Multi-source precipitation product fusion method based on ensemble learning

Abstract

The invention discloses a multi-source precipitation product fusion method based on ensemble learning, and the method employs a plurality of machine learning models and ensemble learning models of different combinations, fuses multi-source precipitation products, considers a plurality of auxiliary factors as the input, and takes a plurality of evaluation indexes as the basis, and generates precipitation products with better quality. And the generated precipitation product is obviously superior to the original precipitation product, has the advantages of different original precipitation products, and can be used for production application and scientific research.

Description

Technical Field

[0001] The present invention belongs to the field of environmental science and meteorological science, and in particular relates to a multi-source precipitation product fusion method based on ensemble learning. Background Art

[0002] Precipitation is an important part of the water cycle and a basic driving factor of surface hydrological processes. It is affected by factors such as topography and climate and has great variability in time and space. How to accurately obtain precipitation information is important for regional water resources management, production activities, disaster prevention and other fields. The traditional method is to observe through meteorological stations. Its disadvantage is that the precipitation information obtained is discrete in space, especially in remote areas with sparse populations and few observation stations. It is difficult to obtain detailed spatially continuous precipitation information. With the development of satellite technology and inversion algorithms, raster precipitation products (abbreviated as precipitation products) began to appear. Precipitation products have the advantage of continuous distribution in space, but their accuracy is not as high as that of station observation data. Since the Tropical Precipitation Meteorological Program TRMM in 1997, a series of satellite precipitation products and reanalysis precipitation products have appeared.

[0003] There are many types of precipitation products today, but different precipitation products have different advantages and disadvantages in different statistical evaluation indicators. It is of great significance to integrate precipitation products with different advantages based on station observation data to obtain precipitation products with comprehensive advantages. Summary of the invention

[0004] Purpose of the invention: In view of the technical problems existing in the prior art, the present invention provides a multi-source precipitation product fusion method based on ensemble learning. Taking the measured precipitation at the site as a benchmark, different precipitation products are fused through machine learning to obtain precipitation products with better accuracy.

[0005] Technical solution: To achieve the above-mentioned invention object, the present invention adopts the following technical solution: a multi-source precipitation product fusion method based on ensemble learning, comprising the following steps: S1, data preparation: prepare meteorological station precipitation data and multi-source precipitation products according to the study area, and extract the raster values ​​of precipitation products corresponding to the meteorological station locations; S2, establish multi-dimensional evaluation indicators: take the station precipitation data as the true value, use the correlation coefficient R, mean absolute relative error MARE, root mean square error RMSE, detection rate POD, false alarm rate FAR, critical success index CSI six evaluation indicators to calculate the evaluation values ​​of multi-source precipitation products; S3, comprehensive score calculation: For the same evaluation index, different precipitation products are ranked from best to worst according to the evaluation value performance, and the ranking is used as the score of the precipitation product under the evaluation index. The comprehensive score SCORE of different precipitation products is calculated by the following formula: j , , In the formula, SCORE j represents the comprehensive score of the j-th type of precipitation product; S kj It represents the score of the j-type precipitation product in the k-th evaluation index, and the value of m is 6; S4, precipitation product optimization: based on the comprehensive scores of different precipitation products obtained in step S3, the different precipitation products are sorted in ascending order according to the comprehensive scores, and the top N precipitation products are selected, where the value of N ranges from 3 to 5; S5, integrated learning model construction (1) Using a variety of machine learning models, taking the measured precipitation data of the station as the benchmark, and taking the precipitation product selected in step S4 and several auxiliary factors as input, the machine learning models are trained and optimized respectively to achieve precipitation fitting; (2) The precipitation products generated by the trained and optimized machine learning model are evaluated according to the evaluation indicators to determine the performance of the machine learning model, and several machine learning models are selected to combine and build an integrated learning model using the stacking method; (3) The ensemble learning model is used to fit the precipitation, and the precipitation product of the study area is generated based on the optimization of the ensemble learning model.

[0006] Furthermore, the multi-source precipitation products in step S1 include satellite precipitation products and reanalysis precipitation products.

[0007] Furthermore, the calculation method of the evaluation values ​​of the six evaluation indicators in step S2 is as follows: (1) Correlation coefficient R j , which is used to measure the correlation between precipitation products and station observation precipitation data. The larger the value, the stronger the linear correlation. It is calculated by the following formula: , In the formula, R j is the correlation coefficient of the j-th type of precipitation product; G i is the observed precipitation value for site i; represents the average value of the observed precipitation at the station, P ij is the j-th type of precipitation product value corresponding to station i, is the average value of the j-th type of precipitation product; n is the number of stations; (2) Root mean square error (RMSE) j, reflects the degree of dispersion of the deviation between the precipitation product and the station observation value. The smaller the root mean square deviation, the smaller the dispersion of the deviation. It is calculated by the following formula: , In the formula, RMSE j is the root mean square deviation of the j-th type of precipitation product; G i is the observed precipitation value at station i, P ij is the j-th type of precipitation product value corresponding to station i; (3) Mean absolute relative deviation (MARE) j , reflects the average deviation between the precipitation product and the station observation value. The closer it is to 0, the smaller the average deviation is, and the more accurate the precipitation product data is. It is calculated by the following formula: , In the formula, MARE j is the average absolute relative deviation of the j-th type of precipitation product; G i is the observed precipitation value at station i, P ij is the j-th type of precipitation product value corresponding to station i; (4) Detection rate POD j , reflects the proportion of precipitation days that the precipitation product can correctly reflect, ranging from [0, 1]. The higher the value, the higher the proportion of precipitation product detection of real precipitation events. It is calculated by the following formula: , In the formula, POD j is the detection rate of the j-th type of precipitation product; H j It represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; M j It indicates the total number of days during the study period when the daily station observation value is less than 0.1 mm, but the corresponding j-type precipitation product value is greater than or equal to 0.1 mm; (5) False Alarm Rate (FAR) j , reflects the proportion of precipitation days that the precipitation product fails to correctly reflect. The value range is [0, 1], and the optimal value is 0. The smaller the value, the lower the error rate of precipitation products in reflecting precipitation events. It is calculated by the following formula: , In the formula, FAR j is the false alarm rate of the jth type of precipitation product; H j represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; F j It represents the total number of days during the study period when the daily station observation value is greater than or equal to 0.1 mm, but the corresponding j-type precipitation product value is less than 0.1 mm. (6) Critical Success Index (CSI) j , which reflects the ratio of successful precipitation detection by precipitation products to rainy days, and its value range is [0, 1]. The closer it is to 1, the more successful the precipitation product detection is and the fewer false alarms are.

[0008] , Where, CSI j is the critical success index of the jth type of precipitation product; H j represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; F j It represents the total number of days during the study period when the daily station observation value is greater than or equal to 0.1 mm, but the corresponding j-th type precipitation product value is less than 0.1 mm; M j It indicates the total number of days during the study period when the daily station observation value is less than 0.1 mm, but the corresponding j-th precipitation product value is greater than or equal to 0.1 mm.

[0009] Furthermore, the several auxiliary factors in step S5 are: longitude, latitude, altitude, distance between land and sea, and cloud cover.

[0010] Furthermore, the machine learning model in step S5 includes RF, CatBoost, KNN, Lasso, DTREE, XGBoost, HGBR and ETREE.

[0011] Furthermore, the training and optimization of the machine learning models respectively described in step S5 is to perform individual training and optimization on each machine learning model.

[0012] Furthermore, the measured precipitation data of the stations in step S5 are divided into a training set, a validation set and a test set in a ratio of 6:2:2 based on the principle of uniform spatial distribution of the stations.

[0013] Furthermore, when the machine learning model described in step S5 is trained and optimized, RandomizedSearchCV is used to find the optimal hyperparameters.

[0014] Beneficial effects: Compared with the prior art, the present invention uses multiple machine learning models and ensemble learning models of different combinations to fuse multi-source precipitation products, and considers multiple auxiliary factors as input, based on multiple evaluation indicators, to generate precipitation products with better quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flowchart of the multi-source precipitation product fusion method based on ensemble learning described in the present invention. DETAILED DESCRIPTION

[0016] The present invention is further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, various equivalent forms of modifications to the present invention by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0017] The multi-source precipitation product fusion method based on ensemble learning of the present invention uses multiple machine learning models and ensemble learning models of different combinations to fuse multi-source precipitation products, and considers multiple auxiliary factors as inputs, based on multiple evaluation indicators, to generate precipitation products with better quality. The specific method steps are described in detail as follows: Figure 1 shown.

[0018] S1. Data preparation for the study area Prepare meteorological station precipitation data and multi-source precipitation products (including satellite precipitation products and reanalysis precipitation products) according to the study area; Extract the raster value of the precipitation product corresponding to the meteorological station location (precipitation product value).

[0019] S2, establish evaluation indicators

[0020] Taking the station precipitation data as the true value, a variety of evaluation indicators are used to evaluate the quality of precipitation products. The evaluation indicators used are: correlation coefficient (R), mean absolute relative error (MARE), root mean square error (RMSE), detection rate (POD), false alarm rate (FAR), critical success index (CSI), and comprehensive score (SCORE), a total of 7 types, the meaning of each indicator is as follows: (1) Correlation coefficient (R j ), which is used to measure the correlation between precipitation products and station observation precipitation data. The larger the value, the stronger the linear correlation.

[0021] , In the formula, R j is the correlation coefficient of the j-th type of precipitation product; G i is the observed precipitation value for site i; represents the average value of the observed precipitation at the station, P ij is the j-th type of precipitation product value corresponding to station i, is the average value of the j-th type of precipitation product; n is the number of stations.

[0022] (2) Root mean square error (RMSE) j ), reflects the degree of dispersion of the deviation between the precipitation product and the station observation value. The smaller the root mean square deviation, the smaller the dispersion of the deviation.

[0023] , In the formula, RMSEj is the root mean square deviation of the j-th type of precipitation product; G i is the observed precipitation value at station i, P ij is the j-th type of precipitation product value corresponding to station i; (3) Mean absolute relative deviation (MARE j ), reflects the average deviation between precipitation products and station observations. The closer it is to 0, the smaller the average deviation is, and the more accurate the precipitation product data is.

[0024] , In the formula, MARE j is the average absolute relative deviation of the j-th type of precipitation product; G i is the observed precipitation value at station i, P ij is the j-th type of precipitation product value corresponding to station i; Note: Since there are many cases where the daily precipitation is 0 mm, only the cases where the daily precipitation at the station is ≥ 1 mm are used for calculation.

[0025] (4) Detection rate (POD) j ), reflects the proportion of precipitation days that the precipitation product can correctly reflect, ranging from [0, 1]. The higher the value, the higher the proportion of precipitation products detecting real precipitation events.

[0026] , In the formula, POD j is the detection rate of the j-th type of precipitation product; H j It represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; M j It indicates the total number of days during the study period when the daily station observation value is less than 0.1 mm, but the corresponding j-th precipitation product value is greater than or equal to 0.1 mm.

[0027] (5) False Alarm Rate (FAR) j ), reflects the proportion of precipitation days that the precipitation product fails to correctly reflect. The value range is [0, 1], and the optimal value is 0. The smaller the value, the lower the error rate of precipitation products in reflecting precipitation events.

[0028] , In the formula, FAR j is the false alarm rate of the jth type of precipitation product; H j represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; F j It represents the total number of days during the study period when the daily station observation value is greater than or equal to 0.1 mm, but the corresponding j-th precipitation product value is less than 0.1 mm.

[0029] (6) Critical Success Index (CSI) j ), which reflects the ratio of successful precipitation detection by precipitation products in the number of rainy days. Its value range is [0, 1]. The closer it is to 1, the more successful the precipitation product detection is and the fewer false alarms are.

[0030] , Where, CSI j is the critical success index of the jth type of precipitation product; H j represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; F j It represents the total number of days during the study period when the daily station observation value is greater than or equal to 0.1 mm, but the corresponding j-th type precipitation product value is less than 0.1 mm; M j It indicates the total number of days during the study period when the daily station observation value is less than 0.1 mm, but the corresponding j-type precipitation product value is greater than or equal to 0.1 mm.

[0031] S3, comprehensive score (SCORE j ), since different precipitation products may show different advantages in the above multiple evaluation indicators, the Brunke ranking method is used to determine the advantages and disadvantages of different precipitation products for the same evaluation indicator. The product with the best indicator is ranked first, the second is ranked second, and so on. That is, different precipitation products are ranked on the same evaluation indicator, and the ranking is the score. Finally, the scores of multiple evaluation indicators are added together to obtain a comprehensive score (SCORE j ), the lower the comprehensive score of the precipitation product, the better its overall performance.

[0032] , In the formula, SCORE j represents the comprehensive score of the j-th type of precipitation product; S kj It represents the score of the j-th type of precipitation product in the k-th evaluation index, and the value of m is 6; S4, precipitation product optimization, combines station precipitation data, uses a variety of evaluation indicators to evaluate the quality of multi-source precipitation products, and selects precipitation products with good quality.

[0033] S5, integrated learning model construction

[0034] (1) Machine learning model training: A variety of machine learning models are used, with the measured precipitation at the site as the benchmark and a variety of precipitation products plus multiple auxiliary factors as input to achieve precipitation fitting. When training the machine learning model, the site data is divided into a training set, a validation set, and a test set according to certain principles and in a certain proportion.

[0035] Note: There are 8 machine learning models used, including RF, CatBoost, KNN, Lasso, DTREE, XGBoost, HGBR and ETREE. Each machine learning model is trained separately. The input auxiliary factors are: longitude, latitude, elevation, sea-land distance and cloud cover. Based on the principle of uniform spatial distribution, the sites are divided into training set, validation set and test set in a ratio of 6:2:2. The training set is used to train model parameters; the validation set is used to adjust model hyperparameters; and the test set is used to evaluate the model fitting effect. In the parameter optimization of each machine learning model, RandomizedSearchCV is used to find the optimal hyperparameters.

[0036] (2) Construction of ensemble learning model: After the parameters of each machine learning model are optimized, the performance of the precipitation products generated by each machine learning model is evaluated according to the evaluation indicators. Multiple machine learning models are selected for different combinations, and an ensemble learning model is constructed using an ensemble method.

[0037] Note: The integration method uses the Stacking method, which can integrate heterogeneous models to achieve the advantages of integrating different models, and then obtain an integrated learning model with comprehensive advantages.

[0038] (3) Generation of precipitation products: The ensemble learning model is used to fit the precipitation. Based on the optimization of the ensemble learning model, precipitation products for the study area are generated.

[0039] Example:

[0040] Taking the daily precipitation data of 198 meteorological stations in the Yangtze River Basin from 2001 to 2023 and 9 hourly precipitation products TRMM, GPM, CMORPH, CHIRPS, ERA5, ERA5-Land, PERSIANN, PERSIANN-CDR, and PERSIANN-CCS as examples, the implementation process of the present invention is described. The hourly precipitation products are accumulated according to the daily precipitation statistical period, that is, (20:00–20:00] points to obtain daily precipitation products. The performance of the 9 precipitation products is evaluated, and finally 4 precipitation products GPM, CMORPH, ERA5, and ERA5-Land are selected for precipitation fusion and analysis. The performance indicators of these 4 precipitation products are shown in Table 1.

[0041] Table 1 Performance evaluation indicators of various raw precipitation products

[0042] This embodiment adopts 8 machine learning models, including RF, CatBoost, KNN, Lasso, DTREE, XGBoost, HGBR and ETREE, and takes the measured precipitation at the station as the benchmark. It takes the above 4 precipitation products GPM, CMORPH, ERA5, ERA5-Land plus multiple auxiliary factors (longitude, latitude, elevation, land-sea distance and cloud cover) as input to achieve precipitation fitting.

[0043] Each machine learning model is trained separately. When training the machine learning model, based on the principle of spatial uniform distribution, the site data is divided into a training set, a validation set, and a test set in a ratio of 6:2:2. The training set is used to train model parameters; the validation set is used to adjust model hyperparameters; and the test set is used to evaluate the model fitting effect. In the parameter optimization of each machine learning model, RandomizedSearchCV is used to find the optimal hyperparameters.

[0044] The performance of the model-generated precipitation products was evaluated using the test set site data. As can be seen from Table 2, RF and CatBoost have excellent overall performance. After training, each machine learning model can achieve relatively ideal results, and can outperform the original precipitation products in terms of correlation, root mean square deviation, and mean absolute relative deviation. In terms of hit rate, some models can be close to or better than the ERA5 series, but at the same time as the high hit rate, the improvement in false alarm rate is not obvious, and few can reach the level of GPM.

[0045] Table 2 Performance evaluation indicators of precipitation products generated by various machine learning models

[0046] In order to explore a better model combination, an integrated learning model is constructed according to the SCORE scores of each model in Table 2. The model combination is shown in Table 3.

[0047] Table 3 Machine learning model combination

[0048] Table 4 shows the performance evaluation indicators of the integrated learning model for generating precipitation products. The integrated learning model ELM4-S has the same model combination as ELM4, but ELM4-S adopts a seasonal training strategy; ELM4-A adopts a partition training strategy to train the integrated learning model in different climate zones; ELM4-SAA adopts both partition and seasonal training strategies.

[0049] It can be seen from Table 4 that the precipitation product generated by ELM4-S has the best performance evaluation index among all ensemble learning models, which is better than the original precipitation product and also better than the precipitation product generated by a single machine learning model.

[0050] Table 4 Performance evaluation indicators of precipitation products generated by ensemble learning models

[0051] Based on the above ideal embodiments of the present invention, the relevant staff can make various changes and modifications without departing from the technical concept of the present invention through the above description. The technical scope of the present invention is not limited to the contents of the specification, and its technical scope must be determined according to the scope of the claims.

Claims

1. A multi-source precipitation product fusion method based on ensemble learning, characterized in that The following steps are involved: S1, data preparation: prepare meteorological station precipitation data and multi-source precipitation products according to the study area, and extract the raster values ​​of precipitation products corresponding to the meteorological station locations; S2, establish multi-dimensional evaluation indicators: take the station precipitation data as the true value, use the correlation coefficient R, mean absolute relative error MARE, root mean square error RMSE, detection rate POD, false alarm rate FAR, critical success index CSI six evaluation indicators to calculate the evaluation values ​​of multi-source precipitation products; S3, comprehensive score calculation: For the same evaluation index, different precipitation products are ranked from best to worst according to the evaluation value performance, and the ranking is used as the score of the precipitation product under the evaluation index. The comprehensive score SCORE of different precipitation products is calculated by the following formula: j , ， In the formula, SCORE j represents the comprehensive score of the j-th type of precipitation product; S kj It represents the score of the j-type precipitation product in the k-th evaluation index, and the value of m is 6; S4, precipitation product selection: based on the comprehensive scores of different precipitation products obtained in step S3, the different precipitation products are sorted in ascending order according to the comprehensive scores, and the top N precipitation products are selected, where the value of N ranges from 3 to 5; S5, integrated learning model construction (1) Using a variety of machine learning models, taking the measured precipitation data at the site as the benchmark, and taking the precipitation product selected in step S4 and several auxiliary factors as input, the machine learning models are trained and optimized respectively to achieve precipitation fitting; (2) The precipitation products generated by the trained and optimized machine learning model are evaluated according to the evaluation indicators to determine the performance of the machine learning model, and several machine learning models are selected for combination to construct an integrated learning model using the stacking method; (3) The ensemble learning model is used to fit the precipitation, and the precipitation product of the study area is generated based on the optimization of the ensemble learning model.

2. The multi-source precipitation product fusion method based on ensemble learning according to claim 1 is characterized by: The multi-source precipitation products in step S1 include satellite precipitation products and reanalysis precipitation products.

3. A multi-source precipitation product fusion method based on ensemble learning, characterized by: The calculation method of the evaluation values ​​of the six evaluation indicators in step S2 is as follows: (1) Correlation coefficient R j , which is used to measure the correlation between precipitation products and station observation precipitation data. The larger the value, the stronger the linear correlation. It is calculated by the following formula: ， In the formula, R j is the correlation coefficient of the j-th type of precipitation product; G i is the observed precipitation value for site i; represents the average value of the observed precipitation at the station, P ij is the j-th type of precipitation product value corresponding to station i, is the average value of the j-th type of precipitation product; n is the number of stations; (2) Root mean square error (RMSE) j , reflects the degree of dispersion of the deviation between the precipitation product and the station observation value. The smaller the root mean square deviation, the smaller the dispersion of the deviation. It is calculated by the following formula: ， In the formula, RMSE j is the root mean square deviation of the j-th type of precipitation product; G i is the observed precipitation value at station i, P ij is the j-th type of precipitation product value corresponding to station i; (3) Mean absolute relative deviation (MARE) j , reflects the average deviation between the precipitation product and the station observation value. The closer it is to 0, the smaller the average deviation is, and the more accurate the precipitation product data is. It is calculated by the following formula: ， In the formula, MARE j is the average absolute relative deviation of the j-th type of precipitation product; G i is the observed precipitation value at station i, P ij is the j-th type of precipitation product value corresponding to station i; (4) Detection rate POD j , reflects the proportion of precipitation days that the precipitation product can correctly reflect, ranging from [0, 1]. The higher the value, the higher the proportion of precipitation product detection of real precipitation events. It is calculated by the following formula: ， In the formula, POD j is the detection rate of the j-th type of precipitation product; H j It represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; M j It indicates the total number of days during the study period when the daily station observation value is less than 0.1 mm, but the corresponding j-type precipitation product value is greater than or equal to 0.1 mm; (5) False Alarm Rate (FAR) j , reflects the proportion of precipitation days that the precipitation product fails to correctly reflect. The value range is [0, 1], and the optimal value is 0. The smaller the value, the lower the error rate of precipitation products in reflecting precipitation events. It is calculated by the following formula: ， In the formula, FAR j is the false alarm rate of the jth type of precipitation product; H j represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; F j It represents the total number of days during the study period when the daily station observation value is greater than or equal to 0.1 mm, but the corresponding j-type precipitation product value is less than 0.1 mm; (6) Critical Success Index (CSI) j , which reflects the ratio of successful precipitation detection by precipitation products to the number of rainy days. Its value range is [0, 1]. The closer it is to 1, the more successful the precipitation product detection is and the fewer false alarms are. ， Where, CSI j is the critical success index of the jth type of precipitation product; H j represents the total number of days during the study period when the daily station observation value and the corresponding j-th type precipitation product value are both greater than or equal to 0.1 mm; F j It represents the total number of days during the study period when the daily station observation value is greater than or equal to 0.1 mm, but the corresponding j-th type precipitation product value is less than 0.1 mm; M j It indicates the total number of days during the study period when the daily station observation value is less than 0.1 mm, but the corresponding j-th precipitation product value is greater than or equal to 0.1 mm.

4. The multi-source precipitation product fusion method based on ensemble learning according to claim 1 is characterized by: The auxiliary factors in step S5 are: longitude, latitude, altitude, distance between land and sea, and cloud cover.

5. The multi-source precipitation product fusion method based on ensemble learning according to claim 1 is characterized by: The machine learning models described in step S5 include RF, CatBoost, KNN, Lasso, DTREE, XGBoost, HGBR and ETREE.

6. The multi-source precipitation product fusion method based on ensemble learning according to claim 1 is characterized by: The training and optimization of the machine learning models described in step S5 is to perform separate training and optimization on each machine learning model.

7. The multi-source precipitation product fusion method based on ensemble learning according to claim 1 is characterized by: The measured precipitation data of the stations in step S5 are divided into a training set, a validation set and a test set in a ratio of 6:2:2 based on the principle of uniform spatial distribution of the stations.

8. The multi-source precipitation product fusion method based on ensemble learning according to claim 1 is characterized by: When the machine learning model described in step S5 is trained and optimized, RandomizedSearchCV is used to find the optimal hyperparameters.

Images