A short-term precipitation forecasting method based on machine learning and OTS

Abstract

This invention discloses a short-term precipitation forecasting method based on machine learning and OTS, belonging to the field of precipitation forecasting technology. Due to the complexity of the physical mechanism of precipitation processes, the uncertainty of the initial field of NWP and the limitation of computing power, the uncertainty of the NWP physical parameterization scheme leads to its precipitation forecasting performance being far lower than that of forecasts of basic meteorological variables such as temperature, geopotential height, humidity, and wind speed. This invention provides a short-term precipitation forecasting method based on machine learning and OTS, including the following steps: training with preprocessed multiple precipitation forecast data to construct two sets of machine learning-based precipitation forecasting models for 1-hour and 24-hour precipitation respectively; training with preprocessed data to construct two sets of OTS-based precipitation forecasting models for 1-hour and 24-hour precipitation respectively; determining the dynamic fusion weights between different precipitation forecast products; and fusing the precipitation forecast models through the dynamic fusion weights to obtain the final precipitation product.

Description

Technical Field

[0001] This invention relates to the field of precipitation forecasting technology, specifically to a short-term precipitation forecasting method based on machine learning and OTS. Background Technology

[0002] Accurate short-term precipitation forecasts are of great significance for disaster prevention and mitigation, and for promoting economic development.

[0003] Numerical weather prediction (NWP) has developed rapidly in recent years and has become an important tool for short-term precipitation forecasting. However, due to the complexity of the physical mechanisms of precipitation processes, precipitation forecasting is one of the most challenging aspects of NWP. The uncertainty of the NWP physical parameterization scheme leads to its precipitation forecasting performance being far lower than that of forecasts for basic meteorological variables such as temperature, geopotential height, humidity, and wind speed. Furthermore, due to the uncertainty of the NWP initial field and computational limitations, NWP results often have certain defects. Therefore, it is essential to employ post-processing methods to correct systematic errors and quantify the uncertainty of NWP.

[0004] NWP post-processing methods have consistently proven effective in improving precipitation forecast accuracy, even when NWP forecast errors are significant. This data-driven approach holds great promise for correcting NWP forecast results, as evidenced by a growing body of research. Early methods, such as the complete forecast method and model output statistics, were based on linear regression to generate NWP post-processing products. In recent years, statistical and machine learning methods have been increasingly applied to precipitation forecast NWP post-processing.

[0005] Among statistical methods, frequency matching (FM) and optimal threat score (OTS) are the two most common correction methods based on target score thresholds. FM reduces systematic errors in forecasts by adjusting the frequency distribution of precipitation forecasts to match the frequency distribution of observed precipitation. Cao Pingping et al. (2018) used FM to correct 12-hour cumulative precipitation forecasts and found that the correction improved the forecasts more significantly for large-scale precipitation, and that the correction threshold for regional statistics was better than that for point-to-point statistics. Yuan Liang et al. (2024) and Dang et al. (2024) respectively used FM to correct the European Centre for Medium-Range Weather Forecasts Integrated Forecasting System (ECMWF-IFS) and the China Meteorological Administration Mesoscale Model (CMA-MESO), effectively improving the performance of NWP forecasts, with particularly significant correction effects on moderate to heavy rainfall forecasts.

[0006] Building upon the FM method, Wu Qishu et al. (2017) proposed the OTS method, which optimizes the threat score (TS) for each precipitation level by adjusting the precipitation thresholds for each level. This method is simple in design and straightforward in its approach, making it easy for frontline forecasters to understand and implement, and it has been widely adopted by meteorological departments across the country. Since Wu Qishu et al. (2017) proposed the OTS method, many researchers have explored and implemented various methods and techniques to improve its forecasting capabilities in operational practice. Wei Guofei et al. (2020), considering the advantages of different models for different precipitation levels, designed an integrated correction technique for OTS using a global model and a regional mesoscale model. The integrated forecast outperformed single-model OTS forecasts and forecaster subjective forecasts in almost all time-lead times. Zhao Ruixia et al. (2022) combined MOS forecasting with the OTS method, which outperformed either MOS or OTS alone. This is because the OTS method, while improving the TS of large-scale precipitation, reduces the accuracy of weather forecasts, while the MOS method can reasonably optimize precipitation distribution, and the TS of each magnitude is significantly improved when OTS correction is performed. Gao Mengzhu et al. (2025) designed a multi-model fusion precipitation correction algorithm that integrates the performance of precipitation forecasts of different magnitudes. They performed OTS precipitation correction on three models, ECMWF-IFS, CMA-MESO, and CMA-SH9, to form fused source data. The results were all better than the models themselves. Then, they fused the three sources and found that the fused forecast could adjust the area of ​​heavy precipitation that was wrong in the single OTS forecast.

[0007] Therefore, the OTS method essentially seeks the optimal threshold for classifying model-predicted precipitation levels, thereby improving the TS of model precipitation forecasts at each level. This technique can greatly improve the TS of precipitation forecasts at each level, especially for large-scale precipitation. Furthermore, the fusion of multi-source forecast data based on the OTS method can further enhance its correction effect.

[0008] Machine learning-based NWP post-processing methods have developed rapidly in recent years, and more and more researchers are beginning to explore the potential of machine learning methods in precipitation forecasting. Rojas-Campos et al. (2023) applied artificial neural networks to NWP precipitation forecast correction, and found that incorporating information from multiple observation stations can improve the performance of most station-corrected forecasts. Generalizing the learning relationships between stations using information from all stations can improve the effectiveness of precipitation forecasts. Zhang et al. (2023b) performed model precipitation correction based on circulation patterns and deep learning models, and the results were better than methods that did not fuse circulation patterns. The results showed that combining large-scale circulation features with local spatiotemporal information is a feasible and effective post-processing method. To further improve the ability to capture large-scale weather information and local-scale spatiotemporal information, Zhang et al. (2025) proposed a deep learning method based on nested convolutional neural networks and long short-term memory networks. Chen Ziwen (2024) and Esquivel-González et al (2025) respectively conducted NWP precipitation forecast corrections based on deep learning models on the National Centers for Environmental Prediction Global Forecast System (NCEP-GFS) and their self-built Weather Research and Forecasting (WRF) system. Their results showed good performance for light to moderate rainfall, but were slightly insufficient for heavy rainfall. Zhu Wengang et al (2024) used a deep forward neural network (DFNN) to design four modeling schemes to correct ECMWF-IFS precipitation forecasts. They then established an integrated model using the multi-model precipitation classification OTS weighted ensemble method. Experiments showed that the precipitation forecasts of the four DFNN schemes were superior to those of ECMWF-IFS, and the precipitation forecast after multi-model ensemble had a certain advantage for heavy rainfall. Wang Yuhong et al. (2025) designed a weighted function combining TS and mean square error as a loss function for U-Net model modeling. The spatial range of the significantly improved forecasting skills decreased with the increase of precipitation intensity, that is, the TS of small precipitation in most areas was improved, while the TS of large precipitation in some areas was also improved.

[0009] Therefore, machine learning methods have advantages over statistical methods in terms of clear and quantitative precipitation, but they often face greater challenges in large-scale precipitation. This is because when using machine learning methods to directly correct precipitation forecasts, the imbalance between positive and negative samples often leads to poor forecast performance for large-scale precipitation, while statistical methods such as OTS perform better in correcting large-scale precipitation. Therefore, balancing the forecast performance of precipitation at both small and large scales is a major challenge. On the other hand, some studies have designed correction models for 1-hour and 24-hour cumulative precipitation, respectively. However, since 24-hour cumulative precipitation is the sum of 1-hour cumulative precipitation, these two tasks are actually a composite task. Balancing the forecast performance between the two is also a major challenge for model precipitation forecast correction. In addition, integrating multiple precipitation forecast products, including global and regional scale models, is also key to improving the model forecast correction capability. Simple arithmetic averaging often fails to achieve the best results. How to balance the weights of different precipitation products determines the effectiveness of the integrated product. Summary of the Invention

[0010] To address the aforementioned problems, this invention aims to provide a short-term precipitation forecasting method based on machine learning and OTS, which can solve the above "three balance" problems. This invention utilizes multi-model forecasting products to construct a precipitation forecasting process.

[0011] The main idea of ​​the technical solution adopted in this invention is to use multi-source numerical weather prediction data to construct a short-term precipitation forecast model based on machine learning-OTS, and generate 24-hour cumulative precipitation forecast products and 24-hour cumulative precipitation forecast products. The modeling scheme is designed to address the "three balances" problem of difficulty in balancing the accuracy of clear and rainy weather and the threat score of heavy precipitation, difficulty in balancing the 1-hour cumulative precipitation forecast and the 24-hour cumulative precipitation forecast, and difficulty in balancing the fusion of multi-source precipitation forecast data. This effectively solves the problem of the "three balances" in precipitation forecasting.

[0012] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A short-term precipitation forecasting method based on machine learning and OTS includes the following steps: Acquire and preprocess various precipitation forecast data; Two machine learning-based precipitation forecasting models were constructed using preprocessed data for training: one for 1-hour precipitation and the other for 24-hour precipitation. The preprocessed data was used for training to construct two OTS-based precipitation forecast models for 1-hour and 24-hour precipitation, respectively. Based on historical precipitation data from other precipitation forecast products, the dynamic fusion weights among different precipitation forecast products are calculated. The final precipitation product is obtained by fusing the above precipitation forecast models through dynamic fusion weights.

[0013] Through the above technical solutions, further: precipitation forecast data includes no less than 3 years of meteorological station precipitation data, ECMWF-IFS precipitation and other meteorological element forecasts, among which ECMWF-IFS precipitation and other meteorological element forecasts include precipitation forecasts and forecasts of meteorological elements such as surface and upper-air temperature, relative humidity, meridional wind, and zonal wind.

[0014] Based on the above technical solution, the preprocessing steps further include: cleaning the acquired data, dividing the precipitation forecast data for the last year into the validation set, and dividing the remaining precipitation forecast data into the training set.

[0015] Based on the above technical solutions, the training steps for the machine learning-based precipitation forecasting model are as follows: The training set data is added to the machine learning model for training. The presence or absence of precipitation is determined by whether the precipitation reaches 0.1 mm. The hyperparameters of the machine learning model are recorded. The validation set data is then added to the trained machine learning model to obtain the precipitation forecast results. The accuracy of the weather forecast is calculated and recorded. Change the machine learning hyperparameters and repeat the above steps. After repeating the process multiple times, compare the results to find the machine learning hyperparameters that achieve the highest accuracy for both sunny and rainy weather. Use this set of hyperparameters to build a machine learning model.

[0016] Based on the above technical solutions, the training steps for the OTS-based precipitation forecasting model are as follows: ECMWF-IFS forecasts and actual precipitation data from weather stations in the training set data were added to the OTS model for training. The threshold for heavy precipitation when the training set TS reaches its maximum value was determined by using a sliding heavy precipitation level threshold. Using the obtained threshold, the validation set ECMWF-IFS is corrected using the following formula: ; in For the previous forecast value, This is the corrected forecast value. is the heavy precipitation threshold, and F is the heavy precipitation correction threshold.

[0017] Based on the above technical solutions, the steps to further determine the dynamic fusion weights among different precipitation forecast products include: The formula for calculating heavy precipitation (TS) is: ; Where NA represents the number of samples where both the actual and forecast data indicate heavy precipitation, NB represents the number of samples where there is no heavy precipitation in the actual data but heavy precipitation is forecast, NC represents the number of samples where there is heavy precipitation in the actual data but no heavy precipitation is forecast, and ND represents the number of samples where both the actual and forecast data indicate no heavy precipitation. ; in For the first The correction forecast fusion weight of precipitation forecast products For the first The revised forecast for heavy precipitation (TS) of a precipitation forecast product.

[0018] Based on the above technical solutions, the steps to obtain the final precipitation product through the fusion strategy include: (1) Determine whether there will be heavy precipitation in the 24-hour heavy precipitation forecast based on OTS; (2) Determine whether there is heavy precipitation in the 1-hour heavy precipitation forecast based on OTS, and make corrections based on the results of step (1). The part with heavy precipitation is directly generated into the final result. (3) The 1-hour weather forecast based on machine learning is corrected for the part of the result in step (2) where there is no heavy precipitation; (4) Correct the 24-hour weather forecast based on machine learning based on the results of step (3); (5) Combine the results of step (2) which contain heavy precipitation with the results of step (4) to obtain the final result.

[0019] The beneficial effects of this invention are: 1. This implementation case aims to provide a short-term precipitation forecasting method to solve the "three balance" problems existing in the prior art: (1) the accuracy of clear / rainy weather forecasts and the TS (True Surveillance) of heavy precipitation are difficult to balance; (2) the 1-hour precipitation forecast and the 24-hour cumulative precipitation forecast are difficult to balance; and (3) the fusion weight of multi-source NWP (Non-Wide Precipitation Forecast) data is difficult to balance. This invention effectively solves the above "three balance" problems by introducing a scheme combining machine learning and OTS (True Surveillance), and generates more accurate and reliable short-term precipitation forecast products.

[0020] 2. This invention proposes a dynamic fusion weighting scheme based on a symmetrical training period, which can effectively fuse different precipitation forecast products in precipitation cases of different seasons and weather types, and solve the problem of difficulty in balancing the fusion weights between different precipitation forecast products. Attached Figure Description

[0021] Figure 1 This is a schematic diagram illustrating the selection of the data time period for this invention; since the validation set year is 2024, the forecast for June 1, 2025 should be calculated using data from the period from May 2 to July 1, 2024. Figure 2 This is a schematic diagram of the precipitation forecasting process of this invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] The inventors' research found that machine learning methods have advantages over statistical methods in terms of clear and quantitative precipitation, but they often face greater challenges in large-scale precipitation. This is because when using machine learning methods to directly correct precipitation forecasts, the imbalance between positive and negative samples often leads to poor forecast performance for large-scale precipitation, while statistical methods such as OTS perform better in correcting large-scale precipitation. Therefore, balancing the forecast performance of precipitation at both small and large scales is a major challenge. On the other hand, some studies have designed correction models for 1-hour cumulative precipitation and 24-hour cumulative precipitation, respectively. However, since 24-hour cumulative precipitation is the sum of 1-hour cumulative precipitation, these two tasks are actually a composite task, and balancing the forecast performance between them is another major challenge in model precipitation forecast correction. In addition, integrating multiple precipitation forecast products, including global and regional scale models, is also key to improving the model forecast correction capability. Simple arithmetic averaging often fails to achieve the best results. How to balance the weights of different precipitation products determines the effectiveness of the integrated product.

[0024] Based on the above findings, this application proposes a short-term precipitation forecasting method based on machine learning and OTS. Utilizing multi-source numerical weather prediction data, a machine learning-OTS-based short-term precipitation forecasting model is constructed to generate 24-hour cumulative precipitation forecast products and 24-hour cumulative precipitation forecast products. A modeling scheme is designed to address the "three balances" problem in precipitation forecasting: the difficulty in balancing the accuracy of clear / rainy weather forecasts and the heavy precipitation threat score, the difficulty in balancing the 1-hour cumulative precipitation forecast and the 24-hour cumulative precipitation forecast, and the difficulty in merging multi-source precipitation forecast data. This effectively solves the "three balances" problem in precipitation forecasting.

[0025] Example 1: See Figures 1-2This application discloses a short-term precipitation forecasting method based on machine learning and OTS. It constructs a precipitation forecasting process using multi-model forecast products and solves the problem of "three balances" in precipitation forecasting, namely, the balance of precipitation forecasting performance on sunny and rainy and heavy precipitation, the balance of 1-hour cumulative precipitation and 24-hour cumulative precipitation, and the balance of fusion weights of different precipitation products.

[0026] It includes the following steps: Step 1: Obtain precipitation forecast products in multiple models; Step 2: Construct two precipitation forecast models based on machine learning: one for 1-hour precipitation and one for 24-hour precipitation. Step 3: Construct two precipitation forecast models based on OTS for 1-hour and 24-hour precipitation; Step 4: Determine the multi-model, multi-task precipitation forecast fusion scheme and the dynamic fusion weights between different precipitation products; Step 5: Add the real-time data from the model forecast to the model trained in Steps 2–4, and then obtain the final precipitation product through a fusion strategy.

[0027] In step 1 of Example 2, obtaining precipitation forecast products from multiple models includes the following steps: (1) Obtain no less than 3 years of meteorological station precipitation data, ECMWF-IFS precipitation and other meteorological element forecasts and other precipitation forecast products from the meteorological big data platform "Tianqing". Among them, ECMWF-IFS precipitation and other meteorological element forecasts include precipitation forecasts and surface and upper-air (1000 hPa, 925 hPa, 850 hPa, 700 hPa and 500 hPa) temperature, relative humidity, meridional wind, zonal wind and other meteorological element forecasts. Other precipitation forecast products include but are not limited to CMA-GFS, CMA-MESO, CMA-GD, CMA-SH9, SCMOC and other precipitation forecast products. (2) Clean the acquired data. First, set a threshold to clean outliers. Then, use the linear interpolation method to resample the forecast data to 1 h in time resolution. Then, align the real and forecast data, remove the samples of the default real data, and replace the samples of the default forecast data with the forecast data from the previous time period. (3) Allocate the data from the last year to the validation set and the remaining data to the training set. For example, if the data obtained comes from 2020–2024, allocate the data from 2020–2023 to the training set and allocate the data from 2024 to the validation set.

[0028] Step 2 involves constructing two machine learning-based precipitation forecasting models: one for 1-hour precipitation and one for 24-hour precipitation.

[0029] (1) Add the training set data mentioned in step 1 to the machine learning model for training. Determine whether there is precipitation by whether the precipitation reaches 0.1 mm. Record the set hyperparameters of the machine learning model. Then add the validation set data to the trained machine learning model to obtain the precipitation forecast results. Calculate and record the accuracy of the weather forecast. The formula for the accuracy of the weather forecast is: ; in Rainfall amount, in mm. This indicates whether there is precipitation; precipitation less than 0.1 mm is considered no precipitation, and precipitation greater than or equal to 0.1 mm is considered precipitation. For the accuracy of weather forecasts, NA represents the number of samples where both the actual and forecast forecasts indicate precipitation, NB represents the number of samples where the actual forecast indicates no precipitation but the forecast indicates precipitation, NC represents the number of samples where the actual forecast indicates precipitation but the forecast indicates no precipitation, and ND represents the number of samples where both the actual and forecast forecasts indicate no precipitation, as shown in Table 1. (2) Change the machine learning hyperparameters, repeat step (1), and after repeated multiple times, compare and find the machine learning hyperparameters with the highest accuracy of weather forecasting. Use this set of hyperparameters to build a machine learning model. (3) The machine learning model includes two sets of 1-h precipitation and 24-h precipitation, that is, two independent machine learning models, both of which determine hyperparameter modeling through steps (1) and (2).

[0030] Table 1 ; Step 3 involves constructing two OTS-based precipitation forecasting models: one for 1-hour precipitation and one for 24-hour precipitation. (1) Referring to GB / T28592—2012 "Precipitation Grades" and "National Intelligent Forecasting Technology Method Exchange Competition Test Scheme", heavy precipitation is defined as 24-hour cumulative precipitation of 50 mm and 1-hour cumulative precipitation of 20 mm respectively; (2) The ECMWF-IFS forecast and actual precipitation data from meteorological stations in the training set data in step 1 are added to the OTS model for training. The threshold for heavy precipitation when the training set TS reaches its maximum value is determined by sliding the heavy precipitation level threshold. (3) Using the threshold from step (2), correct the validation set ECMWF-IFS. The correction formula is as follows: ; in For the previous forecast value, This is the corrected forecast value. is the heavy precipitation threshold (50 mm for 24 h cumulative precipitation, 20 mm for 1 h cumulative precipitation), and F is the heavy precipitation correction threshold, which is determined by step (2).

[0031] Step 4 determines the multi-model, multi-task precipitation forecast fusion scheme and the dynamic fusion weights among different precipitation products, including: (1) Calculate the heavy precipitation TS from the corrected forecast obtained in step 3 using the following formula: ; Where NA represents the number of samples where both the actual and forecast conditions indicate heavy precipitation, NB represents the number of samples where there is no heavy precipitation in the actual situation but heavy precipitation is forecast, NC represents the number of samples where there is heavy precipitation in the actual situation but no heavy precipitation is forecast, and ND represents the number of samples where both the actual and forecast conditions indicate no heavy precipitation, as shown in Table 2. Table 2 ; (2) The fusion weight of each precipitation correction forecast is calculated using the heavy precipitation TS, and the formula is as follows: ; in For the first The correction forecast fusion weight of precipitation forecast products For the first The revised forecast for heavy precipitation (TS) of a precipitation forecast product.

[0032] To achieve dynamic changes in fusion weights and improve the accuracy of fusion forecasts, a symmetrical training period method is adopted. For a forecast of a specific month and day, the heavy precipitation TS is calculated using data from the 30 days before and after that date in the validation set. This data is then used to calculate the fusion forecast weight for that month and day. For example, if the validation set year is 2024, then the forecast for June 1, 2025, should be calculated using data from May 2 to July 1, 2024. Figure 1 .

[0033] In step 5, the real-time data from the model forecast is added to the model trained in steps 2–4, and then a fusion strategy is used to obtain the final precipitation product, including: (1) Determine whether there will be heavy precipitation in the 24-hour heavy precipitation forecast based on OTS; (2) Determine whether there is heavy precipitation in the 1-hour heavy precipitation forecast based on OTS, and make corrections based on the results of step (1). The part with heavy precipitation is directly generated into the final result. (3) The 1-hour weather forecast based on machine learning is corrected for the part of the result in step (2) where there is no heavy precipitation; (4) Correct the 24-hour weather forecast based on machine learning based on the results of step (3); (5) Combine the results of step (2) which contain the heavy precipitation part with the results of step (4) to obtain the final result; The entire forecasting process and strategy are as follows: Figure 2 .

[0034] Example 3: This example selects historical and real-time data from four numerical weather prediction (NWP) models: ECMWF-IFS, CMA-GFS, CMA-GD, and CMA-SH9. It also collects hourly precipitation observation data from 1912 surface meteorological stations in Hunan Province from 2020 to 2023. This implementation uses the LightGBM algorithm as a machine learning method to capture the nonlinear mapping relationship of future precipitation from multi-source NWP features.

[0035] To verify the effectiveness of the methodology presented in this case study, an independent test period from April to September 2024 was selected for forecast verification and comparison with the NWP model forecast. Evaluation indicators included: weather accuracy and heavy precipitation TS score (thresholds were 1-hour cumulative precipitation ≥ 20 mm and 24-hour cumulative precipitation ≥ 50 mm).

[0036] This invention and the evaluation of product results in various modes ; The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A short-term precipitation forecasting method based on machine learning and OTS, characterized in that, Includes the following steps: Step 1: Acquire and preprocess various precipitation forecast data; Step 2: Use the preprocessed data for training to build two precipitation forecast models based on machine learning, one for 1-hour precipitation and the other for 24-hour precipitation. Step 3: Use the preprocessed data for training to build two OTS-based precipitation forecast models for 1-hour and 24-hour precipitation, respectively. Step 4: Calculate the dynamic fusion weights between different precipitation forecast products based on historical precipitation data from other precipitation forecast products; This includes: The formula for calculating heavy precipitation TS is: ； Where NA represents the number of samples where both the actual and forecast data indicate heavy precipitation, NB represents the number of samples where there is no heavy precipitation in the actual data but heavy precipitation is forecast, NC represents the number of samples where there is heavy precipitation in the actual data but no heavy precipitation is forecast, and ND represents the number of samples where both the actual and forecast data indicate no heavy precipitation. Using the symmetric training period method, for a forecast of a specific month and day, the TS (Several Time Tolerance) of heavy precipitation is calculated using data from the 30 days before and after that day in the validation set. The fusion forecast weights for that month and day are then calculated. The formula for calculating the fusion weights of each precipitation correction forecast using the heavy precipitation TS is as follows: ； in For the first The correction forecast fusion weight of precipitation forecast products For the first The revised forecast for heavy precipitation (TS) of a precipitation forecast product; Step 5: The above precipitation forecast models are fused using dynamic fusion weights to obtain the final precipitation product; including: The real-time data from the model forecasts are added to the model trained in steps 2–4, and then a fusion strategy is used to obtain the final precipitation product, including: (1) Determine whether there will be heavy precipitation in the 24-hour heavy precipitation forecast based on OTS; (2) Determine whether there is heavy precipitation in the 1-hour heavy precipitation forecast based on OTS, and make corrections based on the results of step (1). The part with heavy precipitation is directly generated into the final result. (3) The 1-hour weather forecast based on machine learning is corrected for the part of the result in step (2) where there is no heavy precipitation; (4) Correct the 24-hour weather forecast based on machine learning based on the results of step (3); (5) Combine the results of step (2) which contain heavy precipitation with the results of step (4) to obtain the final result.

2. The short-term precipitation forecasting method based on machine learning and OTS according to claim 1, characterized in that: Precipitation forecast data includes at least 3 years of actual precipitation data from meteorological stations, ECMWF-IFS precipitation forecasts, and other meteorological element forecasts. Among them, ECMWF-IFS precipitation forecasts and other meteorological element forecasts include precipitation forecasts and forecasts of surface and upper-air temperatures, relative humidity, meridional winds, and zonal winds.

3. The short-term precipitation forecasting method based on machine learning and OTS according to claim 2, characterized in that, The preprocessing steps include: cleaning the acquired data, allocating the precipitation forecast data for the last year to the validation set, and allocating the remaining precipitation forecast data to the training set.

4. The short-term precipitation forecasting method based on machine learning and OTS according to claim 3, characterized in that, The training steps for a machine learning-based precipitation forecasting model are as follows: The training set data is added to the machine learning model for training. The presence or absence of precipitation is determined by whether the precipitation reaches 0.1 mm. The hyperparameters of the machine learning model are recorded. The validation set data is then added to the trained machine learning model to obtain the precipitation forecast results. The accuracy of the weather forecast is calculated and recorded. Change the machine learning hyperparameters and repeat the above steps. After repeating the process multiple times, compare the results to find the machine learning hyperparameters that achieve the highest accuracy for both sunny and rainy weather. Use this set of hyperparameters to build a machine learning model.

5. The short-term precipitation forecasting method based on machine learning and OTS according to claim 4, characterized in that, The training steps for the OTS-based precipitation forecasting model are as follows: ECMWF-IFS forecasts and actual precipitation data from weather stations in the training set data were added to the OTS model for training. The threshold for heavy precipitation when the training set TS reaches its maximum value was determined by using a sliding heavy precipitation level threshold. Using the obtained threshold, the validation set ECMWF-IFS is corrected using the following formula: ； in For the previous forecast value, This is the corrected forecast value. is the heavy precipitation threshold, and F is the heavy precipitation correction threshold.

Images