Plain area fog potential forecasting method based on weather typing and meteorological element change

Abstract

The invention relates to the technical field of fog potential forecasting, and discloses a plain area fog potential forecasting method based on weather typing and meteorological element change, which comprises the following steps: constructing a low-visibility event instance library, and establishing a high-altitude circulation and ground situation typing library; and then matching the real-time forecast field with the typing library to determine a weather background, and extracting meteorological elements such as temperature inversion and quiet wind. The method is characterized in that a movable boundary conceptual model is integrated, an air mass boundary is identified by calculating a horizontal temperature gradient and is quantified into discrete state characteristics representing cold and warm air mass effects, and finally, a weather background, conventional elements and the discrete state characteristics are jointly input into a Bayesian probability model, so that the air mass effect is obtained. And calculating and outputting the posterior probability of each fog potential level. According to the method, the key physical process is quantified and then fused into the statistical model, so that the capturing capability of a specific fogging mechanism is improved, and the forecasting result has high accuracy, strong interpretability and refined probability characteristics.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of fog potential prediction, in particular to a fog potential prediction method for plain areas based on weather typing and meteorological element changes. BACKGROUND

[0002] Fog is a weather phenomenon that has a serious impact on social and economic activities, especially aviation, navigation and land transportation. Therefore, providing accurate fog potential prediction is of great practical significance for ensuring public safety, improving transportation efficiency and effectively scheduling social production. Currently, fog prediction mainly relies on numerical weather prediction models and various statistical prediction methods.

[0003] However, the existing technical solutions still have certain limitations in practical application. On the one hand, numerical prediction models are relatively accurate in simulating large-scale circulation patterns, but are limited by their spatial resolution and uncertainty in parameterizing boundary layer physical processes, and are not fine enough in depicting local fog processes that evolve rapidly. On the other hand, although traditional statistical prediction methods can establish statistical relationships between local meteorological elements and fog phenomena, they are difficult to integrate macro weather background information, and lack direct and effective quantitative means for key physical mechanisms of fog formation, such as stable stratification formed by cold and warm air mass interaction in plain areas (i.e. cold pad effect). Therefore, they can only rely on some indirect or static physical quantities for characterization, which limits the ability of the prediction model to capture specific fog formation mechanisms.

[0004] The existing technology fails to organically combine macro weather typing, dynamic quantification of key physical processes, and interpretable probabilistic statistical models, and cannot effectively quantify key thermal and dynamic processes that affect fog formation and dissipation under clear weather background classification, and output prediction results in an interpretable and probabilistic manner. Therefore, the present application proposes a fog potential prediction method for plain areas based on weather typing and meteorological element changes to solve the deficiencies of the existing technology. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a fog potential prediction method for plain areas based on weather typing and meteorological element changes, which solves the deficiencies of the existing technical solutions in integrating large-scale circulation background, surface weather situation and local key physical processes, especially the lack of quantitative means for thermal and dynamic boundaries that have a significant impact on fog formation and dissipation in plain areas, and the lack of strong interpretability in some prediction models, which affects the accuracy of the prediction.

[0006] To achieve the above purpose, the present application is implemented by the following technical solution: a fog potential prediction method for plain areas based on weather typing and meteorological element changes, comprising the following steps: S1, performing a data preparation stage: obtaining and processing historical meteorological data and real-time forecast data including multi-level temperature field grid point data, and constructing a low-visibility event case library; S2, performing a weather typing stage: the weather typing stage includes: based on a historical 500 hPa geopotential height field data set, establishing a historical circulation objective typing library by an objective typing method; and based on historical and real-time ground meteorological elements and sea level pressure grid point data, classifying ground weather situations; S3, performing real-time weather typing matching: performing similarity calculation on real-time mode forecast 500 hPa geopotential height field grid point data and the historical circulation objective typing library to determine the circulation type to which the current forecast period belongs; S4, performing a feature engineering stage: extracting and quantifying key meteorological elements related to fog potential formation from historical meteorological data and real-time forecast data; S5, performing a fog potential probability forecast stage: based on the determined circulation type, key meteorological elements in the weather typing stage, constructing and running a Bayesian probability model to calculate the posterior probability of the fog potential level as the forecast result.

[0007] Preferably, in the S1 step, constructing the low-visibility event case library specifically includes: By comparing the historical visibility data with the preset visibility threshold, identifying the continuous time period that meets the low-visibility condition, and defining this continuous time period as a low-visibility event; Defining a complete time window for the low-visibility event including a preset time length before the event occurs; Extracting all relevant meteorological element data within the complete time window from the historical meteorological data, and storing the extracted all relevant meteorological element data in association with the label of the low-visibility event.

[0008] Preferably, the S1 step further includes performing quality control on the obtained original data, specifically including: Removing error values and abnormal values in the historical meteorological data and real-time forecast data, and performing missing value filling processing on the error values and abnormal values; The missing value filling processing includes: judging the missing sequence length, when the missing sequence length is less than or equal to a preset short-time missing threshold, using a time interpolation method for filling, and when the missing sequence length is greater than the preset short-time missing threshold, using a spatial nearest neighbor interpolation method for filling.

[0009] Preferably, in the S2 step, establishing the historical circulation objective typing library specifically includes: Flattening and converting the historical 500 hPa geopotential height field grid point data at each time point into a high-dimensional feature vector; The K-means clustering algorithm is used to perform clustering analysis on the set of all historical high-dimensional feature vectors to obtain a preset number of clusters. The cluster center vectors of all clusters are stored, and the set of cluster center vectors of all clusters constitutes a historical circulation objective classification library.

[0010] Preferably, in the S2 step, the classification of the surface weather pattern specifically includes: According to the preset meteorological definition, the surface weather pattern is classified into at least one of the following: isallobaric field, cold front, high pressure bottom, high pressure rear, or inverted trough.

[0011] Preferably, on the basis of the historical circulation objective classification library established by the above-mentioned K-means clustering algorithm, in the S3 step, the real-time weather classification matching specifically includes: The 500 hPa potential height field grid point data of the real-time model prediction is converted into a high-dimensional real-time feature vector. The cosine similarity method is used to calculate the similarity values between the high-dimensional real-time feature vector and the cluster center vectors of all clusters in the historical circulation objective classification library. The cluster center vector with the maximum similarity value is determined, and the circulation type represented by the cluster center vector is determined as the classification result of the current prediction period.

[0012] Preferably, in the S4 step, extracting and quantifying the key meteorological elements related to fog potential formation specifically includes: From the historical meteorological data and real-time prediction data, the inversion layer bottom height and inversion intensity identified by the vertical temperature profile are analyzed; From the historical meteorological data and real-time prediction data, the ground wind speed data and the ground calm condition or local recirculation phenomenon identified by the wind field vertical profile are analyzed.

[0013] Preferably, the S4 step further includes integration of a movable boundary concept model, specifically including: The temperature field grid point data obtained in the data preparation stage is read, the horizontal temperature gradient of the target plain area and its periphery is calculated by comparing the temperature values of adjacent grid points in the horizontal direction of the temperature field grid point data, and the calculated horizontal temperature gradient is compared with a preset frontal gradient threshold value. The area with a horizontal temperature gradient greater than the frontal gradient threshold value is identified as the air mass boundary; According to the relative position and movement trend of the air mass boundary relative to the target plain area, the discrete state feature representing the interaction state of cold and warm air masses is determined and output, which is input into the Bayesian probability model in the fog potential probability prediction stage as a key meteorological element.

[0014] Preferably, in the S5 step, constructing and running the Bayesian probability model specifically includes: In the model training stage, the prior probability of the fog potential level is counted and stored by traversing the low visibility event case library, and the conditional probability of each feature under the condition of a given fog potential level is stored.

[0015] Preferably, on the basis of the above model training stage, the posterior probability of the fog potential level calculated in the S5 step specifically includes: receiving the current real-time forecast feature vector in the real-time forecast stage; The posterior probability of each fog potential level under the condition of observing the current real-time forecast feature vector is calculated by applying Bayes' theorem and combining the prior probability and the conditional probability.

[0016] The present application provides a plain area fog potential forecast method based on weather typing and meteorological element changes. It has the following beneficial effects: 1. The present application provides a data basis and macro weather background for the subsequent probability forecast model by establishing a systematic low visibility event case library and combining upper air circulation objective typing and surface weather situation classification, so that the forecast can first determine the current macro weather type before fine analysis, thereby improving the pertinence and accuracy of the forecast.

[0017] 2. The present application integrates a movable boundary concept model, which identifies and quantifies the air mass boundary, a physical process crucial to the formation and dissipation of fog, by calculating the horizontal temperature gradient, and inputs it as a discrete state feature into the model. This method overcomes the defects of traditional methods that are difficult to quantify the influence of thermal structure such as front and cold pad, and improves the model's ability to capture specific physical cause fog processes.

[0018] 3. The present application uses a Bayesian probability model, which can transparently integrate large-scale circulation background, surface situation and fine local physical features including air mass boundary. The final output is a posterior probability with clear statistical significance. Compared with deterministic forecast conclusions, this quantitative result provides more rich risk information and strong interpretability, providing stronger support for the decision and risk assessment of forecasters. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 The flowchart of the plain area fog potential forecast method of the present application; Figure 2 The data preparation and low visibility event case library construction flowchart of the present application; Figure 3 The schematic diagram of the typical atmospheric boundary layer structure conducive to fog formation of the present application; Figure 4 The flowchart of the plain area fog potential forecast method of the present application; DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the present application.

[0021] With reference to the drawings Figure 1 The present application provides a fog potential forecast method based on weather typing and meteorological element changes in plain areas, which can be executed by one or more data processing devices (such as a meteorological service server).

[0022] In one specific embodiment, the method first executes a data preparation phase, which is used to build a data basis required for subsequent model training, verification and real-time forecasting. This phase is completed by configuring one or more data collection modules, specifically including: A historical reanalysis data collection module is configured, which is used to obtain and store historical meteorological data from one or more designated data centers. The time range of the historical meteorological data covers 3 years to 3 winter half years, which is defined as October of each year to February of the next year.

[0023] The data content obtained by the historical reanalysis data collection module specifically includes: historical sounding data, historical ground station observation data (whose elements at least include temperature, humidity, air pressure, wind speed), and NEPFNL reanalysis data. The spatial resolution of the NEPFNL reanalysis data is 1°x1°.

[0024] A real-time forecast data access module is configured, which is used to access and obtain numerical weather prediction products published by the Central Meteorological Observatory in real time. The numerical weather prediction products at least include 500 hPa potential height field forecast data and ground meteorological element forecast data.

[0025] A fog condition observation data collection module is configured, which is used to collect and store fog day records and visibility observation data in the North China Plain and the Huanghuai Plain region, for the construction of a low-visibility event case library and the verification of model effects.

[0026] In one specific embodiment, the real-time forecast data access module is further used to obtain grid data as an atmospheric circulation background field input. The time resolution of the atmospheric circulation background field input data is 1 hour, which is used to support short-term forecast analysis.

[0027] The input data of the atmospheric circulation background field specifically includes grid data of temperature, pressure, humidity and wind field elements on the 500 hPa, 700 hPa, 850 hPa and 925 hPa standard isobaric layers.

[0028] In one embodiment of the present application, when performing the data preparation phase, the method further comprises performing quality control on the original data obtained by the historical reanalysis data collection module and the fog observation data collection module.

[0029] The quality control is performed by a data quality control module, which is configured to identify and handle erroneous values, abnormal values and missing values in the data.

[0030] The data quality control module first performs abnormal value detection on the original data. In one specific embodiment, the data quality control module is configured with a physical threshold rule library, which predefines the physically valid value range for each meteorological element (e.g. temperature, humidity, wind speed, pressure) in the historical data. For example, for the relative humidity element, its valid value range is limited to ; for the wind speed element, its valid value range is limited to , where is the maximum wind speed value based on local climate statistics.

[0031] When the data quality control module detects that the value of any data point (e.g. the relative humidity of a certain station is 102% or -2%) exceeds its corresponding valid value range in the physical threshold rule library, the data point is determined to be an erroneous value.

[0032] The data quality control module is further configured with time consistency checking logic to evaluate whether the rate of change of a certain meteorological element within adjacent time steps exceeds a pre-set reasonable change threshold. For example, a maximum allowed hourly temperature change threshold (e.g. 15°C / hour) is set, and when it is detected that the temperature data of a certain station changes more than the threshold within an adjacent hour, the data point is determined to be an abnormal value.

[0033] All data points determined to be erroneous values or abnormal values are excluded from the dataset by the data quality control module and marked as missing.

[0034] After performing abnormal value exclusion, the data quality control module then performs missing value filling processing on the data points marked as missing. The data quality control module is configured with missing sequence length discrimination logic to first calculate the time steps of consecutive missing in a certain data sequence .

[0035] When the data quality control module determines that the missing sequence length is less than or equal to a pre-set short-term missing threshold (For example When set to 2 or 3 time steps, the short-term missing test is defined as a single missing test event.

[0036] For a single missing event, the data quality control module initiates a time interpolation submodule. This submodule fills in the missing data using values ​​from both the preceding and following time series. In one implementation, this filling method is linear interpolation, meaning it uses the values ​​of the nearest valid data point before and after the missing data point to estimate and fill in the missing data point through linear calculation.

[0037] When the data quality control module determines the length of the missing sequence Greater than the short-term missing detection threshold In this case, the missing test is defined as a long-term missing test.

[0038] For long-term missing data, the data quality control module initiates a spatial interpolation submodule. This submodule uses spatial nearest neighbor interpolation to fill in the gaps.

[0039] Specifically, for the target site At the point of time To determine the missing measurement value, the spatial interpolation submodule first accesses a pre-stored site geographic information database, which contains the geographic coordinates of all sites. The spatial interpolation submodule then calculates the geographic distance to determine the distance. The nearest neighboring site .

[0040] The spatial interpolation submodule then queries neighboring sites. At the same time point The observation data. If the data is valid, then that valid value is assigned to the target station. As a point in time The fill value, if At the point of time The data is also missing. The spatial interpolation submodule will query the nearest neighboring sites in order of geographical distance from near to far until the first site with valid data is found and the data is filled.

[0041] In one embodiment of the present invention, the method further includes performing data normalization processing on the quality-controlled valid data output by the data quality control module.

[0042] The data normalization processing is performed by a data normalization module, which is configured to address the technical problem of the non-uniform physical dimension and the great difference in numerical range among the collected multi-source meteorological element data (e.g. temperature, air pressure, wind speed, humidity, etc.). Without the processing, the difference will result in an out-of-proportion influence of the features with a larger numerical range (e.g. the value of the geopotential height field) on the result in the subsequent model calculation.

[0043] The data normalization module is configured with a feature extreme value calculation submodule. The feature extreme value calculation submodule is configured to traverse all the historical reanalysis data sets output by the data quality control module.

[0044] For each independent meteorological element (e.g. 500 hPa geopotential height, ground relative humidity) in the data set, the feature extreme value calculation submodule calculates and determines the global minimum value and the global maximum value of the element in the entire historical data set.

[0045] The values of all the elements and are stored in the feature extreme value parameter library by the feature extreme value calculation submodule for subsequent data transformation calling.

[0046] The data normalization module is further configured with a data transformation submodule. The data transformation submodule adopts a discrete standardization method to linearly transform the data value to a closed interval.

[0047] When the data transformation submodule processes an original data value to be normalized , it first queries the global minimum value and the global maximum value of the meteorological element corresponding to the data value from the feature extreme value parameter library.

[0048] Subsequently, the data transformation submodule calculates the normalized value by applying the following formula: ; wherein: is the original data value to be normalized, is the global minimum value of the meteorological element corresponding to the data value in the historical data set, is the global maximum value of the meteorological element corresponding to the data value in the historical data set, is the normalized data value calculated and located in the interval .

[0049] ​For the real-time forecast data acquired by the real-time forecast data access module, the data transformation sub-module also applies this method to process, specifically, the data transformation sub-module retrieves the values of the historical data set calculated from the feature extreme value parameter library and to the real-time forecast data to perform transformation.

[0050] Through transformation, it can be ensured that all historical data and real-time data input into the subsequent fog potential forecast model are converted into a unified scale space, thereby meeting the requirement of consistency of input data for subsequent model calculation.

[0051] Referring to the accompanying Figure 2 In an embodiment of the present application, the method further comprises constructing a case library specially used for storing and managing low-visibility event data, and the low-visibility event case library construction module is configured with an event screening and time window definition sub-module. The event screening and time window definition sub-module is used to traverse the historical visibility time series data acquired by the fog observation data collection module and subjected to quality control.

[0052] The event screening and time window definition sub-module is preset with one or more visibility threshold values (for example, a 1000-meter threshold value defining heavy fog, or a 1000-meter to 10000-meter interval threshold value defining light fog). The event screening and time window definition sub-module identifies all continuous time periods meeting the low-visibility condition by comparing the historical visibility data with the visibility threshold value.

[0053] For each identified continuous time period, the event screening and time window definition sub-module defines it as an independent low-visibility event.

[0054] The event screening and time window definition sub-module further defines a complete time window for each low-visibility event. The starting time point of the time window is set as the time point when the event first meets the low-visibility condition, and is pushed forward by a preset time length (according to the preferred configuration of the embodiment, the preset time length is 1 hour) to contain the precursor weather information before the event occurs. The ending time point of the time window is set as the time when the event no longer meets the low-visibility condition.

[0055] The low-visibility event case library construction module is further configured with an associated data extraction sub-module. For each determined low-visibility event and its time window The correlation data extraction submodule retrieves and extracts all relevant meteorological element data of the same geographical region within the time window from the database obtained by the historical reanalysis data collection module. The relevant meteorological element data include, but are not limited to, ground temperature, dew point temperature, relative humidity, wind direction and speed, sea level pressure, and high-altitude (such as 925 hPa, 850 hPa, 700 hPa, and 500 hPa) temperature, humidity, and wind pressure grid point data corresponding to the period.

[0056] The low-visibility event instance library construction module associates the extracted meteorological element data with the corresponding low-visibility event label and stores them in a structured manner, forming a long sequence of low-visibility event instance library. The instance library serves as a direct input data source for subsequent analysis of meteorological element variation rules and training of artificial intelligence models.

[0057] In an embodiment of the present application, the data preparation stage further includes preparation of atmospheric circulation background field data. The circulation background field data preparation module processes the numerical weather prediction grid point data obtained by the real-time forecast data access module and constructs a unified, multi-level atmospheric circulation background field for subsequent weather typing and meteorological element analysis steps.

[0058] The circulation background field data preparation module is configured with a data extraction submodule. The data extraction submodule is responsible for extracting specified meteorological element grid point data on a preset isobaric surface from the original numerical weather prediction product received by the real-time forecast data access module.

[0059] According to a specific implementation of the present embodiment, the preset isobaric surface includes 500 hPa, 700 hPa, 850 hPa, and 925 hPa, and the specified meteorological elements include temperature, pressure (or geopotential height), humidity, and wind field (including U and V wind components). The data extraction submodule ensures that only the data necessary for subsequent analysis is sent into the processing flow.

[0060] The circulation background field data preparation module is further configured with a spatial processing submodule. The spatial processing submodule is used to perform spatial range cropping on the extracted grid point data. Specifically, the spatial processing submodule crops the original grid point data covering a large area to a preset geographical coordinate frame that can completely cover the target plain region (such as the North China Plain or the Huanghuai Plain) concerned by the present application. This step reduces the data processing amount and focuses the analysis on the target region.

[0061] The processed data is structured and integrated by a data integration sub-module. The spatial processing sub-module organizes the grid data of different elements and different levels according to time sequence to form a unified multi-dimensional data array (for example, a data structure including time, level, latitude and longitude dimensions). The structured multi-dimensional data array is the atmospheric circulation background field required by the present application.

[0062] In an embodiment of the present application, after the data preparation stage is completed, the method enters a weather typing and real-time matching stage, which aims to establish a historical circulation pattern reference library and provide classification basis for subsequent real-time prediction. The stage first performs establishment of a historical circulation objective typing library.

[0063] The establishment of the historical circulation objective typing library is performed by a historical circulation objective typing library construction module. The input data of the historical circulation objective typing library construction module is the historical 500 hPa geopotential height field data set obtained by the historical reanalysis data collection module and processed by the data quality control module and the data normalization module in the data preparation stage.

[0064] The historical circulation objective typing library construction module is configured with a data vectorization sub-module. Since the historical 500 hPa geopotential height field data is a two-dimensional or three-dimensional grid matrix, the data vectorization sub-module first flattens the geopotential height field grid data at each time point according to a predetermined order (for example, row first and column second) and converts it into a high-dimensional feature vector If the dimension of the original grid field is , the dimension of the converted high-dimensional feature vector is .

[0065] The module is further configured with an objective typing clustering sub-module. The sub-module is used to perform clustering analysis on the entire set of historical feature vectors output by the data vectorization sub-module , where is the total number of historical samples, to identify statistically representative circulation patterns.

[0066] In a specific embodiment, the objective typing clustering sub-module adopts a K-mens (K-means) clustering algorithm. The number of clusters of the algorithm is pre-set to 5 to correspond to 5 circulation types.

[0067] The K-mens clustering algorithm divides the set of historical feature vectors into 5 mutually exclusive clusters through iterative calculation and calculates a cluster center (i.e., a centroid) for each cluster, where . The cluster center is a feature vector have the same dimension , represent the average state of all vectors within the cluster.

[0068] The module is further configured with a classification library storage submodule. The classification library storage submodule stores the five cluster center vectors output by the objective classification clustering submodule in a database.

[0069] The five cluster center vectors constitute a historical circulation objective classification library.

[0070] Each cluster center is associated and labeled as one of the five circulation types (for example, a zonal circulation type), and the five circulation types are: a zonal circulation type, a two-groove-one-ridge type, a one-groove-one-ridge type, a two-ridge-one-groove type, and a uniform potential height type.

[0071] The historical circulation objective classification library (i.e., the stored five center vectors ) is output to a subsequent real-time classification matching module as a reference for similarity comparison of real-time prediction fields.

[0072] In an embodiment of the present application, the weather classification and real-time matching stage, after the establishment of the historical circulation objective classification library, further includes classification of the ground weather situation, which is parallel to the upper air circulation classification and together provides meteorological background constraints for the subsequent fog potential prediction.

[0073] The classification of the ground weather situation is performed by a ground weather situation classification module, and the input data of the ground weather situation classification module is the historical and real-time ground meteorological element and sea level pressure grid data obtained in the data preparation stage and after processing.

[0074] The ground weather situation classification module is configured with one or more classification rule submodules for classifying complex ground weather patterns into several typical situations according to preset meteorological definitions.

[0075] In a specific embodiment, the ground weather situation classification module is configured with a pressure gradient calculation submodule for reading sea level pressure (SLP) grid data and calculating the horizontal pressure gradient between any two grid points in the target area (for example, the North China Plain) .

[0076] The pressure gradient calculation submodule is preset with an isobaric field threshold , and the pressure gradient calculation submodule determines whether the average horizontal pressure gradient in the target area is lower than the threshold ​​At this time, the surface weather pattern of the region is classified as an equalized pressure field (EQP).

[0077] The surface weather pattern classification module is further configured with a front and weather system identification submodule, which is used to comprehensively analyze the sea level pressure field, ground temperature field, ground wind field and humidity field data to identify the location of the front and high and low pressure system.

[0078] When the front and weather system identification submodule identifies a pressure trough on the sea level pressure field, and detects a sharp temperature drop (cold advection), clockwise change of wind direction (for example, from southwest wind to northwest wind) and humidity change near the pressure trough, the pressure trough is identified as a cold front. The region located in front of the moving direction of the cold front (usually with warm and wet advection characteristics) is classified as a cold front edge (F).

[0079] When the front and weather system identification submodule identifies a closed high pressure center (anticyclone) on the sea level pressure field, the front and weather system identification submodule further classifies according to the relative position of the target region and the high pressure center.

[0080] If the target region is located in the southern edge of the high pressure center (usually corresponding to the easterly flow in the Northern Hemisphere), the region is classified as the bottom of the high pressure (OH).

[0081] If the target region is located in the west of the high pressure center (i.e. after the high pressure system passes, usually corresponding to the northerly flow and subsidence motion), the region is classified as the rear of the high pressure (H).

[0082] When the submodule identifies a pressure trough line on the pressure field, and the axis direction of the trough line is north-south or extends from southeast to northwest (i.e. different from the conventional trough line form moving from west to east), the pattern is classified as an inverted trough (INT).

[0083] The surface weather pattern classification module labels the above classification results (for example, one of EQP, F, OH, H, INT) and outputs them to the subsequent fog potential forecast model as a key classification feature input.

[0084] In an embodiment of the present application, the weather typing and real-time matching stage further includes performing real-time weather typing matching, the purpose of which is to quantitatively compare the current real-time weather forecast pattern with the reference mode in the historical circulation objective typing library to determine the atmospheric circulation type to which the current forecast period belongs.

[0085] Real-time weather typing matching is performed by a real-time weather typing matching module, and the input data of the real-time weather typing matching module includes: The real-time model forecast 500hPa geopotential height field grid data is obtained from the real-time forecast data access module and processed by the data normalization module.

[0086] The historical circulation objective fractal library, generated by the historical circulation objective fractal library construction module, stores five cluster center vectors. .

[0087] The real-time weather classification matching module is configured with a real-time data vectorization submodule, which is used to process the input real-time 500hPa geopotential height field grid data.

[0088] The real-time data vectorization submodule employs the same vectorization transformation method as the data vectorization submodule. Specifically, it flattens the real-time two-dimensional grid matrix into a high-dimensional real-time feature vector according to a predetermined "row-first, column-later" order. High-dimensional real-time feature vectors Dimensions (i.e., the total number of grid points) and the cluster center vector in the historical circulation objective typing database Dimensions same.

[0089] The real-time weather classification matching module is further configured with a similarity calculation submodule. This submodule calculates the similarity between the real-time feature vector and the cluster center vector in the historical circulation objective classification database. The similarity between them.

[0090] In one specific implementation, the similarity is calculated using the cosine similarity method, and the similarity calculation submodule performs the calculation for each... The value is calculated using the following formula: ; in, These are high-dimensional real-time feature vectors. The first in the historical circulation objective typing database Cluster center vectors, Real-time feature vector The One portion, For the first Cluster center vectors The One portion, This represents the total dimension of the feature vector. The calculated real-time vector With the historical center vector The cosine similarity value between them ranges from between.

[0091] The real-time weather classification matching module is further configured with a best match determination submodule, which receives the five similarity values output by the similarity calculation submodule .

[0092] The best match determination submodule determines the similarity value with the maximum value among the five similarity values by comparing the five similarity values .

[0093] The best match determination submodule then outputs the index of the cluster center corresponding to the maximum similarity value, and the historical circulation type corresponding to the index is determined as the classification result of the current real-time forecast weather.

[0094] The classification result (for example, a trough-ridge pattern) is output to the subsequent fog potential forecast model construction module as a key background field classification feature input.

[0095] Referring to the accompanying Figure 3 , in an embodiment of the present application, after the weather classification and real-time matching stage is completed, the method enters the feature engineering and association rule analysis stage, aiming to extract and quantify specific physical quantities and phenomena with the highest correlation to the formation of fog potential from the meteorological data obtained in the data preparation stage and processed.

[0096] The feature engineering and association rule analysis stage first performs key meteorological element extraction, which is performed by a key meteorological element extraction module, and the input data of the key meteorological element extraction module is the historical reanalysis data, sounding data, ground observation data obtained in the data preparation stage and processed, and real-time numerical model prediction data.

[0097] The key meteorological element extraction module is configured with an atmospheric boundary layer parameter calculation submodule, which is used to analyze the vertical structure of the atmosphere.

[0098] Specifically, the atmospheric boundary layer parameter calculation submodule reads the temperature data at different height layers output by the sounding data or numerical model to construct a vertical temperature profile , the atmospheric boundary layer parameter calculation submodule identifies an inversion layer by calculating the vertical temperature gradient , when and between satisfy , the is identified as the inversion layer bottom height, and the is identified as the inversion intensity, and the atmospheric boundary layer parameter calculation submodule outputs the inversion layer bottom height as a feature. ​

[0099] The atmospheric boundary layer parameter calculation submodule further directly extracts the grid data of the boundary layer height from the diagnostic output variables of the FNL reanalysis data or numerical model prediction products.

[0100] The key meteorological element extraction module is further configured with a wind field analysis submodule. The wind field analysis submodule reads the ground wind speed data and compares it with a preset low wind speed threshold (e.g. = 3 m / s). The wind field analysis submodule outputs a Boolean value or binary feature (e.g., 1 when ≤ 3 m / s, otherwise 0), representing the ground calm condition.

[0101] The wind field analysis submodule is also used to analyze the vertical profile of the wind field. The wind field analysis submodule reads the wind direction and speed data at different levels such as the ground, 925 hPa, 850 hPa, etc. By comparing the wind vector differences between adjacent layers, the wind field analysis submodule calculates the vertical shear of the wind (including wind speed shear and wind direction shear). The wind field analysis submodule also detects whether there is a 180-degree reversal of the wind direction between the lower layer (such as the ground) and the middle layer (such as 850 hPa) to identify whether there is a local recirculation phenomenon, and outputs this phenomenon as a classification feature.

[0102] The key meteorological element extraction module is further configured with a temperature, humidity, and pressure analysis submodule. The temperature, humidity, and pressure analysis submodule reads the daily average relative humidity data and compares it with a preset high humidity threshold (e.g. = 80%). A binary feature of high humidity condition is generated.

[0103] The temperature, humidity, and pressure analysis submodule is also used to calculate the sea level pressure change. The temperature, humidity, and pressure analysis submodule reads the sea level pressure at the current time and a sea level pressure at a historical time point (e.g. hours), and calculates the sea level pressure tendency value . This value (e.g., continuously decreasing ) is output as a feature.

[0104] The key meteorological element extraction module combines all the elements calculated and extracted above (e.g., inversion layer height, boundary layer height, low wind speed binary feature, wind shear value, high humidity binary feature, pressure tendency value, etc.) into a multi-dimensional feature vector and outputs it to the subsequent association rule establishment module and fog potential prediction model construction module. ​

[0105] In one embodiment of the present application, the feature engineering and association rule analysis stage, after performing the key meteorological element extraction, further includes the establishment of association rules, aiming to construct the logical relationship between the weather background, the key element combination and the fog potential level based on the historical data statistics and the meteorological principles.

[0106] The establishment of association rules is performed by an association rule establishment module, and the input data of the association rule establishment module includes: The case library generated by the low-visibility event case library construction module, which provides the fog potential level (e.g. no fog, light fog, dense fog) of the historical events as the "result" label.

[0107] The historical weather classification results corresponding to each event in the case library (obtained by applying the logic of the historical circulation objective classification library construction module and the surface weather situation classification module to the historical data).

[0108] The key meteorological element feature vector corresponding to each event in the case library (obtained by applying the logic of the key meteorological element extraction module to the historical data).

[0109] The association rule establishment module is configured with a feature result association sub-module, which is used to integrate the above three kinds of input data, and for each time point in the historical case library, mark the corresponding "weather classification" label, "surface situation" label, "key element feature vector" and "fog potential result" label, forming a complete data set for rule mining.

[0110] The association rule establishment module is further configured with a statistical rule mining sub-module, which is used to perform an association rule mining algorithm on the complete data set.

[0111] In a specific embodiment, the statistical rule mining sub-module discovers high-frequency patterns by calculating the support and confidence of different item sets (i.e. the combination of weather classification, element condition and fog potential result), and the statistical rule mining sub-module outputs all candidate association rules that meet the preset support threshold and confidence threshold.

[0112] The association rule establishment module is further configured with a meteorological principle verification sub-module, which is used to filter the candidate association rules output by the statistical rule mining sub-module.

[0113] The meteorological principle verification sub-module pre-stores a set of meteorological mechanism constraints defined by meteorological expert knowledge (e.g. strong surface wind speed is negatively correlated with the occurrence of fog), and the meteorological principle verification sub-module compares the candidate association rules with the set of constraints, and eliminates all statistical rules that are contrary to known meteorological physical principles.

[0114] The rules verified by the meteorological principle verification sub-module are defined as final association rules and stored in an association rule base.

[0115] The rules in the association rule base have an "IF-THEN" logical structure, for example, a rule can be expressed as: IF [weather type = zonal circulation type] AND [ground weather situation = isobaric field (EQP)] AND [inversion layer height < 950 hPa] AND [ground wind speed <= 2 m / s] AND [relative humidity >= 85%] THEN [fog potential = high].

[0116] The association rule base is output to a subsequent fog potential prediction model construction module to provide logical judgment basis or as an auxiliary feature input for the model.

[0117] In an embodiment of the present application, after the feature engineering and association rule analysis stage is completed, the method enters a fog potential probability prediction model construction stage, which aims to realize quantitative probability calculation of the fog potential grade by fusing a meteorological concept model and a probability statistical model.

[0118] The fog potential probability prediction model construction stage first performs integration of a movable boundary concept model, which is performed by a movable boundary concept model integration module, and the input data of the movable boundary concept model integration module includes: multi-level temperature field, humidity field and wind field grid point data prepared by the circulation background field data preparation module and processed.

[0119] The movable boundary concept model integration module is used to simulate and quantify the interaction state between the target plain area where fog occurs and the surrounding cold and warm air masses, especially the movement and position of the air mass boundary (such as cold front, warm front or stationary front).

[0120] The movable boundary concept model integration module is configured with an air mass boundary identification sub-module. The air mass boundary identification sub-module is used to read the temperature field grid point data and calculate the horizontal temperature gradient of the target plain area and its surrounding area .

[0121] The air mass boundary identification sub-module compares the calculated horizontal temperature gradient with a preset front zone gradient threshold value When the in a region is greater than the front zone gradient threshold value , the region is identified as an air mass boundary (i.e. a front zone).

[0122] The movable boundary conceptual model integration module is further configured with a boundary state determination submodule, which is used to analyze the relative position of the frontal zone with respect to the target plain region, and the moving trend of the frontal zone. The moving trend is determined by comparing the frontal zone position at consecutive time steps.

[0123] A state rule set is pre-stored in the boundary state determination submodule. The rule set is used to map the different position and moving trend of the frontal zone into a discrete state feature.

[0124] For example, in the state rule set, it is defined that when the boundary state determination submodule identifies that the target plain region is located on the warm side of a slowly moving or stationary frontal zone, and the wind field data indicates that there is a warm and wet advection from low latitudes, the state is determined as a high potential state.

[0125] For another example, in the state rule set, it is defined that when the boundary state determination submodule identifies that the cold front has completely passed through the target plain region, the region is under the control of a cold air mass, and the wind field data indicates that there is a strong northerly wind and dry and cold advection, the state is determined as a low potential state.

[0126] If there is no obvious frontal zone affecting the target region, the state is determined as a neutral state.

[0127] The movable boundary conceptual model integration module is further configured with a state quantization encoding submodule, which is used to convert the discrete state (e.g. high potential state, low potential state, neutral state) output by the boundary state determination submodule into a numerical feature vector for subsequent calling by the probability model.

[0128] In a specific embodiment, the state quantization encoding submodule adopts a one-hot encoding method. For example, the high potential state is encoded as vector , the low potential state is encoded as vector , and the neutral state is encoded as vector .

[0129] Referring to the accompanying drawings Figure 4 In an embodiment of the present application, the fog potential probability forecast model construction stage further includes the construction and calculation of a Bayesian probability model, which is performed by a Bayesian probability model construction module, and is used to comprehensively utilize all the extracted features to calculate the posterior probability of occurrence of a specific fog potential level under the current meteorological conditions based on the principle of probability statistics.

[0130] The Bayesian probability model construction module is configured with a training data integration submodule. The training data integration submodule is used to obtain the fog potential level labels of all historical events from the low-visibility event case library (For example, =1, 2, 3 represent no fog, light fog, and dense fog, respectively.

[0131] The training data integration submodule simultaneously acquires the complete set of feature vectors corresponding to each historical event, which were processed by the previous module. The complete set of feature vectors Contains one or more feature elements Specifically, it includes: The logic of the real-time weather classification matching module is applied to the 500hPa circulation type (e.g., zonal circulation type) determined by historical data. The logic of the surface weather pattern classification module is applied to the surface weather patterns (e.g., isobaric fields) determined from historical data. The logic of the key meteorological element extraction module is applied to all key elements extracted from historical data (e.g., inversion layer height value, high humidity binary feature value). The logic of the movable boundary conceptual model integration module is applied to the boundary state encoding vector output from historical data (e.g., ).

[0132] The Bayesian probabilistic model building module is further configured with a prior probability and conditional probability calculation submodule. This submodule is used during the model training phase to traverse all historical training data integrated by the prior probability and conditional probability calculation submodule, calculate and store the two types of probability values: First, calculate each fog potential level. Prior probability This calculation is performed by statistically analyzing the level. This is accomplished by analyzing the frequency of occurrences in historical training data.

[0133] Second, calculate the potential fog level. Under certain conditions, specific feature vectors The conditional probability of occurrence, i.e., the likelihood .

[0134] In one specific implementation, to simplify computational complexity, the prior probability and conditional probability calculation submodule adopts the Naive Bayes assumption, which is defined as: given a fog potential level Under the condition that all characteristic elements (in They are independent of each other.

[0135] Based on this Naive Bayes assumption, the likelihood It is decomposed into the product of the conditional probabilities of all individual feature elements, as shown in the formula: where: is the complete feature vector, is the th fog potential level, is the th feature element in the feature vector , is the total number of feature elements, is the th feature element under the fog potential level .

[0136] The prior probability and conditional probability calculation submodule traverses the historical training data, calculates and stores the conditional probability of each feature element and each fog potential level , forming a conditional probability table. For continuous features (such as temperature values), the prior probability and conditional probability calculation submodule first discretizes them (for example, binning), or fits a probability density function (such as a Gaussian distribution) to calculate .

[0137] The Bayesian probability model construction module is further configured with a posterior probability calculation submodule, which is used in the real-time prediction phase.

[0138] The posterior probability calculation submodule receives the current real-time prediction feature vector . The posterior probability calculation submodule applies Bayes' theorem, combining the prior probability and the conditional probability table stored in the prior probability and conditional probability calculation submodule, to calculate the posterior probability of each fog potential level , as shown in the formula: ; where: is the posterior probability, i.e., the probability of the fog potential level occurring under the condition of observing the real-time feature ; is the likelihood, retrieved and multiplied from the conditional probability table according to the formula; is the prior probability; is the probability of the evidence, which serves as a normalization factor, calculated as follows: where: denotes the summation over all possible fog potential levels .

[0139] The posterior probability calculation sub-module outputs a set of probability values, for example: (no fog = 0.1, (light fog = 0.3, (thick fog = 0.6). The set of probability values is the final output of the fog potential probability forecast result of the present application.

[0140] In an embodiment of the present application, after the method completes the fog potential probability forecast model construction stage, it enters the forecast output and business application stage, aiming to convert the quantitative probability result calculated by the model into a qualitative forecast product that can be understood by users, and to perform business applications.

[0141] The forecast output and business application stage is executed by a forecast output and business application module, which is configured with a probability level conversion sub-module. The input data of the probability level conversion sub-module is the set of posterior probabilities of the fog potential level for the current forecast period output by the Bayesian probability model construction module .

[0142] The probability level conversion sub-module is preset with a set of discrimination rules for mapping continuous probability values to discrete optimal forecast levels.

[0143] In a specific embodiment, the discrimination rules first apply the maximum posterior probability principle, that is: ; The rule takes the fog potential level corresponding to the maximum probability value in the probability set (for example, if is the maximum, thick fog is output) as the dominant forecast level.

[0144] The probability level conversion sub-module is also configured with a secondary level discrimination logic, aiming to solve the case where the probability value of the dominant level is high but does not reach an absolute advantage (for example, lower than the preset confidence threshold , such as ). When the maximum probability value , the logic will simultaneously output the two levels corresponding to the maximum probability value and the second largest probability value (for example, thick fog and light fog are output), to prompt the forecaster that the result is uncertain.

[0145] The forecast output and business application module is further configured with a product generation and visualization sub-module, which is used to convert the qualitative forecast result output by the probability level conversion sub-module into a graphical, standard meteorological forecast product.

[0146] Specifically, for the prediction based on the grid data, the product generation and visualization submodule colors the prediction level (or probability value) at each grid point and superimposes it on the geographical base map of the target plain area to generate a fog potential level distribution map, showing the potential spatial distribution of different levels of fog in different geographical areas in the future period.

[0147] The product generation and visualization submodule also generates a time series prediction curve to show the trend of the probability value of different fog potential levels at a specific site over time, and to grasp the evolution process of the fog.

[0148] The prediction output and business application module is finally configured with a business decision support submodule, which pushes the generated graphical prediction products to the downstream business system through a standardized data interface.

[0149] The downstream business system includes but is not limited to the warning system of the traffic management department (for starting the highway, airport, and port traffic control plan), the air quality assessment system of the environmental protection department, and the disaster prevention system of the agricultural production department. The business decision support submodule realizes the closed loop of the method from meteorological data processing to actual business application.

[0150] In an embodiment of the present application, after the prediction output and business application stage is completed, the method further includes a prediction effect evaluation and model optimization stage, which aims to establish a continuous feedback and iteration mechanism. By quantitatively comparing the prediction results of the model with the subsequent real-time observations, the performance of the model is evaluated, and the basis for parameter adjustment and structure optimization of the model is provided. The prediction effect evaluation and model optimization stage is executed by a prediction effect evaluation and model optimization module.

[0151] The prediction effect evaluation and model optimization stage is configured with a prediction real-time matching submodule. The prediction real-time matching submodule is used to automatically obtain the real-time visibility observation data of the corresponding time and place from the fog observation data collection module after the prediction time limit is reached.

[0152] The prediction real-time matching submodule accurately matches each prediction event (including the time, place, and predicted fog potential level) with the corresponding real-time observation visibility data to form a prediction real-time data pair, which is the basis for all subsequent evaluation calculations.

[0153] The prediction effect evaluation and model optimization module is further configured with a multi-classification evaluation index calculation submodule. Since the fog potential prediction is a multi-classification problem (e.g., no fog, light fog, and dense fog), the multi-classification evaluation index calculation submodule first constructs a confusion matrix for each fog potential level based on the prediction real-time data pair.

[0154] Based on the confusion matrix, the multi-classification evaluation index calculation submodule calculates a set of professional classification forecast verification indexes widely used in the meteorological field for each fog potentiality grade. These indexes include but are not limited to: Hit rate: measures the ability of the model to successfully predict a certain grade of fog event.

[0155] False alarm rate: measures the tendency of the model to falsely predict a certain grade of fog event for non- that grade of fog event.

[0156] Critical success index: a comprehensive measure of both hit and false alarm, it is a commonly used index to measure the overall prediction skill of the model for a certain grade of fog event.

[0157] The multi-classification evaluation index calculation submodule also calculates the overall accuracy of the model to evaluate the comprehensive prediction performance of the model for all grades of events.

[0158] The forecast effect evaluation and model optimization module is further configured with a model parameter optimization submodule, which receives the performance evaluation report output by the multi-classification evaluation index calculation submodule.

[0159] A set of hyperparameter optimization mechanisms is preset in the model parameter optimization submodule, which is triggered when the evaluation report shows that a key indicator of the model (e.g. the SI index of dense fog) is below a preset threshold of business usability.

[0160] This mechanism can automatically or under human guidance fine-tune the key parameters involved in the process of the method of the present application. Examples of parameters that can be adjusted include: The number of categories of the clustering algorithm in the historical circulation objective typing library construction module .

[0161] The threshold of physical quantities used to define conditions such as calm wind and high humidity in the key meteorological element extraction module.

[0162] The confidence threshold used to determine the uncertainty of the prediction results in the prediction output and business application module .

[0163] After adjusting the parameters, the model will be retrained and backtested using historical data, and evaluated again by the multi-classification evaluation index calculation submodule to determine the optimal combination of parameters.

[0164] The pre-forecast effect evaluation and model optimization module is also configured with a feature importance analysis submodule, which is used to in-depth analyze the cause of the prediction error. By analyzing historical cases of erroneous predictions, the feature importance analysis submodule uses a specific algorithm (e.g., analyzing the conditional probability contribution of different features in a Bayesian model, or using the permutation importance method) to evaluate the influence weight of each input feature (e.g., 500hPaa circulation type, inversion intensity, boundary layer height, etc.) on prediction accuracy.

[0165] The analysis results can guide researchers to identify features that contribute less or even have a negative impact on model performance, providing scientific basis for subsequent model version iteration (e.g., introducing new, more indicative meteorological elements, or eliminating redundant features), thereby achieving continuous optimization and evolution of the model.

Claims

1. A method for predicting the fog potential in plain areas based on weather typing and changes in meteorological elements, characterized by, The method comprises the following steps: S1, performing a data preparation stage: obtaining and processing historical meteorological data and real-time forecast data including multi-level temperature field grid point data, and constructing a low-visibility event case library; S2, performing a weather typing stage: the weather typing stage comprises: based on historical 500 hPa potential height field data set, establishing a historical circulation objective typing library by an objective typing method; classifying the ground weather situation based on historical and real-time ground meteorological element and sea level pressure grid point data; S3, performing real-time weather typing matching: calculating the similarity of real-time model forecast 500 hPa potential height field grid point data and the historical circulation objective typing library to determine the circulation type to which the current forecast period belongs; S4, performing a feature engineering stage: extracting and quantifying key meteorological elements related to fog potential formation from historical meteorological data and real-time forecast data; S5, performing a fog potential probability forecast stage: based on the determined circulation type, key meteorological elements in the weather typing stage, constructing and running a Bayesian probability model to calculate the posterior probability of the fog potential level as the forecast result.

2. The plain area fog potential forecast method based on weather typing and meteorological element changes according to claim 1, characterized by, In the S1 step, the construction of the low-visibility event case library comprises: By comparing the historical visibility data with the preset visibility threshold, the continuous time period meeting the low-visibility condition is identified, and the continuous time period is defined as a low-visibility event; Defining a complete time window for a low-visibility event including a preset time period before the event occurs; Extracting all relevant meteorological element data within the complete time window from the historical meteorological data, and storing the extracted all relevant meteorological element data in association with the label of the low-visibility event.

3. The method for predicting the potential of fog in plain areas based on weather typing and meteorological element changes according to claim 1, characterized in that, In the S2 step, the establishment of the historical circulation objective typing library comprises: flattening and converting the historical 500 hPa potential height field grid point data at each time point into a high-dimensional feature vector; Using K-means clustering algorithm, clustering analysis is performed on the set of all historical high-dimensional feature vectors to obtain a preset number of clusters; Store the cluster center vectors of all clusters, and the set of cluster center vectors of all clusters constitutes the historical circulation objective typing library.

4. The plain area fog potential forecast method based on weather typing and meteorological element change according to claim 3, characterized by, In the S3 step, real-time weather typing matching comprises: Converting the real-time model forecast 500 hPa potential height field grid point data into a high-dimensional real-time feature vector; Using cosine similarity method, the similarity values between the high-dimensional real-time feature vector and the cluster center vectors of all clusters in the historical circulation objective typing library are calculated; Determine the cluster center vector with the maximum similarity value, and determine the circulation type represented by the cluster center vector as the typing result of the current forecast period.

5. The method for predicting the potential of fog in plain areas based on weather typing and meteorological element changes according to claim 1, characterized in that, In the S2 step, the classification of the ground weather situation comprises: according to the preset meteorological definition, the ground weather pattern is classified into at least one of the following: isallobaric field, cold front, high pressure bottom, high pressure back or inverted trough.

6. The method for predicting the potential of fog in plain areas based on weather typing and meteorological element changes according to claim 1, characterized in that, In the S4 step, extracting and quantifying key meteorological elements related to fog potential formation comprises: From the historical meteorological data and real-time forecast data, analyzing the inversion layer bottom height and inversion intensity identified by the vertical temperature profile; From historical meteorological data and real-time forecast data, ground wind speed data and wind field vertical profile identification ground static wind conditions or local recirculation phenomenon are analyzed.

7. The method for predicting the potential of fog in plain areas based on weather typing and meteorological element changes according to claim 1, characterized in that, The S4 step further includes integration of a movable boundary concept model: Read the temperature field grid point data obtained in the data preparation stage, calculate the horizontal temperature gradient of the target plain area and its periphery by comparing the temperature values of adjacent grid points in the horizontal direction of the temperature field grid point data, and compare the calculated horizontal temperature gradient with the preset frontal zone gradient threshold value, identify the area with a horizontal temperature gradient greater than the frontal zone gradient threshold value as the air mass boundary; According to the relative position and movement trend of the air mass boundary relative to the target plain area, the discrete state feature representing the interaction state of cold and warm air masses is determined and output, which is input as a key meteorological element into the Bayesian probability model of the fog potential probability prediction stage.

8. The method for predicting the potential of fog in plain areas based on weather typing and meteorological element changes according to claim 1, characterized in that, In the S5 step, the Bayesian probability model is constructed and run, including: In the model training stage, the prior probability of the fog potential grade is counted and stored by traversing the low-visibility event case library, and the conditional probability of each feature appearing under the condition of a given fog potential grade is counted and stored.

9. The method for predicting the fog potential in the plain area based on the weather typing and the meteorological element change according to claim 8, characterized in that, In the S5 step, the posterior probability of the fog potential grade is calculated, including: In the real-time prediction stage, the current real-time prediction feature vector is received; Applying Bayes' theorem, the prior probability and the conditional probability are combined to calculate the posterior probability of each fog potential grade under the condition of observing the current real-time prediction feature vector.

10. The method for predicting the potential of fog in plain areas based on weather typing and meteorological element changes according to claim 1, characterized in that, The S1 step further includes performing quality control on the obtained raw data: The error values and abnormal values in the historical meteorological data and real-time forecast data are removed, and the missing value filling processing is performed on the error values and abnormal values; The missing value filling processing includes: judging the length of the missing sequence, when the length of the missing sequence is less than or equal to the preset short-time missing threshold, the time interpolation method is used for filling; When the length of the missing sequence is greater than the preset short-time missing threshold, the spatial nearest neighbor interpolation method is used for filling.

Images