Precipitation data observation quality control method fusing multiple machine learning models

Abstract

The invention provides a rainfall data observation quality control method fusing multiple machine learning models, and relates to the technical field of meteorological observation. The method comprises the steps of obtaining rainfall data of a pre-observation area, and performing observation quality evaluation to obtain an abnormal part; adopting a generative adversarial network to repair an abnormal part in the obtained rainfall data to obtain repaired rainfall data; based on the restored rainfall data, meteorological factors used for complexity analysis are extracted; complexity analysis is carried out based on the extracted meteorological factors, and complexity scores of the meteorological factors are generated; and determining a discrimination path of the current rainfall data based on the complexity score. According to the method, through a series of steps of meteorological factor extraction, generative adversarial network restoration, complexity analysis and path dynamic switching judgment, high-precision observation and real-time adjustment of rainfall data are realized under different meteorological conditions, and reliable data support is provided for meteorological prediction.

Description

Technical Field

[0001] The present invention relates to the field of meteorological observation technology, and in particular to a method for controlling the observation quality of precipitation data by integrating multiple machine learning models. Background Art

[0002] In the field of meteorological observation, accurate observation and prediction of precipitation data are of great significance for disaster prevention and mitigation, agricultural management and water resources allocation.

[0003] However, due to the complexity of precipitation observations and the dynamic changes in the external environment, existing meteorological observations have many problems in the quality control of precipitation data.

[0004] First, precipitation observation data are often missing or abnormal due to equipment failure, signal interference or insufficient collection frequency, which affects the reliability of the data.

[0005] In addition, as meteorological conditions become more complicated, using only a single model to analyze and predict precipitation data is difficult to adapt to changeable weather conditions, resulting in insufficient accuracy of prediction results.

[0006] At present, some precipitation observation methods enhance data analysis capabilities by introducing machine learning models, but there are still deficiencies in model switching and tuning.

[0007] Especially when dealing with complex and extreme weather conditions, there is a lack of real-time feedback mechanisms to adjust collection strategies and model parameters in real time according to the size of errors and data fluctuations.

[0008] Therefore, it is urgent to propose a control method for the precipitation data observation quality that integrates multiple machine learning models. Summary of the invention

[0009] The present invention provides a method for controlling the quality of precipitation data observation by integrating multiple machine learning models, aiming to solve the technical problem that a single model is used to analyze and predict precipitation data in related technologies, which is difficult to adapt to changeable weather conditions, avoid subsequent problems in model switching and tuning, and solve the practical problem of lack of real-time feedback mechanism for regulating means when dealing with complex and extreme weather.

[0010] In order to achieve the above purpose, the embodiments of the present application are implemented in the following manner:

[0011] In a first aspect, the present application provides a method for controlling the quality of precipitation data observation by integrating multiple machine learning models, including the following:

[0012] Obtain precipitation data in the pre-observation area, conduct observation quality assessment, and determine the abnormal parts in the precipitation data based on the quality assessment results;

[0013] Generative adversarial networks are used to repair the abnormal parts of the obtained precipitation data to obtain the repaired precipitation data;

[0014] Based on the restored precipitation data, meteorological factors for complexity analysis are extracted;

[0015] Perform complexity analysis based on the extracted meteorological factors and generate complexity scores for the meteorological factors;

[0016] Based on the complexity score, determining the discrimination path of the current precipitation data, the discrimination path includes a first discrimination path, a second discrimination path and a third discrimination path;

[0017] The first discriminant path: the meteorological factor complexity score is in the low complexity range;

[0018] The second discriminant path: the meteorological factor complexity score is in the medium complexity range;

[0019] The third discriminant path: the meteorological factor complexity score is in the high complexity range;

[0020] Dynamically switch to the corresponding machine learning model based on the discriminant path, and predict the precipitation data based on the switched machine learning model to generate a prediction result;

[0021] Compare the prediction results with the restored precipitation data and calculate the prediction error;

[0022] A feedback optimization mechanism is generated based on the obtained prediction error, and an optimization scheme is generated according to the feedback optimization mechanism to optimize the future precipitation observation and prediction process;

[0023] Based on the error discrimination path, error discrimination optimization adjustment is performed, and the optimization adjustment is applied to the next round of observations. After the next round of observations is completed, the error is calculated again to form a closed loop.

[0024] Furthermore, by obtaining precipitation data in the pre-observation area, an observation quality assessment is performed to obtain the abnormal part, and the abnormal part in the precipitation data is determined based on the quality assessment result; the generative adversarial network is used to repair the abnormal part in the obtained precipitation data to obtain the repaired precipitation data; based on the repaired precipitation data, the meteorological factors used for complexity analysis are extracted; complexity analysis is performed based on the extracted meteorological factors, and a complexity score of the meteorological factors is generated; based on the complexity score, the discrimination path of the current precipitation data is determined; based on the discrimination path, the corresponding machine learning model is dynamically switched, and the precipitation data is predicted based on the switched machine learning model to generate a prediction result; based on the prediction result and the repaired precipitation data, the prediction error is calculated; based on the obtained prediction error, a feedback optimization mechanism is generated, and an optimization plan is generated according to the feedback optimization mechanism to optimize the future precipitation observation and prediction process; based on the error discrimination path, the error discrimination optimization adjustment is performed, and the optimization adjustment is applied to the next round of observations. After the next round of observations is completed, the error is calculated again to form a closed loop. This method effectively improves the quality control level of precipitation data observation through a series of steps including meteorological factor extraction, generative adversarial network repair, complexity analysis, and dynamic switching of discriminant paths. It can achieve high-precision observation and real-time adjustment of precipitation data under different meteorological conditions, providing reliable data support for meteorological forecasting.

[0025] In some specific implementations, obtaining precipitation data in a pre-observation area, performing observation quality assessment, and determining abnormal parts in the precipitation data based on the quality assessment results specifically include:

[0026] Observation quality assessment determines the quality of precipitation data by calculating the degree of deviation of the observed values ​​of meteorological factors in precipitation data. Specifically:

[0027]

[0028] In the formula, Q i represents the quality assessment value of the i-th meteorological factor, X i represents the observed value of the ith meteorological factor, μ i represents the historical mean of the i-th meteorological factor, σ i represents the historical standard deviation of the i-th meteorological factor, and i represents the i-th meteorological factor.

[0029] Based on the quality assessment results, the abnormal part of the precipitation data is determined, and the abnormal threshold τ is set. i When the data exceeds the set abnormal threshold, it is considered abnormal data, specifically:

[0030]

[0031] In the formula, |Q i| represents the absolute quality assessment value of the i-th meteorological factor, τ represents the set abnormal threshold, output 1 indicates that the data is judged to be abnormal, and output 0 indicates that the data is normal.

[0032] Furthermore, the observation quality assessment determines the overall quality of precipitation data by calculating the degree of deviation between the observed values ​​of meteorological factors and their historical means, and then identifies anomalies in the data. This assessment method uses standardized deviation calculation, so that the current observed values ​​of meteorological factors can more clearly show the degree of difference compared with the long-term historical characteristics. In this way, anomalies in the data that may be caused by sensor errors or environmental changes can be quickly identified, providing a reliable basis for subsequent data repair and analysis.

[0033] In some specific implementations, a generative adversarial network is used to repair the abnormal part of the obtained precipitation data to obtain the repaired precipitation data, and the repair formula is:

[0034]

[0035] In the formula, represents the precipitation data of the repaired i-th meteorological factor, represents the repair value of the i-th meteorological factor generated by the generator, λ is the fusion parameter, Represents the original observation data of the i-th meteorological factor.

[0036] Among them, 0<λ<1 is used to control the weight distribution of generated data and original data.

[0037] Furthermore, by repairing the precipitation data, this method can effectively reduce the impact of abnormal data on the quality of precipitation observations. The repair process uses a generative adversarial network to generate reasonable repair values, and realizes the weighted integration of the generated data and the original observation data through fusion parameters, so that the repaired data not only retains the characteristics of the original observations, but also can make up for the deficiencies caused by abnormal values ​​through generated data. This repair method can not only ensure the continuity and stability of the observation data when the data fluctuates abnormally, but also significantly improve the reliability of the data, providing more accurate basic data for subsequent complexity analysis and meteorological forecasting.

[0038] In some specific implementations, based on the restored precipitation data, meteorological factors for complexity analysis are extracted, and the extraction formula is:

[0039]

[0040] In the formula, represents the i-th meteorological factor X i and the correlation between precipitation intensity R, cov(X i , R) represents X iThe covariance between and R, and σ R Respectively represent X i and the standard deviation of R. Based on the calculation results, the meteorological factors used for complexity analysis were extracted.

[0041] Furthermore, by extracting meteorological factors that are highly correlated with precipitation intensity, this method can screen out the most valuable key factors for complexity analysis from the repaired precipitation data, and use correlation analysis combining covariance and standard deviation to effectively ensure that the extracted meteorological factors are not only closely related to precipitation intensity, but also can more intuitively reflect the changing patterns of precipitation data under different meteorological conditions. This extraction method greatly reduces the interference of redundant factors on complexity scoring and subsequent model predictions, and improves the efficiency of data processing and the accuracy of analysis.

[0042] In some specific implementations, complexity analysis is performed based on the extracted meteorological factors, and a complexity score of the meteorological factors is generated, and the calculation formula is:

[0043]

[0044] Where C is the complexity score of the meteorological factor, X i represents the current observed value of the i-th meteorological factor, μ i represents the historical mean of the i-th meteorological factor, σ i represents the historical standard deviation of the ith meteorological factor, w i is the weight of each meteorological factor, i represents the i-th meteorological factor, i traverses all meteorological factors from 1 to n, and each i corresponds to a specific meteorological factor; n represents the total number of samples.

[0045] Furthermore, by performing complexity analysis on the extracted meteorological factors and generating complexity scores, this method can accurately evaluate the complexity of meteorological data under multi-factor conditions. By introducing the historical mean and standard deviation to measure the deviation of the current observation value, and then combining the weights of each meteorological factor to comprehensively calculate the complexity score, the scoring result can more intuitively reflect the volatility and uncertainty of the current meteorological conditions. This complexity analysis based on deviation and weight effectively improves the adaptability under complex meteorological conditions and provides a scientific basis for the dynamic switching of the model.

[0046] In some specific implementations, determining a discrimination path for current precipitation data based on the complexity score specifically includes:

[0047] The first discriminant path indicates that when the complexity score of the meteorological factor is in the low complexity range, linear regression is selected as the machine learning model according to the first discriminant path, and the prediction results are generated based on the selected linear regression model.

[0048] Furthermore, the low-complexity interval indicates that the current meteorological conditions change stably, with little precipitation and no precipitation. Under this interval state, precipitation weather phenomena exhibit linear characteristics, and a linear regression model for quickly processing linear characteristics is used for prediction. The definition of the low-complexity interval is: the complexity score C of the obtained meteorological factors ≤ 30.

[0049] The second discrimination path indicates that the complexity score of the current meteorological factors is in the medium-complexity interval. When in the medium-complexity interval, a random forest is selected as the machine learning model according to the second determination path, and a prediction result is generated based on the selected random forest model.

[0050] Furthermore, the medium-complexity interval indicates that there are certain changes in the current meteorological conditions, but they have not reached extreme conditions. Under this interval state, precipitation weather phenomena will exhibit non-linear characteristics, and a random forest model capable of processing non-linear relationships is used for prediction processing. The definition of the medium-complexity interval is: 30 < C ≤ 70.

[0051] The third discrimination path indicates that the complexity score of the current meteorological factors is in the high-complexity interval. When in the high-complexity interval, a long short-term memory network is selected as the machine learning model according to the third discrimination path, and a prediction result is generated based on the selected long short-term memory network model.

[0052] Furthermore, the high-complexity interval indicates that the meteorological conditions are very complex, with large and rapidly changing precipitation intensity. Under this interval state, while the precipitation intensity presents non-linear characteristics, it also changes rapidly. A long short-term memory network model capable of processing complex non-linear relationships and adapting to rapidly changing weather conditions is used for prediction processing. The high-complexity interval is defined as: C > 70.

[0053] In some specific embodiments, based on the discrimination path, it is dynamically switched to the corresponding machine learning model, and based on the switched machine learning model, precipitation data is predicted to generate a prediction result, specifically including:

[0054] A discrimination interval is set for the discrimination path, the complexity score C is set to vary within the range of [0, 100], and switching between different models is based on the score. The discrimination intervals include:

[0055] C ≤ 30 represents the discrimination interval of the first discrimination path; 30 < C ≤ 70 represents the discrimination interval of the second discrimination path; C > 70 represents the discrimination interval of the third discrimination path.

[0056] The weight W of each model is defined by the complexity score C i (C). Based on the preset meteorological complexity interval, the weights of each model are adjusted to determine the contribution of each model to the final prediction result, and the weight W of each model is defined i(C) A function that varies with the meteorological complexity score C, adjusting the contribution of the model through weighted smooth transition. Let W1(C) be the weight of the linear regression model, W2(C) be the weight of the random forest model, and W3(C) be the weight of the long short-term memory network. The switching formula is as follows:

[0057]

[0058] W1(C) represents the weight of the linear regression model, with a weight of 1 when C ≤ 30 and gradually decreasing to 0; W2(C) represents the weight of the random forest model, gradually increasing when 30 < C ≤ 70, increasing from 0 to 1, and then gradually decreasing to 0 when C > 70; W3(C) represents the weight of the long short-term memory network, gradually increasing when C > 70. C represents the complexity score of the meteorological factor.

[0059] Furthermore, by determining the path based on the complexity score and switching to different machine learning models, this method can effectively adapt to changing meteorological conditions, making the selection of the prediction model more flexible. The setting of the sub-intervals of the complexity score allows for the dynamic selection of the optimal model type under simple, general, and complex meteorological conditions, avoiding the problem that a single model cannot adapt to different meteorological complexities. By assigning different weights to different models, it is possible to preferentially use linear models at low complexity, switch to non-linear models such as random forests at medium complexity, and handle complex meteorological changes through deep models such as long short-term memory networks at high complexity, ensuring the accuracy and real-time nature of the prediction results.

[0060] In some specific embodiments, based on the comparison between the prediction result and the repaired precipitation data, the prediction error is calculated. The error calculation formula is as follows:

[0061]

[0062] In the formula, MSE represents the mean squared error, n represents the total number of samples, represents the prediction error at the t-th moment, represents the precipitation prediction value at the t-th moment, represents the actual precipitation at the t-th moment.

[0063] Furthermore, by comparing the prediction result with the repaired precipitation data, this method calculates the prediction error using the mean squared error to quantify the deviation between the prediction result and the actual observed data. The mean squared error calculation method can not only provide an intuitive measure of the overall error but also emphasize the impact of larger deviations, thus triggering further optimization when the prediction deviation is significant.

[0064] In some specific implementations, a feedback optimization mechanism is generated based on the obtained prediction error, and an optimization scheme is generated according to the feedback optimization mechanism to optimize the future precipitation observation and prediction process, specifically including:

[0065] pass The error E is calculated, and the error threshold ∈ is set based on the error E. The path to which the current error belongs is determined based on the error threshold ∈, specifically:

[0066]

[0067] Wherein, ∈1 represents the slight error threshold, ∈2 represents the significant error threshold, and E represents the error. Path 1: when E<∈1, the error is below the slight error threshold, and no adjustment is performed according to the path to which it belongs. Path 2: when ∈1≤E<∈2, the error is above the slight threshold but has not reached the significant error threshold, and a slight adjustment is performed. A slight adjustment is to fine-tune the parameters of the machine learning model involved in the prediction. Path 3: when E≥∈2, significant adjustment measures are triggered, and significant adjustments include increasing the data collection frequency and sensor calibration.

[0068] Furthermore, by generating a feedback optimization mechanism and performing path discrimination based on the size of the prediction error, this method can flexibly take corresponding adjustment measures for different error levels, thereby achieving dynamic optimization of future precipitation observation and prediction processes. By classifying the error paths, the prediction settings can be kept unchanged when the error is small to avoid unnecessary adjustments; when the error increases to a certain extent, slight or significant adjustments are triggered to optimize the model parameters or data collection frequency respectively. This discrimination and adjustment mechanism ensures computational efficiency while flexibly responding to the uncertainty caused by error fluctuations.

[0069] In some specific implementations, based on the error discrimination path, error discrimination optimization adjustment is performed, and the optimization adjustment is applied to the next round of observations. After the next round of observations is completed, the error is calculated again to form a closed loop, specifically:

[0070] The adjustment strategy selected according to the error path is automatically adjusted, and the adjusted acquisition frequency, model configuration and sensor calibration information are applied to the next round of observations. If the error is reduced and maintained below the slight error threshold, it proves that the current optimization is effective. Based on the effective optimization, the current acquisition frequency and model configuration continue to be used. If the error is still greater than the significant error threshold, the error source is re-evaluated until the error is adjusted below the slight error threshold.

[0071] Furthermore, through optimization adjustment and automated application based on the error discrimination path, this method forms a feedback closed-loop mechanism in the precipitation observation and prediction process, which can gradually improve the prediction accuracy in different observation rounds. After each round of observation, automatic adjustment is performed according to the error path, and the adjusted acquisition frequency, model configuration and sensor calibration parameters are applied to the next round of observation, thereby maintaining a dynamic balance between data quality and prediction accuracy. If the error is reduced after adjustment and maintained below the slight error threshold, the current optimization is judged to be effective, and the current parameter settings continue to be used to ensure the stability and efficiency of the observation process.

[0072] The beneficial effects of the technical solution of the present invention are as follows: using a generative adversarial network to repair abnormal observations in precipitation data, and by fusing the generated data with the original data, ensuring the reliability of the data, and providing high-quality observation data for subsequent analysis; performing complexity scoring based on the extracted meteorological factors, dynamically identifying changes in meteorological conditions, and automatically switching to a machine learning model suitable for the current complexity. This dynamic switching mechanism improves the adaptability of the model, responds quickly under low-complexity conditions, and ensures prediction accuracy under high-complexity conditions; and through an adaptive feedback optimization closed loop, calculating the prediction error based on the mean square error, and making graded adjustments based on the error size, forming an adaptive feedback closed loop. The closed-loop feedback mechanism adopts a hierarchical response strategy to select appropriate optimization schemes according to the error changes to avoid instability caused by excessive adjustment. At the same time, when the error reaches below the slight threshold, the current configuration can be maintained to ensure the stability and resource efficiency of the observation process. Moreover, under complex meteorological conditions, the collection and prediction strategies can be automatically adjusted based on real-time feedback, so that the quality of observation data and the accuracy of prediction results are continuously optimized in a dynamic environment, providing more accurate support for meteorological forecasting.

[0073] In order to make the above-mentioned objects, features and advantages of the present application more obvious and easy to understand, preferred embodiments are specifically cited below and described in detail with reference to the attached drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments of the present application will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without paying creative work.

[0075] Figure 1A flowchart of a method for controlling the quality of precipitation data observations by integrating multiple machine learning models provided in an embodiment of the present application. DETAILED DESCRIPTION

[0076] The technical solutions in the embodiments of the present application will be described below in conjunction with the drawings in the embodiments of the present application.

[0077] See also Figure 1 , Figure 1 A flowchart of a method for controlling the quality of precipitation data observations by integrating multiple machine learning models is provided for an embodiment of the present application.

[0078] In this embodiment, the method for controlling the precipitation data observation quality by integrating multiple machine learning models includes step S100, step S200, step S300, step S400, step S500, step S600, step S700, step S800 and step S900.

[0079] Step S100: Obtain precipitation data in the pre-observation area, perform observation quality assessment, and determine abnormal parts in the precipitation data based on the quality assessment result.

[0080] Here, specifically include:

[0081] Observation quality assessment determines the quality of precipitation data by calculating the degree of deviation of the observed values ​​of meteorological factors in precipitation data. Specifically:

[0082]

[0083] In the formula, Q i represents the quality assessment value of the i-th meteorological factor, X i represents the observed value of the ith meteorological factor, μ i represents the historical mean of the i-th meteorological factor, σ i represents the historical standard deviation of the i-th meteorological factor, and i represents the i-th meteorological factor.

[0084] Based on the quality assessment results, the abnormal part of the precipitation data is determined, and the abnormal threshold τ is set. i When the data exceeds the set abnormal threshold, it is considered abnormal data, specifically:

[0085]

[0086] In the formula, |Q i | represents the absolute quality assessment value of the i-th meteorological factor, τ represents the set abnormal threshold, output 1 indicates that the data is judged to be abnormal, and output 0 indicates that the data is normal.

[0087] It should be noted that in the anomaly judgment, the deviation value of the observation data is judged according to the set anomaly threshold. When the degree of deviation exceeds the threshold, it is judged as an abnormal observation value. This method ensures the validity and accuracy of precipitation data, effectively reduces the interference caused by factors such as data noise and observation errors, ensures high-quality data input, and provides a more accurate data basis for predictive analysis.

[0088] Step S200: Use a generative adversarial network to repair the abnormal part of the obtained precipitation data to obtain repaired precipitation data.

[0089] Here, the repair formula is:

[0090]

[0091] In the formula, represents the precipitation data of the repaired i-th meteorological factor, represents the repair value of the i-th meteorological factor generated by the generator, λ is the fusion parameter, Represents the original observation data of the i-th meteorological factor.

[0092] Among them, 0<λ<1 is used to control the weight distribution of generated data and original data.

[0093] To avoid misunderstanding, we briefly explain how to repair the abnormal parts of precipitation data through generative adversarial networks:

[0094] First, quality assessment is performed to identify abnormal observations in precipitation data, which are caused by sensor failure, data noise or acquisition errors.

[0095] The abnormal data is passed into the generative adversarial network, where the generator model receives the characteristic information of the observed data and generates reasonable repair values ​​similar to the characteristics of the historical data. The goal of the generator is to generate "repair values" that conform to the normal data distribution to replace the abnormal parts in the observed data.

[0096] After generating the data, the discriminator model judges the authenticity of the repaired data generated by the generator to determine whether it conforms to the characteristics of the real data. The discriminator compares the data generated by the generator with the normal observed data and continuously optimizes the generator until the generated data is close enough to the real data.

[0097] When the repaired data generated by the generator passes the verification of the discriminator, it is weightedly fused with the original observed data. The fusion parameters are used to adjust the ratio of the generated data to the original data to ensure that the repaired data retains the actual characteristics of the observed data and can supplement the abnormal data.

[0098] After the data is repaired and fused by the generative adversarial network, the repaired precipitation data is obtained, which is used for subsequent complexity analysis and precipitation prediction.

[0099] It should be noted that the setting of fusion parameters makes the repair process have a certain flexibility, and the fusion ratio of generated data and original data is dynamically adjusted based on the real-time situation of meteorological data, so as to optimize the effect of data repair.

[0100] Step S300: extracting meteorological factors for complexity analysis based on the restored precipitation data.

[0101] Here, the extraction formula is:

[0102]

[0103] In the formula, represents the i-th meteorological factor X i and the correlation between precipitation intensity R, cov(X i , R) represents X i The covariance between and R, and σ R Respectively represent X i and the standard deviation of R. Based on the calculation results, the meteorological factors used for complexity analysis were extracted.

[0104] It should be noted that by setting the correlation threshold, the extracted meteorological factors can be dynamically adjusted according to different meteorological conditions, ensuring that the complexity analysis is always based on the most representative observation factors, so as to provide accurate data support when the meteorological conditions become complicated.

[0105] Step S400: Perform complexity analysis based on the extracted meteorological factors and generate complexity scores for the meteorological factors.

[0106] Here, the calculation formula is:

[0107]

[0108] Where C is the complexity score of the meteorological factor, X i represents the current observed value of the i-th meteorological factor, μ i represents the historical mean of the i-th meteorological factor, σ i represents the historical standard deviation of the ith meteorological factor, w i is the weight of each meteorological factor, i represents the i-th meteorological factor, i traverses all meteorological factors from 1 to n, and each i corresponds to a specific meteorological factor; n represents the total number of samples.

[0109] It should be noted that the weight parameters in the complexity score can be adjusted according to the impact of meteorological factors on precipitation, making the scoring method more flexible and accurate. In different weather scenarios, the weights can be adjusted according to the actual impact of each meteorological factor to achieve real-time and adaptability of complexity analysis, thus providing a more effective reference for subsequent model selection and prediction strategies.

[0110] Step S500: Determine a discrimination path for current precipitation data based on the complexity score.

[0111] Here, specifically include:

[0112] The first discriminant path indicates that when the complexity score of the meteorological factor is in the low complexity range, linear regression is selected as the machine learning model according to the first discriminant path, and the prediction results are generated based on the selected linear regression model.

[0113] Furthermore, the low complexity interval indicates that the current meteorological conditions are stable, with little or no precipitation. In this interval, the precipitation weather phenomenon shows linear characteristics, and a linear regression model that quickly processes linear characteristics is used for prediction. The low complexity interval is defined as: the complexity score C of the obtained meteorological factor is ≤30.

[0114] The second discriminant path indicates that the complexity score of the current meteorological factor is in the medium complexity range. If it is in the medium complexity range, the random forest is selected as the machine learning model according to the second discriminant path, and the prediction results are generated based on the selected random forest model.

[0115] Furthermore, the medium-complexity interval indicates that the current meteorological conditions have changed to a certain extent, but have not yet reached extreme conditions. In this interval, the precipitation weather phenomenon will show nonlinear characteristics. The random forest model that can handle nonlinear relationships is used for prediction. The definition of the medium-complexity interval is: 30 <C≤70。

[0116] The third discriminant path indicates that the complexity score of the current meteorological factor is in the high complexity range. If it is in the high complexity range, the long short-term memory network is selected as the machine learning model according to the third discriminant path, and the prediction results are generated based on the selected long short-term memory network model.

[0117] Furthermore, the high complexity interval indicates that the meteorological conditions are very complex, the precipitation intensity is large and changes rapidly. In this interval, the precipitation intensity shows nonlinear characteristics and changes rapidly. A long short-term memory network model that can handle complex nonlinear relationships and adapt to rapidly changing weather conditions is used for forecasting. The high complexity interval is defined as: C>70.

[0118] Step S600: Dynamically switch to the corresponding machine learning model based on the discrimination path, and predict the precipitation data based on the switched machine learning model to generate a prediction result.

[0119] Herein, it specifically includes:

[0120] Set a discrimination interval for the discrimination path, set the complexity score C to vary within the range of [0, 100], and switch between different models based on the score. The discrimination intervals include:

[0121] C ≤ 30 represents the discrimination interval of the first discrimination path; 30 < C ≤ 70 represents the discrimination interval of the second discrimination path; C > 70 represents the discrimination interval of the third discrimination path.

[0122] Define the weight W of each model through the complexity score C i (C), based on the preset meteorological complexity interval, adjust the weights of each model to determine the contribution of each model to the final prediction result, and define the weight W of each model i (C) as a function that changes with the meteorological complexity score C. Adjust the contribution of the model through smooth weight transition. Let W1(C) be the weight of the linear regression model, W2(C) be the weight of the random forest model, and W3(C) be the weight of the long short-term memory network. The switching formula is:

[0123]

[0124] W1(C) represents the weight of the linear regression model, with a weight of 1 when C ≤ 30 and gradually decreasing to 0; W2(C) represents the weight of the random forest model, gradually increasing when 30 < C ≤ 70, increasing from 0 to 1, and then gradually decreasing to 0 when C > 70; W3(C) represents the weight of the long short-term memory network, gradually increasing when C > 70, and C represents the complexity score of the meteorological factor.

[0125] It should be noted that the smooth change mechanism of the model weights realizes the seamless transition of each model. Within different intervals of the complexity score, the weight ratio of the model is automatically adjusted to determine the contribution of each model to the final prediction result. The smooth switching design can reduce the instability of the prediction when the precipitation intensity fluctuates greatly, prevent the prediction result from mutating due to model switching, and ensure the coherence and reliability of the prediction process. In addition, the mechanism of dynamically adjusting the weight based on the complexity score of the meteorological factor significantly improves the adaptive ability, especially suitable for precipitation prediction under complex and changeable weather conditions.

[0126] Step S700: Compare the prediction result with the repaired precipitation data to calculate the prediction error.

[0127] Herein, the error calculation formula is:

[0128]

[0129] In the formula, MSE represents mean square error, n represents the total number of samples, represents the prediction error at the tth moment, represents the precipitation forecast value at the tth moment, Represents the actual precipitation at the tth moment.

[0130] It should be noted that the application of mean square error is used to judge the reliability of the current model prediction according to the error size, and flexibly adjust the model configuration based on the feedback optimization mechanism. When the error is small, the existing configuration is maintained. When the error exceeds the preset range, the model parameters can be dynamically adjusted or the model type can be changed, thereby improving the adaptability and prediction accuracy of the method.

[0131] Step S800: generating a feedback optimization mechanism based on the obtained prediction error, and generating an optimization scheme according to the feedback optimization mechanism to optimize the future precipitation observation and prediction process.

[0132] Here, specifically include:

[0133] pass The error E is calculated, and the error threshold ∈ is set based on the error E. The path to which the current error belongs is determined based on the error threshold ∈, specifically:

[0134]

[0135] Wherein, ∈1 represents the slight error threshold, ∈2 represents the significant error threshold, and E represents the error. Path 1: when E<∈1, the error is below the slight error threshold, and no adjustment is performed according to the path to which it belongs. Path 2: when ∈1≤E<∈2, the error is above the slight threshold but has not reached the significant error threshold, and a slight adjustment is performed. A slight adjustment is to fine-tune the parameters of the machine learning model involved in the prediction. Path 3: when E≥∈2, significant adjustment measures are triggered, and significant adjustments include increasing the data collection frequency and sensor calibration.

[0136] It should be noted that the hierarchical setting of the error threshold ensures that accurate adaptive adjustments can be made based on the error changes. When the error exceeds the slight threshold but does not reach the significant threshold, effective compensation can be achieved by fine-tuning the model parameters. When the error reaches the significant threshold, adjustments such as frequency increase and sensor calibration are automatically performed to bring the quality of the data source and the acquisition frequency to higher standards. This optimization scheme ensures stability and adaptability under complex meteorological conditions through a graded response mechanism, effectively improving the accuracy of precipitation observation and prediction.

[0137] Step S900: Based on the error discrimination path, perform error discrimination optimization adjustment, apply the optimization adjustment to the next round of observation, and after the next round of observation is completed, calculate the error again to form a closed loop.

[0138] Here, specifically:

[0139] The adjustment strategy selected according to the error path is automatically adjusted, and the adjusted acquisition frequency, model configuration and sensor calibration information are applied to the next round of observations. If the error is reduced and maintained below the slight error threshold, it proves that the current optimization is effective. Based on the effective optimization, the current acquisition frequency and model configuration continue to be used. If the error is still greater than the significant error threshold, the error source is re-evaluated until the error is adjusted below the slight error threshold.

[0140] It should be noted that this closed-loop mechanism has significantly improved its adaptability and self-correction capabilities under complex meteorological conditions through multiple rounds of feedback evaluation and automated adjustment. When the error does not fall within the target range, it can automatically re-evaluate the source of the error and continue to make fine-tuning until it reaches below the slight error threshold, thereby ensuring the long-term accuracy and reliability of precipitation observations and providing high-quality data support for weather forecasts.

[0141] When implementing this method, precipitation data in the pre-observation area is obtained, and the observation quality assessment is performed to obtain the abnormal part, and the abnormal part in the precipitation data is determined based on the quality assessment result; the abnormal part in the obtained precipitation data is repaired by using a generative adversarial network to obtain the repaired precipitation data; based on the repaired precipitation data, meteorological factors for complexity analysis are extracted; complexity analysis is performed based on the extracted meteorological factors, and complexity scores of the meteorological factors are generated; based on the complexity score, the discrimination path of the current precipitation data is determined; based on the discrimination path, the corresponding machine learning model is dynamically switched, and the precipitation data is predicted based on the switched machine learning model to generate a prediction result; based on the prediction result and the repaired precipitation data, the prediction error is calculated; based on the obtained prediction error, a feedback optimization mechanism is generated, and an optimization plan is generated according to the feedback optimization mechanism to optimize the future precipitation observation and prediction process; based on the error discrimination path, error discrimination optimization adjustment is performed, and the optimization adjustment is applied to the next round of observation, and after the next round of observation is completed, the error is calculated again to form a closed loop. This method effectively improves the quality control level of precipitation data observation through a series of steps including meteorological factor extraction, generative adversarial network repair, complexity analysis, and dynamic switching of discriminant paths. It can achieve high-precision observation and real-time adjustment of precipitation data under different meteorological conditions, providing reliable data support for meteorological forecasting.

[0142] For ease of understanding, we will show the implementation process of a heavy rain state turning into moderate rain, and then turning into light rain from moderate rain state, as follows:

[0143] Initial observation data collection: real-time collection of precipitation and meteorological data in the pre-observation area. The observation data include precipitation intensity, temperature, air pressure, humidity and wind speed meteorological factors. Due to the characteristics of heavy rain, precipitation intensity and wind speed vary greatly in the initial observation stage.

[0144] The quality of the collected data is evaluated, and the degree to which the observed values ​​of each meteorological factor deviate from the historical mean is calculated. For data that are susceptible to environmental interference or collection errors, the outliers are repaired through a generative adversarial network to obtain the repaired precipitation data, providing accurate data input for subsequent complexity analysis.

[0145] Based on the repaired precipitation data, key meteorological factors are extracted and complexity analysis is performed. The complexity score is calculated based on the volatility and nonlinearity of the current observation data. Due to the significant meteorological fluctuations during the rainstorm stage, the initial complexity state is in a high complexity range.

[0146] In the high complexity interval of C>70, the main weights are assigned to the long short-term memory network to capture the nonlinear changes of data under heavy rain conditions, generate precipitation forecast results based on LSTM, monitor the fluctuation of precipitation intensity in real time, and continuously update the model weights according to the actual observation data.

[0147] As the heavy rain gradually weakened into moderate rain, the fluctuations in precipitation intensity and meteorological factors tended to stabilize, the complexity score dropped to the medium range, and the model weights were dynamically switched, reducing the weight of the LSTM and gradually increasing the weight of the random forest model to adapt to the more stable changes in meteorological conditions.

[0148] In moderate rain conditions, the random forest model is relied upon to predict future precipitation, which can capture the nonlinear characteristics during moderate rain but has better adaptability to data fluctuations, helping to improve prediction efficiency when complexity is reduced.

[0149] When the precipitation gradually weakens to light rain, the complexity score drops to the low complexity range. At this time, the main weight is assigned to the linear regression model to reduce the amount of calculation and maintain the prediction accuracy. The linear regression model is switched, and the model weight is mainly based on linear regression. Since the precipitation intensity changes little during light rain, linear regression can achieve efficient prediction under conditions of low data fluctuations.

[0150] The prediction results of each stage are compared with the actual observed data after repair, the mean square error is calculated, and it is determined whether the error exceeds the set error threshold.

[0151] If the error value is below the slight error threshold, the current model configuration and acquisition frequency will be maintained; if the error is on the significant error path, significant adjustments will be triggered, including increasing the acquisition frequency, recalibrating the sensor, and adjusting the model parameters again until the error drops below the slight error threshold.

[0152] The next round of observations is entered based on the optimized collection frequency and model parameters. The above process is repeated after each round of observations to form a feedback loop, continuously optimizing prediction accuracy and data quality control as precipitation conditions gradually change.

[0153] At this point, this embodiment is completed.

[0154] To sum up, in the embodiments of the present invention, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for controlling the quality of precipitation data observation by integrating multiple machine learning models, characterized in that: These include: Obtain precipitation data in the pre-observation area, conduct observation quality assessment, and determine the abnormal parts in the precipitation data based on the quality assessment results; Generative adversarial networks are used to repair the abnormal parts of the obtained precipitation data to obtain the repaired precipitation data. Based on the restored precipitation data, meteorological factors for complexity analysis are extracted; Perform complexity analysis based on the extracted meteorological factors and generate complexity scores for the meteorological factors; Based on the complexity score, determining a discrimination path of current precipitation data, wherein the discrimination path includes a first discrimination path, a second discrimination path, and a third discrimination path; The first discriminant path: the meteorological factor complexity score is in the low complexity range; The second discriminant path: the meteorological factor complexity score is in the medium complexity range; The third judgment path: the meteorological factor complexity score is in the high complexity range; Dynamically switch to the corresponding machine learning model based on the discriminant path, and predict the precipitation data based on the switched machine learning model to generate a prediction result; Compare the prediction results with the restored precipitation data and calculate the prediction error; A feedback optimization mechanism is generated based on the obtained prediction error, and an optimization scheme is generated according to the feedback optimization mechanism to optimize the future precipitation observation and prediction process; Based on the error discrimination path, error discrimination optimization adjustment is performed, and the optimization adjustment is applied to the next round of observations. After the next round of observations is completed, the error is calculated again to form a closed loop.

2. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 1, characterized in that: Obtain precipitation data in the pre-observation area, conduct observation quality assessment, and determine the abnormal parts in the precipitation data based on the quality assessment results, including: Observation quality assessment determines the quality of precipitation data by calculating the degree of deviation of the observed values ​​of meteorological factors in precipitation data. Specifically: In the formula, Q i represents the quality assessment value of the i-th meteorological factor, X i represents the observed value of the ith meteorological factor, μ i represents the historical mean of the i-th meteorological factor, σ i represents the historical standard deviation of the i-th meteorological factor, i represents the i-th meteorological factor; Based on the quality assessment results, the abnormal part of the precipitation data is determined, and the abnormal threshold τ is set. i When the data exceeds the set abnormal threshold, it is considered abnormal data, specifically: In the formula, |Q i | represents the absolute quality assessment value of the i-th meteorological factor, τ represents the set abnormal threshold, output 1 indicates that the data is judged to be abnormal, and output 0 indicates that the data is normal.

3. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 2, characterized in that: The generative adversarial network is used to repair the abnormal part of the obtained precipitation data, and the repaired precipitation data is obtained. The repair formula is: In the formula, represents the precipitation data of the repaired i-th meteorological factor, represents the repair value of the i-th meteorological factor generated by the generator, λ is the fusion parameter, Represents the original observation data of the i-th meteorological factor.

4. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 1, characterized in that: Based on the restored precipitation data, the meteorological factors used for complexity analysis are extracted, and the extraction formula is: In the formula, represents the i-th meteorological factor X i and the correlation between precipitation intensity R, cov(X i , R) represents X i The covariance between and R, and σ R Respectively represent X i and the standard deviation of R. Based on the calculation results, the meteorological factors used for complexity analysis were extracted.

5. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 1, characterized in that: Complexity analysis is performed based on the extracted meteorological factors, and the complexity score of the meteorological factors is generated. The calculation formula is: Where C is the complexity score of the meteorological factor, X i represents the current observed value of the i-th meteorological factor, μ i represents the historical mean of the i-th meteorological factor, σ i represents the historical standard deviation of the ith meteorological factor, w i is the weight of each meteorological factor, i represents the i-th meteorological factor, i traverses all meteorological factors from 1 to n, and each i corresponds to a specific meteorological factor; n represents the total number of samples.

6. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 5, characterized in that: Based on the complexity score, the judgment path of the current precipitation data is determined, including: The first discriminant path indicates that when the complexity score of the meteorological factor is in the low complexity range, the linear regression is selected as the machine learning model according to the first discriminant path, and the prediction result is generated based on the selected linear regression model; The second discriminant path indicates that the complexity score of the current meteorological factor is in the medium complexity range. If it is in the medium complexity range, the random forest is selected as the machine learning model according to the second discriminant path, and the prediction result is generated based on the selected random forest model; The third discriminant path indicates that the complexity score of the current meteorological factor is in the high complexity range. If it is in the high complexity range, the long short-term memory network is selected as the machine learning model according to the third discriminant path, and the prediction results are generated based on the selected long short-term memory network model.

7. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 6, characterized in that: Dynamically switch to the corresponding machine learning model based on the discriminant path, and predict the precipitation data based on the switched machine learning model to generate prediction results, including: Set a discrimination interval for the discrimination path, set the complexity score C to vary within the range of [0, 100], and switch between different models based on the score. The discrimination intervals include: C ≤ 30 represents the discrimination interval of the first discrimination path; 30 < C ≤ 70 represents the discrimination interval of the second discrimination path; C > 70 represents the discrimination interval of the third discrimination path. The weight W of each model is defined by the complexity score C i (C) Based on the preset meteorological complexity range, adjust the weight of each model, determine the contribution of each model to the final prediction result, and define the weight W of each model i (C) As the function of the meteorological complexity score C changes, the contribution of the model is adjusted through weight smoothing transition. Let W1(C) be the weight of the linear regression model, W2(C) be the weight of the random forest model, and W3(C) be the weight of the long short-term memory network. The switching formula is: W1(C) represents the weight of the linear regression model, with a weight of 1 when C ≤ 30 and gradually decreasing to 0; W2(C) represents the weight of the random forest model, gradually increasing from 0 to 1 when 30 < C ≤ 70, and then gradually decreasing to 0 when C > 70; W3(C) represents the weight of the long short-term memory network, gradually increasing when C > 70, and C represents the complexity score of meteorological factors.

8. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 1, characterized in that: Based on the comparison between the prediction result and the repaired precipitation data, calculate the prediction error. The error calculation formula is: In the formula, MSE represents mean square error, n represents the total number of samples, represents the prediction error at the tth moment, represents the precipitation forecast value at the tth moment, Represents the actual precipitation at the tth moment.

9. The method for controlling the observation quality of precipitation data integrating multiple machine learning models as claimed in claim 1, characterized in that: Generate a feedback optimization mechanism based on the obtained prediction error, and generate an optimization plan according to the feedback optimization mechanism to optimize the future precipitation observation and prediction process, specifically including: pass The error E is calculated, and the error threshold ∈ is set based on the error E. The path to which the current error belongs is determined based on the error threshold ∈, specifically: In the formula, ∈1 represents the minor error threshold, ∈2 represents the significant error threshold, E represents the error. For path 1, when E < ∈1, the error is below the minor error threshold and no adjustment is made according to the所属 path. For path 2, when ∈1 ≤ E < ∈2, the error is above the minor threshold but not reaching the significant error threshold, and a minor adjustment is performed. The minor adjustment is to finely tune the parameters of the machine learning model participating in the prediction. For path 3, when E ≥ ∈2, significant adjustment measures are triggered, and the significant adjustment includes increasing the data collection frequency and sensor calibration.

10. The method for controlling the observation quality of precipitation data integrating multiple machine learning models according to claim 1, characterized in that: Based on the error discrimination path, perform error discrimination optimization adjustment, apply the optimization adjustment to the next round of observation. After the next round of observation is completed, calculate the error again to form a closed loop, specifically: Automatically adjust according to the adjustment strategy selected by the error path, and apply the adjusted data collection frequency, model configuration, and sensor calibration information to the next round of observation. If the error decreases and remains below the minor error threshold, it proves that the current optimization is effective, and then continue to use the current data collection frequency and model configuration based on the effective optimization. If the error is still greater than the significant error threshold, re-evaluate the error source until the error is adjusted below the minor error threshold.

Images