Data mining method and system based on meteorological service

Abstract

The application relates to the technical field of meteorological information, and discloses a data mining method and system based on meteorological services. The method comprises the following steps: a dynamic space-time graph fusing a space distance and a terrain feature is constructed; a space-time correlation feature of a meteorological element is extracted by using a graph neural network; a Navier-Stokes equation and a thermodynamic energy equation are embedded into a neural network loss function as physical constraints; and a prediction result is output by jointly optimizing a data-driven loss and a physical constraint loss. The system comprises data acquisition and preprocessing, dynamic graph construction, graph feature extraction, physical constraint embedding, joint training optimization and forecast output modules. Through the dual driving mechanism of "physics + data", the physical consistency, robustness of the prediction and the feature extraction efficiency under complex terrain are improved.

Description

Technical Field

[0001] This invention belongs to the field of meteorological information technology, specifically relating to a data mining method and system based on meteorological services. Background Technology

[0002] Meteorological services serve as a crucial guarantee for socio-economic operation and disaster prevention and mitigation, and their accuracy directly impacts key sectors such as transportation, energy dispatch, and agricultural production. With the continuous advancement of global observation networks and remote sensing technology, the meteorological field has accumulated massive amounts of multi-source heterogeneous data. Deep analysis and trend prediction of meteorological elements through big data mining techniques have become a core research topic and application direction in modern meteorological science.

[0003] Meteorological data mining methods based on spatiotemporal characteristics aim to reveal the evolutionary patterns of weather systems by analyzing the correlation between the spatial distribution and temporal evolution of atmospheric states. This technical approach typically utilizes various statistical models or machine learning algorithms to extract features from meteorological station data distributed in irregular geographical locations, striving to achieve intelligent characterization and evolutionary inference of key indicators such as temperature, air pressure, and fluid motion under complex dynamic environments.

[0004] Current technologies primarily employ purely data-driven deep learning models for meteorological modeling, but these models still face numerous challenges in practical applications. First, in marginal regions where meteorological observation stations are sparse or historical data is lacking, traditional models often exhibit poor generalization ability, failing to capture sudden meteorological changes at extremely small scales. Second, due to a lack of fundamental understanding of atmospheric physics, purely data-driven models are prone to producing distorted predictions that violate energy conservation or thermodynamic laws when facing extreme weather scenarios, leading to predictions that contradict common physical sense. Third, conventional neural network structures struggle to effectively handle station topological relationships in non-Euclidean spaces, resulting in low efficiency in extracting spatial dependencies and a lack of physical interpretability. Finally, existing solutions often suffer from insufficient robustness when dealing with complex meteorological scenarios due to a lack of prior physical constraints. These problems collectively limit the accuracy and reliability of meteorological forecasts. Therefore, developing a data mining solution that integrates physical dynamics mechanisms with spatiotemporal topological features has become a critical technical challenge that needs to be addressed in the industry. Summary of the Invention

[0005] The purpose of this invention is to provide a data mining method and system based on meteorological services, which can solve the problems mentioned in the background art. Specifically, this invention aims to address the problem of insufficient generalization ability of traditional meteorological data mining models when facing areas with sparse meteorological observation stations or missing historical data. It also addresses the problem that pure data-driven models easily produce prediction results that violate physical principles such as atmospheric dynamics and thermodynamics in extreme weather forecasting. By introducing a deep fusion mechanism of physical information neural networks and graph neural networks, the robustness, physical consistency, and feature extraction efficiency of the meteorological forecasting system in complex topological spaces are improved.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: a data mining method based on meteorological services, comprising the following steps: S1, acquiring multi-source meteorological monitoring data of the target area and constructing a dynamic spatiotemporal map structure based on the geographical topological relationship of meteorological stations; S2, processing the dynamic spatiotemporal map structure using a graph neural network to extract the spatiotemporal correlation features of meteorological elements in non-Euclidean space; S3, constructing a physical information neural network constrained by atmospheric physics equations, embedding prior physical knowledge into the loss function of the neural network, and guiding the model learning through physical constraint terms; S4, jointly iteratively optimizing the model by combining data-driven loss and physical constraint loss, and outputting the meteorological element mining and trend prediction results of the target area based on the optimized model.

[0007] Preferably, step S1 specifically includes the following steps: S11, collecting observation data from multiple meteorological stations within a preset time period in the target area, including temperature, air pressure, wind speed, wind direction, and humidity; S12, constructing a static spatial map using each meteorological station as a node and the spatial distance and topographic correlation between stations as edge weights; S13, introducing a time series dimension, stacking the static spatial maps of different time steps vertically to form a dynamic spatiotemporal map structure that represents the evolution of meteorological elements over time.

[0008] Preferably, step S2 specifically includes the following steps: S21, normalizing the node features in the dynamic spatiotemporal graph structure and mapping meteorological elements of different dimensions to a predetermined numerical range; S22, aggregating the feature information of neighboring nodes through graph convolution operations and adaptively allocating the weights of different neighboring stations in spatial interaction using an attention mechanism; S23, using recurrent neural network units to process the aggregated time series features and capture the evolution trend of the meteorological system over a long period of time.

[0009] Preferably, step S3 specifically includes the following steps: S31, based on the principles of atmospheric dynamics, establishing the Navier-Stokes equations and thermodynamic energy equations describing atmospheric motion as the core physical constraint sources; S32, using automatic differentiation techniques to calculate the time derivative and spatial derivative of the neural network output relative to the input coordinates; S33, substituting the calculated derivative terms into the Navier-Stokes equations and thermodynamic energy equations to construct physical residual terms; S34, transforming the physical residual terms into a physical constraint loss function, making it an important component of the model's total loss function.

[0010] Preferably, the multi-source meteorological monitoring data of the target area needs to undergo data quality control before constructing the dynamic spatiotemporal map structure. Specifically, this includes: completing missing observations using a preset interpolation algorithm, such as Kriging interpolation or inverse distance weighted interpolation; detecting and removing outliers; and smoothing outliers using a sliding window mean method to ensure that the data input to the model meets the requirements of atmospheric stability in terms of temporal continuity.

[0011] Preferably, in the construction of the dynamic spatiotemporal graph structure, the edge weight calculation method between stations not only considers Euclidean distance but also incorporates altitude differences and underlying surface roughness features. The edge weights are negatively correlated with spatial distance and positively correlated with terrain similarity. By setting a specific distance threshold, redundant connections with correlation below a preset range are eliminated, reducing the computational complexity of subsequent graph convolution operations.

[0012] Preferably, the graph convolution operation in the graph neural network employs either spectral domain convolution or spatial domain convolution. In spatial domain convolution, long-distance feature diffusion is achieved through a multi-layer information transfer mechanism. Each graph convolution layer is coupled with a nonlinear activation function to enhance the model's ability to fit the nonlinear evolution characteristics of meteorological elements. The attention mechanism measures the correlation between stations by calculating the dot product of the query vector and the key vector, enabling the model to automatically identify the key upstream weather systems that have the greatest impact on the target station.

[0013] Preferably, the recurrent neural network unit adopts a long short-term memory network structure or a gated recurrent unit structure. By introducing forget gate, input gate, and output gate control mechanisms, the gradient vanishing problem in long-sequence meteorological data processing is solved, ensuring that the model can retain the evolution memory features of historical heavy precipitation or cold waves and other sudden meteorological events.

[0014] Preferably, the specific construction process of the physical constraint terms involves the discretization of the laws of conservation of mass, momentum, and energy. Specifically, the continuity equation in the Navier-Stokes equations is configured to constrain the evolution of fluid mass, and the momentum equation is configured to constrain the dynamic balance between the wind field velocity vector and the pressure gradient force, Coriolis force, and frictional force. The thermodynamic energy equation is configured to constrain the relationship between temperature change and heat transport, adiabatic compression, and radiative heating.

[0015] Preferably, the automatic differentiation technique is implemented through the reverse mode of the computation graph. During model training, there is no need to manually derive complex partial differential equations; instead, the partial derivatives of the predicted elements in the spatiotemporal coordinate system are directly obtained through the differential operators of the deep learning framework, and the physical residuals are calculated in real time. The magnitude of the physical residuals directly reflects the degree of deviation between the model's prediction results and known physical laws.

[0016] Preferably, the physical constraint loss function is obtained by taking the sum of squares or the sum of absolute values ​​of the physical residuals over the entire spatiotemporal domain. To balance the influence between the data-driven loss and the physical constraint loss, a dynamic weighting strategy is introduced. In the early stages of training, the data-driven loss is given a higher weight to accelerate the model's fitting to the observed data; in the later stages of training, the weight of the physical constraint loss is gradually increased to force the prediction results to regress to the physically feasible solution space.

[0017] Preferably, the joint iterative optimization process employs a gradient-based optimization algorithm and introduces specific regularization terms to prevent model overfitting. The regularization terms include a weight decay term and a dropout layer mechanism, which continuously updates the connection weights of the neural network during the iteration process, ensuring that the total loss function reaches a preset convergence threshold.

[0018] Preferably, the meteorological element mining and trend prediction results output in step S4 include not only specific numerical forecasts, but also confidence intervals based on physical consistency assessment. For areas with small physical residuals, the system assigns a higher confidence level; for extreme and sudden scenarios with large physical residuals, the system outputs an early warning and initiates an adaptive correction mechanism based on physical priors.

[0019] A data mining system based on meteorological services, used to implement the above-mentioned method, includes: a data acquisition and preprocessing module for acquiring multi-source meteorological monitoring data of the target area and performing denoising, interpolation, and normalization processing; a dynamic graph construction module for constructing a dynamic spatiotemporal graph structure based on the geographical topological relationships and temporal evolution characteristics of meteorological stations; a graph feature extraction module, whose input is connected to the output of the dynamic graph construction module, for extracting spatiotemporal correlation features in non-Euclidean space using a graph neural network; a physical constraint embedding module for storing atmospheric physics equations and using an automatic differentiation tool to transform the physics equations into physical constraint terms of the neural network; a joint training and optimization module for iteratively optimizing the model by combining data-driven loss and physical constraint loss; and a forecast output module, whose input is connected to the output of the joint training and optimization module, for outputting meteorological mining and trend prediction results that conform to physical laws.

[0020] Preferably, the data acquisition and preprocessing module includes multiple sensor interfaces for real-time access to multi-source heterogeneous data from ground automatic weather stations, radiosonde stations, weather radars, and satellite remote sensing. This data acquisition and preprocessing module is internally configured with a parallel processing engine, enabling real-time streaming preprocessing of large-scale meteorological data.

[0021] Preferably, the graph feature extraction module adopts a hierarchical architecture design. The bottom layer consists of multi-layer graph attention networks to capture local spatial correlations; the middle layer consists of bidirectional gated recurrent units to capture bidirectional temporal dependencies; and the top layer consists of fully connected layers to map abstract spatiotemporal features back to the physical element space.

[0022] Preferably, the physical constraint embedding module has a built-in extensible physical equation library, allowing users to flexibly select specific physical constraint operators according to different mining tasks (such as temperature prediction, precipitation estimation, or air quality analysis). The physical constraint embedding module is also equipped with a coordinate mapping unit for converting the geographic coordinates of meteorological stations into standardized spatiotemporal coordinates required for model training.

[0023] Preferably, the joint training and optimization module supports a distributed training architecture, enabling parallel optimization of large-scale dynamic spatiotemporal graphs on high-performance computing clusters. This joint training and optimization module also includes a hyperparameter adaptive adjustment unit, which automatically optimizes the learning rate based on the convergence speed of the physical loss term.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] This invention constructs a novel paradigm for meteorological data mining by deeply integrating physical information neural networks and graph neural networks. Utilizing the graph neural network's ability to process non-Euclidean space, it accurately characterizes the complex spatial dependencies between irregularly distributed meteorological stations, overcoming the information loss and bias caused by the requirement of projecting data onto a regular grid in traditional convolutional neural networks when processing station data.

[0026] By embedding atmospheric physics constraints such as the Navier-Stokes equation into the loss function, the deep physical mechanisms of atmospheric science are transformed into prior knowledge of neural networks. This solves the problem of poor generalization ability of pure data-driven models in sparse data regions, enabling the model to make reasonable logical extrapolations based on physical laws even in areas lacking observational data.

[0027] The "physics plus data" dual-drive mode implemented in this invention eliminates the possibility of AI models generating predictions that violate energy conservation or thermodynamic laws, thus enhancing the robustness and interpretability of meteorological forecasts. In summary, this invention, while ensuring forecast accuracy, improves the system's survivability under extreme weather conditions, producing a synergistic effect that combines the powerful nonlinear fitting capabilities of artificial intelligence with strict adherence to the boundaries of atmospheric science and physics, providing solid technical support for precision meteorological services. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0029] Figure 2 This is a schematic diagram of the core principle framework of the deep fusion mechanism of physical information neural network and graph neural network in this invention;

[0030] Figure 3 This is a flowchart illustrating the logical process of constructing the dynamic spatiotemporal graph structure and extracting non-Euclidean spatiotemporal features in this invention.

[0031] Figure 4 This is a flowchart illustrating the logical process of embedding prior physical knowledge and generating physical constraint terms in this invention.

[0032] Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow of the dual-path joint optimization of data-driven and physical constraints in this invention. Detailed Implementation

[0033] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following description is provided in conjunction with the appendix. Figure 1 Only attached Figure 5 The following is a detailed description of the specific implementation methods, structures, features, and effects of the present invention, as well as preferred embodiments.

[0034] Example 1: This example provides a data mining method based on meteorological services. This method is mainly applied to high-precision spatiotemporal mining and evolution prediction of meteorological elements in complex urban environments. Due to the dense buildings and irregular terrain features in urban environments, traditional meteorological observation stations are often unevenly distributed. This leads to significant forecast biases in areas with insufficient observation station coverage, where traditional meteorological forecasting models lack sufficient initialization information. This example aims to solve the challenge of modeling the nonlinear evolution of meteorological elements in such complex environments by fusing physical information neural networks and graph neural networks.

[0035] In step S1, the system first acquires multi-source meteorological monitoring data for the target area. In this embodiment, the target area is selected as the core urban area of ​​a megacity, where 120 automatic ground weather stations and 3 upper-air sounding stations are deployed. The multi-source meteorological monitoring data includes, but is not limited to, real-time observation sequences collected by these stations over the past 48 hours. Specific acquisition parameters cover air temperature, atmospheric pressure, wind speed at a height of 10 meters above the ground, wind direction angle, and relative humidity. To ensure the real-time performance and accuracy of the data, all sensors use industrial-grade high-precision probes, with the temperature sensor achieving a measurement accuracy of 0.1 degrees Celsius, the barometric pressure sensor achieving a measurement accuracy of 0.1 hPa, and the wind speed sensor having an initial wind speed not exceeding 0.5 meters per second.

[0036] After data acquisition, the system automatically enters the data quality control process. Considering that meteorological sensors may be affected by electromagnetic interference or hardware aging during long-term operation, leading to missing or outlier values, this embodiment uses a sliding window mean method to detect outliers in real time. If the difference between an observed value at a certain moment and the mean of its adjacent windows exceeds a preset threshold of three times the standard deviation, that point is determined to be an outlier and removed. For the missing data after removal and for data loss due to network transmission fluctuations, this embodiment uses the Kriging interpolation algorithm to complete the spatial dimension. The Kriging interpolation algorithm utilizes the spatial autocorrelation characteristics between known observation points, calculating the variogram to provide an unbiased optimal estimate for missing locations, ensuring excellent spatial continuity of the data input to the model.

[0037] In constructing the dynamic spatiotemporal graph structure, the system first treats each meteorological station as an independent node in the graph structure. Then, it establishes edge connections between nodes based on the geographical topological relationships of the stations. In this embodiment, the calculation of edge weights not only considers the Euclidean spatial distance between stations but also introduces altitude difference features through a digital elevation model. The magnitude of the edge weight is directly proportional to the reciprocal of the spatial distance and negatively correlated with the absolute value of the altitude difference. This means that the closer two stations are spatially and at similar altitudes, the higher their correlation in the graph structure, and the larger their edge weights. To improve computational efficiency, the system sets a specific correlation threshold, forcibly removing redundant connections with a correlation score below 0.1, transforming the original fully connected graph into a sparse graph, and reducing the complexity of subsequent graph convolution operations. Finally, the system introduces a time series dimension, vertically stacking the static spatial graphs generated at different time steps in chronological order to form a dynamic spatiotemporal graph structure that can simultaneously represent the spatial topological evolution and temporal migration of meteorological elements.

[0038] In step S2, the system uses a graph neural network to extract deep features from the aforementioned dynamic spatiotemporal graph structure. First, normalization is performed on the node features in the graph structure. Because the dimensions of elements such as temperature, air pressure, and wind speed differ significantly, directly inputting them into the network could lead to gradient explosion or training non-convergence. This embodiment uses a minimax mapping algorithm to linearly transform the original values ​​of all elements to a predetermined numerical range between 0 and 1.

[0039] In the feature aggregation stage, the system employs spatial domain graph convolution operations. Each graph convolution layer traverses the neighboring stations of the current node and accumulates the meteorological feature information of the neighboring stations after linear transformation using a preset weight matrix. To achieve adaptive feature allocation, this embodiment introduces an attention mechanism. Specifically, the system calculates the correlation coefficient between the current node and its neighboring nodes, which is measured by calculating the dot product of the query vector and the key vector. For key upstream stations that have a significant impact on the meteorological state of the current station, the attention mechanism automatically assigns them higher weight coefficients, enabling the model to automatically identify and capture the propagation paths of atmospheric fluctuations between different stations.

[0040] The aggregated feature vectors are then fed into a recurrent neural network unit for processing to capture the evolution trend of the meteorological system over a long time span. This embodiment uses a Long Short-Term Memory (LSTM) network structure. This LSM network structure is internally configured with a forget gate, an input gate, and an output gate. The forget gate determines how much historical information to retain based on the current input and the state of the previous time step. For example, in a cold wave scenario, the LSM network can retain the strong cooling memory features from previous time steps through a gating mechanism, thereby significantly improving its ability to capture future large temperature fluctuations and solving the gradient vanishing problem that conventional neural networks face when processing long-sequence data.

[0041] In step S3, this embodiment constructs a physical information neural network constrained by atmospheric physics equations. This is one of the core innovations of this invention. Traditional data-driven models often focus only on the statistical correlations between data, without considering the physical laws that atmospheric motion must follow. In this embodiment, the system establishes the Navier-Stokes equations describing fluid motion and the thermodynamic energy equations describing energy conversion as the core physical constraint sources.

[0042] The specific process of constructing physical constraint terms is as follows: First, the system uses automatic differentiation technology to calculate the derivatives of the meteorological forecast elements output by the neural network with respect to the input time coordinates and spatial three-dimensional coordinates in the backpropagation link of the deep learning framework. For example, for wind field prediction tasks, the system directly obtains the rates of change of the horizontal and vertical velocity components in various spatial directions through differential operators.

[0043] Subsequently, the calculated derivative terms are substituted into the continuity equation of the Navier-Stokes equations. In physics, the continuity equation requires that the mass of the atmosphere, as a fluid, must be conserved. The system measures its physical consistency by calculating the divergence of the velocity vector, i.e., the sum of the derivatives of the horizontal velocity component in the horizontal direction and the vertical velocity component in the vertical direction. If this sum of derivatives is not zero, the resulting residual value is the physical residual term. Similarly, in the momentum equation, the system balances the predicted acceleration term with the pressure gradient force, Coriolis force, and frictional force terms. The thermodynamic energy equation is configured to constrain the rate of temperature change over time to be equal to the sum of contributions from thermal advection, adiabatic compression, and external radiative heating.

[0044] These physical residual terms are transformed into a physical constraint loss function and become an important component of the model's total loss function. Essentially, this approach adds a physical prior regularization term to the neural network's loss function, forcing the model to not only fit historical observation data but also ensure that the prediction results are physically feasible and self-consistent during the search for the optimal solution.

[0045] In step S4, the system performs joint iterative optimization. This embodiment uses a gradient-based adaptive moment estimation algorithm for parameter updates. To balance the weights of data-driven loss and physical constraint loss, a dynamic weighting strategy is introduced. In the initial training phase, the system assigns a higher weight to the data-driven loss, with a weight ratio of 0.9:0.1. This is because the model needs to learn the basic statistical distribution of the data in the early stages of training. As the training epochs increase, the system gradually and smoothly increases the weight of the physical constraint loss, adjusting the weight ratio to 0.4:0.6 in the later stages of training. This dynamic adjustment mechanism ensures that the model can revert to the physical solution space in the later stages, avoiding false predictions that violate atmospheric dynamics.

[0046] After multiple iterations until the total loss function reaches a preset convergence threshold, the optimized model outputs the meteorological element mining and trend prediction results for the target area. The results output in this embodiment not only include specific values ​​for temperature and wind speed, but also a confidence interval based on physical consistency assessment. For areas with minimal physical residuals generated during the calculation process, the system determines that the prediction results highly conform to physical laws and assigns a confidence score of over 95%. For areas with complex terrain leading to large physical residuals, the system automatically triggers an adaptive correction mechanism, using prior physical knowledge to perform logistic regression correction on prediction points with large deviations, outputting forecast products that are both accurate and consistent with atmospheric science common sense.

[0047] Example 2: This example focuses on the specific application of data mining for local circulation and severe convective weather in complex mountainous terrain. Compared with Example 1, this example has made specific technical optimizations in the construction of the dynamic spatiotemporal map structure and the selection of physical constraints to adapt to the highly nonlinear and drastically fluctuating characteristics of meteorological elements in mountainous environments.

[0048] In the construction of the dynamic spatiotemporal map in step S1, due to the complexity of the mountainous terrain, the straight-line spatial distance between stations is insufficient to characterize the correlation of meteorological elements. Therefore, this embodiment specifically introduces the underlying surface roughness characteristics and the consistency parameter of slope aspect when calculating edge weights. If two meteorological stations are located on the same slope of the same ridgeline, even if their physical distance is far, the system will assign them a higher edge weight based on the consistency of slope aspect. In addition, by calculating the topographic relief between two stations, if there is a large mountain barrier in the path, the system will reduce its correlation weight. This graph structure construction method based on topographic dynamics features can more accurately simulate the transmission path of local circulation in mountainous areas at different altitudes.

[0049] In the feature extraction stage of step S2, considering the suddenness of elements such as mountain precipitation, this embodiment introduces a multi-layer information transfer mechanism in the graph convolutional network. Each graph convolutional layer is coupled with a modified linear unit activation function to enhance the model's ability to fit nonlinear step features. Furthermore, the attention mechanism is configured as a multi-head attention mode, that is, simultaneously calculating the correlation between stations in multiple different subspaces. For example, one set of attention heads is specifically responsible for capturing the topographically forced lifting features of the lower atmosphere, while another set of attention heads is responsible for capturing the circulation guidance features of the upper atmosphere. Finally, feature concatenation achieves a deep characterization of the complex meteorological background.

[0050] In the physical constraint embedding step S3, this embodiment adds a buoyancy constraint term under the Businesk approximation for mountainous environments. The atmospheric thermodynamic equations are reinforced to address the thermal differences caused by topography. Because different slopes of a mountain receive varying amounts of radiation over different times, sensible heat fluxes differ significantly. The system embeds the surface energy balance equation into a physical information neural network, constraining changes in surface temperature to satisfy the conservation relationships of shortwave radiation, longwave radiation, sensible heat flux, and latent heat flux. Through this refined physical constraint, this embodiment can correct the inversion prediction failure problem caused by the lack of observational data in traditional models in valley areas.

[0051] In the iterative optimization of step S4, this embodiment employs a distributed training architecture. Since mountain meteorological data contains massive amounts of terrain raster information, single-machine computing power is insufficient. The system divides the target area into multiple sub-computation domains and synchronously updates model parameters using a parallel gradient descent algorithm on a high-performance computing cluster. Simultaneously, considering the sparsity of extreme mountain weather, a weight decay term and a dropout layer mechanism are introduced during the optimization process. In each iteration, 20% of neuron connections are randomly disabled to prevent overfitting to specific historical extreme cases, thereby improving the system's generalization and transfer capabilities across different mountain environments.

[0052] Example 3: This example describes a data mining system based on meteorological services, used to implement the methods described in the above examples. The system is designed as a highly integrated industrial software platform capable of accessing various heterogeneous meteorological observation resources and outputting customized forecast services.

[0053] The system includes a data acquisition and preprocessing module. This module contains multiple standardized sensor interfaces, supporting real-time access to multi-source heterogeneous data from ground automatic weather stations, radiosonde stations, Doppler weather radars, and polar-orbiting weather satellites. In terms of hardware architecture, the module incorporates a parallel processing engine, enabling it to utilize the concurrency capabilities of multi-core processors to perform synchronous denoising, interpolation, and normalization on real-time streaming data. For image-type data such as radar echoes, the module first converts them into a feature field in polar coordinates, and then projects them onto a geographic coordinate system consistent with the station data through a coordinate mapping unit.

[0054] The system also includes a dynamic graph construction module. This module is responsible for calculating the geographic topological relationships between stations and generating spatiotemporal maps. Internally, this module stores high-precision electronic map data of the target area and a terrain elevation database, enabling real-time calculation of terrain similarity between any two observation nodes. Users can flexibly adjust the edge weight calculation rules through a graphical interface; for example, when handling cold wave tasks, they can increase the weight of latitude differences, and when handling typhoon tasks, they can add a reference to offshore distance.

[0055] The input of the graph feature extraction module is connected to the output of the dynamic graph construction module. The graph feature extraction module adopts a hierarchical architecture. The bottom layer consists of multi-layer graph attention networks, used to capture the local spatial correlation of sites and identify spatial structures such as local shear lines or small- to medium-scale vortices; the middle layer consists of bidirectional gated recurrent units, which can simultaneously utilize past and future time-series information to enhance the feature characteristics of the current moment; the top layer consists of fully connected layers, responsible for mapping the multi-layered abstracted spatiotemporal feature vectors back to the real physical feature space.

[0056] The physical constraint embedding module is the core logical unit of the system. This module has a built-in, extensible library of physical equations, storing partial differential equations covering multiple fields such as fluid mechanics, thermodynamics, and cloud physics. Users can select specific physical constraint operators based on the specific mining task. For example, when performing air quality analysis, users can call the convection-diffusion equation from the library as a constraint term; when performing precipitation estimation, they can call the water vapor continuity equation. The physical constraint embedding module is also equipped with an automatic differentiation tool, which can automatically transform abstract physical equations into gradient residual terms that can be recognized by neural networks, eliminating the need for manual formula derivation by the user.

[0057] The joint training and optimization module is responsible for the overall evolution of the model. This module supports distributed storage management, maintaining extremely high computational efficiency when dealing with large-scale dynamic spatiotemporal graphs. Internally, it includes a hyperparameter adaptive adjustment unit that monitors the convergence speed of the physical loss term in real time. If the physical residual is found to be fluctuating at a high level for an extended period, this hyperparameter adaptive adjustment unit automatically lowers the learning rate and increases the penalty factor for physical constraints, forcing the model to revert to physical laws.

[0058] The input of the forecast output module is connected to the output of the joint training and optimization module. This module can output not only traditional contour maps and time-series forecast curves, but also physically consistent heatmaps. In these heatmaps, the system uses different color depths to represent the degree to which the prediction results satisfy the physical laws. For risk areas with large physical residuals, the forecast output module automatically initiates early warning alerts and recommends that meteorological decision-makers increase mobile observation resources in the area.

[0059] The system in this embodiment also possesses strong scalability. Through a standard application programming interface (API), the system can seamlessly interface with downstream industry application models. For example, in traffic meteorological services, the high-precision wind field forecast output by the system can directly drive the vehicle crosswind stability assessment model; in energy dispatch services, the radiation forecast output by the system can serve as a key input parameter for photovoltaic power generation prediction models.

[0060] To further enhance system robustness, each module in this embodiment is deployed using a microservice architecture. Core processes such as data acquisition, feature extraction, and physical constraint calculation all run in independent containers. When a module experiences response delays due to excessive load or hardware failure, the system can automatically perform dynamic horizontal scaling of resources, ensuring that the meteorological service system maintains stable data mining and forecasting capabilities even during large-scale sudden weather disasters.

[0061] In any of the above embodiments, the textual processing of physical constraints follows this logic: When dealing with momentum conservation constraints, the system obtains the components of the fluid velocity vector output by the neural network in three spatial dimensions. Then, it automatically differentiates and calculates the local rate of change of each velocity component with respect to time, representing the fluid acceleration. Simultaneously, it calculates the gradient of the pressure field in the corresponding direction, as well as the Coriolis force term due to the Earth's rotation. The system requires that the acceleration term, pressure gradient force term, Coriolis force term, and frictional force term due to air viscosity reach dynamic equilibrium in the vector space. The remaining values ​​after these terms cancel each other out are calculated, which is the physical residual of the momentum equation. This physical constraint constructed through pure textual logic allows the present invention to accurately convert physical laws into supervisory signals for the neural network without relying on mathematical symbols.

[0062] The verbal description of the energy conservation constraint is as follows: the system extracts the predicted temperature change over time. According to the first law of thermodynamics, this change should equal the sum of the heat advection term driven by wind speed, the work done by adiabatic expansion or compression due to air pressure changes, and the latent heat released due to atmospheric moisture phase change. In each training iteration, the system calculates the algebraic difference between these terms. The smaller the difference, the more the model's prediction conforms to the law of energy conservation. In this way, this invention successfully integrates the profound physical mechanisms of atmospheric science into the end-to-end training process of deep learning.

[0063] In summary, this invention combines the powerful processing capabilities of graph neural networks for non-Euclidean spatial topologies with the hard constraint capabilities of physical information neural networks on physical laws, constructing a novel paradigm for meteorological data mining. This paradigm effectively compensates for the generalization defects of pure data-driven models in sparse data regions, ensuring the physical reliability of prediction results under extreme weather scenarios. Through deep fusion of multi-source heterogeneous data and precise extraction of spatiotemporal correlation features, this invention significantly improves the robustness and interpretability of the system while ensuring prediction accuracy, providing reliable technical support and decision-making basis for modern precision meteorological services.

[0064] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A data mining method based on meteorological services, characterized in that, Includes the following steps: S1. Acquire multi-source meteorological monitoring data of the target area and construct a dynamic spatiotemporal map structure based on the geographical topological relationship of meteorological stations; The construction process of the dynamic spatiotemporal graph structure includes: taking each meteorological station as an independent node in the graph structure, establishing edge connections between nodes based on the geographical topological relationship of the meteorological stations, calculating the spatial distance, altitude difference, and underlying surface roughness characteristics between stations, and determining the edge weights accordingly to construct a static spatial graph; introducing a time series dimension, stacking the static spatial graphs generated at different time steps vertically in chronological order to form a dynamic spatiotemporal graph structure that represents the evolution of meteorological elements over time; S2. The dynamic spatiotemporal graph structure is processed using a graph neural network to extract the spatiotemporal correlation features of meteorological elements in non-Euclidean space; the feature information of neighboring nodes is aggregated through graph convolution operations, and the dot product of the query vector and the key vector is calculated using an attention mechanism to measure the correlation between stations, thereby adaptively allocating the weights of different neighboring stations in spatial interaction; the aggregated time series features are processed using recurrent neural network units to capture the evolution trend of the meteorological system over a long period of time. In the process of constructing the dynamic spatiotemporal graph structure, the calculation logic of the edge weight is as follows: the magnitude of the edge weight is negatively correlated with the spatial distance between the meteorological stations and positively correlated with the terrain similarity. The calculation of the terrain similarity takes into account the differences in altitude, the roughness characteristics of the underlying surface, and the consistency parameters of the slope aspect. When two meteorological stations are located on the same slope of the same ridgeline, the edge weight is increased based on the consistency of slope aspect; When there are preset terrain undulations between two meteorological stations, the edge weights are reduced to simulate the transmission path of local circulation at different altitudes. By setting a correlation threshold, redundant connections with a correlation degree lower than the threshold are eliminated, and the fully connected graph is transformed into a sparse graph, thereby reducing the computational complexity of subsequent graph convolution operations. S3. Construct a physical information neural network constrained by atmospheric physics equations, embed the atmospheric physics equations into the loss function of the physical information neural network, and guide the model learning through physical constraint terms; wherein, automatic differentiation technology is used to calculate the time derivative and spatial derivative of the neural network output relative to the input coordinates; the calculated derivative terms are substituted into the dynamic equations and thermodynamic energy equations describing atmospheric motion to construct physical residual terms; the physical residual terms are transformed into the physical constraint loss function; S4. Combine data-driven loss and physical constraint loss to perform joint iterative optimization of the physical information neural network, and output the results of meteorological element mining and trend prediction of the target area based on the optimized model.

2. The data mining method based on meteorological services according to claim 1, characterized in that, After acquiring multi-source meteorological monitoring data of the target area, the method further includes a step of performing data quality control on the multi-source meteorological monitoring data: The sliding window mean method is used to detect outliers in the multi-source meteorological monitoring data. When the difference between the observed value and the mean in the adjacent window exceeds the preset deviation threshold, the observed value is determined to be an outlier and is removed. For the vacant spaces after removal and the original missing values, an interpolation algorithm is used to complete the spatial dimensions; the interpolation algorithm includes Kriging interpolation or inverse distance weighted interpolation. The Kriging interpolation process utilizes the spatial autocorrelation characteristics between known observation points and calculates the variogram to provide unbiased optimal estimates for missing locations, ensuring that the data input to the model has spatial continuity. The node features in the multi-source meteorological monitoring data are normalized, and the original values ​​of meteorological elements with different dimensions are linearly transformed into a predetermined numerical range through a mapping algorithm to eliminate the impact of dimension differences on the stability of model training.

3. The data mining method based on meteorological services according to claim 1, characterized in that, The process of the graph neural network extracting spatiotemporal correlation features includes: the graph convolution operation adopts spatial domain convolution, and the long-distance diffusion of features is achieved through a multi-layer information transmission mechanism. Each layer of graph convolution is combined with a nonlinear activation function to enhance the model's ability to fit the nonlinear evolution features of meteorological elements. The attention mechanism is configured as a multi-head attention mode, which simultaneously calculates the correlation between stations in multiple different subspaces. One set of attention heads is responsible for capturing the topographic forced lifting features of the lower atmosphere, while another set of attention heads is responsible for capturing the circulation guidance features of the upper atmosphere. The complex meteorological background is characterized by feature stitching. The recurrent neural network unit adopts a long short-term memory network structure or a gated recurrent unit structure, and is internally configured with forget gate, input gate and output gate control mechanism. Through the gating mechanism, the evolution memory features of previous time steps are retained, long sequence meteorological data are processed and the propagation path of atmospheric fluctuations between different stations is captured.

4. The data mining method based on meteorological services according to claim 3, characterized in that, The specific construction process of the physical constraint terms involves the discretization of the laws of conservation of mass, momentum, and energy: the dynamic equations describing atmospheric motion include the Navier-Stokes equations, in which the continuity equation is configured to constrain the evolution of fluid mass, and physical consistency is measured by calculating the divergence of the velocity vector. The momentum equation is configured to constrain the dynamic equilibrium of the wind field velocity vector with the pressure gradient force, Coriolis force, and friction force, requiring the acceleration term, pressure gradient force term, Coriolis force term, and friction force term due to air viscosity to reach equilibrium in the vector space. The thermodynamic energy equation is configured to constrain the conservation relationship between the rate of change of temperature over time and the heat advection term, the work done by adiabatic expansion or compression term, and the external radiation heating term, and the physical residual term is constructed by calculating the algebraic difference between each term.

5. The data mining method based on meteorological services according to claim 4, characterized in that, The construction of the physical constraints also includes a correction step for the terrain environment: in the mountainous environment scenario, a buoyancy constraint term under the Businsk approximation is added to the physical constraints; the thermal difference constraint caused by the terrain is strengthened in the thermodynamic energy equation, and the surface energy balance equation is embedded into the physical information neural network to constrain the change of surface temperature to satisfy the conservation relationship of shortwave radiation, longwave radiation, sensible heat flux and latent heat flux, so as to correct the prediction bias caused by the lack of observation data in the valley area.

6. The data mining method based on meteorological services according to claim 5, characterized in that, The joint iterative optimization process is implemented as follows: A gradient-based adaptive moment estimation algorithm is used for parameter updates, and a regularization term is introduced into the loss function to prevent overfitting. This regularization term includes a weight decay term and a dropout layer mechanism. A dynamic weighting strategy is introduced to balance the data-driven loss and the physical constraint loss. In the initial stage of model training, the data-driven loss is given a high weight proportion, enabling the model to learn the basic statistical distribution in the observed data. In the middle and later stages of model training, the weight proportion of the physical constraint loss is gradually increased, setting the prediction results to regress to a physically feasible solution space. The model convergence is guided by dynamically adjusting the penalty factor of the physical constraint loss.

7. The data mining method based on meteorological services according to claim 6, characterized in that, The output results of meteorological element mining and trend prediction for the target area include: outputting specific predicted values ​​of forecast elements, and simultaneously outputting confidence intervals based on physical consistency assessment; the generation logic of the confidence interval is as follows: using the physical residual term to measure the degree of deviation between the model prediction results and physical laws, assigning a high confidence score to areas where the physical residual is lower than a preset residual threshold; for extreme and sudden scenarios where the physical residual exceeds the preset residual threshold, the system outputs an early warning prompt and initiates an adaptive correction mechanism based on physical priors, using physical equations to perform logistic regression correction on prediction points with large deviations.

8. A data mining system based on meteorological services, characterized in that, The data mining method based on meteorological services as described in any one of claims 1 to 7 includes: The data acquisition and preprocessing module is used to acquire multi-source meteorological monitoring data of the target area and perform noise reduction, interpolation and normalization processing. The dynamic graph construction module is used to construct dynamic spatiotemporal graph structures based on the geographic topological relationships and temporal evolution characteristics of meteorological stations. The dynamic map construction module stores high-precision electronic map data of the target area and a terrain elevation database, which is used to calculate the terrain similarity between observation nodes and generate a spatiotemporal map. The graph feature extraction module has its input end connected to the output end of the dynamic graph construction module, and is used to extract spatiotemporal correlation features in non-Euclidean space using graph neural networks. The graph feature extraction module adopts a hierarchical architecture, including a graph attention network layer that captures local spatial correlations and a recurrent neural network unit layer that captures temporal dependencies. The physical constraint embedding module is used to store the atmospheric physics equation library and use an automatic differentiation tool to transform the physical equations into physical constraint terms of the physical information neural network. The physical residual is obtained by calculating the partial derivatives of the predicted elements in the spatiotemporal coordinate system. A joint training and optimization module is used to iteratively optimize the physical information neural network by combining data-driven loss and physical constraint loss. The joint training and optimization module includes a hyperparameter adaptive adjustment unit, which is used to automatically optimize the learning rate according to the convergence speed of the physical residual term. The forecast output module, whose input is connected to the output of the joint training and optimization module, is used to output meteorological element mining and trend prediction results that conform to atmospheric physical laws, and generate a heat map characterizing the degree of physical consistency of the prediction results.

9. The data mining system based on meteorological services according to claim 8, characterized in that, The system also possesses the following technical features: The data acquisition and preprocessing module includes multiple standardized sensor interfaces for real-time access to multi-source heterogeneous data from ground automatic weather stations, radiosonde stations, weather radars, and satellite remote sensing. It is also equipped with a parallel processing engine to achieve real-time streaming preprocessing of large-scale meteorological data. The physical constraint embedding module has an extensible physical equation library that supports flexible invocation of specific physical constraint operators according to different mining tasks. These physical constraint operators include convection-diffusion equations for air quality analysis and runoff balance equations for hydro-meteorological analysis. The data mining system is deployed using a distributed training architecture and a microservice architecture. The data acquisition, feature extraction, and physical constraint calculation stages run in independent containers, supporting parallel optimization and dynamic horizontal scaling of resources on large-scale dynamic spatiotemporal maps on high-performance computing clusters.

Images