Short and temporary rainfall forecasting method fusing multi-source meteorological data
Patent Information
- Application Number
- CN202511710064.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-20
- Publication Date
- 2026-02-10
Smart Images
Figure CN121502223A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of weather forecasting, in particular to a short-term rainfall forecasting method based on global navigation satellite system (GNSS) and ground meteorological observation data, and more particularly to a method for rainfall prediction using machine learning technology. BACKGROUND
[0002] Although water vapor accounts for less than 5% in the atmosphere, it plays a key role in the formation of clouds, rainfall, and snow weather processes, and is an important factor in global energy exchange and climate change. Abnormal accumulation of water vapor often leads to extreme rainstorm events, which in turn cause basin floods, landslides, urban waterlogging and other disasters, and seriously threaten social sustainable development. Accurate and timely short-term rainfall forecasting is crucial for disaster warning and emergency response.
[0003] Global Navigation Satellite System (GNSS) technology for retrieving atmospheric precipitable water vapor (PWV) has the advantages of high precision, all-weather, high time resolution, and has become an important means of rainfall monitoring and short-term forecasting. Existing research usually uses GNSS-retrieved PWV and its rate of change to establish a rainfall prediction model.
[0004] However, the existing technology has the following disadvantages: Rainfall is affected by many meteorological factors, and only setting thresholds for PWV and its rate of change to determine rainfall has limited prediction accuracy; existing machine learning-based rainfall models lack interpretable analysis of input features and prediction results; most models do not introduce the rate of change and anomaly information of meteorological parameters, making it difficult for the model to capture the transient and abnormal state characteristics of the weather system. SUMMARY
[0005] The purpose of the present application is to provide a short-term rainfall forecasting method that integrates multi-source meteorological data, to solve the problems of low rainfall prediction accuracy, low parameter tuning efficiency, poor model interpretability, and insufficient feature utilization in the prior art.
[0006] To achieve the above-mentioned purpose, the present application provides a short-term rainfall forecasting method that integrates multi-source meteorological data, comprising the following steps: S1, data acquisition: meteorological parameters and GNSS observation data are obtained by meteorological stations and GNSS stations respectively, and sites with a horizontal distance of not more than 5 kilometers and a height difference of not more than 100 meters are selected for data matching, referred to as co-located stations. Meteorological parameters include: temperature, relative humidity, air pressure.
[0007] S2. Water Vapor Retrieval: The acquired GNSS observation data is processed using the existing Gamit 10.71 software to obtain the zenith total delay (ZTD). The zenith hydrostatic delay (ZHD) is calculated based on meteorological parameters such as air pressure, and then subtracted from the ZTD to obtain the zenith wet delay (ZWD). Finally, a conversion factor is calculated using a known weighted average temperature model for the North China region. This conversion factor is then used to convert the ZWD to PWV, thus reflecting the overall water vapor content in the atmosphere.
[0008] S3. Feature Construction: Calculate the rate of change and anomaly value of each parameter, including the rate of change parameters: temperature rate of change, relative humidity rate of change, air pressure rate of change, and PWV rate of change; and the anomaly parameters: temperature anomaly, relative humidity anomaly, air pressure anomaly, and PWV anomaly. Combine the PWV and meteorological parameters to form a total of 12 features, which are used as the input parameters of the model.
[0009] S4. Model Construction: The GA-XGBoost-SHAP model is constructed by combining the Genetic Algorithm (GA), the Extreme Gradient Boosting (XGBoost) method, and the SHAP (Shapley Additive exPlanations) analysis method.
[0010] S5. Sample Balancing: Sample balancing is performed for each GNSS station. All rainfall events are extracted from the station, and the same number of samples as rainfall samples are randomly drawn from non-rainfall events. Non-rainfall samples are evenly distributed throughout the observation period to avoid time concentration, ensuring the representativeness of the sample distribution and the stability of model training.
[0011] S6. Results Output and Evaluation: The GA-XGBoost-SHAP model is used to output rainfall forecast results, and the overall performance is evaluated based on indicators such as Threat Score, accuracy, false alarm rate, and missed alarm rate. Simultaneously, SHAP values are used to analyze the importance of each input parameter (PWV, meteorological parameters and their rates of change, anomalies) to the forecast results.
[0012] Furthermore, as described in S2: the tropospheric wet delay is converted to PWV, and the formula is:
[0013] in, The conversion factor is calculated using the following formula:
[0014] in, For conversion factors, The density of liquid water, Let be the gas constant of water vapor. and It is a constant. This is the weighted average temperature.
[0015] Furthermore, as described in S3: calculate the rate of change and anomaly of each parameter, using the following formula;
[0016]
[0017] in, Represents the individual values and average values of each parameter. The difference, i.e., the parameter anomaly; This represents the increment of each parameter over time. The rate of change within the parameter; Representing the One observation value.
[0018] Furthermore, as described in S4: GA is an intelligent optimization algorithm based on natural selection and genetic mechanisms, capable of globally searching for optimal solutions in complex multidimensional spaces; XGBoost is an efficient machine learning algorithm based on Gradient Boosting Decision Tree (GBDT). GA is used to automatically optimize the hyperparameters of XGBoost; XGBoost is used to build a rainfall classification model; SHAP is used to calculate the importance of each input feature to the model output, enabling interpretable analysis.
[0019] Compared with related technologies, the short-term precipitation forecasting method that integrates multi-source meteorological data provided by this invention has the following beneficial results: This invention employs a genetic algorithm to achieve global automatic optimization of XGBoost hyperparameters. Compared to traditional grid search methods, it shortens parameter tuning time and can escape local optima to find better parameter combinations, significantly improving modeling efficiency and forecast timeliness. Experimental results show a prediction accuracy exceeding 90% and a TS score above 0.8, significantly outperforming traditional threshold-based methods and single machine learning models. Furthermore, the SHAP method is introduced to analyze the importance of various features in rainfall prediction, revealing the physical meaning of key features such as PWV change rate, relative humidity anomaly, and temperature change rate, providing meteorologists with reliable decision-making support and enhancing the model's acceptance in operational applications. Attached Figure Description
[0020] Figure 1 This is a schematic diagram of the XGBoost algorithm framework. Figure 2This is a flowchart of the GA-XGBoost-SHAP short-term precipitation forecasting method of the present invention; Figure 3 This is a schematic diagram of the SHAP feature importance analysis results; Detailed Implementation
[0021] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments. Regarding the "23.7" Hebei torrential rain event, this embodiment takes 52 pairs of GNSS and meteorological stations located together in Hebei Province as the research object, referring to... Figure 2 The method flow shown below, with specific steps as follows: S1. Data Acquisition: Meteorological stations and GNSS stations are two types of observation stations, geographically independent and non-overlapping. Stations with a horizontal distance of no more than 5 kilometers and an elevation difference of no more than 100 meters were selected as co-located stations, resulting in 52 pairs of co-located stations. The observation period was from July to August 2023, totaling 60 days of observation data. GNSS stations acquired raw observation data at 30-second intervals, which were then resampled to 1-hour intervals to match meteorological data. Meteorological stations observed temperature, relative humidity, air pressure, and rainfall at 1-hour intervals.
[0022] S2 and PWV Inversion: The acquired GNSS observation data were processed using the existing open-source Gamit 10.71 software to obtain the ZTD. The ZHD was calculated based on the existing Saastamoinen model and subtracted from the ZTD to obtain the zenith wet delay ZWD. Finally, the conversion coefficient was calculated using the known weighted average temperature model for the North China region, and the ZWD was converted to PWV based on the conversion coefficient.
[0023] S3. Multidimensional Feature Construction: To demonstrate the beneficial results of this invention, five different parameter combination schemes were set up, with each scheme based on GA-RF serving as a comparison. All models are shown in Table 1 below, where C represents meteorological parameters (temperature, relative humidity, air pressure), P represents PWV, A represents the anomaly value of meteorological parameters and PWV, and R represents the rate of change of meteorological parameters and PWV. To ensure the stability and reliability of the model, ten-fold cross-validation was used to determine the optimal model based on accuracy indicators.
[0024] Table 1: Qualitative Prediction Model ; S4. GA-XGBoost-SHAP Model Construction: Construct the GA-XGBoost-SHAP model, referring to... Figure 1 The XGBoost framework shown employs a genetic algorithm to automatically optimize XGBoost hyperparameters. Specific settings are shown in Tables 2 and 3.
[0025] Table 2: Genetic Algorithm Parameters ; Table 3: XGBoost Hyperparameter Settings ; The XGBoost model is trained using this hyperparameter configuration, with a 12-dimensional feature vector as input and a binary classification result of rainfall or no rainfall as output.
[0026] S5. Sample Balancing: To address the issue of sparse rainfall samples, a station-level balanced sampling strategy is adopted: all rainfall samples are extracted, and an equal number of samples are randomly drawn from the non-rainfall samples at each station, ensuring that the drawn non-rainfall samples are evenly distributed throughout the observation period and avoiding time concentration. Finally, a balanced training set with a rainfall to non-rainfall ratio of 1:1 is constructed.
[0027] S6. Model Validation and Performance Evaluation: As shown in Table 4, the prediction models generally performed well. Among them, the accuracy of the five prediction models based on GA-XGBoost was greater than 92.80%, the false positive rate was less than 13%, and the missed positive rate was less than 7.2%. Compared with the performance of the five prediction models based on GA-RF, the GA-XGBoost method showed a significant improvement in accuracy and a significant reduction in false positive and missed positive rates. The results indicate that the GA-XGBoost model can accurately reflect the influence of PWV and meteorological parameters on precipitation forecasts.
[0028] Table 4: Overall Performance Indicators of Each Model ; Combined with Table 4 Figure 3 After incorporating the rate of change and anomaly values of various parameters, the false alarm rates of the other four models decreased compared to the GXB_PC model, which had a false alarm rate of 12.93%. Among them, GXB_PAR had the lowest false alarm rate at 11.20%. This is likely mainly because the rate of change of air pressure, PWV anomaly value, and relative humidity anomaly are used as... Figure 3 The three most important factors affecting rainfall forecasting, as shown, can significantly impact the false alarm rate. The results indicate that the rate of change and anomaly values effectively extract information about parameter variations, helping the GA-XGBoost model reduce misclassifications of non-rainfall events and improve model stability. Furthermore, while the GXB_PAR model has the lowest false alarm rate, its accuracy is the lowest at 92.88%, and its false negative rate is the highest at 7.12%. This is because temperature and relative humidity are relatively important factors in rainfall forecasting, reflecting that the GXB_PAR model reduces the input meteorological parameters. It sacrifices the ability to capture true rainfall events, weakening the model's ability to identify actual rainfall.
[0029] Considering multiple models, the method provided in this invention achieves the best Ts score of 0.84, demonstrating the best predictive performance. The results show that fusing multidimensional features (original meteorological parameters, PWV, anomaly, and parameter change rate) into a feature set as input parameters can more comprehensively describe the rainfall triggering mechanism, thereby improving the reliability of short-term rainfall forecasts.
Claims
1. A method for short-term precipitation forecasting that integrates multi-source meteorological data, characterized in that, Includes the following steps: S1. Data Acquisition: Meteorological parameters such as temperature, relative humidity and air pressure are acquired through ground meteorological stations, raw observation data are acquired through GNSS continuously operating reference stations, and the observation data are unified to a one-hour time resolution. S2. Water vapor inversion: The total zenith delay (ZTD) is obtained by solving the raw GNSS observation data. The static zenith delay (ZHD) is calculated based on meteorological parameters. The wet zenith delay (ZWD) is obtained by subtracting ZHD from ZTD. The ZWD is then converted into atmospheric precipitable water volume (PWV) by a weighted average temperature model. S3. Feature Construction: Calculate the rate of change and anomaly value of the meteorological parameters and the PWV. The rate of change includes the rate of change of temperature, the rate of change of relative humidity, the rate of change of air pressure, and the rate of change of PWV. The anomaly value includes the temperature anomaly, the relative humidity anomaly, the air pressure anomaly, and the PWV anomaly. S4. Model Construction: Set the parameters of the genetic algorithm, optimize the hyperparameters of the XGBoost model by optimizing the extreme gradients through the genetic algorithm, and introduce the SHAP value to analyze its importance, and establish a rainfall forecast model; S5. Sample Balance: For each co-located station, all rainfall samples are extracted, and an equal number of samples are randomly drawn from non-rainfall events to ensure that non-rainfall samples are evenly distributed during the observation period, so as to form a balanced training set with a rainfall to non-rainfall ratio of 1:
1. S6. Model Training and Result Output: The rainfall prediction model is trained using the PWV, meteorological parameters, rate of change, and anomaly value as input feature sets; the optimized model is used to train and predict the input feature set, outputting rainfall or non-rainfall results, and the importance of each feature to the prediction results is analyzed based on the SHAP method.
2. The short-term precipitation forecasting method based on multi-source meteorological data according to claim 1, characterized in that, In S1, the conditions for a meteorological station and a GNSS station to be co-located are: the horizontal distance does not exceed 5 kilometers and the elevation difference does not exceed 100 meters.
3. The short-term precipitation forecasting method based on multi-source meteorological data according to claim 1, characterized in that, The water vapor inversion in S2 includes converting the tropospheric wet delay to PWV, and the formula is as follows: in, The conversion factor is calculated using the following formula: ,in, The density of liquid water, Let be the gas constant of water vapor. and It is a constant. This is the weighted average temperature.
4. The short-term precipitation forecasting method based on multi-source meteorological data according to claim 1, characterized in that, In S3, the input features include the raw values, rates of change, and anomalies of meteorological parameters and PWV, totaling 12 dimensions. The formulas for calculating the rates of change and anomalies of each parameter are as follows: , ,in, Represents the individual values and average values of each parameter. The difference, i.e., the parameter anomaly; This represents the increment of each parameter over time. The rate of change within the parameter; Representing the One observation value.
5. The short-term precipitation forecasting method based on multi-source meteorological data according to claim 1, characterized in that, In step S4, the genetic algorithm parameters include population size, number of generations, crossover probability, and mutation probability; the genetic algorithm is used to optimize the hyperparameters of the XGBoost model, including the maximum tree depth, learning rate, sample sampling ratio, and column sampling ratio.
6. The short-term precipitation forecasting method based on multi-source meteorological data according to claim 1, characterized in that, In S5, the non-rainfall samples are evenly distributed throughout the observation period to avoid time concentration, and finally a balanced training set with a ratio of 1:1 between rainfall and non-rainfall samples is constructed.
7. The short-term precipitation forecasting method based on multi-source meteorological data according to claim 1, characterized in that, In step S6, ten-fold cross-validation is used to determine the optimal combination of model parameters; and the TreeExplainer algorithm of SHAP is used to calculate the SHAP value of each input feature to quantify the importance of each feature to the model prediction result.