A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy

Abstract

The present invention discloses a rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy, which is characterized by the following steps: Step 1, extract historical effective observation values and calculate the reference atmospheric water vapor content; Step 2, construct and train a rainfall event prediction model based on the LSTM machine learning method; Step 3, construct and train a rainfall amount prediction model using a fusion machine learning method based on SVR and GMDH, and optimize the model parameters and their combined weights using the HS algorithm; Step 4, multiply the rainfall time prediction result obtained in Step 2 by the rainfall amount result predicted in Step 3 to obtain the final effective rainfall amount result. This method overcomes the problems of low prediction success rate and high error rate of traditional rainfall event prediction algorithms due to ignoring time correlation.

Description

Technical Field

[0001] The present invention relates to the fields of machine learning and global satellite navigation meteorology, and specifically to a rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy. Background Art

[0002] Accurately predicting rainfall is crucial in various weather events. Whether it is light rain or heavy rain, the amount of rainfall is directly related to people's daily life and work arrangements. Slight rainfall, although it may not cause serious disasters like heavy rainstorms, can still affect aspects such as travel, transportation, and agricultural production. Especially in some specific seasonal rainfall, accurately predicting rainfall can also help people make preparations in advance and avoid unnecessary troubles. Urban waterlogging and floods caused by heavy rain or rainstorms may have a significant impact on urban safety and people's lives and property. Therefore, accurately predicting rainfall not only helps the normal operation of daily life but also provides strong data support for disaster warning and flood control work.

[0003] Currently, there are mainly five types of methods for real-time rainfall prediction. The first type of method is the NWP (Numerical Weather Prediction) model. This type of method is based on the principles of atmospheric dynamics and physics, uses computer to simulate various physical processes of the atmosphere, and then predicts the rainfall amount. These models are suitable for medium- and long-term weather forecasts and can provide precipitation predictions over a large area. Their advantage lies in comprehensively and systematically simulating atmospheric processes and being able to provide relatively accurate precipitation trends globally or regionally. However, the disadvantages of numerical weather prediction models are also very obvious. In high-resolution regional forecasts, the accuracy will be limited by the initial conditions and the simulation of physical processes. For accurate predictions of short-term or local precipitation, the model may show deviations, and the prediction effect is not as good as that of methods with short time scales. The second type is the statistical regression model. This method analyzes the relationship between historical data and meteorological variables and establishes a mathematical model for precipitation prediction. Its advantage is high computational efficiency and suitability for simple rainfall predictions, especially when historical data is relatively sufficient, the model can quickly calculate the prediction results. However, the limitation of the statistical regression model is that it cannot capture the complex non-linear relationships in the atmosphere. Especially when atmospheric conditions change drastically, its prediction ability will be limited. The statistical regression model depends more on the quality and quantity of historical data, and the prediction effect will be greatly reduced under abnormal or complex weather conditions. The third type is machine learning and deep learning methods. This method has been widely used in rainfall prediction in recent years, especially in dealing with complex non-linear weather patterns. Through the training of a large amount of meteorological data, machine learning and deep learning models can self-adjust and identify potential patterns in the data, thereby improving the accuracy of precipitation prediction. For example, methods such as neural networks, support vector machines, and random forests have strong flexibility and generalization ability in dealing with large-scale meteorological data. Although these methods show high effectiveness under complex meteorological conditions, overfitting may occur during the model training process, which may lead to inaccurate predictions of new data. In addition, the performance of different machine learning methods in rainfall prediction may vary, and there may be a bias in predicting heavy rain or light rain. The fourth type is the precipitation radar observation method. This method uses meteorological radar to monitor precipitation in the atmosphere in real time, can accurately track the spatial distribution and intensity of precipitation, and is particularly suitable for the prediction of short-term heavy precipitation and local rainfall (such as thunderstorms and showers). The advantage of radar technology is high spatial resolution, which can provide detailed information such as precipitation intensity and range, and can be updated in real time. However, radar technology also has limitations. It can only cover the range that the radar signal can reach and is greatly affected by weather conditions (such as clouds, winds, etc.). Therefore, in some cases, it may lead to missed or false alarms of precipitation. For the prediction of large-scale or complex weather systems, the effect of radar technology is not as good as other methods.The fifth category is satellite remote sensing technology, which uses satellite sensors to monitor global meteorological phenomena, especially in areas where data cannot be obtained through ground equipment (such as oceans and remote areas). Satellites can help improve precipitation forecasts by monitoring meteorological factors such as cloud changes and aerosols. This method provides valuable precipitation monitoring data worldwide, especially in areas that cannot be covered by traditional ground stations. However, since this type of satellite is a low-orbit satellite with a long repetition period interval, the temporal resolution is low and it is easily affected by environmental interference, resulting in reduced prediction accuracy.

[0004] Therefore, studying accurate and efficient real-time rainfall event and rainfall prediction methods to provide strong support for the meteorological department is a hot and difficult issue that needs to be solved urgently. Summary of the invention

[0005] The purpose of the present invention is to overcome the deficiencies of the prior art and propose a rainfall prediction method based on rainfall event constraints and hybrid machine learning strategies. This method overcomes the problems of low prediction success rate and high error rate caused by ignoring time correlation in traditional rainfall event prediction algorithms, and solves the problems of large rainfall prediction errors in rainless conditions. At the same time, the rainfall prediction algorithm using hybrid machine learning strategies eliminates the problems of low rainfall prediction accuracy caused by differences in machine learning algorithm characteristics. This method has high practical application value and is suitable for real-time precipitation monitoring of different rainfall types.

[0006] In order to achieve the above object, the technical solution specifically adopted by the present invention is as follows:

[0007] A rainfall prediction method based on rainfall event constraints and hybrid machine learning strategy includes the following steps:

[0008] Step 1: Extract effective PWV and related elements involved in prediction, mainly including PWV, corresponding rainfall, temperature, air pressure and related time series.

[0009] First, the original observations of GPS are received by a geodetic receiver, mainly including observation files and navigation files. At the same time, corresponding precise ephemeris, precise clock offset and other precise point positioning related error correction products are downloaded in real time from the IGS (International GNSS Service) official website. The above files can all be obtained from official channels and are common knowledge in the field. Secondly, the relevant files received and downloaded are solved to extract accurate tropospheric delay parameters, denoted here as ZTD. Finally, using ZTD, the conversion relationships between sea level pressure and ZHD, and between surface temperature and conversion factor, the accurate PWV prediction is calculated. The calculation method of the tropospheric hydrostatic delay can be obtained by using the Sa algorithm or the latest GPT3 algorithm. It should be noted that both the Sa algorithm and the GPT3 algorithm are classical algorithms in the field of tropospheric hydrostatic delay, and there are specific formulas for both algorithms for reference, which belong to the common knowledge of the field and will not be elaborated here. For the extraction of rainfall, it needs to be downloaded from the official website of the meteorological bureau or the websites of some meteorological product companies. In addition, the corresponding time series can be directly obtained from the observation files, and the time information of each group of data is recorded in the observation files, from which the time series information corresponding to PWV can be obtained.

[0010] Second step: The technical solution for rainfall event prediction based on the LSTM machine learning method. The PWV, temperature and rainfall data obtained in the first step are used for rainfall prediction. To better capture the changes in PWV, this solution uses the difference between the current PWV and the reference PWV as the input feature, called PWV-DIF. When selecting the reference time, the following criteria are adopted: if there is no rainfall at the current time, the previous time is selected as the reference time; if rainfall starts at the current time, the time of the first rainfall is selected as the reference time; if the rainfall lasts for more than five hours, the time of the sixth hour is taken as the new reference time. All input parameters are normalized before model training to improve the stability and generalization ability of the model. Finally, the data actually participating in the training of the model is the PWV_DIF, temperature and rainfall of the previous time, and the prediction result is the rainfall probability at the current time. The best binary discrimination threshold performance on the training data is selected as the standard (if it is greater than this threshold, it is classified as rainfall, taking 1, otherwise 0), and the trained model is saved for real-time rainfall event prediction. The data types participating in the prediction are consistent with the input parameters of the training model.

[0011] Step 3: A rainfall prediction technical solution based on the fusion machine learning method of SVR and GMDH is adopted, and the HS algorithm is used to optimize the model parameters and their combined weights. The input data set mainly includes the temperature and day of year (DOY) in the previous hour, PWV and rainfall in the previous two hours, and all input variables are normalized before training to improve the stability and convergence speed of the model. The fusion prediction model consists of two parallel sub-models: the SVR model and the GMDH model. The SVR model is optimized by adjusting the regularization parameter, epsilon-insensitive loss function, and kernel function, while the GMDH model captures non-linear features by selecting the best polynomial combination. The HS algorithm is introduced to achieve the complementarity of the two by optimizing the model parameters of SVR and GMDH. The trained model is saved for real-time rainfall prediction, and the data types involved in the prediction are consistent with the input parameters of the training model.

[0012] Step 4: Multiply the rainfall event prediction result (rainy is 1, non-rainy is 0) obtained in Step 2 directly by the rainfall amount prediction result in Step 3 to obtain the final effective rainfall amount result.

[0013] The present invention has the following characteristics and beneficial effects:

[0014] It can effectively solve the problem of high spatio-temporal resolution prediction of precipitation amount and precipitation events, and provide strong support for the decision-making of meteorological departments and related industries. By integrating satellite remote sensing technology, meteorological data and LSTM machine learning method, it can accurately identify whether rainfall occurs, overcome the problems of low prediction success rate and high error rate caused by traditional rainfall event prediction algorithms ignoring time correlation, and solve the problem of large prediction error of rainfall amount in the case of no rain. At the same time, this method combines the advantages of two machine learning methods, SVR and GMDH, and uses the HS algorithm to optimize parameter adjustment, effectively overcoming the bias problem of traditional machine learning methods for heavy rain or light rain prediction, eliminating the problem of low accuracy of rainfall prediction amount caused by the difference in the characteristics of machine learning algorithms, and improving the accuracy and reliability of prediction. Among them, the PWV data obtained by satellite remote sensing technology makes up for the deficiency of traditional methods in capturing local or small-scale rainfall changes, making the precipitation amount prediction more accurate in practical applications. By combining the predicted precipitation event with the precipitation amount, the present invention can effectively improve the accuracy of precipitation amount prediction, has stronger practical application value and wide applicability, and promotes the further development of precipitation amount prediction technology. Description of the Drawings

[0015] Figure 1 It is a schematic flow chart of obtaining GPS-PWV in this embodiment.

[0016] Figure 2Schematic flowchart of the real-time rainfall event prediction method based on GPS-PWV in this embodiment.

[0017] Figure 3 Schematic flowchart of the real-time rainfall amount prediction method based on GPS-PWV in this embodiment.

[0018] Figure 4 Schematic general flowchart of the rainfall amount prediction algorithm based on rainfall event constraints and hybrid machine learning strategies.

[0019] Figure 5 Comparison results of the rainfall event constraint strategy proposed in this embodiment and a single model without constraint strategy at the HKSC station in Hong Kong, China. The actual rainfall amount is given in the upper left subfigure (represented by the black line).

[0020] Figure 6 Comparison results of the integrated rainfall prediction model proposed in this embodiment and single prediction models such as SVR, GMDH, and ANN at the HKSC station in Hong Kong, China.

[0021] Figure 7 Comparison results of the integrated rainfall prediction model with rainfall constraints, SVR with rainfall constraints, and the integrated model without rainfall constraints proposed in this embodiment at the HKSC station in Hong Kong, China. Detailed implementation manners

[0022] The present invention will be described in detail below in conjunction with specific embodiments. The following embodiments will help those skilled in the art to further understand the present invention, but do not limit the present invention in any form. It should be noted that, without conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0023] A rainfall amount prediction method based on rainfall event constraints and hybrid machine learning strategies, as Figure 1 shown, includes the following steps:

[0024] Step 1: A large number of original observations during GPS PPP positioning and precise products required for solution are collected, mainly including pseudorange, carrier phase observations, precise ephemeris, precise clock offset, and corresponding products such as Earth nutation, etc., to prepare for extracting PWV (atmospheric water vapor content). Among them, the original observation files can be directly downloaded through the official websites of local surveying and mapping bureaus or seismological bureaus, and the navigation files and precise products required for solution can be obtained through the IGS official website. The above contents all belong to the common knowledge in the field of navigation and positioning. In this embodiment, they will not be elaborated in detail.

[0025] Step 2: For the data collected in the first step, use the PPP positioning mode for positioning and calculation. During the parameter calculation process, the classic least squares method or the extended Kalman filter method can be selected for solution. In this case, the weighted least squares method is used for parameter calculation. The weighted least squares method is a commonly used algorithm in parameter estimation, and the relevant details will not be elaborated here. It should be noted that the calculated parameters include the position of the receiver, the clock error, and the corresponding ZTD parameter (tropospheric delay), and only the ZTD parameter is used in this case.

[0026] Step 3: Use the Sa empirical model to calculate ZHD (tropospheric dry delay), and use the relationship among ZTD, ZWD, ZHD, and PWV to calculate PWV. The specific process is as follows:

[0027] First, use the Sa empirical model to calculate ZHD. The Sa empirical model is an empirical model algorithm proposed by Saastamoinen in 1972 for calculating ZHD. This algorithm can simply and accurately calculate ZHD at any location in the world and is still used until now. The calculation formula is as follows:

[0028] (1)

[0029] In the formula, φ represents the latitude where the GPS receiver (the device that collects GPS observation data) is located, H represents the height of the GPS receiver, which refers to the altitude here, not the height relative to the ground. Pg represents the pressure at the sea level where the receiver is located. Through the above formula, the ZHD parameter can be accurately calculated.

[0030] Secondly, use the relationship among ZTD, ZWD, and ZHD to calculate ZWD. Since ZTD is mainly composed of ZHD and ZWD, the relationship among the three can be used to calculate ZWD. The specific calculation formula is as follows:

[0031] (2)

[0032] In the formula, ZTD represents the total zenith delay, which is calculated in Step 2. ZHD is the zenith hydrostatic delay, which is calculated by formula (1). Thus, the accurate ZWD can be obtained.

[0033] Then, to achieve the conversion between ZWD and PWV, the conversion coefficient is needed, and its calculation formula is as follows:

[0034] (3)

[0035] In the formula, is the conversion coefficient, is the density of liquid water, and the value is ​ is the water vapor gas constant, with a value of , , is the atmospheric refractive index constant, and the empirical value is usually , .

[0036] Finally, according to the conversion relationship between ZWD, and PWV, PWV can be obtained. The specific calculation formula is as follows:

[0037] (4)

[0038] In the formula, ZWD is obtained by solving formula 2, is obtained by solving formula 3. Save the large number of calculated PWVs and the corresponding times for direct use in subsequent steps.

[0039] Step 4: Use the LSTM method to predict whether a rainfall event occurs. The LSTM method is one of the commonly used algorithms in the field of machine learning and will not be elaborated here. As Figure 2 shown, the specific process is as follows:

[0040] First, prepare the dataset for training the model. The elements of the dataset involved in predicting rainfall events include rainfall, temperature, PWV_DIF (the difference between the reference atmospheric water vapor content and the current atmospheric water vapor content). Among them, PWV_DIF is obtained by transforming the PWV obtained in step 3, which is obtained by using the difference between the PWV at the current moment and the PWV at the reference moment. When selecting the PWV at the reference moment, the following criteria are adopted: If there is no rainfall at the current moment, select the PWV at the previous moment as the reference moment; if rainfall starts at the current moment, select the PWV at the first rainfall moment as the reference moment; if the rainfall lasts for more than five hours, select the PWV at the sixth hour as the new reference moment.

[0041] It should be noted that to prevent the reference atmospheric water vapor content from not being updated for more than five hours, it will be forced to update every five hours. This is to better track the decrease in the atmospheric water vapor content when the rainfall is about to stop. According to research, the general rainfall time does not exceed five hours, and it is best to choose five hours as the reference. Of course, it can also be adjusted appropriately according to regional characteristics.

[0042] Secondly, use the prepared dataset for training. All input parameters are normalized before model training. Finally, the data actually participating in the training model is the PWV_DIF, temperature, and rainfall at the previous moment. The prediction result is the rainfall probability at the current moment. Select the binary discrimination threshold that performs best on the training dataset as the standard (if greater than this threshold, it is classified as rainfall, taking 1; otherwise, it is 0) for binarization of 0 and 1 when predicting the rainfall probability in real time. Save the trained model for predicting the real-time rainfall occurrence probability.

[0043] Finally, use the saved model and input prediction data of the same type as the training dataset to participate in the prediction, namely PWV_DIF, rainfall, and temperature. Use the above binary discrimination threshold to distinguish between rainy or non-rainy and save as 1 (rainy) and 0 (non-rainy).

[0044] Step Five: Use a hybrid model of SVR and GMDH to predict rainfall. The SVR and GMDH methods are one of the commonly used algorithms in the field of machine learning and will not be elaborated here. As Figure 3 shown, the specific process is as follows:

[0045] First, prepare the dataset for training the model. The elements of this dataset participating in the prediction of rainfall include the temperature (T) and day of the year (DOY) in the previous hour, PWV and rainfall in the previous two hours. Among them, the calculation formula for the day of the year is as follows:

[0046] (5)

[0047] In the formula, represents the day of the current year, and is the hour of the day.

[0048] Secondly, use the prepared dataset for training and normalize all input variables before training. This fusion prediction model consists of two parallel sub-models: the SVR model and the GMDH model. The SVR model is optimized by adjusting the regularization parameter, epsilon-insensitive loss function, and kernel function, while the GMDH model captures non-linear patterns by selecting the best model structure and polynomial combination. Introduce the HS algorithm to simultaneously optimize the model parameters of SVR and GMDH to obtain the best parameter settings for the fusion model. Among them, the HS method is one of the commonly used algorithms in the field of machine learning and will not be elaborated here. Save the proportion of the rainfall predicted by the two models that performs best on the training set in the fusion rainfall.

[0049] Finally, the saved model is used for real-time rainfall prediction. The data types involved in the prediction are consistent with the input parameters of the training model. Prediction data of the same type as the training dataset is used for prediction, namely the temperature (T) and day of year (DOY) in the previous hour, the PWV and rainfall in the previous two hours, and the fused rainfall is obtained using the saved weights mentioned above.

[0050] It should be noted that both LSTM and GMDH are considered to converge when the loss function tends to be stable. By adjusting parameters such as the number of training layers and neurons, it can be made to tend to be stable finally while avoiding overfitting caused by choosing overly complex parameters. Briefly speaking, within the number of training epochs I set, it is fine as long as the loss function on the validation set is stable; SVR determines convergence based on whether the rmse or mae on the validation set is stable. Different from the above two methods, it is not an iterative optimization but a convex optimization. Therefore, for it to converge, mainly the selection of parameters is crucial. If the performance changes little for multiple parameter combinations, it indicates that SVR may have reached the optimal state.

[0051] Step six: As Figure 4 shown, obtain the rainfall prediction constrained by the final rainfall event. Multiply the rainfall event prediction result (0 for no rain, 1 for rain) obtained in the last of Step five directly by the rainfall prediction result obtained in Step six to get the final rainfall prediction result.

[0052] At the HKSC station, the performance was evaluated by comparing single models (SVR, GMDH, and ANN) without the constraint strategy with single models adopting the constraint strategy. The models adopting the constraint strategy are labeled as "method + rainfall event constraint". Figure 5 Shows the prediction results of the original model and the model with the constraint strategy, and provides the actual rainfall as a reference. As can be seen from the red ellipses in the subgraphs, there is no rainfall on some days, but due to the PWV fluctuations, the prediction errors of the three models are relatively large during this period. After adopting the constraint strategy, the errors are significantly reduced (see the red dashed ellipses in each subgraph). GMDH is superior to SVR and ANN in predicting heavy rainfall (green dashed ellipses), while SVR and ANN are more accurate in predicting light and moderate rainfall. Therefore, to improve the prediction accuracy, it is recommended to use the SVR and GMDH models in combination.

[0053] To further evaluate the proposed method, we compared the fused model without the constraint strategy with single prediction models (including SVR, GMDH, and ANN models). Figure 6 Shows the results at the HKSC station. The results of the fused model are represented by the orange line, and the actual rainfall and the results of other single models are represented by the blue line. At Figure 6In it, the subfigure in the upper left corner shows the comparison between the results of the fusion prediction model and the actual rainfall. It can be seen that after adopting the proposed fusion prediction model, most of the rainfall can be accurately predicted, especially moderate rain and heavy rain. However, it is worth noting that since PWV always produces small fluctuations in the time series, the fusion prediction model always predicts rainfall even when there is actually no rainfall. Therefore, in the case of no rainfall, the fusion prediction model will have significant prediction errors. In addition, Figure 6 The comparison results between three single prediction models (SVR, GMDH, and ANN) and the fusion model are also shown. Compared with the single models, the fusion model performs better in predicting moderate rain and heavy rain, which can be observed in the red dashed ellipse. For light rain, the performance of the single models is similar to that of the fusion model. Therefore, it can be concluded that the fusion prediction model can significantly improve the prediction accuracy of moderate rain and heavy rain, but there are certain errors in the case of no rain.

[0054] The proposed method based on the constraint strategy and the fusion strategy is evaluated in detail by comparing the actual rainfall, the single SVR model, the SVR model with the constraint strategy, and the fusion model without the constraint strategy. Figure 7 The rainfall prediction results of the HKSC station are shown. It can be seen that the proposed method is significantly better than other methods, and the prediction results are very close to the actual rainfall, especially in predicting heavy rain, which is significantly better than the single SVR model with or without the constraint strategy (see the green dashed ellipse in each subfigure). In addition, compared with the fusion model, the proposed method effectively reduces the prediction error in the non-rainy period (see the red dashed ellipse in the subfigure).

[0055] According to Figures 5 - 7As shown, it can be seen that the prediction method provided by this embodiment overcomes the problems of low prediction success rate and high error rate in traditional rainfall event prediction algorithms due to ignoring temporal correlation, solves the problem of large rainfall prediction error in the case of no rain, and at the same time uses a rainfall prediction algorithm with a hybrid machine learning strategy to eliminate problems such as low accuracy of rainfall prediction due to differences in the characteristics of machine learning algorithms. Based on the constraint method of rainfall events, it can effectively solve the problem that when directly predicting rainfall using machine learning methods, it is impossible to correctly and effectively distinguish whether there is rainfall at the current moment. At the same time, this method incorporates satellite remote sensing technology. By using the delay of satellite signals passing through the ionosphere and certain meteorological data, PWV can be retrieved, and then a certain transformation is performed on PWV to obtain PWV_DIF. Using this data and some meteorological data can fully consider the influence of the precipitable water in the troposphere on the occurrence of rainfall events, thereby increasing the prediction success probability of rainfall events. In addition, considering that when using a single machine learning method for rainfall prediction, there will be problems such as prediction tendency for heavy rain or light rain, and problems such as low accuracy of rainfall prediction due to differences in the characteristics of machine learning algorithms, the present invention considers using a fusion machine learning method of SVR and GMDH to predict rainfall. SVR pays more attention to light and moderate rain when predicting rainfall, and its prediction effect is not good when heavy rain occurs. On the contrary, GMDH pays more attention to heavy rain and has a poor prediction effect on light and moderate rain. Therefore, we use the HS algorithm as the hub to connect the two machine learning methods, and use the HS algorithm to adjust parameters to obtain the best performance of the fusion machine learning method. This method also incorporates PWV obtained by satellite remote sensing technology and some easily accessible meteorological data to participate in the prediction, so as to overcome the limitations brought by a single machine learning method. Finally, the predicted rainfall event is combined with the rainfall amount to obtain the final rainfall prediction value, so as to overcome the problem that it is impossible to distinguish whether there is rainfall when using machine learning methods to predict rainfall, provide a more practically meaningful precipitation prediction, rather than just staying at the level of theoretical data, thereby improving the practical application value of the prediction. In summary, the purpose of the present invention is to provide a method suitable for GPS to predict rainfall events and rainfall amounts, which solves the problems of prediction tendency for heavy rain or light rain and the difficulty in distinguishing whether rainfall occurs in existing machine learning methods for rainfall prediction. At the same time, it solves the deficiencies of satellite remote sensing technology in capturing small-scale or local rainfall changes, provides more accurate rainfall event and rainfall amount predictions, and provides strong guarantee for accurate predictions of meteorological departments.

[0056] 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 by the above embodiments, and the above embodiments and the descriptions in the specification are only preferred examples of the present invention, and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of the present invention claimed is defined by the appended claims and their equivalents.

Claims

1. A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy, characterized in that, It includes the following steps: Step 1: Extract historical valid observations and GPS data files. The valid observations include temperature, air pressure, and corresponding time series. The GPS data files include observation files, navigation files, clock files, and precise ephemeris files. Calculate the atmospheric water vapor content based on the valid observations and GPS data files; Step 2: Construct and train a rainfall event prediction model based on the LSTM machine learning method. When predicting, use the difference between the atmospheric water vapor content and the current atmospheric water vapor content as one of the input features, and add the temperature and rainfall in the previous hour to predict the rainfall probability at the current moment; Step 3: Construct and train a rainfall amount prediction model based on the fusion machine learning method of SVR and GMDH, and use the HS algorithm to optimize the model parameters and their combined weights. The input data for prediction includes the temperature and day of the year in the previous hour, and the atmospheric water vapor content and rainfall amount in the previous two hours; Step 4: Multiply the rainfall time prediction result obtained in Step 2 by the rainfall amount result predicted in Step 3 to obtain the final effective rainfall amount result.

2. The rainfall prediction method based on rainfall event constraint and hybrid machine learning strategy according to claim 1, wherein The calculation method of the atmospheric water vapor content is as follows: First, obtain the GPS data file through a geodetic receiver and the IGS official website, calculate the tropospheric delay by resolving the GPS data file, and calculate the tropospheric dry delay by resolving the Sa empirical model; Then, calculate the atmospheric water vapor content using the conversion relationship between the tropospheric delay, air pressure, tropospheric dry delay, temperature, and conversion factor.

3. A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy according to claim 1, characterized in that, When training the rainfall event prediction model based on the LSTM machine learning method, all input parameters are normalized before model training.

4. A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy according to claim 3, characterized in that, When training the rainfall event prediction model based on the LSTM machine learning method, use the atmospheric water vapor content as an input parameter, and select the following criteria: If there is no rainfall at the current moment, select the atmospheric water vapor content at the previous moment as the reference atmospheric water vapor content; If rainfall starts at the current moment, select the atmospheric water vapor content at the first rainfall as the reference atmospheric water vapor content; If the rainfall lasts for more than five hours, use the atmospheric water vapor content at the sixth hour as the new reference atmospheric water vapor content.

5. A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy according to claim 1, characterized in that, When training the rainfall amount prediction model based on the fusion machine learning method of SVR and GMDH, the input data needs to be normalized.

6. A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy according to claim 1, characterized in that, The rainfall amount prediction model based on the fusion machine learning method of SVR and GMDH includes two parallel sub-models, namely the SVR model and the GMDH model.

7. A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy according to claim 6, characterized in that, When training the SVR model, optimize it by adjusting the regularization parameter, epsilon-insensitive loss function, and kernel function.

8. A rainfall prediction method based on rainfall event constraints and a hybrid machine learning strategy according to claim 6, characterized in that, The GMDH model captures non-linear features by selecting the best polynomial combination.

Images