Meteorological state prediction method and system based on space-time interpolation coupling model

Abstract

The invention relates to the technical field of data analysis, and provides a meteorological state prediction method and system based on a spatio-temporal interpolation coupling model.According to the embodiment of the invention, after credibility authentication of a meteorological forecast field database is passed, a historical meteorological forecast data set is called, and a dynamic influence area is divided by combining geographic position coordinate information of a target observation point; determining a cloud cover ratio time sequence change trend under the condition of no rainfall in the region, and building a meteorological state prediction model of space-time interpolation coupling by combining the visibility state parameters; and then inputting the current meteorological observation data set into the meteorological state prediction model, carrying out spatial-temporal feature coupling calculation to obtain a current state feature vector, and generating a meteorological state prediction result that the target observation point contains a nepheline state within a preset time period based on the current state feature vector, thereby effectively improving the accuracy and reliability of meteorological prediction.

Description

Technical Field

[0001] This application belongs to the field of data analysis technology, specifically relating to a meteorological state prediction method and system based on a spatiotemporal interpolation coupling model. Background Technology

[0002] In the field of meteorological condition forecasting, traditional forecasting methods mainly rely on single meteorological elements or simple statistical models, which have limited predictive capabilities for complex meteorological conditions. For example, some methods only consider common meteorological elements such as temperature and humidity, ignoring important influencing factors such as cloud cover and visibility that affect special meteorological conditions, such as the appearance of sunset glow.

[0003] Moreover, most existing forecasting models use fixed regional divisions and do not fully consider the spatial propagation characteristics of meteorological elements, which affects the accuracy of forecasts under different geographical locations and meteorological environments.

[0004] Furthermore, the utilization of meteorological data is insufficient, often simply relying on current meteorological observations without delving into the potential information within historical meteorological data. In terms of forecasting the state of afterglow, traditional methods lack effective means and cannot accurately predict key information such as the probability of occurrence, duration, and intensity changes of afterglow, thus failing to meet the needs of practical applications.

[0005] Therefore, a more accurate and reliable method for forecasting weather conditions is needed to address the problems existing in current technologies. Summary of the Invention

[0006] This application provides a meteorological state prediction method and system based on a spatiotemporal interpolation coupling model.

[0007] In a first aspect, embodiments of this application provide a meteorological state prediction method based on a spatiotemporal interpolation coupling model, applied to a meteorological state prediction system, the method comprising: After passing the trusted authentication of the meteorological forecast field database, the historical meteorological forecast dataset stored in the meteorological forecast field database is invoked; The dynamic influence area is divided based on the historical weather forecast dataset and the geographical coordinates of the target observation point. The dynamic influence area is centered on the target observation point and is affected by the spatial propagation characteristics of meteorological elements. Determine the time series variation trend of cloud cover ratio within the dynamic influence area under no precipitation conditions, and build a spatiotemporal interpolation coupled meteorological state prediction model by combining the time series variation trend of cloud cover ratio and the visibility state parameters of the dynamic influence area. The current meteorological observation dataset is input into the meteorological state prediction model. The current state feature vector is obtained by performing spatiotemporal feature coupling calculation through the meteorological state prediction model. Based on the current state feature vector, the meteorological state prediction result of the target observation point within a preset time period is generated. The meteorological state prediction result includes the state of the sunset glow.

[0008] Secondly, embodiments of this application provide a weather condition prediction system, which includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor performs the steps of the above-described method.

[0009] Thirdly, embodiments of this application provide a computer-readable storage medium including a computer program, which, when run on a weather state forecasting system, causes the weather state forecasting system to perform the steps of the above-described method.

[0010] This application significantly improves the accuracy and reliability of meteorological state prediction, especially the prediction of overcast conditions. First, historical meteorological forecast datasets are accessed after trusted authentication of the meteorological forecast field database, ensuring the reliability and authority of the data source while also guaranteeing data security within the meteorological forecast field database. Dynamic influence areas are delineated based on historical meteorological forecast datasets and the geographical coordinates of the target observation point, fully considering the spatial propagation characteristics of meteorological elements. This allows the model to more accurately capture changes in the meteorological environment around the target observation point, avoiding the limitations of fixed-area delineation in traditional methods. The time-series variation trend of cloud cover ratio under no-precipitation conditions within the dynamic influence area is determined, and a spatiotemporal interpolation-coupled meteorological state prediction model is constructed by combining the visibility state parameters of the dynamic influence area. This spatiotemporal coupling method comprehensively considers changes in meteorological elements in both time and space, enabling a more comprehensive and accurate reflection of the evolution of meteorological conditions. The current meteorological observation dataset is input into the meteorological state prediction model for spatiotemporal feature coupling calculation to obtain the current state feature vector, thereby generating a meteorological state prediction result for the target observation point within a preset time period, including overcast conditions. This allows for timely and accurate prediction of overcast conditions, providing strong support for related meteorological services and decision-making. Attached Figure Description

[0011] Figure 1 This is a flowchart illustrating a meteorological state prediction method based on a spatiotemporal interpolation coupling model provided in an embodiment of this application.

[0012] Figure 2 This is a schematic diagram of the structure of a weather condition prediction system provided in an embodiment of this application. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of this application will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this application. Obviously, the described embodiments are only some embodiments of the technical solutions of this application, and not all embodiments. Based on the embodiments recorded in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the technical solutions of this application.

[0014] See Figure 1 This is a meteorological state prediction method based on a spatiotemporal interpolation coupling model provided in the embodiments of this application. This method can be applied to a meteorological state prediction system. The specific process is as follows: steps 110-140.

[0015] Step 110: After passing the trusted authentication of the meteorological forecast field database, call the historical meteorological forecast dataset stored in the meteorological forecast field database.

[0016] In this embodiment, when the meteorological state prediction system needs to obtain historical meteorological forecast data, it sends a data retrieval request to the meteorological forecast field database. As a crucial resource storing a large amount of historical meteorological forecast data, the meteorological forecast field database conducts a trusted authentication of the meteorological state prediction system based on the received data retrieval request to ensure the reliability and security of the data provided. The authentication process involves rigorous review in multiple aspects. The meteorological forecast field database assesses the legality and reliability of the meteorological state prediction system's data source, examines whether its data usage purpose is legitimate, and whether its data processing methods comply with regulations. Simultaneously, the meteorological forecast field database also checks the authority of its own data source, the scientific validity of its data collection methods, and the timeliness of its data updates. Furthermore, it examines its own data storage and management methods to determine whether it can guarantee the integrity and accuracy of the data.

[0017] It is understandable that the meteorological forecasting database will only authorize the meteorological state prediction system to access data after the meteorological state prediction system successfully passes these rigorous authentication processes. The historical meteorological forecast dataset in this application embodiment covers a wealth of meteorological element information, including temperature, humidity, air pressure, cloud cover, precipitation, etc., which record the meteorological conditions at different times and locations.

[0018] Step 120: Divide the dynamic influence area according to the historical weather forecast dataset and the geographical coordinate information of the target observation point. The dynamic influence area is centered on the target observation point and is affected by the spatial propagation characteristics of meteorological elements.

[0019] After successfully acquiring historical weather forecast datasets, the weather state prediction system will combine the geographical coordinates of the target observation point to delineate the dynamic influence area. The target observation point is the specific location where the afterglow state prediction needs to be performed; its geographical coordinates accurately identify its position on the Earth's surface. Meteorological elements have spatial propagation characteristics, and meteorological elements at different locations will influence each other. Therefore, it is not only necessary to consider the meteorological conditions of the target observation point itself, but also the changes in meteorological elements within a certain range around it.

[0020] Centered on the target observation point, the meteorological condition prediction system comprehensively considers the propagation range and degree of influence of meteorological elements to determine the dynamic influence area. For example, wind can bring distant meteorological elements such as clouds and water vapor to the vicinity of the target observation point. Therefore, when delineating the dynamic influence area, factors such as wind direction and intensity need to be considered. The size and shape of this area are not fixed and will dynamically adjust with changes in meteorological elements. When the wind is strong, the dynamic influence area may expand to cover the propagation and influence of meteorological elements from a greater distance; while when the wind is weak, the influence area may shrink relatively. By dynamically adjusting the influence area, the meteorological environment around the target observation point can be more accurately reflected.

[0021] Step 130: Determine the time series variation trend of cloud cover ratio in the dynamic influence area under no precipitation conditions, and build a spatiotemporal interpolation coupled meteorological state prediction model by combining the time series variation trend of cloud cover ratio and the visibility state parameters of the dynamic influence area.

[0022] After delineating the dynamic influence area, the next step for the meteorological state prediction system is to determine the time series trend of cloud cover ratio within that area under precipitation-free conditions. Cloud cover ratio refers to the proportion of cloud cover in the sky, and its changes significantly impact the formation of sunset glow. Different cloud cover ratios affect the scattering and transmission of sunlight, thus influencing the color, intensity, and visibility of the sunset glow. Simultaneously, the visibility state parameter of the dynamic influence area also plays a crucial role in the sunset glow's condition. Visibility reflects the clarity of the atmosphere; a clear atmospheric environment facilitates the propagation and observation of sunset glow, while poor visibility reduces its visibility and intensity. The meteorological state prediction system combines these two important factors to build a spatiotemporal interpolation-coupled meteorological state prediction model. This model comprehensively considers changes in meteorological elements across time and space. Through analysis and learning from historical data, it establishes the correlation between the time series trend of cloud cover ratio, visibility state parameters, and meteorological conditions (including sunset glow conditions), thereby more accurately predicting the meteorological conditions at the target observation point.

[0023] In an optional embodiment, determining the time series variation trend of cloud cover ratio within the dynamically affected area under no-precipitation conditions includes: Step 131: Filter the precipitation status identifiers in the historical weather forecast dataset and extract the meteorological data subset with precipitation status identifiers of no precipitation in the dynamic influence area.

[0024] To accurately determine the trend of cloud cover ratio changes under precipitation-free conditions, the meteorological state prediction system needs to filter precipitation status identifiers in historical weather forecast datasets. Precipitation significantly impacts meteorological elements such as cloud cover and visibility. In the presence of precipitation, cloud morphology and distribution change markedly, potentially masking the normal trend of cloud cover ratio changes. Therefore, the meteorological state prediction system searches for meteorological data within the dynamically affected area in the historical weather forecast dataset and filters them based on precipitation status identifiers. This filtering process is based on precipitation status identifiers; the system checks the precipitation status of each data record individually, selecting those that meet the precipitation-free condition to form a new meteorological data subset.

[0025] Step 132: Extract the cloud cover ratio observation values ​​and corresponding timestamp information from the meteorological data subset, and generate the original cloud cover ratio sequence arranged in ascending order of timestamps.

[0026] After obtaining a subset of meteorological data with no precipitation, the meteorological state prediction system extracts the cloud cover ratio observations and their corresponding timestamps from this subset. The timestamps precisely record the exact time of each cloud cover ratio observation, which is crucial for analyzing the changing patterns of the cloud cover ratio over time. The meteorological state prediction system sorts the cloud cover ratio observations in ascending order of timestamps, generating a raw cloud cover ratio sequence. This sorting process ensures that the cloud cover ratio data is arranged chronologically, allowing the system to clearly see the changes in the cloud cover ratio at different points in time. Through the raw cloud cover ratio sequence, the system can initially observe the approximate trend of the cloud cover ratio within a day or over a period of time, such as the difference in cloud cover ratio changes between day and night, or the characteristics of changes in different seasons.

[0027] Step 133: Perform time continuity verification on the original cloud cover ratio sequence, delete discrete data points with timestamp intervals exceeding a preset threshold, and obtain a time-continuous cloud cover ratio calibration sequence.

[0028] The original cloud cover ratio sequence may contain discontinuities, meaning data may be missing at certain points in time or the time intervals may be too large. This discontinuity can affect the accurate analysis of the cloud cover ratio time series trend. Therefore, meteorological state prediction systems need to perform time continuity verification on the original cloud cover ratio sequence. The system sets a preset threshold to determine if the time intervals are too large. For discrete data points with timestamp intervals exceeding the preset threshold, the system will delete them from the original cloud cover ratio sequence. This is because these discrete data points may be due to anomalies during data acquisition, and their presence can interfere with the judgment of normal trends. By deleting these discrete data points, the system obtains a time-continuous cloud cover ratio calibration sequence, which more accurately reflects the continuous changes in cloud cover ratio over time.

[0029] Step 134: Calculate the mean and standard deviation of the cloud cover ratio calibration sequence within the sliding time window, identify and correct abnormal fluctuation points in the cloud cover ratio calibration sequence based on the mean and standard deviation, and generate a smoothed cloud cover ratio time series.

[0030] After obtaining a time-continuous cloud cover ratio calibration sequence, the meteorological state prediction system uses a sliding time window method to further process the data. The sliding time window is a fixed-length period of time; the system moves this window across the cloud cover ratio calibration sequence, calculating the mean and standard deviation of the cloud cover ratio observations within the window each time. The mean reflects the average level of the cloud cover ratio within the window, while the standard deviation reflects the dispersion of the cloud cover ratio observations relative to the mean. Based on the mean and standard deviation, the system can identify anomalous fluctuation points in the cloud cover ratio calibration sequence. Anomalous fluctuation points are those that differ significantly from surrounding data points, possibly due to sudden meteorological events or data acquisition errors. The system sets a judgment criterion based on the mean and standard deviation; when a data point deviates from the mean by more than a certain multiple of the standard deviation, it is identified as an anomalous fluctuation point. For identified anomalous fluctuation points, the system uses appropriate methods to correct them, such as interpolation based on the values ​​of adjacent data points or smoothing. Through this process, the system generates a smoothed cloud cover ratio time series, which removes the interference of anomalous fluctuations and better reflects the true trend of cloud cover ratio changes.

[0031] Step 135: Perform trend fitting processing on the smoothed cloud cover ratio time series, use a time series decomposition algorithm to separate the steady-state trend component and the periodic fluctuation component, and use the superposition result of the steady-state trend component and the periodic fluctuation component as the cloud cover ratio time series change trend.

[0032] The meteorological state prediction system performs trend fitting on the smoothed cloud cover ratio time series to further analyze its changing trend. Specifically, a time series decomposition algorithm is used to decompose the cloud cover ratio time series into a steady-state trend component and a periodic fluctuation component. The steady-state trend component reflects the overall changing trend of the cloud cover ratio over a longer period, such as whether the cloud cover ratio is gradually increasing or decreasing; while the periodic fluctuation component reflects the regular changes in the cloud cover ratio within a certain period, such as daily, weekly, or yearly periodic changes.

[0033] By decomposing the cloud cover ratio time series, the system can gain a clearer understanding of the different components of cloud cover ratio changes. Finally, the system superimposes the steady-state trend component and the periodic fluctuation component to obtain the cloud cover ratio time series change trend, which can comprehensively reflect the long-term change trend and periodic change characteristics of cloud cover ratio in the time dimension.

[0034] In another optional embodiment, the step of constructing a spatiotemporally interpolated coupled meteorological state prediction model by combining the cloud cover ratio time series variation trend and the visibility state parameters of the dynamically affected area includes: Step 136: Extract visibility observation values ​​within the dynamic influence area from the historical weather forecast dataset, and generate a visibility state parameter sequence aligned with the timestamp of the cloud cover ratio time series change trend.

[0035] Meteorological state prediction systems require extracting visibility observations from historical weather forecast datasets within dynamically affected areas. Visibility observations reflect atmospheric clarity and play a crucial role in predicting the state of afterglow. To effectively combine the cloud cover ratio time-series trend with visibility state parameters, the system generates a sequence of visibility state parameters aligned with the timestamps of the cloud cover ratio time-series trend. This means the system searches for corresponding visibility observations in the historical weather forecast dataset based on the timestamp information in the cloud cover ratio time-series trend and arranges these observations in the same chronological order, forming a new sequence of visibility state parameters. Through timestamp alignment, the system ensures the consistency of these two important meteorological elements—cloud cover ratio and visibility—across the time dimension.

[0036] Step 137: Perform time feature encoding processing on the cloud cover ratio time series change trend, and convert the cloud cover ratio time series change trend into a time-dependent time feature vector through a time series embedding algorithm.

[0037] To better utilize the time-series variation trend of cloud cover ratio for meteorological state forecasting, the meteorological state forecasting system performs time feature encoding on it. A time-series embedding algorithm is employed, which transforms the cloud cover ratio time-series variation trend into a time-dependent time feature vector. This time feature vector captures the variation information of cloud cover ratio over time, reflecting the correlation and dependence between different time points. During processing, the system analyzes the characteristics of the cloud cover ratio time-series variation trend and transforms it into a vector form suitable for machine learning models. This time feature vector contains important information such as the steady-state trend component and periodic fluctuation component of the cloud cover ratio.

[0038] Step 138: Perform spatial feature extraction processing on the visibility state parameter sequence, calculate the visibility correlation degree between different sub-regions based on the geographical coordinate information of the dynamic influence area, and generate a spatial feature vector with spatial distribution characteristics.

[0039] The meteorological state prediction system extracts spatial features from the visibility state parameter sequence. Since the dynamic influence area is a region with a certain spatial range, visibility may vary at different locations, and there may be correlations between the visibility of different sub-regions. The system analyzes the visibility correlation between different sub-regions based on the geographic coordinate information of the dynamic influence area. The geographic coordinate information accurately identifies the spatial location of each location within the dynamic influence area. The system can use this coordinate information to calculate the distance, relative position, and other relationships between different sub-regions, and then analyze the degree of visibility correlation between them. Through the calculation of the visibility correlation, the system converts the visibility state parameter sequence into a spatial feature vector with spatial distribution characteristics. This spatial feature vector reflects the visibility distribution at different locations within the dynamic influence area and the interrelationships between them.

[0040] Step 139: Construct a spatiotemporal feature coupling layer. Input the temporal feature vector and the spatial feature vector into the spatiotemporal feature coupling layer. Calculate the mutual information weights of the temporal feature vector and the spatial feature vector through a cross-attention mechanism to generate a coupled feature vector that integrates spatiotemporal correlation information. Using the coupled feature vector as input, construct a meteorological state prediction model that includes a feature mapping layer and an output prediction layer. The feature mapping layer is used to map the coupled feature vector to a state feature space of a preset dimension. The output prediction layer is used to generate a meteorological state prediction probability distribution based on the feature distribution of the state feature space.

[0041] Optionally, the meteorological state prediction system constructs a spatiotemporal feature coupling layer. This layer couples the temporal and spatial feature vectors to comprehensively consider meteorological element information across both time and space dimensions. After inputting the temporal and spatial feature vectors into the spatiotemporal feature coupling layer, the system calculates their mutual information weights using a cross-attention mechanism. This cross-attention mechanism analyzes the interrelationships between the temporal and spatial feature vectors, determines the importance of each feature vector in different dimensions, and thus assigns them corresponding weights.

[0042] By calculating mutual information weights, the system can effectively fuse information in the time and space dimensions to generate a coupled feature vector that integrates spatiotemporal correlation information. This coupled feature vector contains information on the temporal variation of cloud cover ratio and the spatial distribution of visibility, and is an important result of spatiotemporal feature coupling.

[0043] Using coupled feature vectors as input, the meteorological state prediction system constructs a meteorological state prediction model comprising a feature mapping layer and an output prediction layer. The feature mapping layer maps the coupled feature vectors to a state feature space of a preset dimension. In this process, the feature mapping layer transforms and processes the coupled feature vectors, converting them into a feature representation more suitable for meteorological state prediction. The output prediction layer then generates a meteorological state prediction probability distribution based on the feature distribution in the state feature space. Based on the feature information in the state feature space, the output prediction layer calculates the probability of different meteorological states (including overcast skies) occurring, providing a basis for the final meteorological state prediction.

[0044] As one implementation, the construction of the spatiotemporal feature coupling layer involves inputting the temporal feature vector and the spatial feature vector into the spatiotemporal feature coupling layer, calculating the mutual information weights of the temporal feature vector and the spatial feature vector through a cross-attention mechanism, and generating a coupled feature vector that integrates spatiotemporal correlation information, including: Step 1391: Divide the time feature vector into multiple time segment feature units according to the time step, and divide the spatial feature vector into multiple spatial grid feature units according to the geographical location.

[0045] In this step, the meteorological state prediction system divides the time feature vector into multiple time segment feature units according to the time step. The time step is a preset time period, and the system divides the time feature vector into multiple small segments based on this time step; each small segment is a time segment feature unit. Each time segment feature unit contains information on the temporal change of cloud cover ratio within a specific time period, reflecting the characteristics of cloud cover ratio within that time period.

[0046] Simultaneously, the system divides the spatial feature vector into multiple spatial grid feature units based on geographical location. The dynamically affected area is divided into multiple smaller spatial grids, each corresponding to a spatial grid feature unit. Each spatial grid feature unit contains visibility information for that spatial location and its correlation with other spatial locations, reflecting the spatial distribution characteristics of visibility. Through this division method, the system refines both the temporal and spatial feature vectors.

[0047] Step 1392: Calculate the cosine similarity between each time segment feature unit and each spatial grid feature unit to generate a spatiotemporal similarity matrix; based on the spatiotemporal similarity matrix, calculate the spatial attention weight of each time segment feature unit to different spatial grid feature units through row normalization, and calculate the temporal attention weight of each spatial grid feature unit to different time segment feature units through column normalization.

[0048] As we can understand, the meteorological state prediction system calculates the cosine similarity between each time-segment feature unit and each spatial grid feature unit, generating a spatiotemporal similarity matrix. Cosine similarity is an index that measures the degree of similarity between two vectors. By calculating the cosine similarity between time-segment feature units and spatial grid feature units, the system can understand the degree of correlation between them. The spatiotemporal similarity matrix records the similarity information between all time-segment feature units and spatial grid feature units; each element in the matrix represents the similarity between a time-segment feature unit and a spatial grid feature unit.

[0049] Based on the spatiotemporal similarity matrix, the system performs row normalization and column normalization. Row normalization processes each row of the spatiotemporal similarity matrix, calculating the spatial attention weight of each time-segment feature unit to different spatial grid feature units. Through row normalization, the system can determine the relative importance of each spatial grid feature unit when coupled with different spatial grid feature units. Column normalization processes each column of the spatiotemporal similarity matrix, calculating the temporal attention weight of each spatial grid feature unit to different time-segment feature units. Through column normalization, the system can determine the relative importance of each spatial grid feature unit when coupled with different time-segment feature units.

[0050] Step 1393: The time segment feature unit is weighted and summed with the corresponding spatial attention weight to generate a time-guided spatial feature vector; the spatial grid feature unit is weighted and summed with the corresponding temporal attention weight to generate a space-guided temporal feature vector; the time-guided spatial feature vector and the space-guided temporal feature vector are added element by element to generate a coupled feature vector that integrates spatiotemporal correlation information.

[0051] The meteorological state prediction system performs a weighted summation of time-segment feature units and their corresponding spatial attention weights. Each time-segment feature unit is weighted according to its corresponding spatial attention weight, and then these weighted spatial grid feature units are summed to generate a time-guided spatial feature vector. This vector reflects the guiding role of information in the time dimension on spatial features and reflects the comprehensive situation of spatial features at different time points.

[0052] Simultaneously, the system performs a weighted summation of spatial grid feature units and their corresponding temporal attention weights. Each spatial grid feature unit weights different temporal segment feature units according to its corresponding temporal attention weight, and then sums these weighted temporal segment feature units to generate a spatially guided temporal feature vector. This vector reflects the guiding role of spatial dimension information on temporal features and reflects the comprehensive situation of temporal features at different spatial locations.

[0053] Finally, the system performs element-wise addition of the time-guided spatial feature vector and the space-guided time feature vector. During the addition process, the system adds corresponding elements of the two vectors to generate a coupled feature vector that integrates spatiotemporal correlation information. This coupled feature vector combines information from both the temporal and spatial dimensions, effectively coupling the temporal and spatial features.

[0054] Step 140: Input the current meteorological observation dataset into the meteorological state prediction model, perform spatiotemporal feature coupling calculation through the meteorological state prediction model to obtain the current state feature vector, and generate the meteorological state prediction result of the target observation point within a preset time period based on the current state feature vector. The meteorological state prediction result includes the state of the sunset glow.

[0055] Optionally, the meteorological state prediction system inputs the current meteorological observation dataset into the meteorological state prediction model. The current meteorological observation dataset is real-time acquired meteorological data, containing information on current meteorological elements within the target observation point and the dynamically affected area, such as current cloud cover ratio and visibility. Upon receiving the input, the meteorological state prediction model performs spatiotemporal feature coupling calculations. This calculation process comprehensively considers changes in meteorological elements in both time and space, utilizing previously constructed modules such as the spatiotemporal feature coupling layer to process the input data. Through spatiotemporal feature coupling calculations, the system obtains the current state feature vector. The current state feature vector is a comprehensive feature representation of the current meteorological state, containing important information such as the temporal characteristics of the current cloud cover ratio and the spatial characteristics of the current visibility.

[0056] Based on the current state feature vector, the meteorological state prediction model generates meteorological state predictions for the target observation point within a preset time period. The preset time period is a pre-defined time frame, such as the next few hours or days. The meteorological state prediction results include forecast information for various meteorological states, such as the probability of occurrence, duration, and intensity of sunset glow. In this way, the system can provide users with accurate sunset glow predictions.

[0057] As one embodiment, inputting the current meteorological observation dataset into the meteorological state prediction model includes: Step 1411: Perform spatiotemporal dimension verification on the current meteorological observation dataset to ensure that the current meteorological observation dataset includes a spatial coverage range that matches the dynamic impact area and a time sampling interval consistent with the time granularity of the cloud cover ratio time series change trend.

[0058] In this embodiment, spatiotemporal dimension verification is a crucial step in ensuring that the current meteorological observation dataset matches the previously constructed meteorological state prediction model. Spatially, the system checks whether the spatial coverage of the current meteorological observation dataset matches the dynamic influence area. The dynamic influence area is defined based on the target observation point and the propagation characteristics of meteorological elements; the current meteorological observation dataset needs to cover this area to accurately reflect the meteorological conditions of the target observation point and its surroundings. Temporally, the system checks whether the time sampling interval of the current meteorological observation dataset is consistent with the time granularity of the cloud cover ratio time series trend. The cloud cover ratio time series trend is analyzed at a certain time granularity, and the time sampling interval of the current meteorological observation dataset needs to be the same to ensure data consistency in the temporal dimension.

[0059] Step 1412: Extract the real-time cloud cover ratio and real-time visibility from the current meteorological observation dataset to generate the current cloud cover ratio sequence and the current visibility sequence.

[0060] After completing the spatiotemporal dimension verification, the meteorological condition prediction system extracts real-time cloud cover ratio and visibility observation values ​​from the current meteorological observation dataset. The real-time cloud cover ratio reflects the cloud cover ratio at the target observation point and within the dynamically affected area at the current moment, while the real-time visibility reflects the current atmospheric clarity. The system arranges these real-time observation values ​​in chronological order to generate the current cloud cover ratio sequence and the current visibility sequence.

[0061] Step 1413: Align the current cloud cover ratio sequence with the cloud cover ratio time series change trend using time reference processing, and fill in the missing timestamp data points in the current cloud cover ratio sequence using a linear interpolation method.

[0062] The meteorological condition forecasting system aligns the current cloud cover ratio sequence with the cloud cover ratio time series trend using a time reference. This time reference alignment ensures consistency between the current cloud cover ratio sequence and the cloud cover ratio time series trend across the time dimension, facilitating comparison and analysis. During processing, the system checks for missing timestamp data points in the current cloud cover ratio sequence. Due to data collection and other reasons, the current cloud cover ratio sequence may have missing data points at certain times. The system uses linear interpolation to fill in these missing data points. Linear interpolation estimates the cloud cover ratio at missing times based on cloud cover ratio observations at adjacent time points through linear calculation. In this way, the system obtains a complete current cloud cover ratio sequence.

[0063] Step 1414: Keep the current visibility sequence consistent with the spatial grid division of the dynamic influence area, and use the nearest neighbor interpolation method to map the current visibility sequence to the preset spatial grid feature cell dimension.

[0064] In this step, the current visibility sequence is aligned with the spatial grid division of the dynamically affected area. The dynamically affected area is divided into multiple spatial grids, each corresponding to a spatial grid feature cell. The current visibility sequence needs to match these spatial grid divisions to accurately reflect the visibility conditions at different locations within the dynamically affected area. The system uses the nearest neighbor interpolation method to map the current visibility sequence to a preset spatial grid feature cell dimension. The nearest neighbor interpolation method finds the nearest spatial grid feature cell based on the location of each observation point in the current visibility sequence and assigns the visibility value of that observation point to that spatial grid feature cell. In this way, the system converts the current visibility sequence into a spatial grid feature cell dimension suitable for model processing.

[0065] Step 1415: Combine the aligned current cloud cover ratio sequence and the mapped current visibility sequence into a model input feature vector, and input the model input feature vector into the meteorological state prediction model.

[0066] The meteorological state prediction system combines the aligned current cloud cover ratio sequence and the mapped current visibility sequence into a model input feature vector. This model input feature vector is a feature representation that integrates the current cloud cover ratio and visibility information, containing key information from the current meteorological observation dataset. During the combination process, the system arranges and combines the current cloud cover ratio sequence and the current visibility sequence according to certain rules to form a unified vector. Then, the system inputs this model input feature vector into the meteorological state prediction model, providing the necessary input data for the model's spatiotemporal feature coupling calculation.

[0067] As another embodiment, the step of obtaining the current state feature vector through spatiotemporal feature coupling calculation using the meteorological state prediction model includes: Step 1421: Input the model input feature vector into the time feature extraction branch of the meteorological state prediction model, and perform time dependency modeling on the current cloud cover ratio sequence in the model input feature vector through a gated loop unit to generate the current time feature vector.

[0068] The meteorological state prediction system inputs the model's input feature vector into the time feature extraction branch of the meteorological state prediction model. The main function of this branch is to process the current cloud cover ratio sequence in the model's input feature vector, extracting the time-dependent variation characteristics of the cloud cover ratio. The system uses a gated recurrent unit (GRU) to model the time dependency of the current cloud cover ratio sequence. A GRU is a recurrent neural network structure capable of processing sequential data and capturing time dependencies within it. When processing the current cloud cover ratio sequence, the GRU analyzes the correlation and changes in the cloud cover ratio across different time points, transforming it into a time-dependent feature representation. Through this process, the system generates a current time feature vector, which contains the time-dimensional features of the current cloud cover ratio's temporal variation information.

[0069] In one embodiment, step 1421 includes: Step 14211: Through standardization processing, the real-time observed cloud cover ratio values ​​in the current cloud cover ratio sequence are converted to a preset numerical range to generate a standardized cloud cover ratio sequence.

[0070] Optionally, the meteorological state prediction system standardizes the real-time cloud cover ratio observations in the current cloud cover ratio sequence. Standardization transforms these observations into a preset numerical range, making different cloud cover ratio observations comparable. During this process, the system analyzes the characteristics of the current cloud cover ratio sequence to determine a suitable standardization method. A common standardization method is to subtract the mean from each real-time cloud cover ratio observation, then divide by its standard deviation, converting it to a value with a mean of 0 and a standard deviation of 1. Through standardization, the system generates a standardized cloud cover ratio sequence. The cloud cover ratio observations in the standardized sequence fall within a uniform numerical range, facilitating processing and analysis by the gated loop unit.

[0071] Step 14212: Input the standardized cloud cover ratio sequence into the input gate of the gated recurrent unit in chronological order, and calculate the update weight of the cloud cover ratio observation at the current time through the sigmoid activation function of the input gate.

[0072] The input gate is a crucial component of the gated recurrent unit (ROU), controlling the degree to which the current input information updates the hidden state. During input, the system calculates the update weight of the current cloud cover ratio observation using the sigmoid activation function of the input gate. The sigmoid activation function is a non-linear function that maps the input cloud cover ratio observation to a value between 0 and 1. This value represents the update weight of the current cloud cover ratio observation on the hidden state; a larger weight indicates a greater impact of the current cloud cover ratio observation on the hidden state.

[0073] Step 14213: Input the standardized cloud cover ratio sequence into the forget gate of the gated recurrent unit, and calculate the retention weight of the hidden state at historical time through the sigmoid activation function of the forget gate.

[0074] The forget gate controls the degree to which the hidden states of past times are preserved in relation to the hidden states of the current time. The system calculates the retention weights of the hidden states of past times using the sigmoid activation function of the forget gate. Similarly, the sigmoid activation function maps the input cloud cover ratio observation to a value between 0 and 1, representing the retention weight of the hidden states of past times; a larger weight indicates a greater influence of the hidden states of past times on the current time.

[0075] Step 14214: Based on the updated weight and the retained weight, calculate the candidate hidden state and the current hidden state of the gated recurrent unit. The candidate hidden state is generated by the tanh activation function, and the current hidden state is the product of the retained weight and the historical hidden state plus the product of the updated weight and the candidate hidden state.

[0076] Based on the calculated updated and retained weights, the system calculates the candidate hidden state and the current hidden state of the gated recurrent unit. The candidate hidden state is generated using the tanh activation function, which maps the input cloud cover ratio observation to a value between -1 and 1. The current hidden state is obtained by weighting the retained and updated weights. Specifically, the current hidden state is the product of the retained weights and the historical hidden states, plus the product of the updated weights and the candidate hidden states. This calculation process comprehensively considers the hidden states at historical moments and the input information at the current moment, enabling the gated recurrent unit to effectively capture the changing characteristics of the cloud cover ratio over time.

[0077] Step 14215: Extract the current hidden state of the gated loop unit at the last time step as the current time feature vector, and the dimension of the current time feature vector is consistent with the dimension of the time feature vector.

[0078] Optionally, the system extracts the current hidden state of the gated recurrent unit at the last time step as the current time feature vector. The current hidden state at the last time step contains the temporal change information of the entire current cloud cover ratio sequence, which can represent the comprehensive characteristics of the current cloud cover ratio in the time dimension. The system ensures that the dimension of the current time feature vector is consistent with the dimension of the previously generated time feature vectors, thereby ensuring that subsequent spatiotemporal feature coupling calculations can proceed smoothly. In this way, the system completes the temporal dependency modeling of the current cloud cover ratio sequence and obtains the current time feature vector.

[0079] Step 1422: Input the model input feature vector into the spatial feature extraction branch of the meteorological state prediction model, and use a convolutional neural network to model the spatial correlation of the current visibility sequence in the model input feature vector to generate the current spatial feature vector.

[0080] The spatial feature extraction branch primarily processes the current visibility sequence in the model's input feature vector, uncovering spatial correlation features of visibility. The system uses a convolutional neural network (CNN) to model the spatial correlation of the current visibility sequence. A CNN is a neural network specifically designed for processing data with a grid structure, automatically extracting spatial features from the data. When processing the current visibility sequence, the CNN analyzes the correlations and changes in visibility at different spatial locations, transforming them into a spatially correlated feature representation. Through this process, the system generates a current spatial feature vector. This vector contains the spatial distribution information and spatial correlation features of the current visibility.

[0081] In one embodiment, step 1422 includes: Step 14221: Reshape the current visibility sequence into a two-dimensional spatial grid matrix, wherein the rows and columns of the two-dimensional spatial grid matrix correspond to the number of grid divisions in the longitude and latitude directions of the dynamically affected area, respectively.

[0082] Furthermore, the system reshapes the current visibility sequence into a two-dimensional spatial grid matrix. Since the dynamically affected area is divided into multiple spatial grids, the current visibility sequence needs to be converted into a two-dimensional matrix form suitable for convolutional neural network processing. The system arranges the observations in the current visibility sequence according to their spatial location based on the number of grid divisions along the longitude and latitude directions of the dynamically affected area, forming a two-dimensional spatial grid matrix. The rows of the matrix correspond to the grid divisions along the longitude direction, and the columns correspond to the grid divisions along the latitude direction. Each matrix element represents the visibility value of the corresponding spatial grid. In this way, the system converts the current visibility sequence into a two-dimensional spatial grid matrix.

[0083] Step 14222: Perform edge padding on the two-dimensional spatial grid matrix to maintain the consistency of spatial dimensions before and after the convolution operation.

[0084] Before performing the convolution operation, the system pads the edges of the 2D spatial grid matrix. The purpose of edge padding is to maintain consistency in spatial dimensions before and after the convolution operation. The convolution operation involves a sliding window process on the input matrix, which may cause the output matrix to become smaller. To avoid this, the system adds extra elements to the edges of the 2D spatial grid matrix, ensuring that the output matrix has the same dimensions as the input matrix after the convolution operation. During the padding process, the system determines the padding method and amount based on parameters such as the kernel size and stride, ensuring that the convolution operation can proceed correctly and the output result has appropriate spatial dimensions.

[0085] Step 14223: Input the filled two-dimensional spatial grid matrix into the first convolutional layer of the convolutional neural network, and perform sliding window convolution operation on the two-dimensional spatial grid matrix through a preset number of convolutional kernels to generate the first-level spatial feature map.

[0086] Further, the system inputs the padded 2D spatial grid matrix into the first convolutional layer of the convolutional neural network. The first convolutional layer is the first processing layer of the convolutional neural network, containing a predetermined number of convolutional kernels. A convolutional kernel is a small matrix that performs a sliding window convolution operation on the 2D spatial grid matrix. During the sliding process, the convolutional kernel performs element-wise multiplication with local regions of the 2D spatial grid matrix and sums the results to obtain a new value. By performing convolution operations on the entire 2D spatial grid matrix, the system generates a first-level spatial feature map. This first-level spatial feature map is the result of the convolution operation and contains some basic spatial features of the 2D spatial grid matrix, such as edge and texture information.

[0087] Step 14224: Input the first-level spatial feature map into the pooling layer of the convolutional neural network, and reduce the spatial dimension of the first-level spatial feature map through max pooling operation to retain key spatial feature information.

[0088] Understandably, the system inputs the first-level spatial feature map into the pooling layer of the convolutional neural network. The pooling layer reduces the spatial dimensionality of the feature map, decreasing computational cost while preserving key spatial feature information. The system employs max pooling, performing a sliding window operation on the first-level spatial feature map. Within each window, the system selects the maximum value as the output value, resulting in a new feature map. Through max pooling, the system removes redundant information, retaining only the maximum value within each window, making the feature map more compact and representative.

[0089] Step 14225: Input the pooled feature map into the second convolutional layer of the convolutional neural network, and extract deep spatial features from the pooled feature map using a number of convolutional kernels that are more numerous than those in the first convolutional layer, to generate a second-level spatial feature map.

[0090] In this embodiment, the system inputs the pooled feature map into the second convolutional layer of the convolutional neural network. The second convolutional layer contains more convolutional kernels than the first convolutional layer, and its function is to extract deep spatial features from the pooled feature map. More convolutional kernels can extract richer spatial features and capture more complex spatial correlation information. During processing, the convolutional kernels of the second convolutional layer perform sliding window convolution operations on the pooled feature map to generate a second-level spatial feature map. The second-level spatial feature map contains higher-level spatial features than the first-level spatial feature map, and can more accurately reflect the distribution and correlation of visibility in spatial dimensions.

[0091] Step 14226: Perform global average pooling on the second-level spatial feature map to convert the two-dimensional feature map into a one-dimensional feature vector, which is used as the current spatial feature vector. The dimension of the current spatial feature vector is consistent with the dimension of the spatial feature vector.

[0092] In this embodiment, the system performs global average pooling on the second-level spatial feature map. Global average pooling is a method to convert a two-dimensional feature map into a one-dimensional feature vector. During processing, the system averages each channel of the second-level spatial feature map, converting the two-dimensional matrix of each channel into a scalar value. By processing all channels, the system obtains a one-dimensional feature vector, which is the current spatial feature vector, containing all spatial feature information of the second-level spatial feature map. The system ensures that the dimension of the current spatial feature vector is consistent with the dimension of the previously generated spatial feature vectors to ensure smooth subsequent spatiotemporal feature coupling calculations. In this way, the system completes the spatial correlation modeling of the current visibility sequence and obtains the current spatial feature vector.

[0093] Step 1423: Input the current time feature vector and the current spatial feature vector into the spatiotemporal coupling layer of the meteorological state prediction model, call the pre-trained cross-attention weight parameters in the spatiotemporal coupling layer, and calculate the mutual information weight matrix of the current time feature and the current spatial feature.

[0094] It's understandable that the meteorological state prediction system inputs the current temporal feature vector and the current spatial feature vector into the spatiotemporal coupling layer of the meteorological state prediction model. This spatiotemporal coupling layer is a crucial layer in the model for fusing temporal and spatial features, and it has already learned pre-trained cross-attention weight parameters during previous training. The system calls these pre-trained cross-attention weight parameters to calculate the mutual information weight matrix between the current temporal and spatial features. This mutual information weight matrix reflects the relationship and importance between the current temporal and spatial features, helping the system determine the weight allocation of each feature across different dimensions during the spatiotemporal feature coupling process. By calculating the mutual information weight matrix, the system can better fuse temporal and spatial features and uncover their potential correlations.

[0095] Step 1424: Perform weighted fusion processing on the current time feature vector and the current spatial feature vector based on the mutual information weight matrix to generate the current coupled feature vector; input the current coupled feature vector into the feature mapping layer of the meteorological state prediction model, and map the current coupled feature vector to the state feature space through a multilayer perceptron to generate the current state feature vector.

[0096] Optionally, the system performs weighted fusion processing on the current temporal feature vector and the current spatial feature vector based on the mutual information weight matrix. During the weighted fusion process, the system weights each element of the current temporal feature vector and the current spatial feature vector according to the weight values ​​in the mutual information weight matrix, and then combines them to generate the current coupled feature vector. The current coupled feature vector integrates information from both the current temporal and spatial features, reflecting feature coupling in both time and space dimensions.

[0097] Next, the system inputs the current coupled feature vector into the feature mapping layer of the meteorological state prediction model. The feature mapping layer is composed of a multilayer perceptron (MLP), a neural network with multiple hidden layers capable of performing nonlinear transformations and mappings on the input features. In the feature mapping layer, the MLP processes the current coupled feature vector, mapping it to a state feature space. This state feature space is a pre-defined feature space, better suited for meteorological state prediction. Through the mapping by the MLP, the system generates a current state feature vector, which is a comprehensive feature representation of the current meteorological state.

[0098] In yet another embodiment, generating a meteorological state prediction result for the target observation point within a preset time period based on the current state feature vector includes: Step 1431: Input the current state feature vector into the output prediction layer of the meteorological state prediction model, calculate the probability distribution of the current state feature vector in the meteorological state category space through the softmax activation function, and generate a meteorological state probability vector.

[0099] In this embodiment, the meteorological state prediction system inputs the current state feature vector into the output prediction layer of the meteorological state prediction model. The main function of the output prediction layer is to generate a prediction result of the meteorological state based on the current state feature vector. The system calculates the probability distribution of the current state feature vector in the meteorological state category space using a softmax activation function. The meteorological state category space contains various possible meteorological states, such as sunny, cloudy, and overcast. The softmax activation function converts each element of the current state feature vector into a probability value, such that the sum of all probability values ​​is 1. In this way, the system generates a meteorological state probability vector, where each element represents the probability of the corresponding meteorological state occurring.

[0100] Step 1432: Extract the probability value corresponding to the glowing sky state from the meteorological state probability vector, as the probability of the glowing sky state occurring.

[0101] Optionally, the system extracts the probability value corresponding to the sunset glow state from the meteorological state probability vector. The meteorological state probability vector contains the probability of various meteorological states occurring. The system finds the probability value corresponding to the sunset glow state based on the category identifier of the meteorological state. This probability value is the probability of the sunset glow state occurring, reflecting the likelihood that the target observation point will experience a sunset glow state within a preset time period under the current meteorological conditions. By extracting the probability of the sunset glow state occurring, the system can provide users with specific prediction information about the sunset glow state.

[0102] Step 1433: Based on the probability of the occurrence of the glowing sky, combined with the changing trend of the current cloud cover ratio sequence and the spatial distribution characteristics of the current visibility sequence, determine the duration range and intensity changing trend of the glowing sky.

[0103] In this embodiment, the system determines the duration range and intensity trend of the glow phenomenon based on the probability of its occurrence, combined with the changing trend of the current cloud cover ratio sequence and the spatial distribution characteristics of the current visibility sequence. The changing trend of the current cloud cover ratio sequence reflects the change of the cloud cover ratio over time, and different trends in cloud cover ratio will affect the duration and intensity of the glow phenomenon. The spatial distribution characteristics of the current visibility sequence reflect the spatial distribution of atmospheric clarity, which will also affect the propagation and visibility of the glow phenomenon. The system comprehensively considers these three factors and determines the duration range and intensity trend of the glow phenomenon by analyzing their interrelationships. For example, when the cloud cover ratio is in a decreasing phase and the visibility is good, the glow phenomenon may last for a longer period and be more intense.

[0104] In a preferred embodiment, step 1433 includes: Step 14330: Calculate the first derivative of the current cloud cover ratio sequence to determine the positive and negative intervals of the cloud cover ratio change rate. The positive interval corresponds to the cloud cover ratio increasing phase, and the negative interval corresponds to the cloud cover ratio decreasing phase. Analyze the correlation between the probability of the occurrence of the sunset glow and the positive and negative intervals of the cloud cover ratio change rate. When the probability of the occurrence of the sunset glow is higher than a preset threshold and the cloud cover ratio is in a decreasing phase, mark it as the potential sunset glow start time. After the potential sunset glow start time, continuously monitor the change trend of the current cloud cover ratio sequence. When the cloud cover ratio change rate changes from negative to positive and the probability of the sunset glow occurs is lower than a preset threshold, mark it as the potential sunset glow end time. Use the time interval between the potential sunset glow start time and the potential sunset glow end time as the duration interval of the sunset glow. Calculate the mean distribution of the spatial distribution characteristics of the current visibility sequence within the duration interval. Compare the mean distribution with a preset visibility intensity level classification standard to determine the intensity change trend of the sunset glow. The intensity change trend increases with the increase of the mean visibility and decreases with the decrease of the mean visibility.

[0105] First, the system calculates the first derivative of the current cloud cover ratio sequence's changing trend. The first derivative reflects the rate of change of the cloud cover ratio; by calculating the first derivative, the system can determine the positive and negative intervals of the cloud cover ratio's changing rate. A positive interval indicates that the cloud cover ratio is in an increasing phase, and a negative interval indicates that the cloud cover ratio is in a decreasing phase. Then, the system performs a correlation analysis between the probability of a sunset glow and the positive and negative intervals of the cloud cover ratio's changing rate. When the probability of a sunset glow occurring is higher than a preset threshold and the cloud cover ratio is in a decreasing phase, the system marks this moment as the potential start time of a sunset glow. After the potential start time of a sunset glow, the system continuously monitors the changing trend of the current cloud cover ratio sequence. When the rate of change of the cloud cover ratio turns from negative to positive and the probability of a sunset glow occurring is lower than a preset threshold, the system marks this moment as the potential end time of a sunset glow. The time interval between the potential start time and the potential end time of a sunset glow is the duration interval of the sunset glow.

[0106] Secondly, the system calculates the mean of the spatial distribution characteristics of the current visibility sequence within the duration interval. The mean reflects the average clarity of the atmosphere within the duration interval. The system compares the mean with a preset visibility intensity level classification standard and determines the intensity change trend of the sunset glow based on the comparison results. Generally speaking, the higher the mean visibility, the clearer the atmosphere, the better the propagation and visibility of the sunset glow, and the stronger the intensity of the sunset glow; conversely, the lower the mean visibility, the more blurred the atmosphere, and the weaker the intensity of the sunset glow. In this way, the system can accurately determine the duration interval and intensity change trend of the sunset glow.

[0107] Step 1434: Combine the probability of occurrence, duration range, and intensity change trend of the above-mentioned glow state into glow state description information.

[0108] It is understandable that the system combines the probability of the occurrence of a sunset glow, its duration range, and its intensity variation trend into a sunset glow state description. This description is a comprehensive overview of the sunset glow state, containing important information such as the likelihood of its occurrence, its duration, and intensity variations. During the combination process, the system organizes and records this information according to a specific format, forming a complete description. By providing this sunset glow state description, the system can offer users more detailed and accurate sunset glow state predictions, helping them better understand the sunset glow conditions at the target observation point within a preset time period.

[0109] Step 1435: Integrate the description information of the glow state with the probability value corresponding to the meteorological state category of the same period to generate a meteorological state prediction result sequence containing timestamps. The time granularity of the meteorological state prediction result sequence is consistent with the time division of the preset time period.

[0110] Optionally, the system integrates the description information of the afterglow condition with the probability values ​​corresponding to the meteorological state categories of the same period. The probability values ​​corresponding to the meteorological state categories of the same period are the probability values ​​of other meteorological states besides the afterglow condition in the meteorological state probability vector. The system will uniformly organize the afterglow condition description information and these probability values ​​to form a complete meteorological state prediction information. At the same time, the system will add a timestamp to this meteorological state prediction information, and the timestamp will accurately record the time corresponding to each prediction result. The system will ensure that the time granularity of the meteorological state prediction result sequence is consistent with the time division of the preset time period. The preset time period is a pre-set time period. The system will arrange the meteorological state prediction results according to the time division method of this time period to generate a meteorological state prediction result sequence containing timestamps. In this way, the system provides users with a comprehensive and detailed meteorological state prediction result. Users can understand the various meteorological conditions of the target observation point within the preset time period, including the specific situation of the afterglow condition, based on this sequence.

[0111] As a non-limiting embodiment, after generating the meteorological state prediction result of the target observation point within a preset time period based on the current state feature vector, the method further includes: obtaining the actual meteorological observation dataset of the target observation point and the dynamic influence area after the preset time period ends; aligning the actual meteorological observation dataset with the meteorological state prediction result in a spatiotemporal dimension to generate an error feature vector containing prediction bias; performing spatiotemporal distribution pattern recognition on the error feature vector; determining the concentrated fluctuation range in the time dimension and the frequent deviation region in the spatial dimension of the error feature vector using a density clustering algorithm; using the concentrated fluctuation range and the frequent deviation region as weak links in model optimization; adjusting the cross-attention mechanism weight parameters of the spatiotemporal feature coupling layer in the meteorological state prediction model based on the weak link identification; increasing the attention weight ratio of the time segment feature units corresponding to the concentrated fluctuation range; and improving the feature extraction weight of the spatial grid feature units corresponding to the frequent deviation region; updating the adjusted cross-attention mechanism weight parameters to the meteorological state prediction model; evaluating the performance of the updated meteorological state prediction model using a preset validation dataset; and completing the iterative optimization update of the model when the evaluation index meets the preset convergence condition.

[0112] After the preset time period ends, the meteorological state prediction system acquires the actual meteorological observation dataset for the target observation point and the dynamically affected area. This dataset contains actual meteorological data that occurred within the preset time period, reflecting the true meteorological conditions. The system then performs spatiotemporal alignment processing between the actual meteorological observation dataset and the meteorological state prediction results. This alignment is a crucial step in ensuring the comparability of the actual meteorological observation data and the prediction results in both time and space. Through this alignment process, the system identifies the differences between the prediction results and the actual situation, generating an error feature vector containing the prediction bias. Each element in the error feature vector represents the magnitude of the prediction bias at the corresponding time and spatial location.

[0113] The system performs spatiotemporal distribution pattern recognition on the error feature vector. Through density clustering algorithms, the system can identify the concentrated fluctuation ranges of the error feature vector in the time dimension and the frequent deviation regions in the spatial dimension. The concentrated fluctuation ranges refer to the time periods in which the error fluctuates significantly, while the frequent deviation regions refer to the areas in which the error frequently deviates significantly in the spatial dimension. These concentrated fluctuation ranges and frequent deviation regions reflect the inaccuracy of the model's predictions at certain time and spatial locations, and they identify the weak links in model optimization.

[0114] Based on the identification of weak links, the system adjusts the weight parameters of the cross-attention mechanism in the spatiotemporal feature coupling layer of the meteorological state prediction model. Specifically, it increases the attention weight ratio of feature units in time segments corresponding to concentrated fluctuation intervals, enabling the model to pay more attention to relevant features when processing meteorological data in these time periods. At the same time, the system enhances the feature extraction weight of spatial grid feature units corresponding to areas of frequent deviation, allowing the model to extract features more accurately when processing meteorological data in these spatial locations.

[0115] The system updates the weather state prediction model with the adjusted cross-attention mechanism weights. Then, it evaluates the updated model's performance using a pre-defined validation dataset. The validation dataset consists of a set of known real weather conditions; the system compares the updated model's predictions on the validation dataset with the actual conditions and calculates evaluation metrics. When the evaluation metrics meet pre-defined convergence conditions, it indicates that the model's performance has been effectively improved, and the system completes iterative optimization and update of the model. In this way, the system continuously improves the weather state prediction model, enhancing its accuracy and reliability.

[0116] As a non-limiting embodiment, after generating the meteorological state prediction result of the target observation point within a preset time period based on the current state feature vector, the method further includes: extracting spatial propagation characteristic parameters of meteorological elements related to the cloud cover state from the meteorological state prediction result, wherein the spatial propagation characteristic parameters of meteorological elements include propagation direction vector and diffusion rate features; adjusting the spatial grid division density of the dynamic influence area based on the spatial propagation characteristic parameters of meteorological elements, increasing the number of grid divisions along the extension path of the propagation direction vector, and increasing the grid resolution in areas where the diffusion rate feature is higher than a preset threshold; re-extracting the visibility state parameter sequence of the dynamic influence area from the historical meteorological forecast dataset according to the adjusted spatial grid division, and generating an updated spatial feature vector by aligning it with the timestamp of the cloud cover ratio time series change trend; inputting the updated spatial feature vector into the spatial feature extraction branch of the meteorological state prediction model, replacing the original spatial feature vector to update the spatial feature mapping relationship, and dynamically iteratively updating the meteorological state prediction model by comparing the deviation of the model before and after the update for the same input meteorological state prediction results.

[0117] In this embodiment, the meteorological state prediction system extracts spatial propagation characteristic parameters of meteorological elements related to the state of the sunset from the meteorological state prediction results. These parameters include propagation direction vectors and diffusion rate characteristics. The propagation direction vector represents the propagation direction of meteorological elements (such as clouds, water vapor, etc.), and the diffusion rate characteristic represents the diffusion speed of meteorological elements. These parameters are very important for understanding the propagation and changes of meteorological elements in the spatial dimension.

[0118] Based on the spatial propagation characteristics of meteorological elements, the system adjusts the spatial grid density of the dynamically affected area. Along the extension path of the propagation direction vector, the system increases the number of grid cells to capture more detailed changes in meteorological elements along the propagation path. In areas where the diffusion rate characteristics exceed a preset threshold, the system increases the grid resolution, enabling the model to process meteorological data from these areas more accurately. By adjusting the spatial grid density, the system can more precisely describe the meteorological environment within the dynamically affected area.

[0119] Furthermore, the system re-extracts the visibility state parameter sequence for the dynamically affected area from the historical weather forecast dataset based on the adjusted spatial grid division. Because the spatial grid division has changed, the corresponding visibility data needs to be re-extracted. The system timestamps the newly extracted visibility state parameter sequence with the cloud cover ratio time series trend, generating an updated spatial feature vector. The updated spatial feature vector contains more accurate spatial dimensional information and can better reflect the visibility distribution within the dynamically affected area.

[0120] Furthermore, the system inputs the updated spatial feature vector into the spatial feature extraction branch of the meteorological state prediction model, replacing the original spatial feature vector. By replacing the spatial feature vector, the system updates the spatial feature mapping relationship, enabling the model to better handle meteorological data under the new spatial grid division. The system compares the deviation of the model's prediction results for the same input meteorological state before and after the update. If the deviation is large, it indicates that the model needs further adjustment; if the deviation is small, it indicates that the model update is effective. In this way, the system dynamically iterates and updates the meteorological state prediction model, continuously improving the model's performance and prediction accuracy to adapt to the needs of predicting the state of sunset under different meteorological conditions.

[0121] In summary, the embodiments of this application significantly improve the accuracy and reliability of weather condition forecasting, especially the forecasting of sunset conditions. First, historical weather forecast datasets are accessed after credible authentication through the meteorological forecast field database, ensuring the reliability and authority of the data source while also guaranteeing data security within the meteorological forecast field database. Dynamic influence areas are delineated based on historical weather forecast datasets and the geographical coordinates of the target observation point, fully considering the spatial propagation characteristics of meteorological elements. This allows the model to more accurately capture changes in the meteorological environment around the target observation point, avoiding the limitations of fixed-area delineation in traditional methods. The time-series variation trend of cloud cover ratio under precipitation-free conditions within the dynamic influence area is determined, and a spatiotemporal interpolation-coupled meteorological state prediction model is constructed by combining the visibility state parameters of the dynamic influence area. This spatiotemporal coupling method comprehensively considers changes in meteorological elements in both time and space, enabling a more comprehensive and accurate reflection of the evolution of meteorological states. The current meteorological observation dataset is input into the meteorological state prediction model for spatiotemporal feature coupling calculation to obtain the current state feature vector, thereby generating a meteorological state prediction result for the target observation point within a preset time period, including the occurrence of overcast skies. This allows for timely and accurate prediction of overcast skies, providing strong support for related meteorological services and decision-making.

[0122] Based on the same inventive concept, embodiments of this application also provide a weather condition prediction system. See also... Figure 2 As shown, this is a schematic diagram of a possible weather condition prediction system provided in an embodiment of this application. Figure 2 In the meteorological state prediction system 200, there are a processor 210 and a memory 220. The memory 220 stores a computer program that can be executed by the processor 210. By executing the instructions stored in the memory 220, the processor 210 can perform the steps of the meteorological state prediction method based on the spatiotemporal interpolation coupling model described above.

[0123] Based on the same inventive concept, embodiments of this application provide a computer-readable storage medium including a computer program. When the computer program is run on a meteorological state prediction system, it causes the meteorological state prediction system to perform the steps of the aforementioned meteorological state prediction method based on a spatiotemporal interpolation coupling model. In some possible implementations, various aspects of the meteorological state prediction method based on a spatiotemporal interpolation coupling model provided in this application can also be implemented as a program product, including a computer program. When the program product is run on a meteorological state prediction system, the computer program causes the meteorological state prediction system to perform the steps of the aforementioned meteorological state prediction method based on a spatiotemporal interpolation coupling model. For example, the meteorological state prediction system can perform actions such as... Figure 1 The steps are shown in the figure.

[0124] In the technical solutions involved in the above embodiments of the present invention, whether performing comparison calculations of multi-dimensional features or constructing composite parameters, if there are problems caused by significant differences in the number of dimensions, units of measurement, and semantic meanings of different features, those skilled in the art, based on their professional knowledge and past practical experience, can fully understand that these differences need to be properly handled so that the calculation results are accurate and comparable, and to avoid situations such as logical confusion and unclear mathematical meaning.

[0125] In detail, when faced with features of different numbers of dimensions, those skilled in the art can employ various strategies, including but not limited to feature selection, feature extraction, and kernel function processing, in order to accurately calculate the similarity, matching degree, or feature distance between different features.

[0126] In order to achieve comparability alignment of feature spaces when processing the comparison of multidimensional features, those skilled in the art can use a variety of existing common technical means, including but not limited to standardization preprocessing, mapping transformation, and spatial projection.

[0127] In the process of constructing composite parameters (such as loss function values), different parameter terms often have different dimensions. Those skilled in the art can use normalization processing or an adaptive weight allocation mechanism based on distribution characteristics.

[0128] The aforementioned general techniques for solving feature matching and loss balance problems are all common knowledge in this field. These techniques have been fully verified and widely used in numerous practical applications, and those skilled in the art can skillfully and flexibly apply these methods to handle similar problems involving differences in dimensions.

[0129] The formulas and calculation processes involved in the embodiments of this application, whether used for multidimensional feature comparison or composite loss function construction, strictly adhere to the principle of dimensional correspondence. Each variable in the formula has a clear and explicit physical meaning, and its operational logic fully conforms to basic mathematical and physical logic. The calculation results are necessarily the reasonable results expected by this application. Those skilled in the art are capable of effectively solving various problems arising from the number of dimensions, dimensional differences, etc., in the multidimensional feature comparison calculation and composite loss function construction in the embodiments, based on specific data conditions and business needs, by comprehensively utilizing the above-mentioned general technical means, thus ensuring the accuracy, reliability, and implementability of the technical solution of this invention.

Claims

1. A meteorological state prediction method based on a spatiotemporal interpolation coupling model, characterized in that, include: After passing the trusted authentication of the meteorological forecast field database, the historical meteorological forecast dataset stored in the meteorological forecast field database is invoked; The dynamic influence area is divided based on the historical weather forecast dataset and the geographical coordinates of the target observation point. The dynamic influence area is centered on the target observation point and is affected by the spatial propagation characteristics of meteorological elements. Determine the time series variation trend of cloud cover ratio within the dynamic influence area under no precipitation conditions, and build a spatiotemporal interpolation coupled meteorological state prediction model by combining the time series variation trend of cloud cover ratio and the visibility state parameters of the dynamic influence area. The current meteorological observation dataset is input into the meteorological state prediction model. The current state feature vector is obtained by performing spatiotemporal feature coupling calculation through the meteorological state prediction model. Based on the current state feature vector, the meteorological state prediction result of the target observation point within a preset time period is generated. The meteorological state prediction result includes the state of the sunset glow.

2. The method according to claim 1, characterized in that, Determining the time series variation trend of cloud cover ratio within the dynamically affected area under no precipitation conditions includes: The precipitation status identifiers in the historical weather forecast dataset are filtered to extract the meteorological data subset with no precipitation status identifiers within the dynamic influence area. The cloud cover ratio observations and corresponding timestamp information are extracted from the meteorological data subset to generate the original cloud cover ratio sequence arranged in ascending order of timestamps; The original cloud cover ratio sequence is subjected to time continuity verification, and discrete data points with timestamp intervals exceeding a preset threshold are deleted to obtain a time-continuous cloud cover ratio calibration sequence. Calculate the mean and standard deviation of the cloud cover ratio calibration sequence within a sliding time window, identify and correct abnormal fluctuation points in the cloud cover ratio calibration sequence based on the mean and standard deviation, and generate a smoothed cloud cover ratio time series; The smoothed cloud cover ratio time series is subjected to trend fitting processing. A time series decomposition algorithm is used to separate the steady-state trend component and the periodic fluctuation component. The superposition result of the steady-state trend component and the periodic fluctuation component is taken as the change trend of the cloud cover ratio time series.

3. The method according to claim 1, characterized in that, The method of constructing a spatiotemporally interpolated coupled meteorological state prediction model by combining the time series variation trend of cloud cover ratio and the visibility state parameters of the dynamically affected area includes: The visibility observation values ​​within the dynamic influence area are extracted from the historical weather forecast dataset, and a visibility state parameter sequence aligned with the timestamp of the cloud cover ratio time series change trend is generated. The cloud cover ratio time series change trend is processed by time feature encoding, and the cloud cover ratio time series change trend is converted into a time-dependent time feature vector through a time series embedding algorithm; Spatial feature extraction processing is performed on the visibility state parameter sequence. Based on the geographical location coordinate information of the dynamic influence area, the visibility correlation degree between different sub-regions is calculated to generate a spatial feature vector with spatial distribution characteristics. A spatiotemporal feature coupling layer is constructed. The temporal feature vector and the spatial feature vector are input into the spatiotemporal feature coupling layer. The mutual information weights of the temporal feature vector and the spatial feature vector are calculated through a cross-attention mechanism to generate a coupled feature vector that integrates spatiotemporal correlation information. Using the coupled feature vector as input, a meteorological state prediction model is constructed, which includes a feature mapping layer and an output prediction layer. The feature mapping layer is used to map the coupled feature vector to a state feature space of a preset dimension, and the output prediction layer is used to generate a meteorological state prediction probability distribution based on the feature distribution of the state feature space.

4. The method according to claim 3, characterized in that, The construction of the spatiotemporal feature coupling layer involves inputting the temporal feature vector and the spatial feature vector into the spatiotemporal feature coupling layer, calculating the mutual information weights of the temporal feature vector and the spatial feature vector through a cross-attention mechanism, and generating a coupled feature vector that fuses spatiotemporal correlation information, including: The time feature vector is divided into multiple time segment feature units according to the time step, and the spatial feature vector is divided into multiple spatial grid feature units according to the geographical location; Calculate the cosine similarity between each time segment feature unit and each spatial grid feature unit to generate a spatiotemporal similarity matrix; Based on the spatiotemporal similarity matrix, the spatial attention weight of each time segment feature unit to different spatial grid feature units is calculated by row normalization, and the temporal attention weight of each spatial grid feature unit to different time segment feature units is calculated by column normalization. The time segment feature units are weighted and summed with their corresponding spatial attention weights to generate a time-guided spatial feature vector; The spatial grid feature units are weighted and summed with their corresponding temporal attention weights to generate a spatially guided temporal feature vector. The time-guided spatial feature vector and the space-guided time feature vector are added element by element to generate a coupled feature vector that integrates spatiotemporal correlation information.

5. The method according to claim 1, characterized in that, The step of inputting the current meteorological observation dataset into the meteorological state prediction model includes: The current meteorological observation dataset is validated in terms of its spatiotemporal dimensions to ensure that the current meteorological observation dataset contains a spatial coverage range that matches the dynamic impact area and a temporal sampling interval that is consistent with the temporal granularity of the cloud cover ratio time series change trend. Extract real-time cloud cover ratio and real-time visibility from the current meteorological observation dataset to generate the current cloud cover ratio sequence and the current visibility sequence; The current cloud cover ratio sequence is aligned with the cloud cover ratio time series change trend using time reference processing, and the missing timestamp data points in the current cloud cover ratio sequence are filled in using a linear interpolation method. The current visibility sequence is kept consistent with the spatial grid division of the dynamic influence area, and the current visibility sequence is mapped to the preset spatial grid feature unit dimension using the nearest neighbor interpolation method; The aligned current cloud cover ratio sequence and the mapped current visibility sequence are combined into a model input feature vector, which is then input into the meteorological state prediction model.

6. The method according to claim 5, characterized in that, The process of obtaining the current state feature vector through spatiotemporal feature coupling calculation using the meteorological state prediction model includes: The model input feature vector is input into the time feature extraction branch of the meteorological state prediction model. The time dependency model of the current cloud cover ratio sequence in the model input feature vector is performed through a gated loop unit to generate the current time feature vector. The model input feature vector is input into the spatial feature extraction branch of the meteorological state prediction model. The current visibility sequence in the model input feature vector is spatially correlated and modeled by a convolutional neural network to generate the current spatial feature vector. The current time feature vector and the current spatial feature vector are input into the spatiotemporal coupling layer of the meteorological state prediction model. The pre-trained cross-attention weight parameters in the spatiotemporal coupling layer are called to calculate the mutual information weight matrix between the current time feature and the current spatial feature. Based on the mutual information weight matrix, the current time feature vector and the current spatial feature vector are weighted and fused to generate the current coupled feature vector; The current coupled feature vector is input into the feature mapping layer of the meteorological state prediction model, and the current coupled feature vector is mapped to the state feature space through a multilayer perceptron to generate the current state feature vector.

7. The method according to claim 6, characterized in that, The step of inputting the model input feature vector into the time feature extraction branch of the meteorological state prediction model, and performing time dependency modeling on the current cloud cover ratio sequence in the model input feature vector through a gated recurrent unit to generate the current time feature vector, includes: Through standardization processing, the real-time observed cloud cover ratio values ​​in the current cloud cover ratio sequence are converted to a preset numerical range to generate a standardized cloud cover ratio sequence. The standardized cloud cover ratio sequence is input into the input gate of the gated recurrent unit in chronological order, and the update weight of the cloud cover ratio observation at the current time is calculated through the sigmoid activation function of the input gate. The standardized cloud cover ratio sequence is input into the forget gate of the gated recurrent unit, and the retention weight of the hidden state at historical time is calculated through the sigmoid activation function of the forget gate. Based on the updated weights and the retained weights, the candidate hidden state and the current hidden state of the gated recurrent unit are calculated. The candidate hidden state is generated by the tanh activation function, and the current hidden state is the product of the retained weights and the historical hidden state plus the product of the updated weights and the candidate hidden state. The current hidden state of the gated loop unit at the last time step is extracted as the current time feature vector, and the dimension of the current time feature vector is consistent with the dimension of the time feature vector.

8. The method according to claim 6, characterized in that, The step of inputting the model input feature vector into the spatial feature extraction branch of the meteorological state prediction model, and using a convolutional neural network to perform spatial correlation modeling on the current visibility sequence in the model input feature vector to generate the current spatial feature vector includes: The current visibility sequence is reshaped into a two-dimensional spatial grid matrix, where the rows and columns of the two-dimensional spatial grid matrix correspond to the number of grid divisions in the longitude and latitude directions of the dynamically affected area, respectively. Edge padding is applied to the two-dimensional spatial grid matrix to maintain the consistency of spatial dimensions before and after the convolution operation; The filled two-dimensional spatial grid matrix is ​​input into the first convolutional layer of the convolutional neural network. A sliding window convolution operation is performed on the two-dimensional spatial grid matrix through a preset number of convolutional kernels to generate a first-level spatial feature map. The first-level spatial feature map is input into the pooling layer of the convolutional neural network. The spatial dimension of the first-level spatial feature map is reduced by the max pooling operation, while retaining key spatial feature information. The pooled feature map is input into the second convolutional layer of the convolutional neural network. The pooled feature map is then subjected to depth space feature extraction by a number of convolutional kernels that are more numerous than those in the first convolutional layer, generating a second-level spatial feature map. The second-level spatial feature map is subjected to global average pooling to convert the two-dimensional feature map into a one-dimensional feature vector, which is used as the current spatial feature vector. The dimension of the current spatial feature vector is consistent with the dimension of the spatial feature vector.

9. The method according to claim 1, characterized in that, The step of generating meteorological state prediction results for the target observation point within a preset time period based on the current state feature vector includes: The current state feature vector is input into the output prediction layer of the meteorological state prediction model, and the probability distribution of the current state feature vector in the meteorological state category space is calculated by the softmax activation function to generate a meteorological state probability vector. Extract the probability value corresponding to the state of the glowing clouds from the meteorological state probability vector, and use it as the probability of the glowing clouds occurring. Based on the probability of the occurrence of the glowing sky, combined with the changing trend of the current cloud cover ratio sequence and the spatial distribution characteristics of the current visibility sequence, the duration range and intensity changing trend of the glowing sky are determined. The probability of occurrence, duration range, and intensity variation trend of the above-mentioned glow state are combined into glow state description information; The description information of the glow state is integrated with the probability value corresponding to the meteorological state category of the same period to generate a meteorological state prediction result sequence containing timestamps. The time granularity of the meteorological state prediction result sequence is consistent with the time division of the preset time period. The determination of the duration range and intensity variation trend of the overcast state based on the probability of its occurrence, combined with the changing trend of the current cloud cover ratio sequence and the spatial distribution characteristics of the current visibility sequence, includes: The first derivative of the current cloud cover ratio sequence is calculated to determine the positive and negative intervals of the rate of change of cloud cover ratio. The positive interval corresponds to the stage of cloud cover ratio increase, and the negative interval corresponds to the stage of cloud cover ratio decrease. The probability of the occurrence of the glowing sky is correlated with the positive and negative ranges of the rate of change of cloud cover ratio. When the probability of the occurrence of the glowing sky is higher than a preset threshold and the cloud cover ratio is in a decreasing phase, it is marked as the potential glowing sky start time. After the potential sunrise start time, the changing trend of the current cloud cover ratio sequence is continuously monitored. When the cloud cover ratio change rate changes from negative to positive and the probability of sunrise occurs is lower than a preset threshold, it is marked as the potential sunrise end time. The time interval between the start and end times of the potential glow is taken as the duration range of the glow state; Calculate the mean value of the spatial distribution characteristics of the current visibility sequence within the duration interval, compare the mean value with the preset visibility intensity level classification standard, and determine the intensity change trend of the rosy dawn state. The intensity change trend increases with the increase of the mean value of visibility and decreases with the decrease of the mean value of visibility.

10. A meteorological condition prediction system, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when executed by the processor, causes the processor to perform the steps of any one of the methods described in claims 1 to 9.

Images