

[0117] Step 1, data collection: collect historical CMIP multi-model daily precipitation data and daily observed precipitation data in the study area, and interpolate precipitation data with different spatial resolutions to a unified spatial resolution. The detailed information of the CMIP6 model collected in this embodiment is shown in Table 1. The multi-model precipitation data with different spatial resolutions are interpolated to a unified spatial resolution of 0.25° using the bilinear interpolation method, as shown in Table 1. Figure 3 shown.

[0118] Table 1 CMIP6 model information

[0119]

[0120] Step 2, construct training samples: bias-correct the CMIP multi-model precipitation data according to historical observed precipitation data, and sample the precipitation data to the monthly scale; extract the annual distribution characteristics of precipitation grid by grid based on the monthly observed precipitation data; for each grid, regard it as the central grid, calculate the correlation between the central grid and its surrounding grids, and select the surrounding grids with a correlation coefficient greater than 0.95; use the CMIP multi-model monthly precipitation data and annual distribution characteristics of precipitation of the central grid and the selected surrounding grids as training sample features, and the corresponding monthly observed precipitation data as training sample labels to construct training samples.

[0121] like Figure 4 As shown, the box plots of annual precipitation in the four basin units before and after bias correction are shown. Before bias correction, the original CMIP models seriously underestimated the precipitation in the NTM, ETM, and MTM regions. After bias correction, the simulation effect was greatly improved, and its mean was close to the observed value.

[0122] Step 3, model training: The random forest regression model is trained grid by grid with the training samples constructed in step 2. Based on the trained model, the monthly precipitation under different scenarios in the future is calculated. The random forest classification model is trained grid by grid with the historical daily-scale rain data as training samples, and the daily rain data under different scenarios in the future is calculated.

[0123] Randomly select a coordinate point to display the tree structure of the trained random forest regression model, such as Figure 5 shown. Figure 6 The Taylor diagrams of the CMIP models and model ensembles of precipitation in the four basin units of ETM, NTM, STM, and CTM are shown. The correlation coefficients of each partition before the ensemble are around 0.4, with the maximum value not exceeding 0.6, and the variance of each model is close to the variance of the observed data, indicating that the constructed model can better capture the variability and fluctuation range of the observed data. After the model ensemble, the correlation coefficient is increased to above 0.6, which can greatly reduce the uncertainty of precipitation simulation.

[0124] Step 4, generate a daily-scale precipitation data pattern set: based on the daily rainy or not data under different scenarios in the future, calculate the monthly precipitation weight of each rainy day; based on the monthly-scale precipitation under different scenarios in the future and the monthly precipitation weight of rainy days, generate the daily-scale precipitation under different scenarios in the future, and then obtain the daily-scale precipitation data pattern set.

[0125] Figure 7 The first 25 columns are the rainy and rainless conditions in different modes, and the last column is the rainy day judgment after random forest classification and regression. The rainy days and rainless days judgment results are shown in Figure 2. There are 15 rainy days in this month, and the precipitation weights of rainy days are as follows: Figure 8 As shown, it is between 0.008 and 0.12. Then, a daily precipitation data pattern set is generated based on the precipitation weight and monthly precipitation.

[0126] Finally, it should be noted that the above is only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred arrangement scheme, a person skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for generating a CMIP daily-scale precipitation data pattern set based on machine learning, characterized in that: The method comprises the following steps: Step 1: Data collection: Collect historical CMIP multi-model daily precipitation data and daily observation precipitation data in the study area, and interpolate precipitation data of different spatial resolutions to a unified spatial resolution; Step 2, construct training samples: bias correction is performed on CMIP multi-model precipitation data according to historical observed precipitation data, and precipitation data is sampled to the monthly scale; the annual distribution characteristics of precipitation are extracted grid by grid based on the monthly observed precipitation data; for each grid, it is regarded as the central grid, the correlation between the central grid and its surrounding grids is calculated, and the surrounding grids with correlation coefficients greater than 0.95 are selected; the CMIP multi-model monthly precipitation data and annual distribution characteristics of precipitation of the central grid and the selected surrounding grids are used as training sample features, and the corresponding monthly observed precipitation data are used as training sample labels to construct training samples; Step 3, model training: The random forest regression model is trained grid by grid with the training samples constructed in step 2. Based on the trained model, the monthly precipitation under different scenarios in the future is calculated. The random forest classification model is trained grid by grid with the historical daily-scale rain data as training samples, and the daily rain data under different scenarios in the future are calculated. Step 4, generate a daily-scale precipitation data pattern set: based on the daily rainy or not data under different scenarios in the future, calculate the monthly precipitation weight of each rainy day; based on the monthly-scale precipitation under different scenarios in the future and the monthly precipitation weight of rainy days, generate the daily-scale precipitation under different scenarios in the future, and then obtain the daily-scale precipitation data pattern set.

2. The method for generating a CMIP daily-scale precipitation data pattern set based on machine learning according to claim 1, characterized in that: Step 1 is to collect historical CMIP multi-model daily precipitation data and daily observed precipitation data in the study area, and interpolate precipitation data of different spatial resolutions to a unified spatial resolution. Specifically, the collected data include CMIP6 multi-model precipitation data and CN05.1 grid observed precipitation data; The CMIP6 multi-model precipitation data is obtained from the Internet and interpolated to a uniform spatial resolution of 0.25° using a bilinear interpolation method. The formula for the bilinear interpolation method is: Where: f(x,y) is the coordinate point to be interpolated and its corresponding value; Q 11 (x1,y1),Q 21 (x2,y1),Q 21 (x2,y1),Q 12 (x1, y2) are the coordinates of four known points and their corresponding values; The CN05.1 grid observation precipitation data is a daily observation data with a spatial resolution of 0.25° obtained by interpolation and superposition of more than 2,400 national stations of the National Meteorological Information Center using the thin disk spline function method and the angular distance weight method.

3. The method for generating a CMIP daily-scale precipitation data pattern set based on machine learning according to claim 1, characterized in that: The specific method of bias correction of CMIP multi-model precipitation data according to historical observed precipitation data in step 2 is as follows: bias correction is performed using the transfer cumulative probability step-by-step method. Assuming that there is a transfer function T in the historical period, the cumulative probability distribution function of historical observed precipitation data can be related to the cumulative probability distribution function of historical CMIP multi-model precipitation data, as shown in the following formula: T(F MH (X))=F PH (X) (2) First, define u = F MH (X), then The value range of u is [0,1]. Substituting X into the equation, the transfer function T is: Assuming that the transfer function still holds true in the future, the corrected result is: T(F MH (X))=F OF (X) (4) Where: F OH (X) is the cumulative probability distribution function of historical observed precipitation data; F MH (X) is the cumulative probability distribution function of historical CMIP multi-model precipitation data; F OF (X) is the cumulative probability distribution function of future observed precipitation; F CF (X) is the cumulative probability distribution function of future CMIP multi-model precipitation data after bias correction.

4. The method for generating a CMIP daily-scale precipitation data pattern set based on machine learning according to claim 1, characterized in that: The specific method of extracting the annual distribution characteristics of precipitation based on the monthly observed precipitation data grid by grid in step 2 is as follows: given the monthly observed precipitation data P m (x, y), where m represents the month m = 1, 2, ..., 12, (x, y) represents the grid position, for each grid (x, y), calculate the multi-year average monthly precipitation To determine the month with the most precipitation and mark the distribution characteristics within the year; Define the feature labeling matrix T(x,y), whose size is the same as the grid data; extract and label the annual distribution characteristics of precipitation grid by grid according to the following rules: label the month with the maximum precipitation as 4, the second and third months as 3, the fourth and fifth months as 2, and the remaining months as 0; Average monthly precipitation over the years The calculation formula is: Where: It represents the precipitation in the mth month of the nth year; N is the number of observation years.

5. The method for generating a CMIP daily-scale precipitation data pattern set based on machine learning according to claim 1, characterized in that: The calculation of the correlation between the central grid and its surrounding grids in step 2 is specifically as follows: for each grid (x, y), it is regarded as the central grid, and its correlation with the surrounding 8 grids is calculated. The correlation calculation uses the Pearson correlation coefficient PCC, and the calculation formula is: Where: X is the monthly precipitation data of the central grid; Y is the monthly precipitation data of the surrounding grids.

6. The method for generating a CMIP daily-scale precipitation data pattern set based on machine learning according to claim 1, characterized in that: In step 3, the random forest regression model is trained grid by grid with the training samples constructed in step 2. Based on the trained model, the monthly precipitation under different scenarios in the future is calculated as follows: the training samples constructed in step 2 are used, the CMIP multi-mode monthly precipitation data and the annual distribution characteristics of precipitation of the central grid and the screened surrounding grids are used as training sample features, the corresponding monthly observed precipitation data are used as training sample labels, the random forest regression model is trained grid by grid, and the grid search algorithm is used for hyperparameter optimization; The random forest regression model is: Where: represents the predicted value of the random forest model in the grid (x, y); f k (x, y; θ) represents the kth decision tree model; θ represents the model parameter set; The grid search algorithm formula is: Where: θ * is the best hyperparameter combination; GridNum is the total number of grids; Represents the predicted value of the model at the grid (x, y); P obs (x,y) represents the actual observed value of the grid (x,y); Based on the trained random forest regression model for each grid, the monthly precipitation under different scenarios in the future is calculated by taking the bias-corrected CMIP multi-model monthly precipitation data and annual distribution characteristics as input.

7. The method for generating a CMIP daily-scale precipitation data pattern set based on machine learning according to claim 1, characterized in that: In step 3, the random forest classification model is trained grid by grid using the historical daily-scale rain data as training samples to calculate the daily rain data under different scenarios in the future. Specifically, the historical CMIP multi-model daily-scale precipitation data and the daily-scale observed precipitation data are classified into two types: rain and no rain, which are marked as 0 and 1 respectively; the multi-model daily-scale precipitation data with or without rain is used as the training sample feature, and the observed precipitation data with or without rain is used as the training sample label, and the random forest classification model is trained grid by grid, and the grid search algorithm is used for hyperparameter optimization; The two-category labeling formula is: Where: τ is the rainfall threshold (0.5 mm), which is used to distinguish between rain and no rain; is the precipitation of the kth mode on the dth day; Based on the trained random forest classification model, the bias-corrected future CMIP multi-model daily-scale rain-or-no rain characteristics are used as input to calculate the daily rain-or-no rain data under different scenarios in the future.

8. The method for generating a CMIP daily-scale precipitation data pattern set based on machine learning according to claim 1, characterized in that: The step 4 of calculating the monthly precipitation weight of each rainy day based on the daily rainy or not data under different scenarios in the future is specifically as follows: based on the calculated daily rainy or not data under different scenarios in the future, the daily precipitation of the rainy day is calculated by a multi-mode average method, and the monthly precipitation weight of each rainy day is calculated; The calculation formula for precipitation weight is: Where: W d,m represents the precipitation weight of day d in month m, reflecting the proportion of the precipitation on that day in the total precipitation of the whole month; P d (x, y) is the model average precipitation on rainy day d; P m (x,y) is the precipitation on rainy days in the month P d The sum of (x,y); Based on the monthly precipitation under different scenarios in the future and the monthly precipitation weights of rainy days, the daily precipitation under different scenarios in the future is generated, and then the daily precipitation data pattern set is obtained as follows: Based on the calculated monthly precipitation under different scenarios in the future and the calculated monthly precipitation weights of rainy days, the monthly precipitation data is spread to the daily scale to generate the daily precipitation under different scenarios in the future, and then the daily precipitation data pattern set is obtained; The daily precipitation calculation formula is: Where: W d,m (x,y) is the precipitation weight of the dth day in the mth month; is the monthly precipitation of the mth month on the grid (x, y); d,ensemble (x,y) is the daily precipitation on the dth day on the grid (x,y) under different scenarios in the future.

Images